| Message ID | 20221207085236.1424424-7-goldstein.w.n@gmail.com |
|---|---|
| State | New |
| Headers | show |
| Series | [v1,01/27] x86/fpu: Create helper file for common data macros | expand |
On Wed, Dec 7, 2022 at 12:52 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote: > > No changes to the logic, just change how rodata is handled. > > 1. Define the rodatas using the new macros so they check that the > offset is correct. > > 2. Use common data where applicable. > --- > .../multiarch/svml_s_tanhf16_core_avx512.S | 450 ++++++++---------- > 1 file changed, 197 insertions(+), 253 deletions(-) > > diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S b/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > index d74fc7731d..765e9ed7f7 100644 > --- a/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > +++ b/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > @@ -70,94 +70,99 @@ > * > */ > > -/* Offsets for data table __svml_stanh_data_internal_avx512. Ordered > - by use in the function. On cold-starts this might help the > - prefetcher. Possibly a better idea is to interleave start/end so > - that the prefetcher is less likely to detect a stream and pull > - irrelivant lines into cache. */ > > -/* Offsets for data table __svml_stanh_data_internal. 4 bytes each. > - */ > + > +#define LOCAL_DATA_NAME __svml_stanh_data_internal > +#define LOCAL_DATA_NAME_UNALIGNED __svml_stanh_data_internal_unaligned > +#include "svml_s_common_evex512_rodata_offsets.h" > + > +/* Offsets for data table __svml_stanh_data_internal_unaligned. > + 4 bytes each. */ > #define _iExpMantMask_UISA 0 > #define _iMinIdxOfsMask_UISA 4 > #define _iMaxIdxMask_UISA 8 > #define _iExpMask 12 > > -/* Offsets for data table __svml_stanh_data_internal_al64. 64 bytes > - each. */ > -#define _sC_lo 0 > -#define _sC_hi 64 > -#define _sP7_lo 128 > -#define _sP7_hi 192 > -#define _sSignMask 256 > -#define _sP6_lo 320 > -#define _sP6_hi 384 > -#define _sP5_lo 448 > -#define _sP5_hi 512 > -#define _sP4_lo 576 > -#define _sP4_hi 640 > -#define _sP3_lo 704 > -#define _sP3_hi 768 > -#define _sP2_lo 832 > -#define _sP2_hi 896 > -#define _sP0_lo 960 > -#define _sP0_hi 1024 > +/* Offsets for data table __svml_stanh_data_internal. Ordered > + by use in the function. On cold-starts this might help the > + prefetcher. Possibly a better idea is to interleave start/end so > + that the prefetcher is less likely to detect a stream and pull > + irrelivant lines into cache. */ > + > +/* Offsets for data table __svml_stanh_data_internal. > + 64 bytes each. */ > +#define _sC_lo 0 > +#define _sC_hi 64 > +#define _sP7_lo 128 > +#define _sP7_hi 192 > +#define _sP6_lo 256 > +#define _sP6_hi 320 > +#define _sP5_lo 384 > +#define _sP5_hi 448 > +#define _sP4_lo 512 > +#define _sP4_hi 576 > +#define _sP3_lo 640 > +#define _sP3_hi 704 > +#define _sP2_lo 768 > +#define _sP2_hi 832 > +#define _sP0_lo 896 > +#define _sP0_hi 960 > + > > #include <sysdep.h> > -#define TANHF_DATA(x) ((x)+__svml_stanh_data_internal_al64) > -#define TANHF_DATA_UNALIGNED(x) ((x)+__svml_stanh_data_internal) > > .section .text.evex512, "ax", @progbits > ENTRY(_ZGVeN16v_tanhf_skx) > - /* Here huge arguments, INF and NaNs are filtered out to callout. */ > - vpandd TANHF_DATA_UNALIGNED(_iExpMantMask_UISA)(%rip){1to16}, %zmm0, %zmm1 > - vpsubd TANHF_DATA_UNALIGNED(_iMinIdxOfsMask_UISA)(%rip){1to16}, %zmm1, %zmm2 > + /* Here huge arguments, INF and NaNs are filtered out to > + callout. */ > + vpandd LOCAL_DATA_UNALIGNED(_iExpMantMask_UISA)(%rip){1to16}, %zmm0, %zmm1 > + vpsubd LOCAL_DATA_UNALIGNED(_iMinIdxOfsMask_UISA)(%rip){1to16}, %zmm1, %zmm2 > > /* Selection arguments between [0, 0x03e00000] into zmm3. */ > vpxord %zmm3, %zmm3, %zmm3 > vpmaxsd %zmm3, %zmm2, %zmm3 > - vpminsd TANHF_DATA_UNALIGNED(_iMaxIdxMask_UISA)(%rip){1to16}, %zmm3, %zmm3 > + vpminsd LOCAL_DATA_UNALIGNED(_iMaxIdxMask_UISA)(%rip){1to16}, %zmm3, %zmm3 > > /* Setup permute indices in zmm3. */ > vpsrld $21, %zmm3, %zmm3 > > /* Store if there are any special cases in k1. */ > - vpcmpd $6, TANHF_DATA_UNALIGNED(_iExpMask)(%rip){1to16}, %zmm1, %k1 > + vpcmpd $6, LOCAL_DATA_UNALIGNED(_iExpMask)(%rip){1to16}, %zmm1, %k1 > > - vmovaps TANHF_DATA(_sC_lo)(%rip), %zmm5 > - vpermt2ps TANHF_DATA(_sC_hi)(%rip), %zmm3, %zmm5 > + vmovaps LOCAL_DATA(_sC_lo)(%rip), %zmm5 > + vpermt2ps LOCAL_DATA(_sC_hi)(%rip), %zmm3, %zmm5 > > - vmovaps TANHF_DATA(_sP7_lo)(%rip), %zmm2 > - vpermt2ps TANHF_DATA(_sP7_hi)(%rip), %zmm3, %zmm2 > + vmovaps LOCAL_DATA(_sP7_lo)(%rip), %zmm2 > + vpermt2ps LOCAL_DATA(_sP7_hi)(%rip), %zmm3, %zmm2 > > /* Store absolute values of inputs in zmm1. */ > - vmovaps TANHF_DATA(_sSignMask)(%rip), %zmm4 > + vmovaps COMMON_DATA(_SignMask)(%rip), %zmm4 > vandnps %zmm0, %zmm4, %zmm1 > vsubps {rn-sae}, %zmm5, %zmm1, %zmm1 > > - vmovaps TANHF_DATA(_sP6_lo)(%rip), %zmm5 > - vpermt2ps TANHF_DATA(_sP6_hi)(%rip), %zmm3, %zmm5 > + vmovaps LOCAL_DATA(_sP6_lo)(%rip), %zmm5 > + vpermt2ps LOCAL_DATA(_sP6_hi)(%rip), %zmm3, %zmm5 > > - vmovaps TANHF_DATA(_sP5_lo)(%rip), %zmm6 > - vpermt2ps TANHF_DATA(_sP5_hi)(%rip), %zmm3, %zmm6 > + vmovaps LOCAL_DATA(_sP5_lo)(%rip), %zmm6 > + vpermt2ps LOCAL_DATA(_sP5_hi)(%rip), %zmm3, %zmm6 > > vfmadd213ps {rn-sae}, %zmm5, %zmm1, %zmm2 > vfmadd213ps {rn-sae}, %zmm6, %zmm1, %zmm2 > > - vmovaps TANHF_DATA(_sP4_lo)(%rip), %zmm7 > - vpermt2ps TANHF_DATA(_sP4_hi)(%rip), %zmm3, %zmm7 > + vmovaps LOCAL_DATA(_sP4_lo)(%rip), %zmm7 > + vpermt2ps LOCAL_DATA(_sP4_hi)(%rip), %zmm3, %zmm7 > > - vmovaps TANHF_DATA(_sP3_lo)(%rip), %zmm8 > - vpermt2ps TANHF_DATA(_sP3_hi)(%rip), %zmm3, %zmm8 > + vmovaps LOCAL_DATA(_sP3_lo)(%rip), %zmm8 > + vpermt2ps LOCAL_DATA(_sP3_hi)(%rip), %zmm3, %zmm8 > > vfmadd213ps {rn-sae}, %zmm7, %zmm1, %zmm2 > vfmadd213ps {rn-sae}, %zmm8, %zmm1, %zmm2 > > - vmovaps TANHF_DATA(_sP2_lo)(%rip), %zmm9 > - vpermt2ps TANHF_DATA(_sP2_hi)(%rip), %zmm3, %zmm9 > + vmovaps LOCAL_DATA(_sP2_lo)(%rip), %zmm9 > + vpermt2ps LOCAL_DATA(_sP2_hi)(%rip), %zmm3, %zmm9 > > - vmovaps TANHF_DATA(_sP0_lo)(%rip), %zmm10 > - vpermt2ps TANHF_DATA(_sP0_hi)(%rip), %zmm3, %zmm10 > + vmovaps LOCAL_DATA(_sP0_lo)(%rip), %zmm10 > + vpermt2ps LOCAL_DATA(_sP0_hi)(%rip), %zmm3, %zmm10 > > vfmadd213ps {rn-sae}, %zmm9, %zmm1, %zmm2 > vfmadd213ps {rn-sae}, %zmm10, %zmm1, %zmm2 > @@ -167,7 +172,7 @@ ENTRY(_ZGVeN16v_tanhf_skx) > > /* Go to special inputs processing branch. */ > jne L(SPECIAL_VALUES_BRANCH) > - # LOE rbx r12 r13 r14 r15 zmm0 zmm2 zmm4 > + > /* Wait until after branch of write over zmm0. */ > vpternlogd $0xec, %zmm4, %zmm2, %zmm0 > > @@ -176,24 +181,24 @@ ENTRY(_ZGVeN16v_tanhf_skx) > > /* Cold case. edx has 1s where there was a special value that > needs to be handled by a tanhf call. Optimize for code size > - more so than speed here. */ > + more so than speed here. */ > L(SPECIAL_VALUES_BRANCH): > - # LOE rbx rdx r12 r13 r14 r15 zmm0 zmm2 zmm4 > - /* Use r13 to save/restore the stack. This allows us to use rbp as > - callee save register saving code size. */ > + > + /* Use r13 to save/restore the stack. This allows us to use rbp > + as callee save register saving code size. */ > pushq %r13 > - cfi_adjust_cfa_offset(8) > - cfi_offset(r13, -16) > - /* Need to callee save registers to preserve state across tanhf calls. > - */ > + cfi_adjust_cfa_offset (8) > + cfi_offset (r13, -16) > + /* Need to callee save registers to preserve state across tanhf > + calls. */ > pushq %rbx > - cfi_adjust_cfa_offset(8) > - cfi_offset(rbx, -24) > + cfi_adjust_cfa_offset (8) > + cfi_offset (rbx, -24) > pushq %rbp > - cfi_adjust_cfa_offset(8) > - cfi_offset(rbp, -32) > + cfi_adjust_cfa_offset (8) > + cfi_offset (rbp, -32) > movq %rsp, %r13 > - cfi_def_cfa_register(r13) > + cfi_def_cfa_register (r13) > > /* Align stack and make room for 2x zmm vectors. */ > andq $-64, %rsp > @@ -207,16 +212,17 @@ L(SPECIAL_VALUES_BRANCH): > > vzeroupper > > - /* edx has 1s where there was a special value that needs to be handled > - by a tanhf call. */ > + /* edx has 1s where there was a special value that needs to be > + handled by a tanhf call. */ > movl %edx, %ebx > L(SPECIAL_VALUES_LOOP): > - # LOE rbx rbp r12 r13 r14 r15 > - /* use rbp as index for special value that is saved across calls to > - tanhf. We technically don't need a callee save register here as offset > - to rsp is always [0, 56] so we can restore rsp by realigning to 64. > - Essentially the tradeoff is 1 extra save/restore vs 2 extra instructions > - in the loop. Realigning also costs more code size. */ > + > + /* use rbp as index for special value that is saved across calls > + to tanhf. We technically don't need a callee save register > + here as offset to rsp is always [0, 56] so we can restore > + rsp by realigning to 64. Essentially the tradeoff is 1 extra > + save/restore vs 2 extra instructions in the loop. Realigning > + also costs more code size. */ > xorl %ebp, %ebp > tzcntl %ebx, %ebp > > @@ -224,203 +230,141 @@ L(SPECIAL_VALUES_LOOP): > vmovss 64(%rsp, %rbp, 4), %xmm0 > call tanhf@PLT > > - /* No good way to avoid the store-forwarding fault this will cause on > - return. `lfence` avoids the SF fault but at greater cost as it > - serialized stack/callee save restoration. */ > + /* No good way to avoid the store-forwarding fault this will > + cause on return. `lfence` avoids the SF fault but at greater > + cost as it serialized stack/callee save restoration. */ > vmovss %xmm0, (%rsp, %rbp, 4) > > - blsrl %ebx, %ebx > + blsrl %ebx, %ebx > jnz L(SPECIAL_VALUES_LOOP) > - # LOE r12 r13 r14 r15 > + > > /* All results have been written to (%rsp). */ > vmovaps (%rsp), %zmm0 > /* Restore rsp. */ > movq %r13, %rsp > - cfi_def_cfa_register(rsp) > + cfi_def_cfa_register (rsp) > /* Restore callee save registers. */ > popq %rbp > - cfi_adjust_cfa_offset(-8) > - cfi_restore(rbp) > + cfi_adjust_cfa_offset (-8) > + cfi_restore (rbp) > popq %rbx > - cfi_adjust_cfa_offset(-8) > - cfi_restore(rbp) > + cfi_adjust_cfa_offset (-8) > + cfi_restore (rbp) > popq %r13 > - cfi_adjust_cfa_offset(-8) > - cfi_restore(r13) > + cfi_adjust_cfa_offset (-8) > + cfi_restore (r13) > ret > END(_ZGVeN16v_tanhf_skx) > > - .section .rodata, "a" > + .section .rodata.evex512, "a" > .align 16 > -#ifdef __svml_stanh_data_internal_typedef > -typedef unsigned int VUINT32; > -typedef struct > - { > - __declspec(align(4)) VUINT32 _iExpMantMask_UISA[1][1]; > - __declspec(align(4)) VUINT32 _iMinIdxOfsMask_UISA[1][1]; > - __declspec(align(4)) VUINT32 _iMaxIdxMask_UISA[1][1]; > - __declspec(align(4)) VUINT32 _iExpMask[1][1]; > - __declspec(align(64)) VUINT32 _sC_lo[16][1]; > - __declspec(align(64)) VUINT32 _sC_hi[16][1]; > - __declspec(align(64)) VUINT32 _sP7_lo[16][1]; > - __declspec(align(64)) VUINT32 _sP7_hi[16][1]; > - __declspec(align(64)) VUINT32 _sSignMask[16][1]; > - __declspec(align(64)) VUINT32 _sP6_lo[16][1]; > - __declspec(align(64)) VUINT32 _sP6_hi[16][1]; > - __declspec(align(64)) VUINT32 _sP5_lo[16][1]; > - __declspec(align(64)) VUINT32 _sP5_hi[16][1]; > - __declspec(align(64)) VUINT32 _sP4_lo[16][1]; > - __declspec(align(64)) VUINT32 _sP4_hi[16][1]; > - __declspec(align(64)) VUINT32 _sP3_lo[16][1]; > - __declspec(align(64)) VUINT32 _sP3_hi[16][1]; > - __declspec(align(64)) VUINT32 _sP2_lo[16][1]; > - __declspec(align(64)) VUINT32 _sP2_hi[16][1]; > - __declspec(align(64)) VUINT32 _sP0_lo[16][1]; > - __declspec(align(64)) VUINT32 _sP0_hi[16][1]; > -} __svml_stanh_data_internal; > -#endif > - > -__svml_stanh_data_internal: > - .align 4 > - /* _iExpMantMask_UISA */ > - .long 0x7fe00000 > - > - .align 4 > - /* _iMinIdxOfsMask_UISA */ > - .long 0x3d400000 > - > - .align 4 > - /* _iMaxIdxMask_UISA */ > - .long 0x03e00000 > - > - .align 4 > - /* _iExpMask */ > - .long 0x7f000000 > - > - .align 64 > -__svml_stanh_data_internal_al64: > - .align 64 > - /* _sC_lo */ > - .long 0x00000000, 0x3d700000, 0x3d900000, 0x3db00000 > - .long 0x3dd00000, 0x3df00000, 0x3e100000, 0x3e300000 > - .long 0x3e500000, 0x3e700000, 0x3e900000, 0x3eb00000 > - .long 0x3ed00000, 0x3ef00000, 0x3f100000, 0x3f300000 > - > - .align 64 > - /* _sC_hi */ > - .long 0x3f500000, 0x3f700000, 0x3f900000, 0x3fb00000 > - .long 0x3fd00000, 0x3ff00000, 0x40100000, 0x40300000 > - .long 0x40500000, 0x40700000, 0x40900000, 0x40b00000 > - .long 0x40d00000, 0x40f00000, 0x41100000, 0x00000000 > - > - .align 64 > - /* _sP7_lo */ > - .long 0xbc0e2f66, 0x460bda12, 0x43d638ef, 0xc3e11c3e > - .long 0xc2baa4e9, 0xc249da2d, 0xc1859b82, 0x40dd5b57 > - .long 0x40494640, 0x40c730a8, 0xbf0f160e, 0x3e30e76f > - .long 0xbea81387, 0xbdb26a1c, 0xbd351e57, 0xbb4c01a0 > - > - .align 64 > - /* _sP7_hi */ > - .long 0x3c1d7bfb, 0x3c722cd1, 0x3c973f1c, 0x3c33a31b > - .long 0x3b862ef4, 0x3a27b3d0, 0xba3b5907, 0xba0efc22 > - .long 0xb97f9f0f, 0xb8c8af50, 0xb7bdddfb, 0xb64f2950 > - .long 0xb4e085b1, 0xb3731dfa, 0xb15a1f04, 0x00000000 > > - .align 64 > - /* _sSignMask */ > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > - > - .align 64 > - /* _sP6_lo */ > - .long 0x3e0910e9, 0x43761143, 0x4165ecdc, 0xc190f756 > - .long 0xc08c097d, 0xc02ba813, 0xbf7f6bda, 0x3f2b1dc0 > - .long 0x3ece105d, 0x3f426a94, 0xbadb0dc4, 0x3da43b17 > - .long 0xbd51ab88, 0xbcaea23d, 0xbd3b6d8d, 0xbd6caaad > - > - .align 64 > - /* _sP6_hi */ > - .long 0xbd795bed, 0xbd5fddda, 0xbd038f3b, 0xbc1cad63 > - .long 0x3abb4766, 0x3b95f10b, 0x3b825873, 0x3afaea66 > - .long 0x3a49f878, 0x39996bf3, 0x388f3e6c, 0x371bb0e3 > - .long 0x35a8a5e6, 0x34369b17, 0x322487b0, 0x00000000 > - > - .align 64 > - /* _sP5_lo */ > - .long 0xb76dd6b9, 0xbe1c276d, 0x3c1dcf2f, 0x3dc1a78d > - .long 0x3d96f985, 0x3da2b61b, 0x3dc13397, 0x3dd2f670 > - .long 0x3df48a0a, 0x3e06c5a8, 0x3e1a3aba, 0x3e27c405 > - .long 0x3e2e78d0, 0x3e2c3e44, 0x3e1d3097, 0x3df4a8f4 > - > - .align 64 > - /* _sP5_hi */ > - .long 0x3da38508, 0x3d31416a, 0x3b562657, 0xbcaeeac9 > - .long 0xbcce9419, 0xbcaaeac4, 0xbc49e7d0, 0xbba71ddd > - .long 0xbb003b0e, 0xba3f9a05, 0xb92c08a7, 0xb7ba9232 > - .long 0xb64a0b0f, 0xb4dac169, 0xb2ab78ac, 0x00000000 > - > - .align 64 > - /* _sP4_lo */ > - .long 0xbeaaaaa5, 0xbeab0612, 0xbea7f01f, 0xbea4e120 > - .long 0xbea387b7, 0xbea15962, 0xbe9d57f7, 0xbe976b5a > - .long 0xbe90230d, 0xbe880dff, 0xbe7479b3, 0xbe4c3d88 > - .long 0xbe212482, 0xbdeb8cba, 0xbd5e78ad, 0x3c6b5e6e > - > - .align 64 > - /* _sP4_hi */ > - .long 0x3d839143, 0x3dc21ee1, 0x3de347af, 0x3dcbec96 > - .long 0x3d99ef2d, 0x3d542ea1, 0x3cdde701, 0x3c2cca67 > - .long 0x3b81cb27, 0x3ac073a1, 0x39ac3032, 0x383a94d9 > - .long 0x36ca081d, 0x355abd4c, 0x332b3cb6, 0x00000000 > - > - .align 64 > - /* _sP3_lo */ > - .long 0xb0343c7b, 0xbd6ee69d, 0xbd8f0da7, 0xbdae477d > - .long 0xbdcd2a1f, 0xbdeba80d, 0xbe0c443b, 0xbe293cf3 > - .long 0xbe44f282, 0xbe5f3651, 0xbe81c7c0, 0xbe96d7ca > - .long 0xbea7fb8e, 0xbeb50e9e, 0xbec12efe, 0xbec4be92 > - > - .align 64 > - /* _sP3_hi */ > - .long 0xbebce070, 0xbead510e, 0xbe8ef7d6, 0xbe4b8704 > - .long 0xbe083237, 0xbdaf7449, 0xbd2e1ec4, 0xbc83bf06 > - .long 0xbbc3e0b5, 0xbb10aadc, 0xba0157db, 0xb88c18f2 > - .long 0xb717b096, 0xb5a43bae, 0xb383012c, 0x00000000 > - > - .align 64 > - /* _sP2_lo */ > - .long 0x3f800000, 0x3f7f1f84, 0x3f7ebd11, 0x3f7e1e5f > - .long 0x3f7d609f, 0x3f7c842d, 0x3f7b00e5, 0x3f789580 > - .long 0x3f75b8ad, 0x3f726fd9, 0x3f6cc59b, 0x3f63fb92 > - .long 0x3f59ff97, 0x3f4f11d7, 0x3f3d7573, 0x3f24f360 > - > - .align 64 > - /* _sP2_hi */ > - .long 0x3f0cbfe7, 0x3eec1a69, 0x3eb0a801, 0x3e6753a2 > - .long 0x3e132f1a, 0x3db7e7d3, 0x3d320845, 0x3c84d3d4 > - .long 0x3bc477b7, 0x3b10d3da, 0x3a01601e, 0x388c1a3b > - .long 0x3717b0da, 0x35a43bce, 0x338306c6, 0x00000000 > - > - .align 64 > - /* _sP0_lo */ > - .long 0x00000000, 0x3d6fb9c9, 0x3d8fc35f, 0x3daf9169 > - .long 0x3dcf49ab, 0x3deee849, 0x3e0f0ee8, 0x3e2e4984 > - .long 0x3e4d2f8e, 0x3e6bb32e, 0x3e8c51cd, 0x3ea96163 > - .long 0x3ec543f1, 0x3edfd735, 0x3f028438, 0x3f18abf0 > - > - .align 64 > - /* _sP0_hi */ > - .long 0x3f2bc480, 0x3f3bec1c, 0x3f4f2e5b, 0x3f613c53 > - .long 0x3f6ce37d, 0x3f743c4f, 0x3f7a5feb, 0x3f7dea85 > - .long 0x3f7f3b3d, 0x3f7fb78c, 0x3f7fefd4, 0x3f7ffdd0 > - .long 0x3f7fffb4, 0x3f7ffff6, 0x3f7fffff, 0x3f800000 > +LOCAL_DATA_NAME_UNALIGNED: > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iExpMantMask_UISA, 0x7fe00000) > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iMinIdxOfsMask_UISA, 0x3d400000) > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iMaxIdxMask_UISA, 0x03e00000) > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iExpMask, 0x7f000000) > + .type LOCAL_DATA_NAME_UNALIGNED, @object > + .size LOCAL_DATA_NAME_UNALIGNED, .-LOCAL_DATA_NAME_UNALIGNED > > .align 64 > - .type __svml_stanh_data_internal_al64, @object > - .size __svml_stanh_data_internal_al64, .-__svml_stanh_data_internal_al64 > - .type __svml_stanh_data_internal, @object > - .size __svml_stanh_data_internal, .-__svml_stanh_data_internal > +LOCAL_DATA_NAME: > + float_block (LOCAL_DATA_NAME, _sC_lo, > + 0x00000000, 0x3d700000, 0x3d900000, 0x3db00000, > + 0x3dd00000, 0x3df00000, 0x3e100000, 0x3e300000, > + 0x3e500000, 0x3e700000, 0x3e900000, 0x3eb00000, > + 0x3ed00000, 0x3ef00000, 0x3f100000, 0x3f300000) > + > + float_block (LOCAL_DATA_NAME, _sC_hi, > + 0x3f500000, 0x3f700000, 0x3f900000, 0x3fb00000, > + 0x3fd00000, 0x3ff00000, 0x40100000, 0x40300000, > + 0x40500000, 0x40700000, 0x40900000, 0x40b00000, > + 0x40d00000, 0x40f00000, 0x41100000, 0x00000000) > + > + float_block (LOCAL_DATA_NAME, _sP7_lo, > + 0xbc0e2f66, 0x460bda12, 0x43d638ef, 0xc3e11c3e, > + 0xc2baa4e9, 0xc249da2d, 0xc1859b82, 0x40dd5b57, > + 0x40494640, 0x40c730a8, 0xbf0f160e, 0x3e30e76f, > + 0xbea81387, 0xbdb26a1c, 0xbd351e57, 0xbb4c01a0) > + > + float_block (LOCAL_DATA_NAME, _sP7_hi, > + 0x3c1d7bfb, 0x3c722cd1, 0x3c973f1c, 0x3c33a31b, > + 0x3b862ef4, 0x3a27b3d0, 0xba3b5907, 0xba0efc22, > + 0xb97f9f0f, 0xb8c8af50, 0xb7bdddfb, 0xb64f2950, > + 0xb4e085b1, 0xb3731dfa, 0xb15a1f04, 0x00000000) > + > + float_block (LOCAL_DATA_NAME, _sP6_lo, > + 0x3e0910e9, 0x43761143, 0x4165ecdc, 0xc190f756, > + 0xc08c097d, 0xc02ba813, 0xbf7f6bda, 0x3f2b1dc0, > + 0x3ece105d, 0x3f426a94, 0xbadb0dc4, 0x3da43b17, > + 0xbd51ab88, 0xbcaea23d, 0xbd3b6d8d, 0xbd6caaad) > + > + float_block (LOCAL_DATA_NAME, _sP6_hi, > + 0xbd795bed, 0xbd5fddda, 0xbd038f3b, 0xbc1cad63, > + 0x3abb4766, 0x3b95f10b, 0x3b825873, 0x3afaea66, > + 0x3a49f878, 0x39996bf3, 0x388f3e6c, 0x371bb0e3, > + 0x35a8a5e6, 0x34369b17, 0x322487b0, 0x00000000) > + > + float_block (LOCAL_DATA_NAME, _sP5_lo, > + 0xb76dd6b9, 0xbe1c276d, 0x3c1dcf2f, 0x3dc1a78d, > + 0x3d96f985, 0x3da2b61b, 0x3dc13397, 0x3dd2f670, > + 0x3df48a0a, 0x3e06c5a8, 0x3e1a3aba, 0x3e27c405, > + 0x3e2e78d0, 0x3e2c3e44, 0x3e1d3097, 0x3df4a8f4) > + > + float_block (LOCAL_DATA_NAME, _sP5_hi, > + 0x3da38508, 0x3d31416a, 0x3b562657, 0xbcaeeac9, > + 0xbcce9419, 0xbcaaeac4, 0xbc49e7d0, 0xbba71ddd, > + 0xbb003b0e, 0xba3f9a05, 0xb92c08a7, 0xb7ba9232, > + 0xb64a0b0f, 0xb4dac169, 0xb2ab78ac, 0x00000000) > + > + float_block (LOCAL_DATA_NAME, _sP4_lo, > + 0xbeaaaaa5, 0xbeab0612, 0xbea7f01f, 0xbea4e120, > + 0xbea387b7, 0xbea15962, 0xbe9d57f7, 0xbe976b5a, > + 0xbe90230d, 0xbe880dff, 0xbe7479b3, 0xbe4c3d88, > + 0xbe212482, 0xbdeb8cba, 0xbd5e78ad, 0x3c6b5e6e) > + > + float_block (LOCAL_DATA_NAME, _sP4_hi, > + 0x3d839143, 0x3dc21ee1, 0x3de347af, 0x3dcbec96, > + 0x3d99ef2d, 0x3d542ea1, 0x3cdde701, 0x3c2cca67, > + 0x3b81cb27, 0x3ac073a1, 0x39ac3032, 0x383a94d9, > + 0x36ca081d, 0x355abd4c, 0x332b3cb6, 0x00000000) > + > + float_block (LOCAL_DATA_NAME, _sP3_lo, > + 0xb0343c7b, 0xbd6ee69d, 0xbd8f0da7, 0xbdae477d, > + 0xbdcd2a1f, 0xbdeba80d, 0xbe0c443b, 0xbe293cf3, > + 0xbe44f282, 0xbe5f3651, 0xbe81c7c0, 0xbe96d7ca, > + 0xbea7fb8e, 0xbeb50e9e, 0xbec12efe, 0xbec4be92) > + > + float_block (LOCAL_DATA_NAME, _sP3_hi, > + 0xbebce070, 0xbead510e, 0xbe8ef7d6, 0xbe4b8704, > + 0xbe083237, 0xbdaf7449, 0xbd2e1ec4, 0xbc83bf06, > + 0xbbc3e0b5, 0xbb10aadc, 0xba0157db, 0xb88c18f2, > + 0xb717b096, 0xb5a43bae, 0xb383012c, 0x00000000) > + > + float_block (LOCAL_DATA_NAME, _sP2_lo, > + 0x3f800000, 0x3f7f1f84, 0x3f7ebd11, 0x3f7e1e5f, > + 0x3f7d609f, 0x3f7c842d, 0x3f7b00e5, 0x3f789580, > + 0x3f75b8ad, 0x3f726fd9, 0x3f6cc59b, 0x3f63fb92, > + 0x3f59ff97, 0x3f4f11d7, 0x3f3d7573, 0x3f24f360) > + > + float_block (LOCAL_DATA_NAME, _sP2_hi, > + 0x3f0cbfe7, 0x3eec1a69, 0x3eb0a801, 0x3e6753a2, > + 0x3e132f1a, 0x3db7e7d3, 0x3d320845, 0x3c84d3d4, > + 0x3bc477b7, 0x3b10d3da, 0x3a01601e, 0x388c1a3b, > + 0x3717b0da, 0x35a43bce, 0x338306c6, 0x00000000) > + > + float_block (LOCAL_DATA_NAME, _sP0_lo, > + 0x00000000, 0x3d6fb9c9, 0x3d8fc35f, 0x3daf9169, > + 0x3dcf49ab, 0x3deee849, 0x3e0f0ee8, 0x3e2e4984, > + 0x3e4d2f8e, 0x3e6bb32e, 0x3e8c51cd, 0x3ea96163, > + 0x3ec543f1, 0x3edfd735, 0x3f028438, 0x3f18abf0) > + > + float_block (LOCAL_DATA_NAME, _sP0_hi, > + 0x3f2bc480, 0x3f3bec1c, 0x3f4f2e5b, 0x3f613c53, > + 0x3f6ce37d, 0x3f743c4f, 0x3f7a5feb, 0x3f7dea85, > + 0x3f7f3b3d, 0x3f7fb78c, 0x3f7fefd4, 0x3f7ffdd0, > + 0x3f7fffb4, 0x3f7ffff6, 0x3f7fffff, 0x3f800000) > + > + .type LOCAL_DATA_NAME, @object > + .size LOCAL_DATA_NAME, .-LOCAL_DATA_NAME > -- > 2.34.1 > The data movement makes the assembler codes much harder to follow. Sunil, what do you think of this patch series?
On Fri, Dec 16, 2022 at 9:06 AM H.J. Lu <hjl.tools@gmail.com> wrote: > > On Wed, Dec 7, 2022 at 12:52 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote: > > > > No changes to the logic, just change how rodata is handled. > > > > 1. Define the rodatas using the new macros so they check that the > > offset is correct. > > > > 2. Use common data where applicable. > > --- > > .../multiarch/svml_s_tanhf16_core_avx512.S | 450 ++++++++---------- > > 1 file changed, 197 insertions(+), 253 deletions(-) > > > > diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S b/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > > index d74fc7731d..765e9ed7f7 100644 > > --- a/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > > +++ b/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > > @@ -70,94 +70,99 @@ > > * > > */ > > > > -/* Offsets for data table __svml_stanh_data_internal_avx512. Ordered > > - by use in the function. On cold-starts this might help the > > - prefetcher. Possibly a better idea is to interleave start/end so > > - that the prefetcher is less likely to detect a stream and pull > > - irrelivant lines into cache. */ > > > > -/* Offsets for data table __svml_stanh_data_internal. 4 bytes each. > > - */ > > + > > +#define LOCAL_DATA_NAME __svml_stanh_data_internal > > +#define LOCAL_DATA_NAME_UNALIGNED __svml_stanh_data_internal_unaligned > > +#include "svml_s_common_evex512_rodata_offsets.h" > > + > > +/* Offsets for data table __svml_stanh_data_internal_unaligned. > > + 4 bytes each. */ > > #define _iExpMantMask_UISA 0 > > #define _iMinIdxOfsMask_UISA 4 > > #define _iMaxIdxMask_UISA 8 > > #define _iExpMask 12 > > > > -/* Offsets for data table __svml_stanh_data_internal_al64. 64 bytes > > - each. */ > > -#define _sC_lo 0 > > -#define _sC_hi 64 > > -#define _sP7_lo 128 > > -#define _sP7_hi 192 > > -#define _sSignMask 256 > > -#define _sP6_lo 320 > > -#define _sP6_hi 384 > > -#define _sP5_lo 448 > > -#define _sP5_hi 512 > > -#define _sP4_lo 576 > > -#define _sP4_hi 640 > > -#define _sP3_lo 704 > > -#define _sP3_hi 768 > > -#define _sP2_lo 832 > > -#define _sP2_hi 896 > > -#define _sP0_lo 960 > > -#define _sP0_hi 1024 > > +/* Offsets for data table __svml_stanh_data_internal. Ordered > > + by use in the function. On cold-starts this might help the > > + prefetcher. Possibly a better idea is to interleave start/end so > > + that the prefetcher is less likely to detect a stream and pull > > + irrelivant lines into cache. */ > > + > > +/* Offsets for data table __svml_stanh_data_internal. > > + 64 bytes each. */ > > +#define _sC_lo 0 > > +#define _sC_hi 64 > > +#define _sP7_lo 128 > > +#define _sP7_hi 192 > > +#define _sP6_lo 256 > > +#define _sP6_hi 320 > > +#define _sP5_lo 384 > > +#define _sP5_hi 448 > > +#define _sP4_lo 512 > > +#define _sP4_hi 576 > > +#define _sP3_lo 640 > > +#define _sP3_hi 704 > > +#define _sP2_lo 768 > > +#define _sP2_hi 832 > > +#define _sP0_lo 896 > > +#define _sP0_hi 960 > > + > > > > #include <sysdep.h> > > -#define TANHF_DATA(x) ((x)+__svml_stanh_data_internal_al64) > > -#define TANHF_DATA_UNALIGNED(x) ((x)+__svml_stanh_data_internal) > > > > .section .text.evex512, "ax", @progbits > > ENTRY(_ZGVeN16v_tanhf_skx) > > - /* Here huge arguments, INF and NaNs are filtered out to callout. */ > > - vpandd TANHF_DATA_UNALIGNED(_iExpMantMask_UISA)(%rip){1to16}, %zmm0, %zmm1 > > - vpsubd TANHF_DATA_UNALIGNED(_iMinIdxOfsMask_UISA)(%rip){1to16}, %zmm1, %zmm2 > > + /* Here huge arguments, INF and NaNs are filtered out to > > + callout. */ > > + vpandd LOCAL_DATA_UNALIGNED(_iExpMantMask_UISA)(%rip){1to16}, %zmm0, %zmm1 > > + vpsubd LOCAL_DATA_UNALIGNED(_iMinIdxOfsMask_UISA)(%rip){1to16}, %zmm1, %zmm2 > > > > /* Selection arguments between [0, 0x03e00000] into zmm3. */ > > vpxord %zmm3, %zmm3, %zmm3 > > vpmaxsd %zmm3, %zmm2, %zmm3 > > - vpminsd TANHF_DATA_UNALIGNED(_iMaxIdxMask_UISA)(%rip){1to16}, %zmm3, %zmm3 > > + vpminsd LOCAL_DATA_UNALIGNED(_iMaxIdxMask_UISA)(%rip){1to16}, %zmm3, %zmm3 > > > > /* Setup permute indices in zmm3. */ > > vpsrld $21, %zmm3, %zmm3 > > > > /* Store if there are any special cases in k1. */ > > - vpcmpd $6, TANHF_DATA_UNALIGNED(_iExpMask)(%rip){1to16}, %zmm1, %k1 > > + vpcmpd $6, LOCAL_DATA_UNALIGNED(_iExpMask)(%rip){1to16}, %zmm1, %k1 > > > > - vmovaps TANHF_DATA(_sC_lo)(%rip), %zmm5 > > - vpermt2ps TANHF_DATA(_sC_hi)(%rip), %zmm3, %zmm5 > > + vmovaps LOCAL_DATA(_sC_lo)(%rip), %zmm5 > > + vpermt2ps LOCAL_DATA(_sC_hi)(%rip), %zmm3, %zmm5 > > > > - vmovaps TANHF_DATA(_sP7_lo)(%rip), %zmm2 > > - vpermt2ps TANHF_DATA(_sP7_hi)(%rip), %zmm3, %zmm2 > > + vmovaps LOCAL_DATA(_sP7_lo)(%rip), %zmm2 > > + vpermt2ps LOCAL_DATA(_sP7_hi)(%rip), %zmm3, %zmm2 > > > > /* Store absolute values of inputs in zmm1. */ > > - vmovaps TANHF_DATA(_sSignMask)(%rip), %zmm4 > > + vmovaps COMMON_DATA(_SignMask)(%rip), %zmm4 > > vandnps %zmm0, %zmm4, %zmm1 > > vsubps {rn-sae}, %zmm5, %zmm1, %zmm1 > > > > - vmovaps TANHF_DATA(_sP6_lo)(%rip), %zmm5 > > - vpermt2ps TANHF_DATA(_sP6_hi)(%rip), %zmm3, %zmm5 > > + vmovaps LOCAL_DATA(_sP6_lo)(%rip), %zmm5 > > + vpermt2ps LOCAL_DATA(_sP6_hi)(%rip), %zmm3, %zmm5 > > > > - vmovaps TANHF_DATA(_sP5_lo)(%rip), %zmm6 > > - vpermt2ps TANHF_DATA(_sP5_hi)(%rip), %zmm3, %zmm6 > > + vmovaps LOCAL_DATA(_sP5_lo)(%rip), %zmm6 > > + vpermt2ps LOCAL_DATA(_sP5_hi)(%rip), %zmm3, %zmm6 > > > > vfmadd213ps {rn-sae}, %zmm5, %zmm1, %zmm2 > > vfmadd213ps {rn-sae}, %zmm6, %zmm1, %zmm2 > > > > - vmovaps TANHF_DATA(_sP4_lo)(%rip), %zmm7 > > - vpermt2ps TANHF_DATA(_sP4_hi)(%rip), %zmm3, %zmm7 > > + vmovaps LOCAL_DATA(_sP4_lo)(%rip), %zmm7 > > + vpermt2ps LOCAL_DATA(_sP4_hi)(%rip), %zmm3, %zmm7 > > > > - vmovaps TANHF_DATA(_sP3_lo)(%rip), %zmm8 > > - vpermt2ps TANHF_DATA(_sP3_hi)(%rip), %zmm3, %zmm8 > > + vmovaps LOCAL_DATA(_sP3_lo)(%rip), %zmm8 > > + vpermt2ps LOCAL_DATA(_sP3_hi)(%rip), %zmm3, %zmm8 > > > > vfmadd213ps {rn-sae}, %zmm7, %zmm1, %zmm2 > > vfmadd213ps {rn-sae}, %zmm8, %zmm1, %zmm2 > > > > - vmovaps TANHF_DATA(_sP2_lo)(%rip), %zmm9 > > - vpermt2ps TANHF_DATA(_sP2_hi)(%rip), %zmm3, %zmm9 > > + vmovaps LOCAL_DATA(_sP2_lo)(%rip), %zmm9 > > + vpermt2ps LOCAL_DATA(_sP2_hi)(%rip), %zmm3, %zmm9 > > > > - vmovaps TANHF_DATA(_sP0_lo)(%rip), %zmm10 > > - vpermt2ps TANHF_DATA(_sP0_hi)(%rip), %zmm3, %zmm10 > > + vmovaps LOCAL_DATA(_sP0_lo)(%rip), %zmm10 > > + vpermt2ps LOCAL_DATA(_sP0_hi)(%rip), %zmm3, %zmm10 > > > > vfmadd213ps {rn-sae}, %zmm9, %zmm1, %zmm2 > > vfmadd213ps {rn-sae}, %zmm10, %zmm1, %zmm2 > > @@ -167,7 +172,7 @@ ENTRY(_ZGVeN16v_tanhf_skx) > > > > /* Go to special inputs processing branch. */ > > jne L(SPECIAL_VALUES_BRANCH) > > - # LOE rbx r12 r13 r14 r15 zmm0 zmm2 zmm4 > > + > > /* Wait until after branch of write over zmm0. */ > > vpternlogd $0xec, %zmm4, %zmm2, %zmm0 > > > > @@ -176,24 +181,24 @@ ENTRY(_ZGVeN16v_tanhf_skx) > > > > /* Cold case. edx has 1s where there was a special value that > > needs to be handled by a tanhf call. Optimize for code size > > - more so than speed here. */ > > + more so than speed here. */ > > L(SPECIAL_VALUES_BRANCH): > > - # LOE rbx rdx r12 r13 r14 r15 zmm0 zmm2 zmm4 > > - /* Use r13 to save/restore the stack. This allows us to use rbp as > > - callee save register saving code size. */ > > + > > + /* Use r13 to save/restore the stack. This allows us to use rbp > > + as callee save register saving code size. */ > > pushq %r13 > > - cfi_adjust_cfa_offset(8) > > - cfi_offset(r13, -16) > > - /* Need to callee save registers to preserve state across tanhf calls. > > - */ > > + cfi_adjust_cfa_offset (8) > > + cfi_offset (r13, -16) > > + /* Need to callee save registers to preserve state across tanhf > > + calls. */ > > pushq %rbx > > - cfi_adjust_cfa_offset(8) > > - cfi_offset(rbx, -24) > > + cfi_adjust_cfa_offset (8) > > + cfi_offset (rbx, -24) > > pushq %rbp > > - cfi_adjust_cfa_offset(8) > > - cfi_offset(rbp, -32) > > + cfi_adjust_cfa_offset (8) > > + cfi_offset (rbp, -32) > > movq %rsp, %r13 > > - cfi_def_cfa_register(r13) > > + cfi_def_cfa_register (r13) > > > > /* Align stack and make room for 2x zmm vectors. */ > > andq $-64, %rsp > > @@ -207,16 +212,17 @@ L(SPECIAL_VALUES_BRANCH): > > > > vzeroupper > > > > - /* edx has 1s where there was a special value that needs to be handled > > - by a tanhf call. */ > > + /* edx has 1s where there was a special value that needs to be > > + handled by a tanhf call. */ > > movl %edx, %ebx > > L(SPECIAL_VALUES_LOOP): > > - # LOE rbx rbp r12 r13 r14 r15 > > - /* use rbp as index for special value that is saved across calls to > > - tanhf. We technically don't need a callee save register here as offset > > - to rsp is always [0, 56] so we can restore rsp by realigning to 64. > > - Essentially the tradeoff is 1 extra save/restore vs 2 extra instructions > > - in the loop. Realigning also costs more code size. */ > > + > > + /* use rbp as index for special value that is saved across calls > > + to tanhf. We technically don't need a callee save register > > + here as offset to rsp is always [0, 56] so we can restore > > + rsp by realigning to 64. Essentially the tradeoff is 1 extra > > + save/restore vs 2 extra instructions in the loop. Realigning > > + also costs more code size. */ > > xorl %ebp, %ebp > > tzcntl %ebx, %ebp > > > > @@ -224,203 +230,141 @@ L(SPECIAL_VALUES_LOOP): > > vmovss 64(%rsp, %rbp, 4), %xmm0 > > call tanhf@PLT > > > > - /* No good way to avoid the store-forwarding fault this will cause on > > - return. `lfence` avoids the SF fault but at greater cost as it > > - serialized stack/callee save restoration. */ > > + /* No good way to avoid the store-forwarding fault this will > > + cause on return. `lfence` avoids the SF fault but at greater > > + cost as it serialized stack/callee save restoration. */ > > vmovss %xmm0, (%rsp, %rbp, 4) > > > > - blsrl %ebx, %ebx > > + blsrl %ebx, %ebx > > jnz L(SPECIAL_VALUES_LOOP) > > - # LOE r12 r13 r14 r15 > > + > > > > /* All results have been written to (%rsp). */ > > vmovaps (%rsp), %zmm0 > > /* Restore rsp. */ > > movq %r13, %rsp > > - cfi_def_cfa_register(rsp) > > + cfi_def_cfa_register (rsp) > > /* Restore callee save registers. */ > > popq %rbp > > - cfi_adjust_cfa_offset(-8) > > - cfi_restore(rbp) > > + cfi_adjust_cfa_offset (-8) > > + cfi_restore (rbp) > > popq %rbx > > - cfi_adjust_cfa_offset(-8) > > - cfi_restore(rbp) > > + cfi_adjust_cfa_offset (-8) > > + cfi_restore (rbp) > > popq %r13 > > - cfi_adjust_cfa_offset(-8) > > - cfi_restore(r13) > > + cfi_adjust_cfa_offset (-8) > > + cfi_restore (r13) > > ret > > END(_ZGVeN16v_tanhf_skx) > > > > - .section .rodata, "a" > > + .section .rodata.evex512, "a" > > .align 16 > > -#ifdef __svml_stanh_data_internal_typedef > > -typedef unsigned int VUINT32; > > -typedef struct > > - { > > - __declspec(align(4)) VUINT32 _iExpMantMask_UISA[1][1]; > > - __declspec(align(4)) VUINT32 _iMinIdxOfsMask_UISA[1][1]; > > - __declspec(align(4)) VUINT32 _iMaxIdxMask_UISA[1][1]; > > - __declspec(align(4)) VUINT32 _iExpMask[1][1]; > > - __declspec(align(64)) VUINT32 _sC_lo[16][1]; > > - __declspec(align(64)) VUINT32 _sC_hi[16][1]; > > - __declspec(align(64)) VUINT32 _sP7_lo[16][1]; > > - __declspec(align(64)) VUINT32 _sP7_hi[16][1]; > > - __declspec(align(64)) VUINT32 _sSignMask[16][1]; > > - __declspec(align(64)) VUINT32 _sP6_lo[16][1]; > > - __declspec(align(64)) VUINT32 _sP6_hi[16][1]; > > - __declspec(align(64)) VUINT32 _sP5_lo[16][1]; > > - __declspec(align(64)) VUINT32 _sP5_hi[16][1]; > > - __declspec(align(64)) VUINT32 _sP4_lo[16][1]; > > - __declspec(align(64)) VUINT32 _sP4_hi[16][1]; > > - __declspec(align(64)) VUINT32 _sP3_lo[16][1]; > > - __declspec(align(64)) VUINT32 _sP3_hi[16][1]; > > - __declspec(align(64)) VUINT32 _sP2_lo[16][1]; > > - __declspec(align(64)) VUINT32 _sP2_hi[16][1]; > > - __declspec(align(64)) VUINT32 _sP0_lo[16][1]; > > - __declspec(align(64)) VUINT32 _sP0_hi[16][1]; > > -} __svml_stanh_data_internal; > > -#endif > > - > > -__svml_stanh_data_internal: > > - .align 4 > > - /* _iExpMantMask_UISA */ > > - .long 0x7fe00000 > > - > > - .align 4 > > - /* _iMinIdxOfsMask_UISA */ > > - .long 0x3d400000 > > - > > - .align 4 > > - /* _iMaxIdxMask_UISA */ > > - .long 0x03e00000 > > - > > - .align 4 > > - /* _iExpMask */ > > - .long 0x7f000000 > > - > > - .align 64 > > -__svml_stanh_data_internal_al64: > > - .align 64 > > - /* _sC_lo */ > > - .long 0x00000000, 0x3d700000, 0x3d900000, 0x3db00000 > > - .long 0x3dd00000, 0x3df00000, 0x3e100000, 0x3e300000 > > - .long 0x3e500000, 0x3e700000, 0x3e900000, 0x3eb00000 > > - .long 0x3ed00000, 0x3ef00000, 0x3f100000, 0x3f300000 > > - > > - .align 64 > > - /* _sC_hi */ > > - .long 0x3f500000, 0x3f700000, 0x3f900000, 0x3fb00000 > > - .long 0x3fd00000, 0x3ff00000, 0x40100000, 0x40300000 > > - .long 0x40500000, 0x40700000, 0x40900000, 0x40b00000 > > - .long 0x40d00000, 0x40f00000, 0x41100000, 0x00000000 > > - > > - .align 64 > > - /* _sP7_lo */ > > - .long 0xbc0e2f66, 0x460bda12, 0x43d638ef, 0xc3e11c3e > > - .long 0xc2baa4e9, 0xc249da2d, 0xc1859b82, 0x40dd5b57 > > - .long 0x40494640, 0x40c730a8, 0xbf0f160e, 0x3e30e76f > > - .long 0xbea81387, 0xbdb26a1c, 0xbd351e57, 0xbb4c01a0 > > - > > - .align 64 > > - /* _sP7_hi */ > > - .long 0x3c1d7bfb, 0x3c722cd1, 0x3c973f1c, 0x3c33a31b > > - .long 0x3b862ef4, 0x3a27b3d0, 0xba3b5907, 0xba0efc22 > > - .long 0xb97f9f0f, 0xb8c8af50, 0xb7bdddfb, 0xb64f2950 > > - .long 0xb4e085b1, 0xb3731dfa, 0xb15a1f04, 0x00000000 > > > > - .align 64 > > - /* _sSignMask */ > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > - > > - .align 64 > > - /* _sP6_lo */ > > - .long 0x3e0910e9, 0x43761143, 0x4165ecdc, 0xc190f756 > > - .long 0xc08c097d, 0xc02ba813, 0xbf7f6bda, 0x3f2b1dc0 > > - .long 0x3ece105d, 0x3f426a94, 0xbadb0dc4, 0x3da43b17 > > - .long 0xbd51ab88, 0xbcaea23d, 0xbd3b6d8d, 0xbd6caaad > > - > > - .align 64 > > - /* _sP6_hi */ > > - .long 0xbd795bed, 0xbd5fddda, 0xbd038f3b, 0xbc1cad63 > > - .long 0x3abb4766, 0x3b95f10b, 0x3b825873, 0x3afaea66 > > - .long 0x3a49f878, 0x39996bf3, 0x388f3e6c, 0x371bb0e3 > > - .long 0x35a8a5e6, 0x34369b17, 0x322487b0, 0x00000000 > > - > > - .align 64 > > - /* _sP5_lo */ > > - .long 0xb76dd6b9, 0xbe1c276d, 0x3c1dcf2f, 0x3dc1a78d > > - .long 0x3d96f985, 0x3da2b61b, 0x3dc13397, 0x3dd2f670 > > - .long 0x3df48a0a, 0x3e06c5a8, 0x3e1a3aba, 0x3e27c405 > > - .long 0x3e2e78d0, 0x3e2c3e44, 0x3e1d3097, 0x3df4a8f4 > > - > > - .align 64 > > - /* _sP5_hi */ > > - .long 0x3da38508, 0x3d31416a, 0x3b562657, 0xbcaeeac9 > > - .long 0xbcce9419, 0xbcaaeac4, 0xbc49e7d0, 0xbba71ddd > > - .long 0xbb003b0e, 0xba3f9a05, 0xb92c08a7, 0xb7ba9232 > > - .long 0xb64a0b0f, 0xb4dac169, 0xb2ab78ac, 0x00000000 > > - > > - .align 64 > > - /* _sP4_lo */ > > - .long 0xbeaaaaa5, 0xbeab0612, 0xbea7f01f, 0xbea4e120 > > - .long 0xbea387b7, 0xbea15962, 0xbe9d57f7, 0xbe976b5a > > - .long 0xbe90230d, 0xbe880dff, 0xbe7479b3, 0xbe4c3d88 > > - .long 0xbe212482, 0xbdeb8cba, 0xbd5e78ad, 0x3c6b5e6e > > - > > - .align 64 > > - /* _sP4_hi */ > > - .long 0x3d839143, 0x3dc21ee1, 0x3de347af, 0x3dcbec96 > > - .long 0x3d99ef2d, 0x3d542ea1, 0x3cdde701, 0x3c2cca67 > > - .long 0x3b81cb27, 0x3ac073a1, 0x39ac3032, 0x383a94d9 > > - .long 0x36ca081d, 0x355abd4c, 0x332b3cb6, 0x00000000 > > - > > - .align 64 > > - /* _sP3_lo */ > > - .long 0xb0343c7b, 0xbd6ee69d, 0xbd8f0da7, 0xbdae477d > > - .long 0xbdcd2a1f, 0xbdeba80d, 0xbe0c443b, 0xbe293cf3 > > - .long 0xbe44f282, 0xbe5f3651, 0xbe81c7c0, 0xbe96d7ca > > - .long 0xbea7fb8e, 0xbeb50e9e, 0xbec12efe, 0xbec4be92 > > - > > - .align 64 > > - /* _sP3_hi */ > > - .long 0xbebce070, 0xbead510e, 0xbe8ef7d6, 0xbe4b8704 > > - .long 0xbe083237, 0xbdaf7449, 0xbd2e1ec4, 0xbc83bf06 > > - .long 0xbbc3e0b5, 0xbb10aadc, 0xba0157db, 0xb88c18f2 > > - .long 0xb717b096, 0xb5a43bae, 0xb383012c, 0x00000000 > > - > > - .align 64 > > - /* _sP2_lo */ > > - .long 0x3f800000, 0x3f7f1f84, 0x3f7ebd11, 0x3f7e1e5f > > - .long 0x3f7d609f, 0x3f7c842d, 0x3f7b00e5, 0x3f789580 > > - .long 0x3f75b8ad, 0x3f726fd9, 0x3f6cc59b, 0x3f63fb92 > > - .long 0x3f59ff97, 0x3f4f11d7, 0x3f3d7573, 0x3f24f360 > > - > > - .align 64 > > - /* _sP2_hi */ > > - .long 0x3f0cbfe7, 0x3eec1a69, 0x3eb0a801, 0x3e6753a2 > > - .long 0x3e132f1a, 0x3db7e7d3, 0x3d320845, 0x3c84d3d4 > > - .long 0x3bc477b7, 0x3b10d3da, 0x3a01601e, 0x388c1a3b > > - .long 0x3717b0da, 0x35a43bce, 0x338306c6, 0x00000000 > > - > > - .align 64 > > - /* _sP0_lo */ > > - .long 0x00000000, 0x3d6fb9c9, 0x3d8fc35f, 0x3daf9169 > > - .long 0x3dcf49ab, 0x3deee849, 0x3e0f0ee8, 0x3e2e4984 > > - .long 0x3e4d2f8e, 0x3e6bb32e, 0x3e8c51cd, 0x3ea96163 > > - .long 0x3ec543f1, 0x3edfd735, 0x3f028438, 0x3f18abf0 > > - > > - .align 64 > > - /* _sP0_hi */ > > - .long 0x3f2bc480, 0x3f3bec1c, 0x3f4f2e5b, 0x3f613c53 > > - .long 0x3f6ce37d, 0x3f743c4f, 0x3f7a5feb, 0x3f7dea85 > > - .long 0x3f7f3b3d, 0x3f7fb78c, 0x3f7fefd4, 0x3f7ffdd0 > > - .long 0x3f7fffb4, 0x3f7ffff6, 0x3f7fffff, 0x3f800000 > > +LOCAL_DATA_NAME_UNALIGNED: > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iExpMantMask_UISA, 0x7fe00000) > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iMinIdxOfsMask_UISA, 0x3d400000) > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iMaxIdxMask_UISA, 0x03e00000) > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iExpMask, 0x7f000000) > > + .type LOCAL_DATA_NAME_UNALIGNED, @object > > + .size LOCAL_DATA_NAME_UNALIGNED, .-LOCAL_DATA_NAME_UNALIGNED > > > > .align 64 > > - .type __svml_stanh_data_internal_al64, @object > > - .size __svml_stanh_data_internal_al64, .-__svml_stanh_data_internal_al64 > > - .type __svml_stanh_data_internal, @object > > - .size __svml_stanh_data_internal, .-__svml_stanh_data_internal > > +LOCAL_DATA_NAME: > > + float_block (LOCAL_DATA_NAME, _sC_lo, > > + 0x00000000, 0x3d700000, 0x3d900000, 0x3db00000, > > + 0x3dd00000, 0x3df00000, 0x3e100000, 0x3e300000, > > + 0x3e500000, 0x3e700000, 0x3e900000, 0x3eb00000, > > + 0x3ed00000, 0x3ef00000, 0x3f100000, 0x3f300000) > > + > > + float_block (LOCAL_DATA_NAME, _sC_hi, > > + 0x3f500000, 0x3f700000, 0x3f900000, 0x3fb00000, > > + 0x3fd00000, 0x3ff00000, 0x40100000, 0x40300000, > > + 0x40500000, 0x40700000, 0x40900000, 0x40b00000, > > + 0x40d00000, 0x40f00000, 0x41100000, 0x00000000) > > + > > + float_block (LOCAL_DATA_NAME, _sP7_lo, > > + 0xbc0e2f66, 0x460bda12, 0x43d638ef, 0xc3e11c3e, > > + 0xc2baa4e9, 0xc249da2d, 0xc1859b82, 0x40dd5b57, > > + 0x40494640, 0x40c730a8, 0xbf0f160e, 0x3e30e76f, > > + 0xbea81387, 0xbdb26a1c, 0xbd351e57, 0xbb4c01a0) > > + > > + float_block (LOCAL_DATA_NAME, _sP7_hi, > > + 0x3c1d7bfb, 0x3c722cd1, 0x3c973f1c, 0x3c33a31b, > > + 0x3b862ef4, 0x3a27b3d0, 0xba3b5907, 0xba0efc22, > > + 0xb97f9f0f, 0xb8c8af50, 0xb7bdddfb, 0xb64f2950, > > + 0xb4e085b1, 0xb3731dfa, 0xb15a1f04, 0x00000000) > > + > > + float_block (LOCAL_DATA_NAME, _sP6_lo, > > + 0x3e0910e9, 0x43761143, 0x4165ecdc, 0xc190f756, > > + 0xc08c097d, 0xc02ba813, 0xbf7f6bda, 0x3f2b1dc0, > > + 0x3ece105d, 0x3f426a94, 0xbadb0dc4, 0x3da43b17, > > + 0xbd51ab88, 0xbcaea23d, 0xbd3b6d8d, 0xbd6caaad) > > + > > + float_block (LOCAL_DATA_NAME, _sP6_hi, > > + 0xbd795bed, 0xbd5fddda, 0xbd038f3b, 0xbc1cad63, > > + 0x3abb4766, 0x3b95f10b, 0x3b825873, 0x3afaea66, > > + 0x3a49f878, 0x39996bf3, 0x388f3e6c, 0x371bb0e3, > > + 0x35a8a5e6, 0x34369b17, 0x322487b0, 0x00000000) > > + > > + float_block (LOCAL_DATA_NAME, _sP5_lo, > > + 0xb76dd6b9, 0xbe1c276d, 0x3c1dcf2f, 0x3dc1a78d, > > + 0x3d96f985, 0x3da2b61b, 0x3dc13397, 0x3dd2f670, > > + 0x3df48a0a, 0x3e06c5a8, 0x3e1a3aba, 0x3e27c405, > > + 0x3e2e78d0, 0x3e2c3e44, 0x3e1d3097, 0x3df4a8f4) > > + > > + float_block (LOCAL_DATA_NAME, _sP5_hi, > > + 0x3da38508, 0x3d31416a, 0x3b562657, 0xbcaeeac9, > > + 0xbcce9419, 0xbcaaeac4, 0xbc49e7d0, 0xbba71ddd, > > + 0xbb003b0e, 0xba3f9a05, 0xb92c08a7, 0xb7ba9232, > > + 0xb64a0b0f, 0xb4dac169, 0xb2ab78ac, 0x00000000) > > + > > + float_block (LOCAL_DATA_NAME, _sP4_lo, > > + 0xbeaaaaa5, 0xbeab0612, 0xbea7f01f, 0xbea4e120, > > + 0xbea387b7, 0xbea15962, 0xbe9d57f7, 0xbe976b5a, > > + 0xbe90230d, 0xbe880dff, 0xbe7479b3, 0xbe4c3d88, > > + 0xbe212482, 0xbdeb8cba, 0xbd5e78ad, 0x3c6b5e6e) > > + > > + float_block (LOCAL_DATA_NAME, _sP4_hi, > > + 0x3d839143, 0x3dc21ee1, 0x3de347af, 0x3dcbec96, > > + 0x3d99ef2d, 0x3d542ea1, 0x3cdde701, 0x3c2cca67, > > + 0x3b81cb27, 0x3ac073a1, 0x39ac3032, 0x383a94d9, > > + 0x36ca081d, 0x355abd4c, 0x332b3cb6, 0x00000000) > > + > > + float_block (LOCAL_DATA_NAME, _sP3_lo, > > + 0xb0343c7b, 0xbd6ee69d, 0xbd8f0da7, 0xbdae477d, > > + 0xbdcd2a1f, 0xbdeba80d, 0xbe0c443b, 0xbe293cf3, > > + 0xbe44f282, 0xbe5f3651, 0xbe81c7c0, 0xbe96d7ca, > > + 0xbea7fb8e, 0xbeb50e9e, 0xbec12efe, 0xbec4be92) > > + > > + float_block (LOCAL_DATA_NAME, _sP3_hi, > > + 0xbebce070, 0xbead510e, 0xbe8ef7d6, 0xbe4b8704, > > + 0xbe083237, 0xbdaf7449, 0xbd2e1ec4, 0xbc83bf06, > > + 0xbbc3e0b5, 0xbb10aadc, 0xba0157db, 0xb88c18f2, > > + 0xb717b096, 0xb5a43bae, 0xb383012c, 0x00000000) > > + > > + float_block (LOCAL_DATA_NAME, _sP2_lo, > > + 0x3f800000, 0x3f7f1f84, 0x3f7ebd11, 0x3f7e1e5f, > > + 0x3f7d609f, 0x3f7c842d, 0x3f7b00e5, 0x3f789580, > > + 0x3f75b8ad, 0x3f726fd9, 0x3f6cc59b, 0x3f63fb92, > > + 0x3f59ff97, 0x3f4f11d7, 0x3f3d7573, 0x3f24f360) > > + > > + float_block (LOCAL_DATA_NAME, _sP2_hi, > > + 0x3f0cbfe7, 0x3eec1a69, 0x3eb0a801, 0x3e6753a2, > > + 0x3e132f1a, 0x3db7e7d3, 0x3d320845, 0x3c84d3d4, > > + 0x3bc477b7, 0x3b10d3da, 0x3a01601e, 0x388c1a3b, > > + 0x3717b0da, 0x35a43bce, 0x338306c6, 0x00000000) > > + > > + float_block (LOCAL_DATA_NAME, _sP0_lo, > > + 0x00000000, 0x3d6fb9c9, 0x3d8fc35f, 0x3daf9169, > > + 0x3dcf49ab, 0x3deee849, 0x3e0f0ee8, 0x3e2e4984, > > + 0x3e4d2f8e, 0x3e6bb32e, 0x3e8c51cd, 0x3ea96163, > > + 0x3ec543f1, 0x3edfd735, 0x3f028438, 0x3f18abf0) > > + > > + float_block (LOCAL_DATA_NAME, _sP0_hi, > > + 0x3f2bc480, 0x3f3bec1c, 0x3f4f2e5b, 0x3f613c53, > > + 0x3f6ce37d, 0x3f743c4f, 0x3f7a5feb, 0x3f7dea85, > > + 0x3f7f3b3d, 0x3f7fb78c, 0x3f7fefd4, 0x3f7ffdd0, > > + 0x3f7fffb4, 0x3f7ffff6, 0x3f7fffff, 0x3f800000) > > + > > + .type LOCAL_DATA_NAME, @object > > + .size LOCAL_DATA_NAME, .-LOCAL_DATA_NAME > > -- > > 2.34.1 > > > > The data movement makes the assembler codes much harder to follow. > Sunil, what do you think of this patch series? What do you mean? The change on in how we define rodata or the movement to multiple files or something else? > > > -- > H.J.
On Fri, Dec 16, 2022 at 10:18 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote: > > On Fri, Dec 16, 2022 at 9:06 AM H.J. Lu <hjl.tools@gmail.com> wrote: > > > > On Wed, Dec 7, 2022 at 12:52 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote: > > > > > > No changes to the logic, just change how rodata is handled. > > > > > > 1. Define the rodatas using the new macros so they check that the > > > offset is correct. > > > > > > 2. Use common data where applicable. > > > --- > > > .../multiarch/svml_s_tanhf16_core_avx512.S | 450 ++++++++---------- > > > 1 file changed, 197 insertions(+), 253 deletions(-) > > > > > > diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S b/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > > > index d74fc7731d..765e9ed7f7 100644 > > > --- a/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > > > +++ b/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > > > @@ -70,94 +70,99 @@ > > > * > > > */ > > > > > > -/* Offsets for data table __svml_stanh_data_internal_avx512. Ordered > > > - by use in the function. On cold-starts this might help the > > > - prefetcher. Possibly a better idea is to interleave start/end so > > > - that the prefetcher is less likely to detect a stream and pull > > > - irrelivant lines into cache. */ > > > > > > -/* Offsets for data table __svml_stanh_data_internal. 4 bytes each. > > > - */ > > > + > > > +#define LOCAL_DATA_NAME __svml_stanh_data_internal > > > +#define LOCAL_DATA_NAME_UNALIGNED __svml_stanh_data_internal_unaligned > > > +#include "svml_s_common_evex512_rodata_offsets.h" > > > + > > > +/* Offsets for data table __svml_stanh_data_internal_unaligned. > > > + 4 bytes each. */ > > > #define _iExpMantMask_UISA 0 > > > #define _iMinIdxOfsMask_UISA 4 > > > #define _iMaxIdxMask_UISA 8 > > > #define _iExpMask 12 > > > > > > -/* Offsets for data table __svml_stanh_data_internal_al64. 64 bytes > > > - each. */ > > > -#define _sC_lo 0 > > > -#define _sC_hi 64 > > > -#define _sP7_lo 128 > > > -#define _sP7_hi 192 > > > -#define _sSignMask 256 > > > -#define _sP6_lo 320 > > > -#define _sP6_hi 384 > > > -#define _sP5_lo 448 > > > -#define _sP5_hi 512 > > > -#define _sP4_lo 576 > > > -#define _sP4_hi 640 > > > -#define _sP3_lo 704 > > > -#define _sP3_hi 768 > > > -#define _sP2_lo 832 > > > -#define _sP2_hi 896 > > > -#define _sP0_lo 960 > > > -#define _sP0_hi 1024 > > > +/* Offsets for data table __svml_stanh_data_internal. Ordered > > > + by use in the function. On cold-starts this might help the > > > + prefetcher. Possibly a better idea is to interleave start/end so > > > + that the prefetcher is less likely to detect a stream and pull > > > + irrelivant lines into cache. */ > > > + > > > +/* Offsets for data table __svml_stanh_data_internal. > > > + 64 bytes each. */ > > > +#define _sC_lo 0 > > > +#define _sC_hi 64 > > > +#define _sP7_lo 128 > > > +#define _sP7_hi 192 > > > +#define _sP6_lo 256 > > > +#define _sP6_hi 320 > > > +#define _sP5_lo 384 > > > +#define _sP5_hi 448 > > > +#define _sP4_lo 512 > > > +#define _sP4_hi 576 > > > +#define _sP3_lo 640 > > > +#define _sP3_hi 704 > > > +#define _sP2_lo 768 > > > +#define _sP2_hi 832 > > > +#define _sP0_lo 896 > > > +#define _sP0_hi 960 > > > + > > > > > > #include <sysdep.h> > > > -#define TANHF_DATA(x) ((x)+__svml_stanh_data_internal_al64) > > > -#define TANHF_DATA_UNALIGNED(x) ((x)+__svml_stanh_data_internal) > > > > > > .section .text.evex512, "ax", @progbits > > > ENTRY(_ZGVeN16v_tanhf_skx) > > > - /* Here huge arguments, INF and NaNs are filtered out to callout. */ > > > - vpandd TANHF_DATA_UNALIGNED(_iExpMantMask_UISA)(%rip){1to16}, %zmm0, %zmm1 > > > - vpsubd TANHF_DATA_UNALIGNED(_iMinIdxOfsMask_UISA)(%rip){1to16}, %zmm1, %zmm2 > > > + /* Here huge arguments, INF and NaNs are filtered out to > > > + callout. */ > > > + vpandd LOCAL_DATA_UNALIGNED(_iExpMantMask_UISA)(%rip){1to16}, %zmm0, %zmm1 > > > + vpsubd LOCAL_DATA_UNALIGNED(_iMinIdxOfsMask_UISA)(%rip){1to16}, %zmm1, %zmm2 > > > > > > /* Selection arguments between [0, 0x03e00000] into zmm3. */ > > > vpxord %zmm3, %zmm3, %zmm3 > > > vpmaxsd %zmm3, %zmm2, %zmm3 > > > - vpminsd TANHF_DATA_UNALIGNED(_iMaxIdxMask_UISA)(%rip){1to16}, %zmm3, %zmm3 > > > + vpminsd LOCAL_DATA_UNALIGNED(_iMaxIdxMask_UISA)(%rip){1to16}, %zmm3, %zmm3 > > > > > > /* Setup permute indices in zmm3. */ > > > vpsrld $21, %zmm3, %zmm3 > > > > > > /* Store if there are any special cases in k1. */ > > > - vpcmpd $6, TANHF_DATA_UNALIGNED(_iExpMask)(%rip){1to16}, %zmm1, %k1 > > > + vpcmpd $6, LOCAL_DATA_UNALIGNED(_iExpMask)(%rip){1to16}, %zmm1, %k1 > > > > > > - vmovaps TANHF_DATA(_sC_lo)(%rip), %zmm5 > > > - vpermt2ps TANHF_DATA(_sC_hi)(%rip), %zmm3, %zmm5 > > > + vmovaps LOCAL_DATA(_sC_lo)(%rip), %zmm5 > > > + vpermt2ps LOCAL_DATA(_sC_hi)(%rip), %zmm3, %zmm5 > > > > > > - vmovaps TANHF_DATA(_sP7_lo)(%rip), %zmm2 > > > - vpermt2ps TANHF_DATA(_sP7_hi)(%rip), %zmm3, %zmm2 > > > + vmovaps LOCAL_DATA(_sP7_lo)(%rip), %zmm2 > > > + vpermt2ps LOCAL_DATA(_sP7_hi)(%rip), %zmm3, %zmm2 > > > > > > /* Store absolute values of inputs in zmm1. */ > > > - vmovaps TANHF_DATA(_sSignMask)(%rip), %zmm4 > > > + vmovaps COMMON_DATA(_SignMask)(%rip), %zmm4 > > > vandnps %zmm0, %zmm4, %zmm1 > > > vsubps {rn-sae}, %zmm5, %zmm1, %zmm1 > > > > > > - vmovaps TANHF_DATA(_sP6_lo)(%rip), %zmm5 > > > - vpermt2ps TANHF_DATA(_sP6_hi)(%rip), %zmm3, %zmm5 > > > + vmovaps LOCAL_DATA(_sP6_lo)(%rip), %zmm5 > > > + vpermt2ps LOCAL_DATA(_sP6_hi)(%rip), %zmm3, %zmm5 > > > > > > - vmovaps TANHF_DATA(_sP5_lo)(%rip), %zmm6 > > > - vpermt2ps TANHF_DATA(_sP5_hi)(%rip), %zmm3, %zmm6 > > > + vmovaps LOCAL_DATA(_sP5_lo)(%rip), %zmm6 > > > + vpermt2ps LOCAL_DATA(_sP5_hi)(%rip), %zmm3, %zmm6 > > > > > > vfmadd213ps {rn-sae}, %zmm5, %zmm1, %zmm2 > > > vfmadd213ps {rn-sae}, %zmm6, %zmm1, %zmm2 > > > > > > - vmovaps TANHF_DATA(_sP4_lo)(%rip), %zmm7 > > > - vpermt2ps TANHF_DATA(_sP4_hi)(%rip), %zmm3, %zmm7 > > > + vmovaps LOCAL_DATA(_sP4_lo)(%rip), %zmm7 > > > + vpermt2ps LOCAL_DATA(_sP4_hi)(%rip), %zmm3, %zmm7 > > > > > > - vmovaps TANHF_DATA(_sP3_lo)(%rip), %zmm8 > > > - vpermt2ps TANHF_DATA(_sP3_hi)(%rip), %zmm3, %zmm8 > > > + vmovaps LOCAL_DATA(_sP3_lo)(%rip), %zmm8 > > > + vpermt2ps LOCAL_DATA(_sP3_hi)(%rip), %zmm3, %zmm8 > > > > > > vfmadd213ps {rn-sae}, %zmm7, %zmm1, %zmm2 > > > vfmadd213ps {rn-sae}, %zmm8, %zmm1, %zmm2 > > > > > > - vmovaps TANHF_DATA(_sP2_lo)(%rip), %zmm9 > > > - vpermt2ps TANHF_DATA(_sP2_hi)(%rip), %zmm3, %zmm9 > > > + vmovaps LOCAL_DATA(_sP2_lo)(%rip), %zmm9 > > > + vpermt2ps LOCAL_DATA(_sP2_hi)(%rip), %zmm3, %zmm9 > > > > > > - vmovaps TANHF_DATA(_sP0_lo)(%rip), %zmm10 > > > - vpermt2ps TANHF_DATA(_sP0_hi)(%rip), %zmm3, %zmm10 > > > + vmovaps LOCAL_DATA(_sP0_lo)(%rip), %zmm10 > > > + vpermt2ps LOCAL_DATA(_sP0_hi)(%rip), %zmm3, %zmm10 > > > > > > vfmadd213ps {rn-sae}, %zmm9, %zmm1, %zmm2 > > > vfmadd213ps {rn-sae}, %zmm10, %zmm1, %zmm2 > > > @@ -167,7 +172,7 @@ ENTRY(_ZGVeN16v_tanhf_skx) > > > > > > /* Go to special inputs processing branch. */ > > > jne L(SPECIAL_VALUES_BRANCH) > > > - # LOE rbx r12 r13 r14 r15 zmm0 zmm2 zmm4 > > > + > > > /* Wait until after branch of write over zmm0. */ > > > vpternlogd $0xec, %zmm4, %zmm2, %zmm0 > > > > > > @@ -176,24 +181,24 @@ ENTRY(_ZGVeN16v_tanhf_skx) > > > > > > /* Cold case. edx has 1s where there was a special value that > > > needs to be handled by a tanhf call. Optimize for code size > > > - more so than speed here. */ > > > + more so than speed here. */ > > > L(SPECIAL_VALUES_BRANCH): > > > - # LOE rbx rdx r12 r13 r14 r15 zmm0 zmm2 zmm4 > > > - /* Use r13 to save/restore the stack. This allows us to use rbp as > > > - callee save register saving code size. */ > > > + > > > + /* Use r13 to save/restore the stack. This allows us to use rbp > > > + as callee save register saving code size. */ > > > pushq %r13 > > > - cfi_adjust_cfa_offset(8) > > > - cfi_offset(r13, -16) > > > - /* Need to callee save registers to preserve state across tanhf calls. > > > - */ > > > + cfi_adjust_cfa_offset (8) > > > + cfi_offset (r13, -16) > > > + /* Need to callee save registers to preserve state across tanhf > > > + calls. */ > > > pushq %rbx > > > - cfi_adjust_cfa_offset(8) > > > - cfi_offset(rbx, -24) > > > + cfi_adjust_cfa_offset (8) > > > + cfi_offset (rbx, -24) > > > pushq %rbp > > > - cfi_adjust_cfa_offset(8) > > > - cfi_offset(rbp, -32) > > > + cfi_adjust_cfa_offset (8) > > > + cfi_offset (rbp, -32) > > > movq %rsp, %r13 > > > - cfi_def_cfa_register(r13) > > > + cfi_def_cfa_register (r13) > > > > > > /* Align stack and make room for 2x zmm vectors. */ > > > andq $-64, %rsp > > > @@ -207,16 +212,17 @@ L(SPECIAL_VALUES_BRANCH): > > > > > > vzeroupper > > > > > > - /* edx has 1s where there was a special value that needs to be handled > > > - by a tanhf call. */ > > > + /* edx has 1s where there was a special value that needs to be > > > + handled by a tanhf call. */ > > > movl %edx, %ebx > > > L(SPECIAL_VALUES_LOOP): > > > - # LOE rbx rbp r12 r13 r14 r15 > > > - /* use rbp as index for special value that is saved across calls to > > > - tanhf. We technically don't need a callee save register here as offset > > > - to rsp is always [0, 56] so we can restore rsp by realigning to 64. > > > - Essentially the tradeoff is 1 extra save/restore vs 2 extra instructions > > > - in the loop. Realigning also costs more code size. */ > > > + > > > + /* use rbp as index for special value that is saved across calls > > > + to tanhf. We technically don't need a callee save register > > > + here as offset to rsp is always [0, 56] so we can restore > > > + rsp by realigning to 64. Essentially the tradeoff is 1 extra > > > + save/restore vs 2 extra instructions in the loop. Realigning > > > + also costs more code size. */ > > > xorl %ebp, %ebp > > > tzcntl %ebx, %ebp > > > > > > @@ -224,203 +230,141 @@ L(SPECIAL_VALUES_LOOP): > > > vmovss 64(%rsp, %rbp, 4), %xmm0 > > > call tanhf@PLT > > > > > > - /* No good way to avoid the store-forwarding fault this will cause on > > > - return. `lfence` avoids the SF fault but at greater cost as it > > > - serialized stack/callee save restoration. */ > > > + /* No good way to avoid the store-forwarding fault this will > > > + cause on return. `lfence` avoids the SF fault but at greater > > > + cost as it serialized stack/callee save restoration. */ > > > vmovss %xmm0, (%rsp, %rbp, 4) > > > > > > - blsrl %ebx, %ebx > > > + blsrl %ebx, %ebx > > > jnz L(SPECIAL_VALUES_LOOP) > > > - # LOE r12 r13 r14 r15 > > > + > > > > > > /* All results have been written to (%rsp). */ > > > vmovaps (%rsp), %zmm0 > > > /* Restore rsp. */ > > > movq %r13, %rsp > > > - cfi_def_cfa_register(rsp) > > > + cfi_def_cfa_register (rsp) > > > /* Restore callee save registers. */ > > > popq %rbp > > > - cfi_adjust_cfa_offset(-8) > > > - cfi_restore(rbp) > > > + cfi_adjust_cfa_offset (-8) > > > + cfi_restore (rbp) > > > popq %rbx > > > - cfi_adjust_cfa_offset(-8) > > > - cfi_restore(rbp) > > > + cfi_adjust_cfa_offset (-8) > > > + cfi_restore (rbp) > > > popq %r13 > > > - cfi_adjust_cfa_offset(-8) > > > - cfi_restore(r13) > > > + cfi_adjust_cfa_offset (-8) > > > + cfi_restore (r13) > > > ret > > > END(_ZGVeN16v_tanhf_skx) > > > > > > - .section .rodata, "a" > > > + .section .rodata.evex512, "a" > > > .align 16 > > > -#ifdef __svml_stanh_data_internal_typedef > > > -typedef unsigned int VUINT32; > > > -typedef struct > > > - { > > > - __declspec(align(4)) VUINT32 _iExpMantMask_UISA[1][1]; > > > - __declspec(align(4)) VUINT32 _iMinIdxOfsMask_UISA[1][1]; > > > - __declspec(align(4)) VUINT32 _iMaxIdxMask_UISA[1][1]; > > > - __declspec(align(4)) VUINT32 _iExpMask[1][1]; > > > - __declspec(align(64)) VUINT32 _sC_lo[16][1]; > > > - __declspec(align(64)) VUINT32 _sC_hi[16][1]; > > > - __declspec(align(64)) VUINT32 _sP7_lo[16][1]; > > > - __declspec(align(64)) VUINT32 _sP7_hi[16][1]; > > > - __declspec(align(64)) VUINT32 _sSignMask[16][1]; > > > - __declspec(align(64)) VUINT32 _sP6_lo[16][1]; > > > - __declspec(align(64)) VUINT32 _sP6_hi[16][1]; > > > - __declspec(align(64)) VUINT32 _sP5_lo[16][1]; > > > - __declspec(align(64)) VUINT32 _sP5_hi[16][1]; > > > - __declspec(align(64)) VUINT32 _sP4_lo[16][1]; > > > - __declspec(align(64)) VUINT32 _sP4_hi[16][1]; > > > - __declspec(align(64)) VUINT32 _sP3_lo[16][1]; > > > - __declspec(align(64)) VUINT32 _sP3_hi[16][1]; > > > - __declspec(align(64)) VUINT32 _sP2_lo[16][1]; > > > - __declspec(align(64)) VUINT32 _sP2_hi[16][1]; > > > - __declspec(align(64)) VUINT32 _sP0_lo[16][1]; > > > - __declspec(align(64)) VUINT32 _sP0_hi[16][1]; > > > -} __svml_stanh_data_internal; > > > -#endif > > > - > > > -__svml_stanh_data_internal: > > > - .align 4 > > > - /* _iExpMantMask_UISA */ > > > - .long 0x7fe00000 > > > - > > > - .align 4 > > > - /* _iMinIdxOfsMask_UISA */ > > > - .long 0x3d400000 > > > - > > > - .align 4 > > > - /* _iMaxIdxMask_UISA */ > > > - .long 0x03e00000 > > > - > > > - .align 4 > > > - /* _iExpMask */ > > > - .long 0x7f000000 > > > - > > > - .align 64 > > > -__svml_stanh_data_internal_al64: > > > - .align 64 > > > - /* _sC_lo */ > > > - .long 0x00000000, 0x3d700000, 0x3d900000, 0x3db00000 > > > - .long 0x3dd00000, 0x3df00000, 0x3e100000, 0x3e300000 > > > - .long 0x3e500000, 0x3e700000, 0x3e900000, 0x3eb00000 > > > - .long 0x3ed00000, 0x3ef00000, 0x3f100000, 0x3f300000 > > > - > > > - .align 64 > > > - /* _sC_hi */ > > > - .long 0x3f500000, 0x3f700000, 0x3f900000, 0x3fb00000 > > > - .long 0x3fd00000, 0x3ff00000, 0x40100000, 0x40300000 > > > - .long 0x40500000, 0x40700000, 0x40900000, 0x40b00000 > > > - .long 0x40d00000, 0x40f00000, 0x41100000, 0x00000000 > > > - > > > - .align 64 > > > - /* _sP7_lo */ > > > - .long 0xbc0e2f66, 0x460bda12, 0x43d638ef, 0xc3e11c3e > > > - .long 0xc2baa4e9, 0xc249da2d, 0xc1859b82, 0x40dd5b57 > > > - .long 0x40494640, 0x40c730a8, 0xbf0f160e, 0x3e30e76f > > > - .long 0xbea81387, 0xbdb26a1c, 0xbd351e57, 0xbb4c01a0 > > > - > > > - .align 64 > > > - /* _sP7_hi */ > > > - .long 0x3c1d7bfb, 0x3c722cd1, 0x3c973f1c, 0x3c33a31b > > > - .long 0x3b862ef4, 0x3a27b3d0, 0xba3b5907, 0xba0efc22 > > > - .long 0xb97f9f0f, 0xb8c8af50, 0xb7bdddfb, 0xb64f2950 > > > - .long 0xb4e085b1, 0xb3731dfa, 0xb15a1f04, 0x00000000 > > > > > > - .align 64 > > > - /* _sSignMask */ > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > - > > > - .align 64 > > > - /* _sP6_lo */ > > > - .long 0x3e0910e9, 0x43761143, 0x4165ecdc, 0xc190f756 > > > - .long 0xc08c097d, 0xc02ba813, 0xbf7f6bda, 0x3f2b1dc0 > > > - .long 0x3ece105d, 0x3f426a94, 0xbadb0dc4, 0x3da43b17 > > > - .long 0xbd51ab88, 0xbcaea23d, 0xbd3b6d8d, 0xbd6caaad > > > - > > > - .align 64 > > > - /* _sP6_hi */ > > > - .long 0xbd795bed, 0xbd5fddda, 0xbd038f3b, 0xbc1cad63 > > > - .long 0x3abb4766, 0x3b95f10b, 0x3b825873, 0x3afaea66 > > > - .long 0x3a49f878, 0x39996bf3, 0x388f3e6c, 0x371bb0e3 > > > - .long 0x35a8a5e6, 0x34369b17, 0x322487b0, 0x00000000 > > > - > > > - .align 64 > > > - /* _sP5_lo */ > > > - .long 0xb76dd6b9, 0xbe1c276d, 0x3c1dcf2f, 0x3dc1a78d > > > - .long 0x3d96f985, 0x3da2b61b, 0x3dc13397, 0x3dd2f670 > > > - .long 0x3df48a0a, 0x3e06c5a8, 0x3e1a3aba, 0x3e27c405 > > > - .long 0x3e2e78d0, 0x3e2c3e44, 0x3e1d3097, 0x3df4a8f4 > > > - > > > - .align 64 > > > - /* _sP5_hi */ > > > - .long 0x3da38508, 0x3d31416a, 0x3b562657, 0xbcaeeac9 > > > - .long 0xbcce9419, 0xbcaaeac4, 0xbc49e7d0, 0xbba71ddd > > > - .long 0xbb003b0e, 0xba3f9a05, 0xb92c08a7, 0xb7ba9232 > > > - .long 0xb64a0b0f, 0xb4dac169, 0xb2ab78ac, 0x00000000 > > > - > > > - .align 64 > > > - /* _sP4_lo */ > > > - .long 0xbeaaaaa5, 0xbeab0612, 0xbea7f01f, 0xbea4e120 > > > - .long 0xbea387b7, 0xbea15962, 0xbe9d57f7, 0xbe976b5a > > > - .long 0xbe90230d, 0xbe880dff, 0xbe7479b3, 0xbe4c3d88 > > > - .long 0xbe212482, 0xbdeb8cba, 0xbd5e78ad, 0x3c6b5e6e > > > - > > > - .align 64 > > > - /* _sP4_hi */ > > > - .long 0x3d839143, 0x3dc21ee1, 0x3de347af, 0x3dcbec96 > > > - .long 0x3d99ef2d, 0x3d542ea1, 0x3cdde701, 0x3c2cca67 > > > - .long 0x3b81cb27, 0x3ac073a1, 0x39ac3032, 0x383a94d9 > > > - .long 0x36ca081d, 0x355abd4c, 0x332b3cb6, 0x00000000 > > > - > > > - .align 64 > > > - /* _sP3_lo */ > > > - .long 0xb0343c7b, 0xbd6ee69d, 0xbd8f0da7, 0xbdae477d > > > - .long 0xbdcd2a1f, 0xbdeba80d, 0xbe0c443b, 0xbe293cf3 > > > - .long 0xbe44f282, 0xbe5f3651, 0xbe81c7c0, 0xbe96d7ca > > > - .long 0xbea7fb8e, 0xbeb50e9e, 0xbec12efe, 0xbec4be92 > > > - > > > - .align 64 > > > - /* _sP3_hi */ > > > - .long 0xbebce070, 0xbead510e, 0xbe8ef7d6, 0xbe4b8704 > > > - .long 0xbe083237, 0xbdaf7449, 0xbd2e1ec4, 0xbc83bf06 > > > - .long 0xbbc3e0b5, 0xbb10aadc, 0xba0157db, 0xb88c18f2 > > > - .long 0xb717b096, 0xb5a43bae, 0xb383012c, 0x00000000 > > > - > > > - .align 64 > > > - /* _sP2_lo */ > > > - .long 0x3f800000, 0x3f7f1f84, 0x3f7ebd11, 0x3f7e1e5f > > > - .long 0x3f7d609f, 0x3f7c842d, 0x3f7b00e5, 0x3f789580 > > > - .long 0x3f75b8ad, 0x3f726fd9, 0x3f6cc59b, 0x3f63fb92 > > > - .long 0x3f59ff97, 0x3f4f11d7, 0x3f3d7573, 0x3f24f360 > > > - > > > - .align 64 > > > - /* _sP2_hi */ > > > - .long 0x3f0cbfe7, 0x3eec1a69, 0x3eb0a801, 0x3e6753a2 > > > - .long 0x3e132f1a, 0x3db7e7d3, 0x3d320845, 0x3c84d3d4 > > > - .long 0x3bc477b7, 0x3b10d3da, 0x3a01601e, 0x388c1a3b > > > - .long 0x3717b0da, 0x35a43bce, 0x338306c6, 0x00000000 > > > - > > > - .align 64 > > > - /* _sP0_lo */ > > > - .long 0x00000000, 0x3d6fb9c9, 0x3d8fc35f, 0x3daf9169 > > > - .long 0x3dcf49ab, 0x3deee849, 0x3e0f0ee8, 0x3e2e4984 > > > - .long 0x3e4d2f8e, 0x3e6bb32e, 0x3e8c51cd, 0x3ea96163 > > > - .long 0x3ec543f1, 0x3edfd735, 0x3f028438, 0x3f18abf0 > > > - > > > - .align 64 > > > - /* _sP0_hi */ > > > - .long 0x3f2bc480, 0x3f3bec1c, 0x3f4f2e5b, 0x3f613c53 > > > - .long 0x3f6ce37d, 0x3f743c4f, 0x3f7a5feb, 0x3f7dea85 > > > - .long 0x3f7f3b3d, 0x3f7fb78c, 0x3f7fefd4, 0x3f7ffdd0 > > > - .long 0x3f7fffb4, 0x3f7ffff6, 0x3f7fffff, 0x3f800000 > > > +LOCAL_DATA_NAME_UNALIGNED: > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iExpMantMask_UISA, 0x7fe00000) > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iMinIdxOfsMask_UISA, 0x3d400000) > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iMaxIdxMask_UISA, 0x03e00000) > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iExpMask, 0x7f000000) > > > + .type LOCAL_DATA_NAME_UNALIGNED, @object > > > + .size LOCAL_DATA_NAME_UNALIGNED, .-LOCAL_DATA_NAME_UNALIGNED > > > > > > .align 64 > > > - .type __svml_stanh_data_internal_al64, @object > > > - .size __svml_stanh_data_internal_al64, .-__svml_stanh_data_internal_al64 > > > - .type __svml_stanh_data_internal, @object > > > - .size __svml_stanh_data_internal, .-__svml_stanh_data_internal > > > +LOCAL_DATA_NAME: > > > + float_block (LOCAL_DATA_NAME, _sC_lo, > > > + 0x00000000, 0x3d700000, 0x3d900000, 0x3db00000, > > > + 0x3dd00000, 0x3df00000, 0x3e100000, 0x3e300000, > > > + 0x3e500000, 0x3e700000, 0x3e900000, 0x3eb00000, > > > + 0x3ed00000, 0x3ef00000, 0x3f100000, 0x3f300000) > > > + > > > + float_block (LOCAL_DATA_NAME, _sC_hi, > > > + 0x3f500000, 0x3f700000, 0x3f900000, 0x3fb00000, > > > + 0x3fd00000, 0x3ff00000, 0x40100000, 0x40300000, > > > + 0x40500000, 0x40700000, 0x40900000, 0x40b00000, > > > + 0x40d00000, 0x40f00000, 0x41100000, 0x00000000) > > > + > > > + float_block (LOCAL_DATA_NAME, _sP7_lo, > > > + 0xbc0e2f66, 0x460bda12, 0x43d638ef, 0xc3e11c3e, > > > + 0xc2baa4e9, 0xc249da2d, 0xc1859b82, 0x40dd5b57, > > > + 0x40494640, 0x40c730a8, 0xbf0f160e, 0x3e30e76f, > > > + 0xbea81387, 0xbdb26a1c, 0xbd351e57, 0xbb4c01a0) > > > + > > > + float_block (LOCAL_DATA_NAME, _sP7_hi, > > > + 0x3c1d7bfb, 0x3c722cd1, 0x3c973f1c, 0x3c33a31b, > > > + 0x3b862ef4, 0x3a27b3d0, 0xba3b5907, 0xba0efc22, > > > + 0xb97f9f0f, 0xb8c8af50, 0xb7bdddfb, 0xb64f2950, > > > + 0xb4e085b1, 0xb3731dfa, 0xb15a1f04, 0x00000000) > > > + > > > + float_block (LOCAL_DATA_NAME, _sP6_lo, > > > + 0x3e0910e9, 0x43761143, 0x4165ecdc, 0xc190f756, > > > + 0xc08c097d, 0xc02ba813, 0xbf7f6bda, 0x3f2b1dc0, > > > + 0x3ece105d, 0x3f426a94, 0xbadb0dc4, 0x3da43b17, > > > + 0xbd51ab88, 0xbcaea23d, 0xbd3b6d8d, 0xbd6caaad) > > > + > > > + float_block (LOCAL_DATA_NAME, _sP6_hi, > > > + 0xbd795bed, 0xbd5fddda, 0xbd038f3b, 0xbc1cad63, > > > + 0x3abb4766, 0x3b95f10b, 0x3b825873, 0x3afaea66, > > > + 0x3a49f878, 0x39996bf3, 0x388f3e6c, 0x371bb0e3, > > > + 0x35a8a5e6, 0x34369b17, 0x322487b0, 0x00000000) > > > + > > > + float_block (LOCAL_DATA_NAME, _sP5_lo, > > > + 0xb76dd6b9, 0xbe1c276d, 0x3c1dcf2f, 0x3dc1a78d, > > > + 0x3d96f985, 0x3da2b61b, 0x3dc13397, 0x3dd2f670, > > > + 0x3df48a0a, 0x3e06c5a8, 0x3e1a3aba, 0x3e27c405, > > > + 0x3e2e78d0, 0x3e2c3e44, 0x3e1d3097, 0x3df4a8f4) > > > + > > > + float_block (LOCAL_DATA_NAME, _sP5_hi, > > > + 0x3da38508, 0x3d31416a, 0x3b562657, 0xbcaeeac9, > > > + 0xbcce9419, 0xbcaaeac4, 0xbc49e7d0, 0xbba71ddd, > > > + 0xbb003b0e, 0xba3f9a05, 0xb92c08a7, 0xb7ba9232, > > > + 0xb64a0b0f, 0xb4dac169, 0xb2ab78ac, 0x00000000) > > > + > > > + float_block (LOCAL_DATA_NAME, _sP4_lo, > > > + 0xbeaaaaa5, 0xbeab0612, 0xbea7f01f, 0xbea4e120, > > > + 0xbea387b7, 0xbea15962, 0xbe9d57f7, 0xbe976b5a, > > > + 0xbe90230d, 0xbe880dff, 0xbe7479b3, 0xbe4c3d88, > > > + 0xbe212482, 0xbdeb8cba, 0xbd5e78ad, 0x3c6b5e6e) > > > + > > > + float_block (LOCAL_DATA_NAME, _sP4_hi, > > > + 0x3d839143, 0x3dc21ee1, 0x3de347af, 0x3dcbec96, > > > + 0x3d99ef2d, 0x3d542ea1, 0x3cdde701, 0x3c2cca67, > > > + 0x3b81cb27, 0x3ac073a1, 0x39ac3032, 0x383a94d9, > > > + 0x36ca081d, 0x355abd4c, 0x332b3cb6, 0x00000000) > > > + > > > + float_block (LOCAL_DATA_NAME, _sP3_lo, > > > + 0xb0343c7b, 0xbd6ee69d, 0xbd8f0da7, 0xbdae477d, > > > + 0xbdcd2a1f, 0xbdeba80d, 0xbe0c443b, 0xbe293cf3, > > > + 0xbe44f282, 0xbe5f3651, 0xbe81c7c0, 0xbe96d7ca, > > > + 0xbea7fb8e, 0xbeb50e9e, 0xbec12efe, 0xbec4be92) > > > + > > > + float_block (LOCAL_DATA_NAME, _sP3_hi, > > > + 0xbebce070, 0xbead510e, 0xbe8ef7d6, 0xbe4b8704, > > > + 0xbe083237, 0xbdaf7449, 0xbd2e1ec4, 0xbc83bf06, > > > + 0xbbc3e0b5, 0xbb10aadc, 0xba0157db, 0xb88c18f2, > > > + 0xb717b096, 0xb5a43bae, 0xb383012c, 0x00000000) > > > + > > > + float_block (LOCAL_DATA_NAME, _sP2_lo, > > > + 0x3f800000, 0x3f7f1f84, 0x3f7ebd11, 0x3f7e1e5f, > > > + 0x3f7d609f, 0x3f7c842d, 0x3f7b00e5, 0x3f789580, > > > + 0x3f75b8ad, 0x3f726fd9, 0x3f6cc59b, 0x3f63fb92, > > > + 0x3f59ff97, 0x3f4f11d7, 0x3f3d7573, 0x3f24f360) > > > + > > > + float_block (LOCAL_DATA_NAME, _sP2_hi, > > > + 0x3f0cbfe7, 0x3eec1a69, 0x3eb0a801, 0x3e6753a2, > > > + 0x3e132f1a, 0x3db7e7d3, 0x3d320845, 0x3c84d3d4, > > > + 0x3bc477b7, 0x3b10d3da, 0x3a01601e, 0x388c1a3b, > > > + 0x3717b0da, 0x35a43bce, 0x338306c6, 0x00000000) > > > + > > > + float_block (LOCAL_DATA_NAME, _sP0_lo, > > > + 0x00000000, 0x3d6fb9c9, 0x3d8fc35f, 0x3daf9169, > > > + 0x3dcf49ab, 0x3deee849, 0x3e0f0ee8, 0x3e2e4984, > > > + 0x3e4d2f8e, 0x3e6bb32e, 0x3e8c51cd, 0x3ea96163, > > > + 0x3ec543f1, 0x3edfd735, 0x3f028438, 0x3f18abf0) > > > + > > > + float_block (LOCAL_DATA_NAME, _sP0_hi, > > > + 0x3f2bc480, 0x3f3bec1c, 0x3f4f2e5b, 0x3f613c53, > > > + 0x3f6ce37d, 0x3f743c4f, 0x3f7a5feb, 0x3f7dea85, > > > + 0x3f7f3b3d, 0x3f7fb78c, 0x3f7fefd4, 0x3f7ffdd0, > > > + 0x3f7fffb4, 0x3f7ffff6, 0x3f7fffff, 0x3f800000) > > > + > > > + .type LOCAL_DATA_NAME, @object > > > + .size LOCAL_DATA_NAME, .-LOCAL_DATA_NAME > > > -- > > > 2.34.1 > > > > > > > The data movement makes the assembler codes much harder to follow. > > Sunil, what do you think of this patch series? > > What do you mean? The change on in how we define rodata or the movement > to multiple files or something else? The glibc way to support data files for assembly codes is to define data in C and use *.sym to generate offsets for assembly files, like sysdeps/x86/cpu-features-offsets.sym:XSAVE_STATE_SIZE_OFFSET offsetof (struct cpu_features, xsave_state_size) sysdeps/x86_64/dl-trampoline.h: sub _rtld_local_ro+RTLD_GLOBAL_RO_DL_X86_CPU_FEATURES_OFFSET+XSAVE_STATE_SIZE_OFFSET(%rip), %RSP_LP sysdeps/x86_64/dl-trampoline.h: sub _dl_x86_cpu_features+XSAVE_STATE_SIZE_OFFSET(%rip), %RSP_LP
On Fri, Dec 16, 2022 at 1:38 PM H.J. Lu <hjl.tools@gmail.com> wrote: > > On Fri, Dec 16, 2022 at 10:18 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote: > > > > On Fri, Dec 16, 2022 at 9:06 AM H.J. Lu <hjl.tools@gmail.com> wrote: > > > > > > On Wed, Dec 7, 2022 at 12:52 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote: > > > > > > > > No changes to the logic, just change how rodata is handled. > > > > > > > > 1. Define the rodatas using the new macros so they check that the > > > > offset is correct. > > > > > > > > 2. Use common data where applicable. > > > > --- > > > > .../multiarch/svml_s_tanhf16_core_avx512.S | 450 ++++++++---------- > > > > 1 file changed, 197 insertions(+), 253 deletions(-) > > > > > > > > diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S b/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > > > > index d74fc7731d..765e9ed7f7 100644 > > > > --- a/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > > > > +++ b/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > > > > @@ -70,94 +70,99 @@ > > > > * > > > > */ > > > > > > > > -/* Offsets for data table __svml_stanh_data_internal_avx512. Ordered > > > > - by use in the function. On cold-starts this might help the > > > > - prefetcher. Possibly a better idea is to interleave start/end so > > > > - that the prefetcher is less likely to detect a stream and pull > > > > - irrelivant lines into cache. */ > > > > > > > > -/* Offsets for data table __svml_stanh_data_internal. 4 bytes each. > > > > - */ > > > > + > > > > +#define LOCAL_DATA_NAME __svml_stanh_data_internal > > > > +#define LOCAL_DATA_NAME_UNALIGNED __svml_stanh_data_internal_unaligned > > > > +#include "svml_s_common_evex512_rodata_offsets.h" > > > > + > > > > +/* Offsets for data table __svml_stanh_data_internal_unaligned. > > > > + 4 bytes each. */ > > > > #define _iExpMantMask_UISA 0 > > > > #define _iMinIdxOfsMask_UISA 4 > > > > #define _iMaxIdxMask_UISA 8 > > > > #define _iExpMask 12 > > > > > > > > -/* Offsets for data table __svml_stanh_data_internal_al64. 64 bytes > > > > - each. */ > > > > -#define _sC_lo 0 > > > > -#define _sC_hi 64 > > > > -#define _sP7_lo 128 > > > > -#define _sP7_hi 192 > > > > -#define _sSignMask 256 > > > > -#define _sP6_lo 320 > > > > -#define _sP6_hi 384 > > > > -#define _sP5_lo 448 > > > > -#define _sP5_hi 512 > > > > -#define _sP4_lo 576 > > > > -#define _sP4_hi 640 > > > > -#define _sP3_lo 704 > > > > -#define _sP3_hi 768 > > > > -#define _sP2_lo 832 > > > > -#define _sP2_hi 896 > > > > -#define _sP0_lo 960 > > > > -#define _sP0_hi 1024 > > > > +/* Offsets for data table __svml_stanh_data_internal. Ordered > > > > + by use in the function. On cold-starts this might help the > > > > + prefetcher. Possibly a better idea is to interleave start/end so > > > > + that the prefetcher is less likely to detect a stream and pull > > > > + irrelivant lines into cache. */ > > > > + > > > > +/* Offsets for data table __svml_stanh_data_internal. > > > > + 64 bytes each. */ > > > > +#define _sC_lo 0 > > > > +#define _sC_hi 64 > > > > +#define _sP7_lo 128 > > > > +#define _sP7_hi 192 > > > > +#define _sP6_lo 256 > > > > +#define _sP6_hi 320 > > > > +#define _sP5_lo 384 > > > > +#define _sP5_hi 448 > > > > +#define _sP4_lo 512 > > > > +#define _sP4_hi 576 > > > > +#define _sP3_lo 640 > > > > +#define _sP3_hi 704 > > > > +#define _sP2_lo 768 > > > > +#define _sP2_hi 832 > > > > +#define _sP0_lo 896 > > > > +#define _sP0_hi 960 > > > > + > > > > > > > > #include <sysdep.h> > > > > -#define TANHF_DATA(x) ((x)+__svml_stanh_data_internal_al64) > > > > -#define TANHF_DATA_UNALIGNED(x) ((x)+__svml_stanh_data_internal) > > > > > > > > .section .text.evex512, "ax", @progbits > > > > ENTRY(_ZGVeN16v_tanhf_skx) > > > > - /* Here huge arguments, INF and NaNs are filtered out to callout. */ > > > > - vpandd TANHF_DATA_UNALIGNED(_iExpMantMask_UISA)(%rip){1to16}, %zmm0, %zmm1 > > > > - vpsubd TANHF_DATA_UNALIGNED(_iMinIdxOfsMask_UISA)(%rip){1to16}, %zmm1, %zmm2 > > > > + /* Here huge arguments, INF and NaNs are filtered out to > > > > + callout. */ > > > > + vpandd LOCAL_DATA_UNALIGNED(_iExpMantMask_UISA)(%rip){1to16}, %zmm0, %zmm1 > > > > + vpsubd LOCAL_DATA_UNALIGNED(_iMinIdxOfsMask_UISA)(%rip){1to16}, %zmm1, %zmm2 > > > > > > > > /* Selection arguments between [0, 0x03e00000] into zmm3. */ > > > > vpxord %zmm3, %zmm3, %zmm3 > > > > vpmaxsd %zmm3, %zmm2, %zmm3 > > > > - vpminsd TANHF_DATA_UNALIGNED(_iMaxIdxMask_UISA)(%rip){1to16}, %zmm3, %zmm3 > > > > + vpminsd LOCAL_DATA_UNALIGNED(_iMaxIdxMask_UISA)(%rip){1to16}, %zmm3, %zmm3 > > > > > > > > /* Setup permute indices in zmm3. */ > > > > vpsrld $21, %zmm3, %zmm3 > > > > > > > > /* Store if there are any special cases in k1. */ > > > > - vpcmpd $6, TANHF_DATA_UNALIGNED(_iExpMask)(%rip){1to16}, %zmm1, %k1 > > > > + vpcmpd $6, LOCAL_DATA_UNALIGNED(_iExpMask)(%rip){1to16}, %zmm1, %k1 > > > > > > > > - vmovaps TANHF_DATA(_sC_lo)(%rip), %zmm5 > > > > - vpermt2ps TANHF_DATA(_sC_hi)(%rip), %zmm3, %zmm5 > > > > + vmovaps LOCAL_DATA(_sC_lo)(%rip), %zmm5 > > > > + vpermt2ps LOCAL_DATA(_sC_hi)(%rip), %zmm3, %zmm5 > > > > > > > > - vmovaps TANHF_DATA(_sP7_lo)(%rip), %zmm2 > > > > - vpermt2ps TANHF_DATA(_sP7_hi)(%rip), %zmm3, %zmm2 > > > > + vmovaps LOCAL_DATA(_sP7_lo)(%rip), %zmm2 > > > > + vpermt2ps LOCAL_DATA(_sP7_hi)(%rip), %zmm3, %zmm2 > > > > > > > > /* Store absolute values of inputs in zmm1. */ > > > > - vmovaps TANHF_DATA(_sSignMask)(%rip), %zmm4 > > > > + vmovaps COMMON_DATA(_SignMask)(%rip), %zmm4 > > > > vandnps %zmm0, %zmm4, %zmm1 > > > > vsubps {rn-sae}, %zmm5, %zmm1, %zmm1 > > > > > > > > - vmovaps TANHF_DATA(_sP6_lo)(%rip), %zmm5 > > > > - vpermt2ps TANHF_DATA(_sP6_hi)(%rip), %zmm3, %zmm5 > > > > + vmovaps LOCAL_DATA(_sP6_lo)(%rip), %zmm5 > > > > + vpermt2ps LOCAL_DATA(_sP6_hi)(%rip), %zmm3, %zmm5 > > > > > > > > - vmovaps TANHF_DATA(_sP5_lo)(%rip), %zmm6 > > > > - vpermt2ps TANHF_DATA(_sP5_hi)(%rip), %zmm3, %zmm6 > > > > + vmovaps LOCAL_DATA(_sP5_lo)(%rip), %zmm6 > > > > + vpermt2ps LOCAL_DATA(_sP5_hi)(%rip), %zmm3, %zmm6 > > > > > > > > vfmadd213ps {rn-sae}, %zmm5, %zmm1, %zmm2 > > > > vfmadd213ps {rn-sae}, %zmm6, %zmm1, %zmm2 > > > > > > > > - vmovaps TANHF_DATA(_sP4_lo)(%rip), %zmm7 > > > > - vpermt2ps TANHF_DATA(_sP4_hi)(%rip), %zmm3, %zmm7 > > > > + vmovaps LOCAL_DATA(_sP4_lo)(%rip), %zmm7 > > > > + vpermt2ps LOCAL_DATA(_sP4_hi)(%rip), %zmm3, %zmm7 > > > > > > > > - vmovaps TANHF_DATA(_sP3_lo)(%rip), %zmm8 > > > > - vpermt2ps TANHF_DATA(_sP3_hi)(%rip), %zmm3, %zmm8 > > > > + vmovaps LOCAL_DATA(_sP3_lo)(%rip), %zmm8 > > > > + vpermt2ps LOCAL_DATA(_sP3_hi)(%rip), %zmm3, %zmm8 > > > > > > > > vfmadd213ps {rn-sae}, %zmm7, %zmm1, %zmm2 > > > > vfmadd213ps {rn-sae}, %zmm8, %zmm1, %zmm2 > > > > > > > > - vmovaps TANHF_DATA(_sP2_lo)(%rip), %zmm9 > > > > - vpermt2ps TANHF_DATA(_sP2_hi)(%rip), %zmm3, %zmm9 > > > > + vmovaps LOCAL_DATA(_sP2_lo)(%rip), %zmm9 > > > > + vpermt2ps LOCAL_DATA(_sP2_hi)(%rip), %zmm3, %zmm9 > > > > > > > > - vmovaps TANHF_DATA(_sP0_lo)(%rip), %zmm10 > > > > - vpermt2ps TANHF_DATA(_sP0_hi)(%rip), %zmm3, %zmm10 > > > > + vmovaps LOCAL_DATA(_sP0_lo)(%rip), %zmm10 > > > > + vpermt2ps LOCAL_DATA(_sP0_hi)(%rip), %zmm3, %zmm10 > > > > > > > > vfmadd213ps {rn-sae}, %zmm9, %zmm1, %zmm2 > > > > vfmadd213ps {rn-sae}, %zmm10, %zmm1, %zmm2 > > > > @@ -167,7 +172,7 @@ ENTRY(_ZGVeN16v_tanhf_skx) > > > > > > > > /* Go to special inputs processing branch. */ > > > > jne L(SPECIAL_VALUES_BRANCH) > > > > - # LOE rbx r12 r13 r14 r15 zmm0 zmm2 zmm4 > > > > + > > > > /* Wait until after branch of write over zmm0. */ > > > > vpternlogd $0xec, %zmm4, %zmm2, %zmm0 > > > > > > > > @@ -176,24 +181,24 @@ ENTRY(_ZGVeN16v_tanhf_skx) > > > > > > > > /* Cold case. edx has 1s where there was a special value that > > > > needs to be handled by a tanhf call. Optimize for code size > > > > - more so than speed here. */ > > > > + more so than speed here. */ > > > > L(SPECIAL_VALUES_BRANCH): > > > > - # LOE rbx rdx r12 r13 r14 r15 zmm0 zmm2 zmm4 > > > > - /* Use r13 to save/restore the stack. This allows us to use rbp as > > > > - callee save register saving code size. */ > > > > + > > > > + /* Use r13 to save/restore the stack. This allows us to use rbp > > > > + as callee save register saving code size. */ > > > > pushq %r13 > > > > - cfi_adjust_cfa_offset(8) > > > > - cfi_offset(r13, -16) > > > > - /* Need to callee save registers to preserve state across tanhf calls. > > > > - */ > > > > + cfi_adjust_cfa_offset (8) > > > > + cfi_offset (r13, -16) > > > > + /* Need to callee save registers to preserve state across tanhf > > > > + calls. */ > > > > pushq %rbx > > > > - cfi_adjust_cfa_offset(8) > > > > - cfi_offset(rbx, -24) > > > > + cfi_adjust_cfa_offset (8) > > > > + cfi_offset (rbx, -24) > > > > pushq %rbp > > > > - cfi_adjust_cfa_offset(8) > > > > - cfi_offset(rbp, -32) > > > > + cfi_adjust_cfa_offset (8) > > > > + cfi_offset (rbp, -32) > > > > movq %rsp, %r13 > > > > - cfi_def_cfa_register(r13) > > > > + cfi_def_cfa_register (r13) > > > > > > > > /* Align stack and make room for 2x zmm vectors. */ > > > > andq $-64, %rsp > > > > @@ -207,16 +212,17 @@ L(SPECIAL_VALUES_BRANCH): > > > > > > > > vzeroupper > > > > > > > > - /* edx has 1s where there was a special value that needs to be handled > > > > - by a tanhf call. */ > > > > + /* edx has 1s where there was a special value that needs to be > > > > + handled by a tanhf call. */ > > > > movl %edx, %ebx > > > > L(SPECIAL_VALUES_LOOP): > > > > - # LOE rbx rbp r12 r13 r14 r15 > > > > - /* use rbp as index for special value that is saved across calls to > > > > - tanhf. We technically don't need a callee save register here as offset > > > > - to rsp is always [0, 56] so we can restore rsp by realigning to 64. > > > > - Essentially the tradeoff is 1 extra save/restore vs 2 extra instructions > > > > - in the loop. Realigning also costs more code size. */ > > > > + > > > > + /* use rbp as index for special value that is saved across calls > > > > + to tanhf. We technically don't need a callee save register > > > > + here as offset to rsp is always [0, 56] so we can restore > > > > + rsp by realigning to 64. Essentially the tradeoff is 1 extra > > > > + save/restore vs 2 extra instructions in the loop. Realigning > > > > + also costs more code size. */ > > > > xorl %ebp, %ebp > > > > tzcntl %ebx, %ebp > > > > > > > > @@ -224,203 +230,141 @@ L(SPECIAL_VALUES_LOOP): > > > > vmovss 64(%rsp, %rbp, 4), %xmm0 > > > > call tanhf@PLT > > > > > > > > - /* No good way to avoid the store-forwarding fault this will cause on > > > > - return. `lfence` avoids the SF fault but at greater cost as it > > > > - serialized stack/callee save restoration. */ > > > > + /* No good way to avoid the store-forwarding fault this will > > > > + cause on return. `lfence` avoids the SF fault but at greater > > > > + cost as it serialized stack/callee save restoration. */ > > > > vmovss %xmm0, (%rsp, %rbp, 4) > > > > > > > > - blsrl %ebx, %ebx > > > > + blsrl %ebx, %ebx > > > > jnz L(SPECIAL_VALUES_LOOP) > > > > - # LOE r12 r13 r14 r15 > > > > + > > > > > > > > /* All results have been written to (%rsp). */ > > > > vmovaps (%rsp), %zmm0 > > > > /* Restore rsp. */ > > > > movq %r13, %rsp > > > > - cfi_def_cfa_register(rsp) > > > > + cfi_def_cfa_register (rsp) > > > > /* Restore callee save registers. */ > > > > popq %rbp > > > > - cfi_adjust_cfa_offset(-8) > > > > - cfi_restore(rbp) > > > > + cfi_adjust_cfa_offset (-8) > > > > + cfi_restore (rbp) > > > > popq %rbx > > > > - cfi_adjust_cfa_offset(-8) > > > > - cfi_restore(rbp) > > > > + cfi_adjust_cfa_offset (-8) > > > > + cfi_restore (rbp) > > > > popq %r13 > > > > - cfi_adjust_cfa_offset(-8) > > > > - cfi_restore(r13) > > > > + cfi_adjust_cfa_offset (-8) > > > > + cfi_restore (r13) > > > > ret > > > > END(_ZGVeN16v_tanhf_skx) > > > > > > > > - .section .rodata, "a" > > > > + .section .rodata.evex512, "a" > > > > .align 16 > > > > -#ifdef __svml_stanh_data_internal_typedef > > > > -typedef unsigned int VUINT32; > > > > -typedef struct > > > > - { > > > > - __declspec(align(4)) VUINT32 _iExpMantMask_UISA[1][1]; > > > > - __declspec(align(4)) VUINT32 _iMinIdxOfsMask_UISA[1][1]; > > > > - __declspec(align(4)) VUINT32 _iMaxIdxMask_UISA[1][1]; > > > > - __declspec(align(4)) VUINT32 _iExpMask[1][1]; > > > > - __declspec(align(64)) VUINT32 _sC_lo[16][1]; > > > > - __declspec(align(64)) VUINT32 _sC_hi[16][1]; > > > > - __declspec(align(64)) VUINT32 _sP7_lo[16][1]; > > > > - __declspec(align(64)) VUINT32 _sP7_hi[16][1]; > > > > - __declspec(align(64)) VUINT32 _sSignMask[16][1]; > > > > - __declspec(align(64)) VUINT32 _sP6_lo[16][1]; > > > > - __declspec(align(64)) VUINT32 _sP6_hi[16][1]; > > > > - __declspec(align(64)) VUINT32 _sP5_lo[16][1]; > > > > - __declspec(align(64)) VUINT32 _sP5_hi[16][1]; > > > > - __declspec(align(64)) VUINT32 _sP4_lo[16][1]; > > > > - __declspec(align(64)) VUINT32 _sP4_hi[16][1]; > > > > - __declspec(align(64)) VUINT32 _sP3_lo[16][1]; > > > > - __declspec(align(64)) VUINT32 _sP3_hi[16][1]; > > > > - __declspec(align(64)) VUINT32 _sP2_lo[16][1]; > > > > - __declspec(align(64)) VUINT32 _sP2_hi[16][1]; > > > > - __declspec(align(64)) VUINT32 _sP0_lo[16][1]; > > > > - __declspec(align(64)) VUINT32 _sP0_hi[16][1]; > > > > -} __svml_stanh_data_internal; > > > > -#endif > > > > - > > > > -__svml_stanh_data_internal: > > > > - .align 4 > > > > - /* _iExpMantMask_UISA */ > > > > - .long 0x7fe00000 > > > > - > > > > - .align 4 > > > > - /* _iMinIdxOfsMask_UISA */ > > > > - .long 0x3d400000 > > > > - > > > > - .align 4 > > > > - /* _iMaxIdxMask_UISA */ > > > > - .long 0x03e00000 > > > > - > > > > - .align 4 > > > > - /* _iExpMask */ > > > > - .long 0x7f000000 > > > > - > > > > - .align 64 > > > > -__svml_stanh_data_internal_al64: > > > > - .align 64 > > > > - /* _sC_lo */ > > > > - .long 0x00000000, 0x3d700000, 0x3d900000, 0x3db00000 > > > > - .long 0x3dd00000, 0x3df00000, 0x3e100000, 0x3e300000 > > > > - .long 0x3e500000, 0x3e700000, 0x3e900000, 0x3eb00000 > > > > - .long 0x3ed00000, 0x3ef00000, 0x3f100000, 0x3f300000 > > > > - > > > > - .align 64 > > > > - /* _sC_hi */ > > > > - .long 0x3f500000, 0x3f700000, 0x3f900000, 0x3fb00000 > > > > - .long 0x3fd00000, 0x3ff00000, 0x40100000, 0x40300000 > > > > - .long 0x40500000, 0x40700000, 0x40900000, 0x40b00000 > > > > - .long 0x40d00000, 0x40f00000, 0x41100000, 0x00000000 > > > > - > > > > - .align 64 > > > > - /* _sP7_lo */ > > > > - .long 0xbc0e2f66, 0x460bda12, 0x43d638ef, 0xc3e11c3e > > > > - .long 0xc2baa4e9, 0xc249da2d, 0xc1859b82, 0x40dd5b57 > > > > - .long 0x40494640, 0x40c730a8, 0xbf0f160e, 0x3e30e76f > > > > - .long 0xbea81387, 0xbdb26a1c, 0xbd351e57, 0xbb4c01a0 > > > > - > > > > - .align 64 > > > > - /* _sP7_hi */ > > > > - .long 0x3c1d7bfb, 0x3c722cd1, 0x3c973f1c, 0x3c33a31b > > > > - .long 0x3b862ef4, 0x3a27b3d0, 0xba3b5907, 0xba0efc22 > > > > - .long 0xb97f9f0f, 0xb8c8af50, 0xb7bdddfb, 0xb64f2950 > > > > - .long 0xb4e085b1, 0xb3731dfa, 0xb15a1f04, 0x00000000 > > > > > > > > - .align 64 > > > > - /* _sSignMask */ > > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > > - > > > > - .align 64 > > > > - /* _sP6_lo */ > > > > - .long 0x3e0910e9, 0x43761143, 0x4165ecdc, 0xc190f756 > > > > - .long 0xc08c097d, 0xc02ba813, 0xbf7f6bda, 0x3f2b1dc0 > > > > - .long 0x3ece105d, 0x3f426a94, 0xbadb0dc4, 0x3da43b17 > > > > - .long 0xbd51ab88, 0xbcaea23d, 0xbd3b6d8d, 0xbd6caaad > > > > - > > > > - .align 64 > > > > - /* _sP6_hi */ > > > > - .long 0xbd795bed, 0xbd5fddda, 0xbd038f3b, 0xbc1cad63 > > > > - .long 0x3abb4766, 0x3b95f10b, 0x3b825873, 0x3afaea66 > > > > - .long 0x3a49f878, 0x39996bf3, 0x388f3e6c, 0x371bb0e3 > > > > - .long 0x35a8a5e6, 0x34369b17, 0x322487b0, 0x00000000 > > > > - > > > > - .align 64 > > > > - /* _sP5_lo */ > > > > - .long 0xb76dd6b9, 0xbe1c276d, 0x3c1dcf2f, 0x3dc1a78d > > > > - .long 0x3d96f985, 0x3da2b61b, 0x3dc13397, 0x3dd2f670 > > > > - .long 0x3df48a0a, 0x3e06c5a8, 0x3e1a3aba, 0x3e27c405 > > > > - .long 0x3e2e78d0, 0x3e2c3e44, 0x3e1d3097, 0x3df4a8f4 > > > > - > > > > - .align 64 > > > > - /* _sP5_hi */ > > > > - .long 0x3da38508, 0x3d31416a, 0x3b562657, 0xbcaeeac9 > > > > - .long 0xbcce9419, 0xbcaaeac4, 0xbc49e7d0, 0xbba71ddd > > > > - .long 0xbb003b0e, 0xba3f9a05, 0xb92c08a7, 0xb7ba9232 > > > > - .long 0xb64a0b0f, 0xb4dac169, 0xb2ab78ac, 0x00000000 > > > > - > > > > - .align 64 > > > > - /* _sP4_lo */ > > > > - .long 0xbeaaaaa5, 0xbeab0612, 0xbea7f01f, 0xbea4e120 > > > > - .long 0xbea387b7, 0xbea15962, 0xbe9d57f7, 0xbe976b5a > > > > - .long 0xbe90230d, 0xbe880dff, 0xbe7479b3, 0xbe4c3d88 > > > > - .long 0xbe212482, 0xbdeb8cba, 0xbd5e78ad, 0x3c6b5e6e > > > > - > > > > - .align 64 > > > > - /* _sP4_hi */ > > > > - .long 0x3d839143, 0x3dc21ee1, 0x3de347af, 0x3dcbec96 > > > > - .long 0x3d99ef2d, 0x3d542ea1, 0x3cdde701, 0x3c2cca67 > > > > - .long 0x3b81cb27, 0x3ac073a1, 0x39ac3032, 0x383a94d9 > > > > - .long 0x36ca081d, 0x355abd4c, 0x332b3cb6, 0x00000000 > > > > - > > > > - .align 64 > > > > - /* _sP3_lo */ > > > > - .long 0xb0343c7b, 0xbd6ee69d, 0xbd8f0da7, 0xbdae477d > > > > - .long 0xbdcd2a1f, 0xbdeba80d, 0xbe0c443b, 0xbe293cf3 > > > > - .long 0xbe44f282, 0xbe5f3651, 0xbe81c7c0, 0xbe96d7ca > > > > - .long 0xbea7fb8e, 0xbeb50e9e, 0xbec12efe, 0xbec4be92 > > > > - > > > > - .align 64 > > > > - /* _sP3_hi */ > > > > - .long 0xbebce070, 0xbead510e, 0xbe8ef7d6, 0xbe4b8704 > > > > - .long 0xbe083237, 0xbdaf7449, 0xbd2e1ec4, 0xbc83bf06 > > > > - .long 0xbbc3e0b5, 0xbb10aadc, 0xba0157db, 0xb88c18f2 > > > > - .long 0xb717b096, 0xb5a43bae, 0xb383012c, 0x00000000 > > > > - > > > > - .align 64 > > > > - /* _sP2_lo */ > > > > - .long 0x3f800000, 0x3f7f1f84, 0x3f7ebd11, 0x3f7e1e5f > > > > - .long 0x3f7d609f, 0x3f7c842d, 0x3f7b00e5, 0x3f789580 > > > > - .long 0x3f75b8ad, 0x3f726fd9, 0x3f6cc59b, 0x3f63fb92 > > > > - .long 0x3f59ff97, 0x3f4f11d7, 0x3f3d7573, 0x3f24f360 > > > > - > > > > - .align 64 > > > > - /* _sP2_hi */ > > > > - .long 0x3f0cbfe7, 0x3eec1a69, 0x3eb0a801, 0x3e6753a2 > > > > - .long 0x3e132f1a, 0x3db7e7d3, 0x3d320845, 0x3c84d3d4 > > > > - .long 0x3bc477b7, 0x3b10d3da, 0x3a01601e, 0x388c1a3b > > > > - .long 0x3717b0da, 0x35a43bce, 0x338306c6, 0x00000000 > > > > - > > > > - .align 64 > > > > - /* _sP0_lo */ > > > > - .long 0x00000000, 0x3d6fb9c9, 0x3d8fc35f, 0x3daf9169 > > > > - .long 0x3dcf49ab, 0x3deee849, 0x3e0f0ee8, 0x3e2e4984 > > > > - .long 0x3e4d2f8e, 0x3e6bb32e, 0x3e8c51cd, 0x3ea96163 > > > > - .long 0x3ec543f1, 0x3edfd735, 0x3f028438, 0x3f18abf0 > > > > - > > > > - .align 64 > > > > - /* _sP0_hi */ > > > > - .long 0x3f2bc480, 0x3f3bec1c, 0x3f4f2e5b, 0x3f613c53 > > > > - .long 0x3f6ce37d, 0x3f743c4f, 0x3f7a5feb, 0x3f7dea85 > > > > - .long 0x3f7f3b3d, 0x3f7fb78c, 0x3f7fefd4, 0x3f7ffdd0 > > > > - .long 0x3f7fffb4, 0x3f7ffff6, 0x3f7fffff, 0x3f800000 > > > > +LOCAL_DATA_NAME_UNALIGNED: > > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iExpMantMask_UISA, 0x7fe00000) > > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iMinIdxOfsMask_UISA, 0x3d400000) > > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iMaxIdxMask_UISA, 0x03e00000) > > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iExpMask, 0x7f000000) > > > > + .type LOCAL_DATA_NAME_UNALIGNED, @object > > > > + .size LOCAL_DATA_NAME_UNALIGNED, .-LOCAL_DATA_NAME_UNALIGNED > > > > > > > > .align 64 > > > > - .type __svml_stanh_data_internal_al64, @object > > > > - .size __svml_stanh_data_internal_al64, .-__svml_stanh_data_internal_al64 > > > > - .type __svml_stanh_data_internal, @object > > > > - .size __svml_stanh_data_internal, .-__svml_stanh_data_internal > > > > +LOCAL_DATA_NAME: > > > > + float_block (LOCAL_DATA_NAME, _sC_lo, > > > > + 0x00000000, 0x3d700000, 0x3d900000, 0x3db00000, > > > > + 0x3dd00000, 0x3df00000, 0x3e100000, 0x3e300000, > > > > + 0x3e500000, 0x3e700000, 0x3e900000, 0x3eb00000, > > > > + 0x3ed00000, 0x3ef00000, 0x3f100000, 0x3f300000) > > > > + > > > > + float_block (LOCAL_DATA_NAME, _sC_hi, > > > > + 0x3f500000, 0x3f700000, 0x3f900000, 0x3fb00000, > > > > + 0x3fd00000, 0x3ff00000, 0x40100000, 0x40300000, > > > > + 0x40500000, 0x40700000, 0x40900000, 0x40b00000, > > > > + 0x40d00000, 0x40f00000, 0x41100000, 0x00000000) > > > > + > > > > + float_block (LOCAL_DATA_NAME, _sP7_lo, > > > > + 0xbc0e2f66, 0x460bda12, 0x43d638ef, 0xc3e11c3e, > > > > + 0xc2baa4e9, 0xc249da2d, 0xc1859b82, 0x40dd5b57, > > > > + 0x40494640, 0x40c730a8, 0xbf0f160e, 0x3e30e76f, > > > > + 0xbea81387, 0xbdb26a1c, 0xbd351e57, 0xbb4c01a0) > > > > + > > > > + float_block (LOCAL_DATA_NAME, _sP7_hi, > > > > + 0x3c1d7bfb, 0x3c722cd1, 0x3c973f1c, 0x3c33a31b, > > > > + 0x3b862ef4, 0x3a27b3d0, 0xba3b5907, 0xba0efc22, > > > > + 0xb97f9f0f, 0xb8c8af50, 0xb7bdddfb, 0xb64f2950, > > > > + 0xb4e085b1, 0xb3731dfa, 0xb15a1f04, 0x00000000) > > > > + > > > > + float_block (LOCAL_DATA_NAME, _sP6_lo, > > > > + 0x3e0910e9, 0x43761143, 0x4165ecdc, 0xc190f756, > > > > + 0xc08c097d, 0xc02ba813, 0xbf7f6bda, 0x3f2b1dc0, > > > > + 0x3ece105d, 0x3f426a94, 0xbadb0dc4, 0x3da43b17, > > > > + 0xbd51ab88, 0xbcaea23d, 0xbd3b6d8d, 0xbd6caaad) > > > > + > > > > + float_block (LOCAL_DATA_NAME, _sP6_hi, > > > > + 0xbd795bed, 0xbd5fddda, 0xbd038f3b, 0xbc1cad63, > > > > + 0x3abb4766, 0x3b95f10b, 0x3b825873, 0x3afaea66, > > > > + 0x3a49f878, 0x39996bf3, 0x388f3e6c, 0x371bb0e3, > > > > + 0x35a8a5e6, 0x34369b17, 0x322487b0, 0x00000000) > > > > + > > > > + float_block (LOCAL_DATA_NAME, _sP5_lo, > > > > + 0xb76dd6b9, 0xbe1c276d, 0x3c1dcf2f, 0x3dc1a78d, > > > > + 0x3d96f985, 0x3da2b61b, 0x3dc13397, 0x3dd2f670, > > > > + 0x3df48a0a, 0x3e06c5a8, 0x3e1a3aba, 0x3e27c405, > > > > + 0x3e2e78d0, 0x3e2c3e44, 0x3e1d3097, 0x3df4a8f4) > > > > + > > > > + float_block (LOCAL_DATA_NAME, _sP5_hi, > > > > + 0x3da38508, 0x3d31416a, 0x3b562657, 0xbcaeeac9, > > > > + 0xbcce9419, 0xbcaaeac4, 0xbc49e7d0, 0xbba71ddd, > > > > + 0xbb003b0e, 0xba3f9a05, 0xb92c08a7, 0xb7ba9232, > > > > + 0xb64a0b0f, 0xb4dac169, 0xb2ab78ac, 0x00000000) > > > > + > > > > + float_block (LOCAL_DATA_NAME, _sP4_lo, > > > > + 0xbeaaaaa5, 0xbeab0612, 0xbea7f01f, 0xbea4e120, > > > > + 0xbea387b7, 0xbea15962, 0xbe9d57f7, 0xbe976b5a, > > > > + 0xbe90230d, 0xbe880dff, 0xbe7479b3, 0xbe4c3d88, > > > > + 0xbe212482, 0xbdeb8cba, 0xbd5e78ad, 0x3c6b5e6e) > > > > + > > > > + float_block (LOCAL_DATA_NAME, _sP4_hi, > > > > + 0x3d839143, 0x3dc21ee1, 0x3de347af, 0x3dcbec96, > > > > + 0x3d99ef2d, 0x3d542ea1, 0x3cdde701, 0x3c2cca67, > > > > + 0x3b81cb27, 0x3ac073a1, 0x39ac3032, 0x383a94d9, > > > > + 0x36ca081d, 0x355abd4c, 0x332b3cb6, 0x00000000) > > > > + > > > > + float_block (LOCAL_DATA_NAME, _sP3_lo, > > > > + 0xb0343c7b, 0xbd6ee69d, 0xbd8f0da7, 0xbdae477d, > > > > + 0xbdcd2a1f, 0xbdeba80d, 0xbe0c443b, 0xbe293cf3, > > > > + 0xbe44f282, 0xbe5f3651, 0xbe81c7c0, 0xbe96d7ca, > > > > + 0xbea7fb8e, 0xbeb50e9e, 0xbec12efe, 0xbec4be92) > > > > + > > > > + float_block (LOCAL_DATA_NAME, _sP3_hi, > > > > + 0xbebce070, 0xbead510e, 0xbe8ef7d6, 0xbe4b8704, > > > > + 0xbe083237, 0xbdaf7449, 0xbd2e1ec4, 0xbc83bf06, > > > > + 0xbbc3e0b5, 0xbb10aadc, 0xba0157db, 0xb88c18f2, > > > > + 0xb717b096, 0xb5a43bae, 0xb383012c, 0x00000000) > > > > + > > > > + float_block (LOCAL_DATA_NAME, _sP2_lo, > > > > + 0x3f800000, 0x3f7f1f84, 0x3f7ebd11, 0x3f7e1e5f, > > > > + 0x3f7d609f, 0x3f7c842d, 0x3f7b00e5, 0x3f789580, > > > > + 0x3f75b8ad, 0x3f726fd9, 0x3f6cc59b, 0x3f63fb92, > > > > + 0x3f59ff97, 0x3f4f11d7, 0x3f3d7573, 0x3f24f360) > > > > + > > > > + float_block (LOCAL_DATA_NAME, _sP2_hi, > > > > + 0x3f0cbfe7, 0x3eec1a69, 0x3eb0a801, 0x3e6753a2, > > > > + 0x3e132f1a, 0x3db7e7d3, 0x3d320845, 0x3c84d3d4, > > > > + 0x3bc477b7, 0x3b10d3da, 0x3a01601e, 0x388c1a3b, > > > > + 0x3717b0da, 0x35a43bce, 0x338306c6, 0x00000000) > > > > + > > > > + float_block (LOCAL_DATA_NAME, _sP0_lo, > > > > + 0x00000000, 0x3d6fb9c9, 0x3d8fc35f, 0x3daf9169, > > > > + 0x3dcf49ab, 0x3deee849, 0x3e0f0ee8, 0x3e2e4984, > > > > + 0x3e4d2f8e, 0x3e6bb32e, 0x3e8c51cd, 0x3ea96163, > > > > + 0x3ec543f1, 0x3edfd735, 0x3f028438, 0x3f18abf0) > > > > + > > > > + float_block (LOCAL_DATA_NAME, _sP0_hi, > > > > + 0x3f2bc480, 0x3f3bec1c, 0x3f4f2e5b, 0x3f613c53, > > > > + 0x3f6ce37d, 0x3f743c4f, 0x3f7a5feb, 0x3f7dea85, > > > > + 0x3f7f3b3d, 0x3f7fb78c, 0x3f7fefd4, 0x3f7ffdd0, > > > > + 0x3f7fffb4, 0x3f7ffff6, 0x3f7fffff, 0x3f800000) > > > > + > > > > + .type LOCAL_DATA_NAME, @object > > > > + .size LOCAL_DATA_NAME, .-LOCAL_DATA_NAME > > > > -- > > > > 2.34.1 > > > > > > > > > > The data movement makes the assembler codes much harder to follow. > > > Sunil, what do you think of this patch series? > > > > What do you mean? The change on in how we define rodata or the movement > > to multiple files or something else? > > The glibc way to support data files for assembly codes is to define > data in C and use *.sym to generate offsets for assembly files, like I see. Although to be fair the entire SVML codebase bucks that trend. Seems like a more dramatic trend to move all the offsets to C. > > sysdeps/x86/cpu-features-offsets.sym:XSAVE_STATE_SIZE_OFFSET > offsetof (struct cpu_features, xsave_state_size) > sysdeps/x86_64/dl-trampoline.h: sub > _rtld_local_ro+RTLD_GLOBAL_RO_DL_X86_CPU_FEATURES_OFFSET+XSAVE_STATE_SIZE_OFFSET(%rip), > %RSP_LP > sysdeps/x86_64/dl-trampoline.h: sub > _dl_x86_cpu_features+XSAVE_STATE_SIZE_OFFSET(%rip), %RSP_LP > > -- > H.J.
On Fri, Dec 16, 2022 at 1:52 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote: > > On Fri, Dec 16, 2022 at 1:38 PM H.J. Lu <hjl.tools@gmail.com> wrote: > > > > On Fri, Dec 16, 2022 at 10:18 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote: > > > > > > On Fri, Dec 16, 2022 at 9:06 AM H.J. Lu <hjl.tools@gmail.com> wrote: > > > > > > > > On Wed, Dec 7, 2022 at 12:52 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote: > > > > > > > > > > No changes to the logic, just change how rodata is handled. > > > > > > > > > > 1. Define the rodatas using the new macros so they check that the > > > > > offset is correct. > > > > > > > > > > 2. Use common data where applicable. > > > > > --- > > > > > .../multiarch/svml_s_tanhf16_core_avx512.S | 450 ++++++++---------- > > > > > 1 file changed, 197 insertions(+), 253 deletions(-) > > > > > > > > > > diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S b/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > > > > > index d74fc7731d..765e9ed7f7 100644 > > > > > --- a/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > > > > > +++ b/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > > > > > @@ -70,94 +70,99 @@ > > > > > * > > > > > */ > > > > > > > > > > -/* Offsets for data table __svml_stanh_data_internal_avx512. Ordered > > > > > - by use in the function. On cold-starts this might help the > > > > > - prefetcher. Possibly a better idea is to interleave start/end so > > > > > - that the prefetcher is less likely to detect a stream and pull > > > > > - irrelivant lines into cache. */ > > > > > > > > > > -/* Offsets for data table __svml_stanh_data_internal. 4 bytes each. > > > > > - */ > > > > > + > > > > > +#define LOCAL_DATA_NAME __svml_stanh_data_internal > > > > > +#define LOCAL_DATA_NAME_UNALIGNED __svml_stanh_data_internal_unaligned > > > > > +#include "svml_s_common_evex512_rodata_offsets.h" > > > > > + > > > > > +/* Offsets for data table __svml_stanh_data_internal_unaligned. > > > > > + 4 bytes each. */ > > > > > #define _iExpMantMask_UISA 0 > > > > > #define _iMinIdxOfsMask_UISA 4 > > > > > #define _iMaxIdxMask_UISA 8 > > > > > #define _iExpMask 12 > > > > > > > > > > -/* Offsets for data table __svml_stanh_data_internal_al64. 64 bytes > > > > > - each. */ > > > > > -#define _sC_lo 0 > > > > > -#define _sC_hi 64 > > > > > -#define _sP7_lo 128 > > > > > -#define _sP7_hi 192 > > > > > -#define _sSignMask 256 > > > > > -#define _sP6_lo 320 > > > > > -#define _sP6_hi 384 > > > > > -#define _sP5_lo 448 > > > > > -#define _sP5_hi 512 > > > > > -#define _sP4_lo 576 > > > > > -#define _sP4_hi 640 > > > > > -#define _sP3_lo 704 > > > > > -#define _sP3_hi 768 > > > > > -#define _sP2_lo 832 > > > > > -#define _sP2_hi 896 > > > > > -#define _sP0_lo 960 > > > > > -#define _sP0_hi 1024 > > > > > +/* Offsets for data table __svml_stanh_data_internal. Ordered > > > > > + by use in the function. On cold-starts this might help the > > > > > + prefetcher. Possibly a better idea is to interleave start/end so > > > > > + that the prefetcher is less likely to detect a stream and pull > > > > > + irrelivant lines into cache. */ > > > > > + > > > > > +/* Offsets for data table __svml_stanh_data_internal. > > > > > + 64 bytes each. */ > > > > > +#define _sC_lo 0 > > > > > +#define _sC_hi 64 > > > > > +#define _sP7_lo 128 > > > > > +#define _sP7_hi 192 > > > > > +#define _sP6_lo 256 > > > > > +#define _sP6_hi 320 > > > > > +#define _sP5_lo 384 > > > > > +#define _sP5_hi 448 > > > > > +#define _sP4_lo 512 > > > > > +#define _sP4_hi 576 > > > > > +#define _sP3_lo 640 > > > > > +#define _sP3_hi 704 > > > > > +#define _sP2_lo 768 > > > > > +#define _sP2_hi 832 > > > > > +#define _sP0_lo 896 > > > > > +#define _sP0_hi 960 > > > > > + > > > > > > > > > > #include <sysdep.h> > > > > > -#define TANHF_DATA(x) ((x)+__svml_stanh_data_internal_al64) > > > > > -#define TANHF_DATA_UNALIGNED(x) ((x)+__svml_stanh_data_internal) > > > > > > > > > > .section .text.evex512, "ax", @progbits > > > > > ENTRY(_ZGVeN16v_tanhf_skx) > > > > > - /* Here huge arguments, INF and NaNs are filtered out to callout. */ > > > > > - vpandd TANHF_DATA_UNALIGNED(_iExpMantMask_UISA)(%rip){1to16}, %zmm0, %zmm1 > > > > > - vpsubd TANHF_DATA_UNALIGNED(_iMinIdxOfsMask_UISA)(%rip){1to16}, %zmm1, %zmm2 > > > > > + /* Here huge arguments, INF and NaNs are filtered out to > > > > > + callout. */ > > > > > + vpandd LOCAL_DATA_UNALIGNED(_iExpMantMask_UISA)(%rip){1to16}, %zmm0, %zmm1 > > > > > + vpsubd LOCAL_DATA_UNALIGNED(_iMinIdxOfsMask_UISA)(%rip){1to16}, %zmm1, %zmm2 > > > > > > > > > > /* Selection arguments between [0, 0x03e00000] into zmm3. */ > > > > > vpxord %zmm3, %zmm3, %zmm3 > > > > > vpmaxsd %zmm3, %zmm2, %zmm3 > > > > > - vpminsd TANHF_DATA_UNALIGNED(_iMaxIdxMask_UISA)(%rip){1to16}, %zmm3, %zmm3 > > > > > + vpminsd LOCAL_DATA_UNALIGNED(_iMaxIdxMask_UISA)(%rip){1to16}, %zmm3, %zmm3 > > > > > > > > > > /* Setup permute indices in zmm3. */ > > > > > vpsrld $21, %zmm3, %zmm3 > > > > > > > > > > /* Store if there are any special cases in k1. */ > > > > > - vpcmpd $6, TANHF_DATA_UNALIGNED(_iExpMask)(%rip){1to16}, %zmm1, %k1 > > > > > + vpcmpd $6, LOCAL_DATA_UNALIGNED(_iExpMask)(%rip){1to16}, %zmm1, %k1 > > > > > > > > > > - vmovaps TANHF_DATA(_sC_lo)(%rip), %zmm5 > > > > > - vpermt2ps TANHF_DATA(_sC_hi)(%rip), %zmm3, %zmm5 > > > > > + vmovaps LOCAL_DATA(_sC_lo)(%rip), %zmm5 > > > > > + vpermt2ps LOCAL_DATA(_sC_hi)(%rip), %zmm3, %zmm5 > > > > > > > > > > - vmovaps TANHF_DATA(_sP7_lo)(%rip), %zmm2 > > > > > - vpermt2ps TANHF_DATA(_sP7_hi)(%rip), %zmm3, %zmm2 > > > > > + vmovaps LOCAL_DATA(_sP7_lo)(%rip), %zmm2 > > > > > + vpermt2ps LOCAL_DATA(_sP7_hi)(%rip), %zmm3, %zmm2 > > > > > > > > > > /* Store absolute values of inputs in zmm1. */ > > > > > - vmovaps TANHF_DATA(_sSignMask)(%rip), %zmm4 > > > > > + vmovaps COMMON_DATA(_SignMask)(%rip), %zmm4 > > > > > vandnps %zmm0, %zmm4, %zmm1 > > > > > vsubps {rn-sae}, %zmm5, %zmm1, %zmm1 > > > > > > > > > > - vmovaps TANHF_DATA(_sP6_lo)(%rip), %zmm5 > > > > > - vpermt2ps TANHF_DATA(_sP6_hi)(%rip), %zmm3, %zmm5 > > > > > + vmovaps LOCAL_DATA(_sP6_lo)(%rip), %zmm5 > > > > > + vpermt2ps LOCAL_DATA(_sP6_hi)(%rip), %zmm3, %zmm5 > > > > > > > > > > - vmovaps TANHF_DATA(_sP5_lo)(%rip), %zmm6 > > > > > - vpermt2ps TANHF_DATA(_sP5_hi)(%rip), %zmm3, %zmm6 > > > > > + vmovaps LOCAL_DATA(_sP5_lo)(%rip), %zmm6 > > > > > + vpermt2ps LOCAL_DATA(_sP5_hi)(%rip), %zmm3, %zmm6 > > > > > > > > > > vfmadd213ps {rn-sae}, %zmm5, %zmm1, %zmm2 > > > > > vfmadd213ps {rn-sae}, %zmm6, %zmm1, %zmm2 > > > > > > > > > > - vmovaps TANHF_DATA(_sP4_lo)(%rip), %zmm7 > > > > > - vpermt2ps TANHF_DATA(_sP4_hi)(%rip), %zmm3, %zmm7 > > > > > + vmovaps LOCAL_DATA(_sP4_lo)(%rip), %zmm7 > > > > > + vpermt2ps LOCAL_DATA(_sP4_hi)(%rip), %zmm3, %zmm7 > > > > > > > > > > - vmovaps TANHF_DATA(_sP3_lo)(%rip), %zmm8 > > > > > - vpermt2ps TANHF_DATA(_sP3_hi)(%rip), %zmm3, %zmm8 > > > > > + vmovaps LOCAL_DATA(_sP3_lo)(%rip), %zmm8 > > > > > + vpermt2ps LOCAL_DATA(_sP3_hi)(%rip), %zmm3, %zmm8 > > > > > > > > > > vfmadd213ps {rn-sae}, %zmm7, %zmm1, %zmm2 > > > > > vfmadd213ps {rn-sae}, %zmm8, %zmm1, %zmm2 > > > > > > > > > > - vmovaps TANHF_DATA(_sP2_lo)(%rip), %zmm9 > > > > > - vpermt2ps TANHF_DATA(_sP2_hi)(%rip), %zmm3, %zmm9 > > > > > + vmovaps LOCAL_DATA(_sP2_lo)(%rip), %zmm9 > > > > > + vpermt2ps LOCAL_DATA(_sP2_hi)(%rip), %zmm3, %zmm9 > > > > > > > > > > - vmovaps TANHF_DATA(_sP0_lo)(%rip), %zmm10 > > > > > - vpermt2ps TANHF_DATA(_sP0_hi)(%rip), %zmm3, %zmm10 > > > > > + vmovaps LOCAL_DATA(_sP0_lo)(%rip), %zmm10 > > > > > + vpermt2ps LOCAL_DATA(_sP0_hi)(%rip), %zmm3, %zmm10 > > > > > > > > > > vfmadd213ps {rn-sae}, %zmm9, %zmm1, %zmm2 > > > > > vfmadd213ps {rn-sae}, %zmm10, %zmm1, %zmm2 > > > > > @@ -167,7 +172,7 @@ ENTRY(_ZGVeN16v_tanhf_skx) > > > > > > > > > > /* Go to special inputs processing branch. */ > > > > > jne L(SPECIAL_VALUES_BRANCH) > > > > > - # LOE rbx r12 r13 r14 r15 zmm0 zmm2 zmm4 > > > > > + > > > > > /* Wait until after branch of write over zmm0. */ > > > > > vpternlogd $0xec, %zmm4, %zmm2, %zmm0 > > > > > > > > > > @@ -176,24 +181,24 @@ ENTRY(_ZGVeN16v_tanhf_skx) > > > > > > > > > > /* Cold case. edx has 1s where there was a special value that > > > > > needs to be handled by a tanhf call. Optimize for code size > > > > > - more so than speed here. */ > > > > > + more so than speed here. */ > > > > > L(SPECIAL_VALUES_BRANCH): > > > > > - # LOE rbx rdx r12 r13 r14 r15 zmm0 zmm2 zmm4 > > > > > - /* Use r13 to save/restore the stack. This allows us to use rbp as > > > > > - callee save register saving code size. */ > > > > > + > > > > > + /* Use r13 to save/restore the stack. This allows us to use rbp > > > > > + as callee save register saving code size. */ > > > > > pushq %r13 > > > > > - cfi_adjust_cfa_offset(8) > > > > > - cfi_offset(r13, -16) > > > > > - /* Need to callee save registers to preserve state across tanhf calls. > > > > > - */ > > > > > + cfi_adjust_cfa_offset (8) > > > > > + cfi_offset (r13, -16) > > > > > + /* Need to callee save registers to preserve state across tanhf > > > > > + calls. */ > > > > > pushq %rbx > > > > > - cfi_adjust_cfa_offset(8) > > > > > - cfi_offset(rbx, -24) > > > > > + cfi_adjust_cfa_offset (8) > > > > > + cfi_offset (rbx, -24) > > > > > pushq %rbp > > > > > - cfi_adjust_cfa_offset(8) > > > > > - cfi_offset(rbp, -32) > > > > > + cfi_adjust_cfa_offset (8) > > > > > + cfi_offset (rbp, -32) > > > > > movq %rsp, %r13 > > > > > - cfi_def_cfa_register(r13) > > > > > + cfi_def_cfa_register (r13) > > > > > > > > > > /* Align stack and make room for 2x zmm vectors. */ > > > > > andq $-64, %rsp > > > > > @@ -207,16 +212,17 @@ L(SPECIAL_VALUES_BRANCH): > > > > > > > > > > vzeroupper > > > > > > > > > > - /* edx has 1s where there was a special value that needs to be handled > > > > > - by a tanhf call. */ > > > > > + /* edx has 1s where there was a special value that needs to be > > > > > + handled by a tanhf call. */ > > > > > movl %edx, %ebx > > > > > L(SPECIAL_VALUES_LOOP): > > > > > - # LOE rbx rbp r12 r13 r14 r15 > > > > > - /* use rbp as index for special value that is saved across calls to > > > > > - tanhf. We technically don't need a callee save register here as offset > > > > > - to rsp is always [0, 56] so we can restore rsp by realigning to 64. > > > > > - Essentially the tradeoff is 1 extra save/restore vs 2 extra instructions > > > > > - in the loop. Realigning also costs more code size. */ > > > > > + > > > > > + /* use rbp as index for special value that is saved across calls > > > > > + to tanhf. We technically don't need a callee save register > > > > > + here as offset to rsp is always [0, 56] so we can restore > > > > > + rsp by realigning to 64. Essentially the tradeoff is 1 extra > > > > > + save/restore vs 2 extra instructions in the loop. Realigning > > > > > + also costs more code size. */ > > > > > xorl %ebp, %ebp > > > > > tzcntl %ebx, %ebp > > > > > > > > > > @@ -224,203 +230,141 @@ L(SPECIAL_VALUES_LOOP): > > > > > vmovss 64(%rsp, %rbp, 4), %xmm0 > > > > > call tanhf@PLT > > > > > > > > > > - /* No good way to avoid the store-forwarding fault this will cause on > > > > > - return. `lfence` avoids the SF fault but at greater cost as it > > > > > - serialized stack/callee save restoration. */ > > > > > + /* No good way to avoid the store-forwarding fault this will > > > > > + cause on return. `lfence` avoids the SF fault but at greater > > > > > + cost as it serialized stack/callee save restoration. */ > > > > > vmovss %xmm0, (%rsp, %rbp, 4) > > > > > > > > > > - blsrl %ebx, %ebx > > > > > + blsrl %ebx, %ebx > > > > > jnz L(SPECIAL_VALUES_LOOP) > > > > > - # LOE r12 r13 r14 r15 > > > > > + > > > > > > > > > > /* All results have been written to (%rsp). */ > > > > > vmovaps (%rsp), %zmm0 > > > > > /* Restore rsp. */ > > > > > movq %r13, %rsp > > > > > - cfi_def_cfa_register(rsp) > > > > > + cfi_def_cfa_register (rsp) > > > > > /* Restore callee save registers. */ > > > > > popq %rbp > > > > > - cfi_adjust_cfa_offset(-8) > > > > > - cfi_restore(rbp) > > > > > + cfi_adjust_cfa_offset (-8) > > > > > + cfi_restore (rbp) > > > > > popq %rbx > > > > > - cfi_adjust_cfa_offset(-8) > > > > > - cfi_restore(rbp) > > > > > + cfi_adjust_cfa_offset (-8) > > > > > + cfi_restore (rbp) > > > > > popq %r13 > > > > > - cfi_adjust_cfa_offset(-8) > > > > > - cfi_restore(r13) > > > > > + cfi_adjust_cfa_offset (-8) > > > > > + cfi_restore (r13) > > > > > ret > > > > > END(_ZGVeN16v_tanhf_skx) > > > > > > > > > > - .section .rodata, "a" > > > > > + .section .rodata.evex512, "a" > > > > > .align 16 > > > > > -#ifdef __svml_stanh_data_internal_typedef > > > > > -typedef unsigned int VUINT32; > > > > > -typedef struct > > > > > - { > > > > > - __declspec(align(4)) VUINT32 _iExpMantMask_UISA[1][1]; > > > > > - __declspec(align(4)) VUINT32 _iMinIdxOfsMask_UISA[1][1]; > > > > > - __declspec(align(4)) VUINT32 _iMaxIdxMask_UISA[1][1]; > > > > > - __declspec(align(4)) VUINT32 _iExpMask[1][1]; > > > > > - __declspec(align(64)) VUINT32 _sC_lo[16][1]; > > > > > - __declspec(align(64)) VUINT32 _sC_hi[16][1]; > > > > > - __declspec(align(64)) VUINT32 _sP7_lo[16][1]; > > > > > - __declspec(align(64)) VUINT32 _sP7_hi[16][1]; > > > > > - __declspec(align(64)) VUINT32 _sSignMask[16][1]; > > > > > - __declspec(align(64)) VUINT32 _sP6_lo[16][1]; > > > > > - __declspec(align(64)) VUINT32 _sP6_hi[16][1]; > > > > > - __declspec(align(64)) VUINT32 _sP5_lo[16][1]; > > > > > - __declspec(align(64)) VUINT32 _sP5_hi[16][1]; > > > > > - __declspec(align(64)) VUINT32 _sP4_lo[16][1]; > > > > > - __declspec(align(64)) VUINT32 _sP4_hi[16][1]; > > > > > - __declspec(align(64)) VUINT32 _sP3_lo[16][1]; > > > > > - __declspec(align(64)) VUINT32 _sP3_hi[16][1]; > > > > > - __declspec(align(64)) VUINT32 _sP2_lo[16][1]; > > > > > - __declspec(align(64)) VUINT32 _sP2_hi[16][1]; > > > > > - __declspec(align(64)) VUINT32 _sP0_lo[16][1]; > > > > > - __declspec(align(64)) VUINT32 _sP0_hi[16][1]; > > > > > -} __svml_stanh_data_internal; > > > > > -#endif > > > > > - > > > > > -__svml_stanh_data_internal: > > > > > - .align 4 > > > > > - /* _iExpMantMask_UISA */ > > > > > - .long 0x7fe00000 > > > > > - > > > > > - .align 4 > > > > > - /* _iMinIdxOfsMask_UISA */ > > > > > - .long 0x3d400000 > > > > > - > > > > > - .align 4 > > > > > - /* _iMaxIdxMask_UISA */ > > > > > - .long 0x03e00000 > > > > > - > > > > > - .align 4 > > > > > - /* _iExpMask */ > > > > > - .long 0x7f000000 > > > > > - > > > > > - .align 64 > > > > > -__svml_stanh_data_internal_al64: > > > > > - .align 64 > > > > > - /* _sC_lo */ > > > > > - .long 0x00000000, 0x3d700000, 0x3d900000, 0x3db00000 > > > > > - .long 0x3dd00000, 0x3df00000, 0x3e100000, 0x3e300000 > > > > > - .long 0x3e500000, 0x3e700000, 0x3e900000, 0x3eb00000 > > > > > - .long 0x3ed00000, 0x3ef00000, 0x3f100000, 0x3f300000 > > > > > - > > > > > - .align 64 > > > > > - /* _sC_hi */ > > > > > - .long 0x3f500000, 0x3f700000, 0x3f900000, 0x3fb00000 > > > > > - .long 0x3fd00000, 0x3ff00000, 0x40100000, 0x40300000 > > > > > - .long 0x40500000, 0x40700000, 0x40900000, 0x40b00000 > > > > > - .long 0x40d00000, 0x40f00000, 0x41100000, 0x00000000 > > > > > - > > > > > - .align 64 > > > > > - /* _sP7_lo */ > > > > > - .long 0xbc0e2f66, 0x460bda12, 0x43d638ef, 0xc3e11c3e > > > > > - .long 0xc2baa4e9, 0xc249da2d, 0xc1859b82, 0x40dd5b57 > > > > > - .long 0x40494640, 0x40c730a8, 0xbf0f160e, 0x3e30e76f > > > > > - .long 0xbea81387, 0xbdb26a1c, 0xbd351e57, 0xbb4c01a0 > > > > > - > > > > > - .align 64 > > > > > - /* _sP7_hi */ > > > > > - .long 0x3c1d7bfb, 0x3c722cd1, 0x3c973f1c, 0x3c33a31b > > > > > - .long 0x3b862ef4, 0x3a27b3d0, 0xba3b5907, 0xba0efc22 > > > > > - .long 0xb97f9f0f, 0xb8c8af50, 0xb7bdddfb, 0xb64f2950 > > > > > - .long 0xb4e085b1, 0xb3731dfa, 0xb15a1f04, 0x00000000 > > > > > > > > > > - .align 64 > > > > > - /* _sSignMask */ > > > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > > > - > > > > > - .align 64 > > > > > - /* _sP6_lo */ > > > > > - .long 0x3e0910e9, 0x43761143, 0x4165ecdc, 0xc190f756 > > > > > - .long 0xc08c097d, 0xc02ba813, 0xbf7f6bda, 0x3f2b1dc0 > > > > > - .long 0x3ece105d, 0x3f426a94, 0xbadb0dc4, 0x3da43b17 > > > > > - .long 0xbd51ab88, 0xbcaea23d, 0xbd3b6d8d, 0xbd6caaad > > > > > - > > > > > - .align 64 > > > > > - /* _sP6_hi */ > > > > > - .long 0xbd795bed, 0xbd5fddda, 0xbd038f3b, 0xbc1cad63 > > > > > - .long 0x3abb4766, 0x3b95f10b, 0x3b825873, 0x3afaea66 > > > > > - .long 0x3a49f878, 0x39996bf3, 0x388f3e6c, 0x371bb0e3 > > > > > - .long 0x35a8a5e6, 0x34369b17, 0x322487b0, 0x00000000 > > > > > - > > > > > - .align 64 > > > > > - /* _sP5_lo */ > > > > > - .long 0xb76dd6b9, 0xbe1c276d, 0x3c1dcf2f, 0x3dc1a78d > > > > > - .long 0x3d96f985, 0x3da2b61b, 0x3dc13397, 0x3dd2f670 > > > > > - .long 0x3df48a0a, 0x3e06c5a8, 0x3e1a3aba, 0x3e27c405 > > > > > - .long 0x3e2e78d0, 0x3e2c3e44, 0x3e1d3097, 0x3df4a8f4 > > > > > - > > > > > - .align 64 > > > > > - /* _sP5_hi */ > > > > > - .long 0x3da38508, 0x3d31416a, 0x3b562657, 0xbcaeeac9 > > > > > - .long 0xbcce9419, 0xbcaaeac4, 0xbc49e7d0, 0xbba71ddd > > > > > - .long 0xbb003b0e, 0xba3f9a05, 0xb92c08a7, 0xb7ba9232 > > > > > - .long 0xb64a0b0f, 0xb4dac169, 0xb2ab78ac, 0x00000000 > > > > > - > > > > > - .align 64 > > > > > - /* _sP4_lo */ > > > > > - .long 0xbeaaaaa5, 0xbeab0612, 0xbea7f01f, 0xbea4e120 > > > > > - .long 0xbea387b7, 0xbea15962, 0xbe9d57f7, 0xbe976b5a > > > > > - .long 0xbe90230d, 0xbe880dff, 0xbe7479b3, 0xbe4c3d88 > > > > > - .long 0xbe212482, 0xbdeb8cba, 0xbd5e78ad, 0x3c6b5e6e > > > > > - > > > > > - .align 64 > > > > > - /* _sP4_hi */ > > > > > - .long 0x3d839143, 0x3dc21ee1, 0x3de347af, 0x3dcbec96 > > > > > - .long 0x3d99ef2d, 0x3d542ea1, 0x3cdde701, 0x3c2cca67 > > > > > - .long 0x3b81cb27, 0x3ac073a1, 0x39ac3032, 0x383a94d9 > > > > > - .long 0x36ca081d, 0x355abd4c, 0x332b3cb6, 0x00000000 > > > > > - > > > > > - .align 64 > > > > > - /* _sP3_lo */ > > > > > - .long 0xb0343c7b, 0xbd6ee69d, 0xbd8f0da7, 0xbdae477d > > > > > - .long 0xbdcd2a1f, 0xbdeba80d, 0xbe0c443b, 0xbe293cf3 > > > > > - .long 0xbe44f282, 0xbe5f3651, 0xbe81c7c0, 0xbe96d7ca > > > > > - .long 0xbea7fb8e, 0xbeb50e9e, 0xbec12efe, 0xbec4be92 > > > > > - > > > > > - .align 64 > > > > > - /* _sP3_hi */ > > > > > - .long 0xbebce070, 0xbead510e, 0xbe8ef7d6, 0xbe4b8704 > > > > > - .long 0xbe083237, 0xbdaf7449, 0xbd2e1ec4, 0xbc83bf06 > > > > > - .long 0xbbc3e0b5, 0xbb10aadc, 0xba0157db, 0xb88c18f2 > > > > > - .long 0xb717b096, 0xb5a43bae, 0xb383012c, 0x00000000 > > > > > - > > > > > - .align 64 > > > > > - /* _sP2_lo */ > > > > > - .long 0x3f800000, 0x3f7f1f84, 0x3f7ebd11, 0x3f7e1e5f > > > > > - .long 0x3f7d609f, 0x3f7c842d, 0x3f7b00e5, 0x3f789580 > > > > > - .long 0x3f75b8ad, 0x3f726fd9, 0x3f6cc59b, 0x3f63fb92 > > > > > - .long 0x3f59ff97, 0x3f4f11d7, 0x3f3d7573, 0x3f24f360 > > > > > - > > > > > - .align 64 > > > > > - /* _sP2_hi */ > > > > > - .long 0x3f0cbfe7, 0x3eec1a69, 0x3eb0a801, 0x3e6753a2 > > > > > - .long 0x3e132f1a, 0x3db7e7d3, 0x3d320845, 0x3c84d3d4 > > > > > - .long 0x3bc477b7, 0x3b10d3da, 0x3a01601e, 0x388c1a3b > > > > > - .long 0x3717b0da, 0x35a43bce, 0x338306c6, 0x00000000 > > > > > - > > > > > - .align 64 > > > > > - /* _sP0_lo */ > > > > > - .long 0x00000000, 0x3d6fb9c9, 0x3d8fc35f, 0x3daf9169 > > > > > - .long 0x3dcf49ab, 0x3deee849, 0x3e0f0ee8, 0x3e2e4984 > > > > > - .long 0x3e4d2f8e, 0x3e6bb32e, 0x3e8c51cd, 0x3ea96163 > > > > > - .long 0x3ec543f1, 0x3edfd735, 0x3f028438, 0x3f18abf0 > > > > > - > > > > > - .align 64 > > > > > - /* _sP0_hi */ > > > > > - .long 0x3f2bc480, 0x3f3bec1c, 0x3f4f2e5b, 0x3f613c53 > > > > > - .long 0x3f6ce37d, 0x3f743c4f, 0x3f7a5feb, 0x3f7dea85 > > > > > - .long 0x3f7f3b3d, 0x3f7fb78c, 0x3f7fefd4, 0x3f7ffdd0 > > > > > - .long 0x3f7fffb4, 0x3f7ffff6, 0x3f7fffff, 0x3f800000 > > > > > +LOCAL_DATA_NAME_UNALIGNED: > > > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iExpMantMask_UISA, 0x7fe00000) > > > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iMinIdxOfsMask_UISA, 0x3d400000) > > > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iMaxIdxMask_UISA, 0x03e00000) > > > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iExpMask, 0x7f000000) > > > > > + .type LOCAL_DATA_NAME_UNALIGNED, @object > > > > > + .size LOCAL_DATA_NAME_UNALIGNED, .-LOCAL_DATA_NAME_UNALIGNED > > > > > > > > > > .align 64 > > > > > - .type __svml_stanh_data_internal_al64, @object > > > > > - .size __svml_stanh_data_internal_al64, .-__svml_stanh_data_internal_al64 > > > > > - .type __svml_stanh_data_internal, @object > > > > > - .size __svml_stanh_data_internal, .-__svml_stanh_data_internal > > > > > +LOCAL_DATA_NAME: > > > > > + float_block (LOCAL_DATA_NAME, _sC_lo, > > > > > + 0x00000000, 0x3d700000, 0x3d900000, 0x3db00000, > > > > > + 0x3dd00000, 0x3df00000, 0x3e100000, 0x3e300000, > > > > > + 0x3e500000, 0x3e700000, 0x3e900000, 0x3eb00000, > > > > > + 0x3ed00000, 0x3ef00000, 0x3f100000, 0x3f300000) > > > > > + > > > > > + float_block (LOCAL_DATA_NAME, _sC_hi, > > > > > + 0x3f500000, 0x3f700000, 0x3f900000, 0x3fb00000, > > > > > + 0x3fd00000, 0x3ff00000, 0x40100000, 0x40300000, > > > > > + 0x40500000, 0x40700000, 0x40900000, 0x40b00000, > > > > > + 0x40d00000, 0x40f00000, 0x41100000, 0x00000000) > > > > > + > > > > > + float_block (LOCAL_DATA_NAME, _sP7_lo, > > > > > + 0xbc0e2f66, 0x460bda12, 0x43d638ef, 0xc3e11c3e, > > > > > + 0xc2baa4e9, 0xc249da2d, 0xc1859b82, 0x40dd5b57, > > > > > + 0x40494640, 0x40c730a8, 0xbf0f160e, 0x3e30e76f, > > > > > + 0xbea81387, 0xbdb26a1c, 0xbd351e57, 0xbb4c01a0) > > > > > + > > > > > + float_block (LOCAL_DATA_NAME, _sP7_hi, > > > > > + 0x3c1d7bfb, 0x3c722cd1, 0x3c973f1c, 0x3c33a31b, > > > > > + 0x3b862ef4, 0x3a27b3d0, 0xba3b5907, 0xba0efc22, > > > > > + 0xb97f9f0f, 0xb8c8af50, 0xb7bdddfb, 0xb64f2950, > > > > > + 0xb4e085b1, 0xb3731dfa, 0xb15a1f04, 0x00000000) > > > > > + > > > > > + float_block (LOCAL_DATA_NAME, _sP6_lo, > > > > > + 0x3e0910e9, 0x43761143, 0x4165ecdc, 0xc190f756, > > > > > + 0xc08c097d, 0xc02ba813, 0xbf7f6bda, 0x3f2b1dc0, > > > > > + 0x3ece105d, 0x3f426a94, 0xbadb0dc4, 0x3da43b17, > > > > > + 0xbd51ab88, 0xbcaea23d, 0xbd3b6d8d, 0xbd6caaad) > > > > > + > > > > > + float_block (LOCAL_DATA_NAME, _sP6_hi, > > > > > + 0xbd795bed, 0xbd5fddda, 0xbd038f3b, 0xbc1cad63, > > > > > + 0x3abb4766, 0x3b95f10b, 0x3b825873, 0x3afaea66, > > > > > + 0x3a49f878, 0x39996bf3, 0x388f3e6c, 0x371bb0e3, > > > > > + 0x35a8a5e6, 0x34369b17, 0x322487b0, 0x00000000) > > > > > + > > > > > + float_block (LOCAL_DATA_NAME, _sP5_lo, > > > > > + 0xb76dd6b9, 0xbe1c276d, 0x3c1dcf2f, 0x3dc1a78d, > > > > > + 0x3d96f985, 0x3da2b61b, 0x3dc13397, 0x3dd2f670, > > > > > + 0x3df48a0a, 0x3e06c5a8, 0x3e1a3aba, 0x3e27c405, > > > > > + 0x3e2e78d0, 0x3e2c3e44, 0x3e1d3097, 0x3df4a8f4) > > > > > + > > > > > + float_block (LOCAL_DATA_NAME, _sP5_hi, > > > > > + 0x3da38508, 0x3d31416a, 0x3b562657, 0xbcaeeac9, > > > > > + 0xbcce9419, 0xbcaaeac4, 0xbc49e7d0, 0xbba71ddd, > > > > > + 0xbb003b0e, 0xba3f9a05, 0xb92c08a7, 0xb7ba9232, > > > > > + 0xb64a0b0f, 0xb4dac169, 0xb2ab78ac, 0x00000000) > > > > > + > > > > > + float_block (LOCAL_DATA_NAME, _sP4_lo, > > > > > + 0xbeaaaaa5, 0xbeab0612, 0xbea7f01f, 0xbea4e120, > > > > > + 0xbea387b7, 0xbea15962, 0xbe9d57f7, 0xbe976b5a, > > > > > + 0xbe90230d, 0xbe880dff, 0xbe7479b3, 0xbe4c3d88, > > > > > + 0xbe212482, 0xbdeb8cba, 0xbd5e78ad, 0x3c6b5e6e) > > > > > + > > > > > + float_block (LOCAL_DATA_NAME, _sP4_hi, > > > > > + 0x3d839143, 0x3dc21ee1, 0x3de347af, 0x3dcbec96, > > > > > + 0x3d99ef2d, 0x3d542ea1, 0x3cdde701, 0x3c2cca67, > > > > > + 0x3b81cb27, 0x3ac073a1, 0x39ac3032, 0x383a94d9, > > > > > + 0x36ca081d, 0x355abd4c, 0x332b3cb6, 0x00000000) > > > > > + > > > > > + float_block (LOCAL_DATA_NAME, _sP3_lo, > > > > > + 0xb0343c7b, 0xbd6ee69d, 0xbd8f0da7, 0xbdae477d, > > > > > + 0xbdcd2a1f, 0xbdeba80d, 0xbe0c443b, 0xbe293cf3, > > > > > + 0xbe44f282, 0xbe5f3651, 0xbe81c7c0, 0xbe96d7ca, > > > > > + 0xbea7fb8e, 0xbeb50e9e, 0xbec12efe, 0xbec4be92) > > > > > + > > > > > + float_block (LOCAL_DATA_NAME, _sP3_hi, > > > > > + 0xbebce070, 0xbead510e, 0xbe8ef7d6, 0xbe4b8704, > > > > > + 0xbe083237, 0xbdaf7449, 0xbd2e1ec4, 0xbc83bf06, > > > > > + 0xbbc3e0b5, 0xbb10aadc, 0xba0157db, 0xb88c18f2, > > > > > + 0xb717b096, 0xb5a43bae, 0xb383012c, 0x00000000) > > > > > + > > > > > + float_block (LOCAL_DATA_NAME, _sP2_lo, > > > > > + 0x3f800000, 0x3f7f1f84, 0x3f7ebd11, 0x3f7e1e5f, > > > > > + 0x3f7d609f, 0x3f7c842d, 0x3f7b00e5, 0x3f789580, > > > > > + 0x3f75b8ad, 0x3f726fd9, 0x3f6cc59b, 0x3f63fb92, > > > > > + 0x3f59ff97, 0x3f4f11d7, 0x3f3d7573, 0x3f24f360) > > > > > + > > > > > + float_block (LOCAL_DATA_NAME, _sP2_hi, > > > > > + 0x3f0cbfe7, 0x3eec1a69, 0x3eb0a801, 0x3e6753a2, > > > > > + 0x3e132f1a, 0x3db7e7d3, 0x3d320845, 0x3c84d3d4, > > > > > + 0x3bc477b7, 0x3b10d3da, 0x3a01601e, 0x388c1a3b, > > > > > + 0x3717b0da, 0x35a43bce, 0x338306c6, 0x00000000) > > > > > + > > > > > + float_block (LOCAL_DATA_NAME, _sP0_lo, > > > > > + 0x00000000, 0x3d6fb9c9, 0x3d8fc35f, 0x3daf9169, > > > > > + 0x3dcf49ab, 0x3deee849, 0x3e0f0ee8, 0x3e2e4984, > > > > > + 0x3e4d2f8e, 0x3e6bb32e, 0x3e8c51cd, 0x3ea96163, > > > > > + 0x3ec543f1, 0x3edfd735, 0x3f028438, 0x3f18abf0) > > > > > + > > > > > + float_block (LOCAL_DATA_NAME, _sP0_hi, > > > > > + 0x3f2bc480, 0x3f3bec1c, 0x3f4f2e5b, 0x3f613c53, > > > > > + 0x3f6ce37d, 0x3f743c4f, 0x3f7a5feb, 0x3f7dea85, > > > > > + 0x3f7f3b3d, 0x3f7fb78c, 0x3f7fefd4, 0x3f7ffdd0, > > > > > + 0x3f7fffb4, 0x3f7ffff6, 0x3f7fffff, 0x3f800000) > > > > > + > > > > > + .type LOCAL_DATA_NAME, @object > > > > > + .size LOCAL_DATA_NAME, .-LOCAL_DATA_NAME > > > > > -- > > > > > 2.34.1 > > > > > > > > > > > > > The data movement makes the assembler codes much harder to follow. > > > > Sunil, what do you think of this patch series? > > > > > > What do you mean? The change on in how we define rodata or the movement > > > to multiple files or something else? > > > > The glibc way to support data files for assembly codes is to define > > data in C and use *.sym to generate offsets for assembly files, like > > I see. Although to be fair the entire SVML codebase bucks that trend. It is because libmvec codes were generated by ICC and processed by scripts. > Seems like a more dramatic trend to move all the offsets to C. Since you are adding data by hand, you should do it in C. > > > > sysdeps/x86/cpu-features-offsets.sym:XSAVE_STATE_SIZE_OFFSET > > offsetof (struct cpu_features, xsave_state_size) > > sysdeps/x86_64/dl-trampoline.h: sub > > _rtld_local_ro+RTLD_GLOBAL_RO_DL_X86_CPU_FEATURES_OFFSET+XSAVE_STATE_SIZE_OFFSET(%rip), > > %RSP_LP > > sysdeps/x86_64/dl-trampoline.h: sub > > _dl_x86_cpu_features+XSAVE_STATE_SIZE_OFFSET(%rip), %RSP_LP > > > > -- > > H.J.
On Fri, Dec 16, 2022 at 2:01 PM H.J. Lu <hjl.tools@gmail.com> wrote: > > On Fri, Dec 16, 2022 at 1:52 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote: > > > > On Fri, Dec 16, 2022 at 1:38 PM H.J. Lu <hjl.tools@gmail.com> wrote: > > > > > > On Fri, Dec 16, 2022 at 10:18 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote: > > > > > > > > On Fri, Dec 16, 2022 at 9:06 AM H.J. Lu <hjl.tools@gmail.com> wrote: > > > > > > > > > > On Wed, Dec 7, 2022 at 12:52 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote: > > > > > > > > > > > > No changes to the logic, just change how rodata is handled. > > > > > > > > > > > > 1. Define the rodatas using the new macros so they check that the > > > > > > offset is correct. > > > > > > > > > > > > 2. Use common data where applicable. > > > > > > --- > > > > > > .../multiarch/svml_s_tanhf16_core_avx512.S | 450 ++++++++---------- > > > > > > 1 file changed, 197 insertions(+), 253 deletions(-) > > > > > > > > > > > > diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S b/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > > > > > > index d74fc7731d..765e9ed7f7 100644 > > > > > > --- a/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > > > > > > +++ b/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > > > > > > @@ -70,94 +70,99 @@ > > > > > > * > > > > > > */ > > > > > > > > > > > > -/* Offsets for data table __svml_stanh_data_internal_avx512. Ordered > > > > > > - by use in the function. On cold-starts this might help the > > > > > > - prefetcher. Possibly a better idea is to interleave start/end so > > > > > > - that the prefetcher is less likely to detect a stream and pull > > > > > > - irrelivant lines into cache. */ > > > > > > > > > > > > -/* Offsets for data table __svml_stanh_data_internal. 4 bytes each. > > > > > > - */ > > > > > > + > > > > > > +#define LOCAL_DATA_NAME __svml_stanh_data_internal > > > > > > +#define LOCAL_DATA_NAME_UNALIGNED __svml_stanh_data_internal_unaligned > > > > > > +#include "svml_s_common_evex512_rodata_offsets.h" > > > > > > + > > > > > > +/* Offsets for data table __svml_stanh_data_internal_unaligned. > > > > > > + 4 bytes each. */ > > > > > > #define _iExpMantMask_UISA 0 > > > > > > #define _iMinIdxOfsMask_UISA 4 > > > > > > #define _iMaxIdxMask_UISA 8 > > > > > > #define _iExpMask 12 > > > > > > > > > > > > -/* Offsets for data table __svml_stanh_data_internal_al64. 64 bytes > > > > > > - each. */ > > > > > > -#define _sC_lo 0 > > > > > > -#define _sC_hi 64 > > > > > > -#define _sP7_lo 128 > > > > > > -#define _sP7_hi 192 > > > > > > -#define _sSignMask 256 > > > > > > -#define _sP6_lo 320 > > > > > > -#define _sP6_hi 384 > > > > > > -#define _sP5_lo 448 > > > > > > -#define _sP5_hi 512 > > > > > > -#define _sP4_lo 576 > > > > > > -#define _sP4_hi 640 > > > > > > -#define _sP3_lo 704 > > > > > > -#define _sP3_hi 768 > > > > > > -#define _sP2_lo 832 > > > > > > -#define _sP2_hi 896 > > > > > > -#define _sP0_lo 960 > > > > > > -#define _sP0_hi 1024 > > > > > > +/* Offsets for data table __svml_stanh_data_internal. Ordered > > > > > > + by use in the function. On cold-starts this might help the > > > > > > + prefetcher. Possibly a better idea is to interleave start/end so > > > > > > + that the prefetcher is less likely to detect a stream and pull > > > > > > + irrelivant lines into cache. */ > > > > > > + > > > > > > +/* Offsets for data table __svml_stanh_data_internal. > > > > > > + 64 bytes each. */ > > > > > > +#define _sC_lo 0 > > > > > > +#define _sC_hi 64 > > > > > > +#define _sP7_lo 128 > > > > > > +#define _sP7_hi 192 > > > > > > +#define _sP6_lo 256 > > > > > > +#define _sP6_hi 320 > > > > > > +#define _sP5_lo 384 > > > > > > +#define _sP5_hi 448 > > > > > > +#define _sP4_lo 512 > > > > > > +#define _sP4_hi 576 > > > > > > +#define _sP3_lo 640 > > > > > > +#define _sP3_hi 704 > > > > > > +#define _sP2_lo 768 > > > > > > +#define _sP2_hi 832 > > > > > > +#define _sP0_lo 896 > > > > > > +#define _sP0_hi 960 > > > > > > + > > > > > > > > > > > > #include <sysdep.h> > > > > > > -#define TANHF_DATA(x) ((x)+__svml_stanh_data_internal_al64) > > > > > > -#define TANHF_DATA_UNALIGNED(x) ((x)+__svml_stanh_data_internal) > > > > > > > > > > > > .section .text.evex512, "ax", @progbits > > > > > > ENTRY(_ZGVeN16v_tanhf_skx) > > > > > > - /* Here huge arguments, INF and NaNs are filtered out to callout. */ > > > > > > - vpandd TANHF_DATA_UNALIGNED(_iExpMantMask_UISA)(%rip){1to16}, %zmm0, %zmm1 > > > > > > - vpsubd TANHF_DATA_UNALIGNED(_iMinIdxOfsMask_UISA)(%rip){1to16}, %zmm1, %zmm2 > > > > > > + /* Here huge arguments, INF and NaNs are filtered out to > > > > > > + callout. */ > > > > > > + vpandd LOCAL_DATA_UNALIGNED(_iExpMantMask_UISA)(%rip){1to16}, %zmm0, %zmm1 > > > > > > + vpsubd LOCAL_DATA_UNALIGNED(_iMinIdxOfsMask_UISA)(%rip){1to16}, %zmm1, %zmm2 > > > > > > > > > > > > /* Selection arguments between [0, 0x03e00000] into zmm3. */ > > > > > > vpxord %zmm3, %zmm3, %zmm3 > > > > > > vpmaxsd %zmm3, %zmm2, %zmm3 > > > > > > - vpminsd TANHF_DATA_UNALIGNED(_iMaxIdxMask_UISA)(%rip){1to16}, %zmm3, %zmm3 > > > > > > + vpminsd LOCAL_DATA_UNALIGNED(_iMaxIdxMask_UISA)(%rip){1to16}, %zmm3, %zmm3 > > > > > > > > > > > > /* Setup permute indices in zmm3. */ > > > > > > vpsrld $21, %zmm3, %zmm3 > > > > > > > > > > > > /* Store if there are any special cases in k1. */ > > > > > > - vpcmpd $6, TANHF_DATA_UNALIGNED(_iExpMask)(%rip){1to16}, %zmm1, %k1 > > > > > > + vpcmpd $6, LOCAL_DATA_UNALIGNED(_iExpMask)(%rip){1to16}, %zmm1, %k1 > > > > > > > > > > > > - vmovaps TANHF_DATA(_sC_lo)(%rip), %zmm5 > > > > > > - vpermt2ps TANHF_DATA(_sC_hi)(%rip), %zmm3, %zmm5 > > > > > > + vmovaps LOCAL_DATA(_sC_lo)(%rip), %zmm5 > > > > > > + vpermt2ps LOCAL_DATA(_sC_hi)(%rip), %zmm3, %zmm5 > > > > > > > > > > > > - vmovaps TANHF_DATA(_sP7_lo)(%rip), %zmm2 > > > > > > - vpermt2ps TANHF_DATA(_sP7_hi)(%rip), %zmm3, %zmm2 > > > > > > + vmovaps LOCAL_DATA(_sP7_lo)(%rip), %zmm2 > > > > > > + vpermt2ps LOCAL_DATA(_sP7_hi)(%rip), %zmm3, %zmm2 > > > > > > > > > > > > /* Store absolute values of inputs in zmm1. */ > > > > > > - vmovaps TANHF_DATA(_sSignMask)(%rip), %zmm4 > > > > > > + vmovaps COMMON_DATA(_SignMask)(%rip), %zmm4 > > > > > > vandnps %zmm0, %zmm4, %zmm1 > > > > > > vsubps {rn-sae}, %zmm5, %zmm1, %zmm1 > > > > > > > > > > > > - vmovaps TANHF_DATA(_sP6_lo)(%rip), %zmm5 > > > > > > - vpermt2ps TANHF_DATA(_sP6_hi)(%rip), %zmm3, %zmm5 > > > > > > + vmovaps LOCAL_DATA(_sP6_lo)(%rip), %zmm5 > > > > > > + vpermt2ps LOCAL_DATA(_sP6_hi)(%rip), %zmm3, %zmm5 > > > > > > > > > > > > - vmovaps TANHF_DATA(_sP5_lo)(%rip), %zmm6 > > > > > > - vpermt2ps TANHF_DATA(_sP5_hi)(%rip), %zmm3, %zmm6 > > > > > > + vmovaps LOCAL_DATA(_sP5_lo)(%rip), %zmm6 > > > > > > + vpermt2ps LOCAL_DATA(_sP5_hi)(%rip), %zmm3, %zmm6 > > > > > > > > > > > > vfmadd213ps {rn-sae}, %zmm5, %zmm1, %zmm2 > > > > > > vfmadd213ps {rn-sae}, %zmm6, %zmm1, %zmm2 > > > > > > > > > > > > - vmovaps TANHF_DATA(_sP4_lo)(%rip), %zmm7 > > > > > > - vpermt2ps TANHF_DATA(_sP4_hi)(%rip), %zmm3, %zmm7 > > > > > > + vmovaps LOCAL_DATA(_sP4_lo)(%rip), %zmm7 > > > > > > + vpermt2ps LOCAL_DATA(_sP4_hi)(%rip), %zmm3, %zmm7 > > > > > > > > > > > > - vmovaps TANHF_DATA(_sP3_lo)(%rip), %zmm8 > > > > > > - vpermt2ps TANHF_DATA(_sP3_hi)(%rip), %zmm3, %zmm8 > > > > > > + vmovaps LOCAL_DATA(_sP3_lo)(%rip), %zmm8 > > > > > > + vpermt2ps LOCAL_DATA(_sP3_hi)(%rip), %zmm3, %zmm8 > > > > > > > > > > > > vfmadd213ps {rn-sae}, %zmm7, %zmm1, %zmm2 > > > > > > vfmadd213ps {rn-sae}, %zmm8, %zmm1, %zmm2 > > > > > > > > > > > > - vmovaps TANHF_DATA(_sP2_lo)(%rip), %zmm9 > > > > > > - vpermt2ps TANHF_DATA(_sP2_hi)(%rip), %zmm3, %zmm9 > > > > > > + vmovaps LOCAL_DATA(_sP2_lo)(%rip), %zmm9 > > > > > > + vpermt2ps LOCAL_DATA(_sP2_hi)(%rip), %zmm3, %zmm9 > > > > > > > > > > > > - vmovaps TANHF_DATA(_sP0_lo)(%rip), %zmm10 > > > > > > - vpermt2ps TANHF_DATA(_sP0_hi)(%rip), %zmm3, %zmm10 > > > > > > + vmovaps LOCAL_DATA(_sP0_lo)(%rip), %zmm10 > > > > > > + vpermt2ps LOCAL_DATA(_sP0_hi)(%rip), %zmm3, %zmm10 > > > > > > > > > > > > vfmadd213ps {rn-sae}, %zmm9, %zmm1, %zmm2 > > > > > > vfmadd213ps {rn-sae}, %zmm10, %zmm1, %zmm2 > > > > > > @@ -167,7 +172,7 @@ ENTRY(_ZGVeN16v_tanhf_skx) > > > > > > > > > > > > /* Go to special inputs processing branch. */ > > > > > > jne L(SPECIAL_VALUES_BRANCH) > > > > > > - # LOE rbx r12 r13 r14 r15 zmm0 zmm2 zmm4 > > > > > > + > > > > > > /* Wait until after branch of write over zmm0. */ > > > > > > vpternlogd $0xec, %zmm4, %zmm2, %zmm0 > > > > > > > > > > > > @@ -176,24 +181,24 @@ ENTRY(_ZGVeN16v_tanhf_skx) > > > > > > > > > > > > /* Cold case. edx has 1s where there was a special value that > > > > > > needs to be handled by a tanhf call. Optimize for code size > > > > > > - more so than speed here. */ > > > > > > + more so than speed here. */ > > > > > > L(SPECIAL_VALUES_BRANCH): > > > > > > - # LOE rbx rdx r12 r13 r14 r15 zmm0 zmm2 zmm4 > > > > > > - /* Use r13 to save/restore the stack. This allows us to use rbp as > > > > > > - callee save register saving code size. */ > > > > > > + > > > > > > + /* Use r13 to save/restore the stack. This allows us to use rbp > > > > > > + as callee save register saving code size. */ > > > > > > pushq %r13 > > > > > > - cfi_adjust_cfa_offset(8) > > > > > > - cfi_offset(r13, -16) > > > > > > - /* Need to callee save registers to preserve state across tanhf calls. > > > > > > - */ > > > > > > + cfi_adjust_cfa_offset (8) > > > > > > + cfi_offset (r13, -16) > > > > > > + /* Need to callee save registers to preserve state across tanhf > > > > > > + calls. */ > > > > > > pushq %rbx > > > > > > - cfi_adjust_cfa_offset(8) > > > > > > - cfi_offset(rbx, -24) > > > > > > + cfi_adjust_cfa_offset (8) > > > > > > + cfi_offset (rbx, -24) > > > > > > pushq %rbp > > > > > > - cfi_adjust_cfa_offset(8) > > > > > > - cfi_offset(rbp, -32) > > > > > > + cfi_adjust_cfa_offset (8) > > > > > > + cfi_offset (rbp, -32) > > > > > > movq %rsp, %r13 > > > > > > - cfi_def_cfa_register(r13) > > > > > > + cfi_def_cfa_register (r13) > > > > > > > > > > > > /* Align stack and make room for 2x zmm vectors. */ > > > > > > andq $-64, %rsp > > > > > > @@ -207,16 +212,17 @@ L(SPECIAL_VALUES_BRANCH): > > > > > > > > > > > > vzeroupper > > > > > > > > > > > > - /* edx has 1s where there was a special value that needs to be handled > > > > > > - by a tanhf call. */ > > > > > > + /* edx has 1s where there was a special value that needs to be > > > > > > + handled by a tanhf call. */ > > > > > > movl %edx, %ebx > > > > > > L(SPECIAL_VALUES_LOOP): > > > > > > - # LOE rbx rbp r12 r13 r14 r15 > > > > > > - /* use rbp as index for special value that is saved across calls to > > > > > > - tanhf. We technically don't need a callee save register here as offset > > > > > > - to rsp is always [0, 56] so we can restore rsp by realigning to 64. > > > > > > - Essentially the tradeoff is 1 extra save/restore vs 2 extra instructions > > > > > > - in the loop. Realigning also costs more code size. */ > > > > > > + > > > > > > + /* use rbp as index for special value that is saved across calls > > > > > > + to tanhf. We technically don't need a callee save register > > > > > > + here as offset to rsp is always [0, 56] so we can restore > > > > > > + rsp by realigning to 64. Essentially the tradeoff is 1 extra > > > > > > + save/restore vs 2 extra instructions in the loop. Realigning > > > > > > + also costs more code size. */ > > > > > > xorl %ebp, %ebp > > > > > > tzcntl %ebx, %ebp > > > > > > > > > > > > @@ -224,203 +230,141 @@ L(SPECIAL_VALUES_LOOP): > > > > > > vmovss 64(%rsp, %rbp, 4), %xmm0 > > > > > > call tanhf@PLT > > > > > > > > > > > > - /* No good way to avoid the store-forwarding fault this will cause on > > > > > > - return. `lfence` avoids the SF fault but at greater cost as it > > > > > > - serialized stack/callee save restoration. */ > > > > > > + /* No good way to avoid the store-forwarding fault this will > > > > > > + cause on return. `lfence` avoids the SF fault but at greater > > > > > > + cost as it serialized stack/callee save restoration. */ > > > > > > vmovss %xmm0, (%rsp, %rbp, 4) > > > > > > > > > > > > - blsrl %ebx, %ebx > > > > > > + blsrl %ebx, %ebx > > > > > > jnz L(SPECIAL_VALUES_LOOP) > > > > > > - # LOE r12 r13 r14 r15 > > > > > > + > > > > > > > > > > > > /* All results have been written to (%rsp). */ > > > > > > vmovaps (%rsp), %zmm0 > > > > > > /* Restore rsp. */ > > > > > > movq %r13, %rsp > > > > > > - cfi_def_cfa_register(rsp) > > > > > > + cfi_def_cfa_register (rsp) > > > > > > /* Restore callee save registers. */ > > > > > > popq %rbp > > > > > > - cfi_adjust_cfa_offset(-8) > > > > > > - cfi_restore(rbp) > > > > > > + cfi_adjust_cfa_offset (-8) > > > > > > + cfi_restore (rbp) > > > > > > popq %rbx > > > > > > - cfi_adjust_cfa_offset(-8) > > > > > > - cfi_restore(rbp) > > > > > > + cfi_adjust_cfa_offset (-8) > > > > > > + cfi_restore (rbp) > > > > > > popq %r13 > > > > > > - cfi_adjust_cfa_offset(-8) > > > > > > - cfi_restore(r13) > > > > > > + cfi_adjust_cfa_offset (-8) > > > > > > + cfi_restore (r13) > > > > > > ret > > > > > > END(_ZGVeN16v_tanhf_skx) > > > > > > > > > > > > - .section .rodata, "a" > > > > > > + .section .rodata.evex512, "a" > > > > > > .align 16 > > > > > > -#ifdef __svml_stanh_data_internal_typedef > > > > > > -typedef unsigned int VUINT32; > > > > > > -typedef struct > > > > > > - { > > > > > > - __declspec(align(4)) VUINT32 _iExpMantMask_UISA[1][1]; > > > > > > - __declspec(align(4)) VUINT32 _iMinIdxOfsMask_UISA[1][1]; > > > > > > - __declspec(align(4)) VUINT32 _iMaxIdxMask_UISA[1][1]; > > > > > > - __declspec(align(4)) VUINT32 _iExpMask[1][1]; > > > > > > - __declspec(align(64)) VUINT32 _sC_lo[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sC_hi[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP7_lo[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP7_hi[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sSignMask[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP6_lo[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP6_hi[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP5_lo[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP5_hi[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP4_lo[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP4_hi[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP3_lo[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP3_hi[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP2_lo[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP2_hi[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP0_lo[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP0_hi[16][1]; > > > > > > -} __svml_stanh_data_internal; > > > > > > -#endif > > > > > > - > > > > > > -__svml_stanh_data_internal: > > > > > > - .align 4 > > > > > > - /* _iExpMantMask_UISA */ > > > > > > - .long 0x7fe00000 > > > > > > - > > > > > > - .align 4 > > > > > > - /* _iMinIdxOfsMask_UISA */ > > > > > > - .long 0x3d400000 > > > > > > - > > > > > > - .align 4 > > > > > > - /* _iMaxIdxMask_UISA */ > > > > > > - .long 0x03e00000 > > > > > > - > > > > > > - .align 4 > > > > > > - /* _iExpMask */ > > > > > > - .long 0x7f000000 > > > > > > - > > > > > > - .align 64 > > > > > > -__svml_stanh_data_internal_al64: > > > > > > - .align 64 > > > > > > - /* _sC_lo */ > > > > > > - .long 0x00000000, 0x3d700000, 0x3d900000, 0x3db00000 > > > > > > - .long 0x3dd00000, 0x3df00000, 0x3e100000, 0x3e300000 > > > > > > - .long 0x3e500000, 0x3e700000, 0x3e900000, 0x3eb00000 > > > > > > - .long 0x3ed00000, 0x3ef00000, 0x3f100000, 0x3f300000 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sC_hi */ > > > > > > - .long 0x3f500000, 0x3f700000, 0x3f900000, 0x3fb00000 > > > > > > - .long 0x3fd00000, 0x3ff00000, 0x40100000, 0x40300000 > > > > > > - .long 0x40500000, 0x40700000, 0x40900000, 0x40b00000 > > > > > > - .long 0x40d00000, 0x40f00000, 0x41100000, 0x00000000 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP7_lo */ > > > > > > - .long 0xbc0e2f66, 0x460bda12, 0x43d638ef, 0xc3e11c3e > > > > > > - .long 0xc2baa4e9, 0xc249da2d, 0xc1859b82, 0x40dd5b57 > > > > > > - .long 0x40494640, 0x40c730a8, 0xbf0f160e, 0x3e30e76f > > > > > > - .long 0xbea81387, 0xbdb26a1c, 0xbd351e57, 0xbb4c01a0 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP7_hi */ > > > > > > - .long 0x3c1d7bfb, 0x3c722cd1, 0x3c973f1c, 0x3c33a31b > > > > > > - .long 0x3b862ef4, 0x3a27b3d0, 0xba3b5907, 0xba0efc22 > > > > > > - .long 0xb97f9f0f, 0xb8c8af50, 0xb7bdddfb, 0xb64f2950 > > > > > > - .long 0xb4e085b1, 0xb3731dfa, 0xb15a1f04, 0x00000000 > > > > > > > > > > > > - .align 64 > > > > > > - /* _sSignMask */ > > > > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP6_lo */ > > > > > > - .long 0x3e0910e9, 0x43761143, 0x4165ecdc, 0xc190f756 > > > > > > - .long 0xc08c097d, 0xc02ba813, 0xbf7f6bda, 0x3f2b1dc0 > > > > > > - .long 0x3ece105d, 0x3f426a94, 0xbadb0dc4, 0x3da43b17 > > > > > > - .long 0xbd51ab88, 0xbcaea23d, 0xbd3b6d8d, 0xbd6caaad > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP6_hi */ > > > > > > - .long 0xbd795bed, 0xbd5fddda, 0xbd038f3b, 0xbc1cad63 > > > > > > - .long 0x3abb4766, 0x3b95f10b, 0x3b825873, 0x3afaea66 > > > > > > - .long 0x3a49f878, 0x39996bf3, 0x388f3e6c, 0x371bb0e3 > > > > > > - .long 0x35a8a5e6, 0x34369b17, 0x322487b0, 0x00000000 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP5_lo */ > > > > > > - .long 0xb76dd6b9, 0xbe1c276d, 0x3c1dcf2f, 0x3dc1a78d > > > > > > - .long 0x3d96f985, 0x3da2b61b, 0x3dc13397, 0x3dd2f670 > > > > > > - .long 0x3df48a0a, 0x3e06c5a8, 0x3e1a3aba, 0x3e27c405 > > > > > > - .long 0x3e2e78d0, 0x3e2c3e44, 0x3e1d3097, 0x3df4a8f4 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP5_hi */ > > > > > > - .long 0x3da38508, 0x3d31416a, 0x3b562657, 0xbcaeeac9 > > > > > > - .long 0xbcce9419, 0xbcaaeac4, 0xbc49e7d0, 0xbba71ddd > > > > > > - .long 0xbb003b0e, 0xba3f9a05, 0xb92c08a7, 0xb7ba9232 > > > > > > - .long 0xb64a0b0f, 0xb4dac169, 0xb2ab78ac, 0x00000000 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP4_lo */ > > > > > > - .long 0xbeaaaaa5, 0xbeab0612, 0xbea7f01f, 0xbea4e120 > > > > > > - .long 0xbea387b7, 0xbea15962, 0xbe9d57f7, 0xbe976b5a > > > > > > - .long 0xbe90230d, 0xbe880dff, 0xbe7479b3, 0xbe4c3d88 > > > > > > - .long 0xbe212482, 0xbdeb8cba, 0xbd5e78ad, 0x3c6b5e6e > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP4_hi */ > > > > > > - .long 0x3d839143, 0x3dc21ee1, 0x3de347af, 0x3dcbec96 > > > > > > - .long 0x3d99ef2d, 0x3d542ea1, 0x3cdde701, 0x3c2cca67 > > > > > > - .long 0x3b81cb27, 0x3ac073a1, 0x39ac3032, 0x383a94d9 > > > > > > - .long 0x36ca081d, 0x355abd4c, 0x332b3cb6, 0x00000000 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP3_lo */ > > > > > > - .long 0xb0343c7b, 0xbd6ee69d, 0xbd8f0da7, 0xbdae477d > > > > > > - .long 0xbdcd2a1f, 0xbdeba80d, 0xbe0c443b, 0xbe293cf3 > > > > > > - .long 0xbe44f282, 0xbe5f3651, 0xbe81c7c0, 0xbe96d7ca > > > > > > - .long 0xbea7fb8e, 0xbeb50e9e, 0xbec12efe, 0xbec4be92 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP3_hi */ > > > > > > - .long 0xbebce070, 0xbead510e, 0xbe8ef7d6, 0xbe4b8704 > > > > > > - .long 0xbe083237, 0xbdaf7449, 0xbd2e1ec4, 0xbc83bf06 > > > > > > - .long 0xbbc3e0b5, 0xbb10aadc, 0xba0157db, 0xb88c18f2 > > > > > > - .long 0xb717b096, 0xb5a43bae, 0xb383012c, 0x00000000 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP2_lo */ > > > > > > - .long 0x3f800000, 0x3f7f1f84, 0x3f7ebd11, 0x3f7e1e5f > > > > > > - .long 0x3f7d609f, 0x3f7c842d, 0x3f7b00e5, 0x3f789580 > > > > > > - .long 0x3f75b8ad, 0x3f726fd9, 0x3f6cc59b, 0x3f63fb92 > > > > > > - .long 0x3f59ff97, 0x3f4f11d7, 0x3f3d7573, 0x3f24f360 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP2_hi */ > > > > > > - .long 0x3f0cbfe7, 0x3eec1a69, 0x3eb0a801, 0x3e6753a2 > > > > > > - .long 0x3e132f1a, 0x3db7e7d3, 0x3d320845, 0x3c84d3d4 > > > > > > - .long 0x3bc477b7, 0x3b10d3da, 0x3a01601e, 0x388c1a3b > > > > > > - .long 0x3717b0da, 0x35a43bce, 0x338306c6, 0x00000000 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP0_lo */ > > > > > > - .long 0x00000000, 0x3d6fb9c9, 0x3d8fc35f, 0x3daf9169 > > > > > > - .long 0x3dcf49ab, 0x3deee849, 0x3e0f0ee8, 0x3e2e4984 > > > > > > - .long 0x3e4d2f8e, 0x3e6bb32e, 0x3e8c51cd, 0x3ea96163 > > > > > > - .long 0x3ec543f1, 0x3edfd735, 0x3f028438, 0x3f18abf0 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP0_hi */ > > > > > > - .long 0x3f2bc480, 0x3f3bec1c, 0x3f4f2e5b, 0x3f613c53 > > > > > > - .long 0x3f6ce37d, 0x3f743c4f, 0x3f7a5feb, 0x3f7dea85 > > > > > > - .long 0x3f7f3b3d, 0x3f7fb78c, 0x3f7fefd4, 0x3f7ffdd0 > > > > > > - .long 0x3f7fffb4, 0x3f7ffff6, 0x3f7fffff, 0x3f800000 > > > > > > +LOCAL_DATA_NAME_UNALIGNED: > > > > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iExpMantMask_UISA, 0x7fe00000) > > > > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iMinIdxOfsMask_UISA, 0x3d400000) > > > > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iMaxIdxMask_UISA, 0x03e00000) > > > > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iExpMask, 0x7f000000) > > > > > > + .type LOCAL_DATA_NAME_UNALIGNED, @object > > > > > > + .size LOCAL_DATA_NAME_UNALIGNED, .-LOCAL_DATA_NAME_UNALIGNED > > > > > > > > > > > > .align 64 > > > > > > - .type __svml_stanh_data_internal_al64, @object > > > > > > - .size __svml_stanh_data_internal_al64, .-__svml_stanh_data_internal_al64 > > > > > > - .type __svml_stanh_data_internal, @object > > > > > > - .size __svml_stanh_data_internal, .-__svml_stanh_data_internal > > > > > > +LOCAL_DATA_NAME: > > > > > > + float_block (LOCAL_DATA_NAME, _sC_lo, > > > > > > + 0x00000000, 0x3d700000, 0x3d900000, 0x3db00000, > > > > > > + 0x3dd00000, 0x3df00000, 0x3e100000, 0x3e300000, > > > > > > + 0x3e500000, 0x3e700000, 0x3e900000, 0x3eb00000, > > > > > > + 0x3ed00000, 0x3ef00000, 0x3f100000, 0x3f300000) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sC_hi, > > > > > > + 0x3f500000, 0x3f700000, 0x3f900000, 0x3fb00000, > > > > > > + 0x3fd00000, 0x3ff00000, 0x40100000, 0x40300000, > > > > > > + 0x40500000, 0x40700000, 0x40900000, 0x40b00000, > > > > > > + 0x40d00000, 0x40f00000, 0x41100000, 0x00000000) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP7_lo, > > > > > > + 0xbc0e2f66, 0x460bda12, 0x43d638ef, 0xc3e11c3e, > > > > > > + 0xc2baa4e9, 0xc249da2d, 0xc1859b82, 0x40dd5b57, > > > > > > + 0x40494640, 0x40c730a8, 0xbf0f160e, 0x3e30e76f, > > > > > > + 0xbea81387, 0xbdb26a1c, 0xbd351e57, 0xbb4c01a0) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP7_hi, > > > > > > + 0x3c1d7bfb, 0x3c722cd1, 0x3c973f1c, 0x3c33a31b, > > > > > > + 0x3b862ef4, 0x3a27b3d0, 0xba3b5907, 0xba0efc22, > > > > > > + 0xb97f9f0f, 0xb8c8af50, 0xb7bdddfb, 0xb64f2950, > > > > > > + 0xb4e085b1, 0xb3731dfa, 0xb15a1f04, 0x00000000) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP6_lo, > > > > > > + 0x3e0910e9, 0x43761143, 0x4165ecdc, 0xc190f756, > > > > > > + 0xc08c097d, 0xc02ba813, 0xbf7f6bda, 0x3f2b1dc0, > > > > > > + 0x3ece105d, 0x3f426a94, 0xbadb0dc4, 0x3da43b17, > > > > > > + 0xbd51ab88, 0xbcaea23d, 0xbd3b6d8d, 0xbd6caaad) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP6_hi, > > > > > > + 0xbd795bed, 0xbd5fddda, 0xbd038f3b, 0xbc1cad63, > > > > > > + 0x3abb4766, 0x3b95f10b, 0x3b825873, 0x3afaea66, > > > > > > + 0x3a49f878, 0x39996bf3, 0x388f3e6c, 0x371bb0e3, > > > > > > + 0x35a8a5e6, 0x34369b17, 0x322487b0, 0x00000000) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP5_lo, > > > > > > + 0xb76dd6b9, 0xbe1c276d, 0x3c1dcf2f, 0x3dc1a78d, > > > > > > + 0x3d96f985, 0x3da2b61b, 0x3dc13397, 0x3dd2f670, > > > > > > + 0x3df48a0a, 0x3e06c5a8, 0x3e1a3aba, 0x3e27c405, > > > > > > + 0x3e2e78d0, 0x3e2c3e44, 0x3e1d3097, 0x3df4a8f4) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP5_hi, > > > > > > + 0x3da38508, 0x3d31416a, 0x3b562657, 0xbcaeeac9, > > > > > > + 0xbcce9419, 0xbcaaeac4, 0xbc49e7d0, 0xbba71ddd, > > > > > > + 0xbb003b0e, 0xba3f9a05, 0xb92c08a7, 0xb7ba9232, > > > > > > + 0xb64a0b0f, 0xb4dac169, 0xb2ab78ac, 0x00000000) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP4_lo, > > > > > > + 0xbeaaaaa5, 0xbeab0612, 0xbea7f01f, 0xbea4e120, > > > > > > + 0xbea387b7, 0xbea15962, 0xbe9d57f7, 0xbe976b5a, > > > > > > + 0xbe90230d, 0xbe880dff, 0xbe7479b3, 0xbe4c3d88, > > > > > > + 0xbe212482, 0xbdeb8cba, 0xbd5e78ad, 0x3c6b5e6e) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP4_hi, > > > > > > + 0x3d839143, 0x3dc21ee1, 0x3de347af, 0x3dcbec96, > > > > > > + 0x3d99ef2d, 0x3d542ea1, 0x3cdde701, 0x3c2cca67, > > > > > > + 0x3b81cb27, 0x3ac073a1, 0x39ac3032, 0x383a94d9, > > > > > > + 0x36ca081d, 0x355abd4c, 0x332b3cb6, 0x00000000) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP3_lo, > > > > > > + 0xb0343c7b, 0xbd6ee69d, 0xbd8f0da7, 0xbdae477d, > > > > > > + 0xbdcd2a1f, 0xbdeba80d, 0xbe0c443b, 0xbe293cf3, > > > > > > + 0xbe44f282, 0xbe5f3651, 0xbe81c7c0, 0xbe96d7ca, > > > > > > + 0xbea7fb8e, 0xbeb50e9e, 0xbec12efe, 0xbec4be92) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP3_hi, > > > > > > + 0xbebce070, 0xbead510e, 0xbe8ef7d6, 0xbe4b8704, > > > > > > + 0xbe083237, 0xbdaf7449, 0xbd2e1ec4, 0xbc83bf06, > > > > > > + 0xbbc3e0b5, 0xbb10aadc, 0xba0157db, 0xb88c18f2, > > > > > > + 0xb717b096, 0xb5a43bae, 0xb383012c, 0x00000000) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP2_lo, > > > > > > + 0x3f800000, 0x3f7f1f84, 0x3f7ebd11, 0x3f7e1e5f, > > > > > > + 0x3f7d609f, 0x3f7c842d, 0x3f7b00e5, 0x3f789580, > > > > > > + 0x3f75b8ad, 0x3f726fd9, 0x3f6cc59b, 0x3f63fb92, > > > > > > + 0x3f59ff97, 0x3f4f11d7, 0x3f3d7573, 0x3f24f360) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP2_hi, > > > > > > + 0x3f0cbfe7, 0x3eec1a69, 0x3eb0a801, 0x3e6753a2, > > > > > > + 0x3e132f1a, 0x3db7e7d3, 0x3d320845, 0x3c84d3d4, > > > > > > + 0x3bc477b7, 0x3b10d3da, 0x3a01601e, 0x388c1a3b, > > > > > > + 0x3717b0da, 0x35a43bce, 0x338306c6, 0x00000000) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP0_lo, > > > > > > + 0x00000000, 0x3d6fb9c9, 0x3d8fc35f, 0x3daf9169, > > > > > > + 0x3dcf49ab, 0x3deee849, 0x3e0f0ee8, 0x3e2e4984, > > > > > > + 0x3e4d2f8e, 0x3e6bb32e, 0x3e8c51cd, 0x3ea96163, > > > > > > + 0x3ec543f1, 0x3edfd735, 0x3f028438, 0x3f18abf0) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP0_hi, > > > > > > + 0x3f2bc480, 0x3f3bec1c, 0x3f4f2e5b, 0x3f613c53, > > > > > > + 0x3f6ce37d, 0x3f743c4f, 0x3f7a5feb, 0x3f7dea85, > > > > > > + 0x3f7f3b3d, 0x3f7fb78c, 0x3f7fefd4, 0x3f7ffdd0, > > > > > > + 0x3f7fffb4, 0x3f7ffff6, 0x3f7fffff, 0x3f800000) > > > > > > + > > > > > > + .type LOCAL_DATA_NAME, @object > > > > > > + .size LOCAL_DATA_NAME, .-LOCAL_DATA_NAME > > > > > > -- > > > > > > 2.34.1 > > > > > > > > > > > > > > > > The data movement makes the assembler codes much harder to follow. > > > > > Sunil, what do you think of this patch series? > > > > > > > > What do you mean? The change on in how we define rodata or the movement > > > > to multiple files or something else? > > > > > > The glibc way to support data files for assembly codes is to define > > > data in C and use *.sym to generate offsets for assembly files, like > > > > I see. Although to be fair the entire SVML codebase bucks that trend. > > It is because libmvec codes were generated by ICC and processed > by scripts. > > > Seems like a more dramatic trend to move all the offsets to C. > > Since you are adding data by hand, you should do it in C. > > > > > > > sysdeps/x86/cpu-features-offsets.sym:XSAVE_STATE_SIZE_OFFSET > > > offsetof (struct cpu_features, xsave_state_size) > > > sysdeps/x86_64/dl-trampoline.h: sub > > > _rtld_local_ro+RTLD_GLOBAL_RO_DL_X86_CPU_FEATURES_OFFSET+XSAVE_STATE_SIZE_OFFSET(%rip), > > > %RSP_LP > > > sysdeps/x86_64/dl-trampoline.h: sub > > > _dl_x86_cpu_features+XSAVE_STATE_SIZE_OFFSET(%rip), %RSP_LP > > > > > > -- > > > H.J. > > Does this restructuring provide any performance benefit as measured by libmvec microbenchmark? > > -- > H.J.
On Fri, Dec 16, 2022 at 4:01 PM H.J. Lu <hjl.tools@gmail.com> wrote: > > On Fri, Dec 16, 2022 at 1:52 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote: > > > > On Fri, Dec 16, 2022 at 1:38 PM H.J. Lu <hjl.tools@gmail.com> wrote: > > > > > > On Fri, Dec 16, 2022 at 10:18 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote: > > > > > > > > On Fri, Dec 16, 2022 at 9:06 AM H.J. Lu <hjl.tools@gmail.com> wrote: > > > > > > > > > > On Wed, Dec 7, 2022 at 12:52 AM Noah Goldstein <goldstein.w.n@gmail.com> wrote: > > > > > > > > > > > > No changes to the logic, just change how rodata is handled. > > > > > > > > > > > > 1. Define the rodatas using the new macros so they check that the > > > > > > offset is correct. > > > > > > > > > > > > 2. Use common data where applicable. > > > > > > --- > > > > > > .../multiarch/svml_s_tanhf16_core_avx512.S | 450 ++++++++---------- > > > > > > 1 file changed, 197 insertions(+), 253 deletions(-) > > > > > > > > > > > > diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S b/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > > > > > > index d74fc7731d..765e9ed7f7 100644 > > > > > > --- a/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > > > > > > +++ b/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S > > > > > > @@ -70,94 +70,99 @@ > > > > > > * > > > > > > */ > > > > > > > > > > > > -/* Offsets for data table __svml_stanh_data_internal_avx512. Ordered > > > > > > - by use in the function. On cold-starts this might help the > > > > > > - prefetcher. Possibly a better idea is to interleave start/end so > > > > > > - that the prefetcher is less likely to detect a stream and pull > > > > > > - irrelivant lines into cache. */ > > > > > > > > > > > > -/* Offsets for data table __svml_stanh_data_internal. 4 bytes each. > > > > > > - */ > > > > > > + > > > > > > +#define LOCAL_DATA_NAME __svml_stanh_data_internal > > > > > > +#define LOCAL_DATA_NAME_UNALIGNED __svml_stanh_data_internal_unaligned > > > > > > +#include "svml_s_common_evex512_rodata_offsets.h" > > > > > > + > > > > > > +/* Offsets for data table __svml_stanh_data_internal_unaligned. > > > > > > + 4 bytes each. */ > > > > > > #define _iExpMantMask_UISA 0 > > > > > > #define _iMinIdxOfsMask_UISA 4 > > > > > > #define _iMaxIdxMask_UISA 8 > > > > > > #define _iExpMask 12 > > > > > > > > > > > > -/* Offsets for data table __svml_stanh_data_internal_al64. 64 bytes > > > > > > - each. */ > > > > > > -#define _sC_lo 0 > > > > > > -#define _sC_hi 64 > > > > > > -#define _sP7_lo 128 > > > > > > -#define _sP7_hi 192 > > > > > > -#define _sSignMask 256 > > > > > > -#define _sP6_lo 320 > > > > > > -#define _sP6_hi 384 > > > > > > -#define _sP5_lo 448 > > > > > > -#define _sP5_hi 512 > > > > > > -#define _sP4_lo 576 > > > > > > -#define _sP4_hi 640 > > > > > > -#define _sP3_lo 704 > > > > > > -#define _sP3_hi 768 > > > > > > -#define _sP2_lo 832 > > > > > > -#define _sP2_hi 896 > > > > > > -#define _sP0_lo 960 > > > > > > -#define _sP0_hi 1024 > > > > > > +/* Offsets for data table __svml_stanh_data_internal. Ordered > > > > > > + by use in the function. On cold-starts this might help the > > > > > > + prefetcher. Possibly a better idea is to interleave start/end so > > > > > > + that the prefetcher is less likely to detect a stream and pull > > > > > > + irrelivant lines into cache. */ > > > > > > + > > > > > > +/* Offsets for data table __svml_stanh_data_internal. > > > > > > + 64 bytes each. */ > > > > > > +#define _sC_lo 0 > > > > > > +#define _sC_hi 64 > > > > > > +#define _sP7_lo 128 > > > > > > +#define _sP7_hi 192 > > > > > > +#define _sP6_lo 256 > > > > > > +#define _sP6_hi 320 > > > > > > +#define _sP5_lo 384 > > > > > > +#define _sP5_hi 448 > > > > > > +#define _sP4_lo 512 > > > > > > +#define _sP4_hi 576 > > > > > > +#define _sP3_lo 640 > > > > > > +#define _sP3_hi 704 > > > > > > +#define _sP2_lo 768 > > > > > > +#define _sP2_hi 832 > > > > > > +#define _sP0_lo 896 > > > > > > +#define _sP0_hi 960 > > > > > > + > > > > > > > > > > > > #include <sysdep.h> > > > > > > -#define TANHF_DATA(x) ((x)+__svml_stanh_data_internal_al64) > > > > > > -#define TANHF_DATA_UNALIGNED(x) ((x)+__svml_stanh_data_internal) > > > > > > > > > > > > .section .text.evex512, "ax", @progbits > > > > > > ENTRY(_ZGVeN16v_tanhf_skx) > > > > > > - /* Here huge arguments, INF and NaNs are filtered out to callout. */ > > > > > > - vpandd TANHF_DATA_UNALIGNED(_iExpMantMask_UISA)(%rip){1to16}, %zmm0, %zmm1 > > > > > > - vpsubd TANHF_DATA_UNALIGNED(_iMinIdxOfsMask_UISA)(%rip){1to16}, %zmm1, %zmm2 > > > > > > + /* Here huge arguments, INF and NaNs are filtered out to > > > > > > + callout. */ > > > > > > + vpandd LOCAL_DATA_UNALIGNED(_iExpMantMask_UISA)(%rip){1to16}, %zmm0, %zmm1 > > > > > > + vpsubd LOCAL_DATA_UNALIGNED(_iMinIdxOfsMask_UISA)(%rip){1to16}, %zmm1, %zmm2 > > > > > > > > > > > > /* Selection arguments between [0, 0x03e00000] into zmm3. */ > > > > > > vpxord %zmm3, %zmm3, %zmm3 > > > > > > vpmaxsd %zmm3, %zmm2, %zmm3 > > > > > > - vpminsd TANHF_DATA_UNALIGNED(_iMaxIdxMask_UISA)(%rip){1to16}, %zmm3, %zmm3 > > > > > > + vpminsd LOCAL_DATA_UNALIGNED(_iMaxIdxMask_UISA)(%rip){1to16}, %zmm3, %zmm3 > > > > > > > > > > > > /* Setup permute indices in zmm3. */ > > > > > > vpsrld $21, %zmm3, %zmm3 > > > > > > > > > > > > /* Store if there are any special cases in k1. */ > > > > > > - vpcmpd $6, TANHF_DATA_UNALIGNED(_iExpMask)(%rip){1to16}, %zmm1, %k1 > > > > > > + vpcmpd $6, LOCAL_DATA_UNALIGNED(_iExpMask)(%rip){1to16}, %zmm1, %k1 > > > > > > > > > > > > - vmovaps TANHF_DATA(_sC_lo)(%rip), %zmm5 > > > > > > - vpermt2ps TANHF_DATA(_sC_hi)(%rip), %zmm3, %zmm5 > > > > > > + vmovaps LOCAL_DATA(_sC_lo)(%rip), %zmm5 > > > > > > + vpermt2ps LOCAL_DATA(_sC_hi)(%rip), %zmm3, %zmm5 > > > > > > > > > > > > - vmovaps TANHF_DATA(_sP7_lo)(%rip), %zmm2 > > > > > > - vpermt2ps TANHF_DATA(_sP7_hi)(%rip), %zmm3, %zmm2 > > > > > > + vmovaps LOCAL_DATA(_sP7_lo)(%rip), %zmm2 > > > > > > + vpermt2ps LOCAL_DATA(_sP7_hi)(%rip), %zmm3, %zmm2 > > > > > > > > > > > > /* Store absolute values of inputs in zmm1. */ > > > > > > - vmovaps TANHF_DATA(_sSignMask)(%rip), %zmm4 > > > > > > + vmovaps COMMON_DATA(_SignMask)(%rip), %zmm4 > > > > > > vandnps %zmm0, %zmm4, %zmm1 > > > > > > vsubps {rn-sae}, %zmm5, %zmm1, %zmm1 > > > > > > > > > > > > - vmovaps TANHF_DATA(_sP6_lo)(%rip), %zmm5 > > > > > > - vpermt2ps TANHF_DATA(_sP6_hi)(%rip), %zmm3, %zmm5 > > > > > > + vmovaps LOCAL_DATA(_sP6_lo)(%rip), %zmm5 > > > > > > + vpermt2ps LOCAL_DATA(_sP6_hi)(%rip), %zmm3, %zmm5 > > > > > > > > > > > > - vmovaps TANHF_DATA(_sP5_lo)(%rip), %zmm6 > > > > > > - vpermt2ps TANHF_DATA(_sP5_hi)(%rip), %zmm3, %zmm6 > > > > > > + vmovaps LOCAL_DATA(_sP5_lo)(%rip), %zmm6 > > > > > > + vpermt2ps LOCAL_DATA(_sP5_hi)(%rip), %zmm3, %zmm6 > > > > > > > > > > > > vfmadd213ps {rn-sae}, %zmm5, %zmm1, %zmm2 > > > > > > vfmadd213ps {rn-sae}, %zmm6, %zmm1, %zmm2 > > > > > > > > > > > > - vmovaps TANHF_DATA(_sP4_lo)(%rip), %zmm7 > > > > > > - vpermt2ps TANHF_DATA(_sP4_hi)(%rip), %zmm3, %zmm7 > > > > > > + vmovaps LOCAL_DATA(_sP4_lo)(%rip), %zmm7 > > > > > > + vpermt2ps LOCAL_DATA(_sP4_hi)(%rip), %zmm3, %zmm7 > > > > > > > > > > > > - vmovaps TANHF_DATA(_sP3_lo)(%rip), %zmm8 > > > > > > - vpermt2ps TANHF_DATA(_sP3_hi)(%rip), %zmm3, %zmm8 > > > > > > + vmovaps LOCAL_DATA(_sP3_lo)(%rip), %zmm8 > > > > > > + vpermt2ps LOCAL_DATA(_sP3_hi)(%rip), %zmm3, %zmm8 > > > > > > > > > > > > vfmadd213ps {rn-sae}, %zmm7, %zmm1, %zmm2 > > > > > > vfmadd213ps {rn-sae}, %zmm8, %zmm1, %zmm2 > > > > > > > > > > > > - vmovaps TANHF_DATA(_sP2_lo)(%rip), %zmm9 > > > > > > - vpermt2ps TANHF_DATA(_sP2_hi)(%rip), %zmm3, %zmm9 > > > > > > + vmovaps LOCAL_DATA(_sP2_lo)(%rip), %zmm9 > > > > > > + vpermt2ps LOCAL_DATA(_sP2_hi)(%rip), %zmm3, %zmm9 > > > > > > > > > > > > - vmovaps TANHF_DATA(_sP0_lo)(%rip), %zmm10 > > > > > > - vpermt2ps TANHF_DATA(_sP0_hi)(%rip), %zmm3, %zmm10 > > > > > > + vmovaps LOCAL_DATA(_sP0_lo)(%rip), %zmm10 > > > > > > + vpermt2ps LOCAL_DATA(_sP0_hi)(%rip), %zmm3, %zmm10 > > > > > > > > > > > > vfmadd213ps {rn-sae}, %zmm9, %zmm1, %zmm2 > > > > > > vfmadd213ps {rn-sae}, %zmm10, %zmm1, %zmm2 > > > > > > @@ -167,7 +172,7 @@ ENTRY(_ZGVeN16v_tanhf_skx) > > > > > > > > > > > > /* Go to special inputs processing branch. */ > > > > > > jne L(SPECIAL_VALUES_BRANCH) > > > > > > - # LOE rbx r12 r13 r14 r15 zmm0 zmm2 zmm4 > > > > > > + > > > > > > /* Wait until after branch of write over zmm0. */ > > > > > > vpternlogd $0xec, %zmm4, %zmm2, %zmm0 > > > > > > > > > > > > @@ -176,24 +181,24 @@ ENTRY(_ZGVeN16v_tanhf_skx) > > > > > > > > > > > > /* Cold case. edx has 1s where there was a special value that > > > > > > needs to be handled by a tanhf call. Optimize for code size > > > > > > - more so than speed here. */ > > > > > > + more so than speed here. */ > > > > > > L(SPECIAL_VALUES_BRANCH): > > > > > > - # LOE rbx rdx r12 r13 r14 r15 zmm0 zmm2 zmm4 > > > > > > - /* Use r13 to save/restore the stack. This allows us to use rbp as > > > > > > - callee save register saving code size. */ > > > > > > + > > > > > > + /* Use r13 to save/restore the stack. This allows us to use rbp > > > > > > + as callee save register saving code size. */ > > > > > > pushq %r13 > > > > > > - cfi_adjust_cfa_offset(8) > > > > > > - cfi_offset(r13, -16) > > > > > > - /* Need to callee save registers to preserve state across tanhf calls. > > > > > > - */ > > > > > > + cfi_adjust_cfa_offset (8) > > > > > > + cfi_offset (r13, -16) > > > > > > + /* Need to callee save registers to preserve state across tanhf > > > > > > + calls. */ > > > > > > pushq %rbx > > > > > > - cfi_adjust_cfa_offset(8) > > > > > > - cfi_offset(rbx, -24) > > > > > > + cfi_adjust_cfa_offset (8) > > > > > > + cfi_offset (rbx, -24) > > > > > > pushq %rbp > > > > > > - cfi_adjust_cfa_offset(8) > > > > > > - cfi_offset(rbp, -32) > > > > > > + cfi_adjust_cfa_offset (8) > > > > > > + cfi_offset (rbp, -32) > > > > > > movq %rsp, %r13 > > > > > > - cfi_def_cfa_register(r13) > > > > > > + cfi_def_cfa_register (r13) > > > > > > > > > > > > /* Align stack and make room for 2x zmm vectors. */ > > > > > > andq $-64, %rsp > > > > > > @@ -207,16 +212,17 @@ L(SPECIAL_VALUES_BRANCH): > > > > > > > > > > > > vzeroupper > > > > > > > > > > > > - /* edx has 1s where there was a special value that needs to be handled > > > > > > - by a tanhf call. */ > > > > > > + /* edx has 1s where there was a special value that needs to be > > > > > > + handled by a tanhf call. */ > > > > > > movl %edx, %ebx > > > > > > L(SPECIAL_VALUES_LOOP): > > > > > > - # LOE rbx rbp r12 r13 r14 r15 > > > > > > - /* use rbp as index for special value that is saved across calls to > > > > > > - tanhf. We technically don't need a callee save register here as offset > > > > > > - to rsp is always [0, 56] so we can restore rsp by realigning to 64. > > > > > > - Essentially the tradeoff is 1 extra save/restore vs 2 extra instructions > > > > > > - in the loop. Realigning also costs more code size. */ > > > > > > + > > > > > > + /* use rbp as index for special value that is saved across calls > > > > > > + to tanhf. We technically don't need a callee save register > > > > > > + here as offset to rsp is always [0, 56] so we can restore > > > > > > + rsp by realigning to 64. Essentially the tradeoff is 1 extra > > > > > > + save/restore vs 2 extra instructions in the loop. Realigning > > > > > > + also costs more code size. */ > > > > > > xorl %ebp, %ebp > > > > > > tzcntl %ebx, %ebp > > > > > > > > > > > > @@ -224,203 +230,141 @@ L(SPECIAL_VALUES_LOOP): > > > > > > vmovss 64(%rsp, %rbp, 4), %xmm0 > > > > > > call tanhf@PLT > > > > > > > > > > > > - /* No good way to avoid the store-forwarding fault this will cause on > > > > > > - return. `lfence` avoids the SF fault but at greater cost as it > > > > > > - serialized stack/callee save restoration. */ > > > > > > + /* No good way to avoid the store-forwarding fault this will > > > > > > + cause on return. `lfence` avoids the SF fault but at greater > > > > > > + cost as it serialized stack/callee save restoration. */ > > > > > > vmovss %xmm0, (%rsp, %rbp, 4) > > > > > > > > > > > > - blsrl %ebx, %ebx > > > > > > + blsrl %ebx, %ebx > > > > > > jnz L(SPECIAL_VALUES_LOOP) > > > > > > - # LOE r12 r13 r14 r15 > > > > > > + > > > > > > > > > > > > /* All results have been written to (%rsp). */ > > > > > > vmovaps (%rsp), %zmm0 > > > > > > /* Restore rsp. */ > > > > > > movq %r13, %rsp > > > > > > - cfi_def_cfa_register(rsp) > > > > > > + cfi_def_cfa_register (rsp) > > > > > > /* Restore callee save registers. */ > > > > > > popq %rbp > > > > > > - cfi_adjust_cfa_offset(-8) > > > > > > - cfi_restore(rbp) > > > > > > + cfi_adjust_cfa_offset (-8) > > > > > > + cfi_restore (rbp) > > > > > > popq %rbx > > > > > > - cfi_adjust_cfa_offset(-8) > > > > > > - cfi_restore(rbp) > > > > > > + cfi_adjust_cfa_offset (-8) > > > > > > + cfi_restore (rbp) > > > > > > popq %r13 > > > > > > - cfi_adjust_cfa_offset(-8) > > > > > > - cfi_restore(r13) > > > > > > + cfi_adjust_cfa_offset (-8) > > > > > > + cfi_restore (r13) > > > > > > ret > > > > > > END(_ZGVeN16v_tanhf_skx) > > > > > > > > > > > > - .section .rodata, "a" > > > > > > + .section .rodata.evex512, "a" > > > > > > .align 16 > > > > > > -#ifdef __svml_stanh_data_internal_typedef > > > > > > -typedef unsigned int VUINT32; > > > > > > -typedef struct > > > > > > - { > > > > > > - __declspec(align(4)) VUINT32 _iExpMantMask_UISA[1][1]; > > > > > > - __declspec(align(4)) VUINT32 _iMinIdxOfsMask_UISA[1][1]; > > > > > > - __declspec(align(4)) VUINT32 _iMaxIdxMask_UISA[1][1]; > > > > > > - __declspec(align(4)) VUINT32 _iExpMask[1][1]; > > > > > > - __declspec(align(64)) VUINT32 _sC_lo[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sC_hi[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP7_lo[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP7_hi[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sSignMask[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP6_lo[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP6_hi[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP5_lo[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP5_hi[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP4_lo[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP4_hi[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP3_lo[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP3_hi[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP2_lo[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP2_hi[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP0_lo[16][1]; > > > > > > - __declspec(align(64)) VUINT32 _sP0_hi[16][1]; > > > > > > -} __svml_stanh_data_internal; > > > > > > -#endif > > > > > > - > > > > > > -__svml_stanh_data_internal: > > > > > > - .align 4 > > > > > > - /* _iExpMantMask_UISA */ > > > > > > - .long 0x7fe00000 > > > > > > - > > > > > > - .align 4 > > > > > > - /* _iMinIdxOfsMask_UISA */ > > > > > > - .long 0x3d400000 > > > > > > - > > > > > > - .align 4 > > > > > > - /* _iMaxIdxMask_UISA */ > > > > > > - .long 0x03e00000 > > > > > > - > > > > > > - .align 4 > > > > > > - /* _iExpMask */ > > > > > > - .long 0x7f000000 > > > > > > - > > > > > > - .align 64 > > > > > > -__svml_stanh_data_internal_al64: > > > > > > - .align 64 > > > > > > - /* _sC_lo */ > > > > > > - .long 0x00000000, 0x3d700000, 0x3d900000, 0x3db00000 > > > > > > - .long 0x3dd00000, 0x3df00000, 0x3e100000, 0x3e300000 > > > > > > - .long 0x3e500000, 0x3e700000, 0x3e900000, 0x3eb00000 > > > > > > - .long 0x3ed00000, 0x3ef00000, 0x3f100000, 0x3f300000 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sC_hi */ > > > > > > - .long 0x3f500000, 0x3f700000, 0x3f900000, 0x3fb00000 > > > > > > - .long 0x3fd00000, 0x3ff00000, 0x40100000, 0x40300000 > > > > > > - .long 0x40500000, 0x40700000, 0x40900000, 0x40b00000 > > > > > > - .long 0x40d00000, 0x40f00000, 0x41100000, 0x00000000 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP7_lo */ > > > > > > - .long 0xbc0e2f66, 0x460bda12, 0x43d638ef, 0xc3e11c3e > > > > > > - .long 0xc2baa4e9, 0xc249da2d, 0xc1859b82, 0x40dd5b57 > > > > > > - .long 0x40494640, 0x40c730a8, 0xbf0f160e, 0x3e30e76f > > > > > > - .long 0xbea81387, 0xbdb26a1c, 0xbd351e57, 0xbb4c01a0 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP7_hi */ > > > > > > - .long 0x3c1d7bfb, 0x3c722cd1, 0x3c973f1c, 0x3c33a31b > > > > > > - .long 0x3b862ef4, 0x3a27b3d0, 0xba3b5907, 0xba0efc22 > > > > > > - .long 0xb97f9f0f, 0xb8c8af50, 0xb7bdddfb, 0xb64f2950 > > > > > > - .long 0xb4e085b1, 0xb3731dfa, 0xb15a1f04, 0x00000000 > > > > > > > > > > > > - .align 64 > > > > > > - /* _sSignMask */ > > > > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > > > > - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP6_lo */ > > > > > > - .long 0x3e0910e9, 0x43761143, 0x4165ecdc, 0xc190f756 > > > > > > - .long 0xc08c097d, 0xc02ba813, 0xbf7f6bda, 0x3f2b1dc0 > > > > > > - .long 0x3ece105d, 0x3f426a94, 0xbadb0dc4, 0x3da43b17 > > > > > > - .long 0xbd51ab88, 0xbcaea23d, 0xbd3b6d8d, 0xbd6caaad > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP6_hi */ > > > > > > - .long 0xbd795bed, 0xbd5fddda, 0xbd038f3b, 0xbc1cad63 > > > > > > - .long 0x3abb4766, 0x3b95f10b, 0x3b825873, 0x3afaea66 > > > > > > - .long 0x3a49f878, 0x39996bf3, 0x388f3e6c, 0x371bb0e3 > > > > > > - .long 0x35a8a5e6, 0x34369b17, 0x322487b0, 0x00000000 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP5_lo */ > > > > > > - .long 0xb76dd6b9, 0xbe1c276d, 0x3c1dcf2f, 0x3dc1a78d > > > > > > - .long 0x3d96f985, 0x3da2b61b, 0x3dc13397, 0x3dd2f670 > > > > > > - .long 0x3df48a0a, 0x3e06c5a8, 0x3e1a3aba, 0x3e27c405 > > > > > > - .long 0x3e2e78d0, 0x3e2c3e44, 0x3e1d3097, 0x3df4a8f4 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP5_hi */ > > > > > > - .long 0x3da38508, 0x3d31416a, 0x3b562657, 0xbcaeeac9 > > > > > > - .long 0xbcce9419, 0xbcaaeac4, 0xbc49e7d0, 0xbba71ddd > > > > > > - .long 0xbb003b0e, 0xba3f9a05, 0xb92c08a7, 0xb7ba9232 > > > > > > - .long 0xb64a0b0f, 0xb4dac169, 0xb2ab78ac, 0x00000000 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP4_lo */ > > > > > > - .long 0xbeaaaaa5, 0xbeab0612, 0xbea7f01f, 0xbea4e120 > > > > > > - .long 0xbea387b7, 0xbea15962, 0xbe9d57f7, 0xbe976b5a > > > > > > - .long 0xbe90230d, 0xbe880dff, 0xbe7479b3, 0xbe4c3d88 > > > > > > - .long 0xbe212482, 0xbdeb8cba, 0xbd5e78ad, 0x3c6b5e6e > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP4_hi */ > > > > > > - .long 0x3d839143, 0x3dc21ee1, 0x3de347af, 0x3dcbec96 > > > > > > - .long 0x3d99ef2d, 0x3d542ea1, 0x3cdde701, 0x3c2cca67 > > > > > > - .long 0x3b81cb27, 0x3ac073a1, 0x39ac3032, 0x383a94d9 > > > > > > - .long 0x36ca081d, 0x355abd4c, 0x332b3cb6, 0x00000000 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP3_lo */ > > > > > > - .long 0xb0343c7b, 0xbd6ee69d, 0xbd8f0da7, 0xbdae477d > > > > > > - .long 0xbdcd2a1f, 0xbdeba80d, 0xbe0c443b, 0xbe293cf3 > > > > > > - .long 0xbe44f282, 0xbe5f3651, 0xbe81c7c0, 0xbe96d7ca > > > > > > - .long 0xbea7fb8e, 0xbeb50e9e, 0xbec12efe, 0xbec4be92 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP3_hi */ > > > > > > - .long 0xbebce070, 0xbead510e, 0xbe8ef7d6, 0xbe4b8704 > > > > > > - .long 0xbe083237, 0xbdaf7449, 0xbd2e1ec4, 0xbc83bf06 > > > > > > - .long 0xbbc3e0b5, 0xbb10aadc, 0xba0157db, 0xb88c18f2 > > > > > > - .long 0xb717b096, 0xb5a43bae, 0xb383012c, 0x00000000 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP2_lo */ > > > > > > - .long 0x3f800000, 0x3f7f1f84, 0x3f7ebd11, 0x3f7e1e5f > > > > > > - .long 0x3f7d609f, 0x3f7c842d, 0x3f7b00e5, 0x3f789580 > > > > > > - .long 0x3f75b8ad, 0x3f726fd9, 0x3f6cc59b, 0x3f63fb92 > > > > > > - .long 0x3f59ff97, 0x3f4f11d7, 0x3f3d7573, 0x3f24f360 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP2_hi */ > > > > > > - .long 0x3f0cbfe7, 0x3eec1a69, 0x3eb0a801, 0x3e6753a2 > > > > > > - .long 0x3e132f1a, 0x3db7e7d3, 0x3d320845, 0x3c84d3d4 > > > > > > - .long 0x3bc477b7, 0x3b10d3da, 0x3a01601e, 0x388c1a3b > > > > > > - .long 0x3717b0da, 0x35a43bce, 0x338306c6, 0x00000000 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP0_lo */ > > > > > > - .long 0x00000000, 0x3d6fb9c9, 0x3d8fc35f, 0x3daf9169 > > > > > > - .long 0x3dcf49ab, 0x3deee849, 0x3e0f0ee8, 0x3e2e4984 > > > > > > - .long 0x3e4d2f8e, 0x3e6bb32e, 0x3e8c51cd, 0x3ea96163 > > > > > > - .long 0x3ec543f1, 0x3edfd735, 0x3f028438, 0x3f18abf0 > > > > > > - > > > > > > - .align 64 > > > > > > - /* _sP0_hi */ > > > > > > - .long 0x3f2bc480, 0x3f3bec1c, 0x3f4f2e5b, 0x3f613c53 > > > > > > - .long 0x3f6ce37d, 0x3f743c4f, 0x3f7a5feb, 0x3f7dea85 > > > > > > - .long 0x3f7f3b3d, 0x3f7fb78c, 0x3f7fefd4, 0x3f7ffdd0 > > > > > > - .long 0x3f7fffb4, 0x3f7ffff6, 0x3f7fffff, 0x3f800000 > > > > > > +LOCAL_DATA_NAME_UNALIGNED: > > > > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iExpMantMask_UISA, 0x7fe00000) > > > > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iMinIdxOfsMask_UISA, 0x3d400000) > > > > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iMaxIdxMask_UISA, 0x03e00000) > > > > > > + float_block (LOCAL_DATA_NAME_UNALIGNED, _iExpMask, 0x7f000000) > > > > > > + .type LOCAL_DATA_NAME_UNALIGNED, @object > > > > > > + .size LOCAL_DATA_NAME_UNALIGNED, .-LOCAL_DATA_NAME_UNALIGNED > > > > > > > > > > > > .align 64 > > > > > > - .type __svml_stanh_data_internal_al64, @object > > > > > > - .size __svml_stanh_data_internal_al64, .-__svml_stanh_data_internal_al64 > > > > > > - .type __svml_stanh_data_internal, @object > > > > > > - .size __svml_stanh_data_internal, .-__svml_stanh_data_internal > > > > > > +LOCAL_DATA_NAME: > > > > > > + float_block (LOCAL_DATA_NAME, _sC_lo, > > > > > > + 0x00000000, 0x3d700000, 0x3d900000, 0x3db00000, > > > > > > + 0x3dd00000, 0x3df00000, 0x3e100000, 0x3e300000, > > > > > > + 0x3e500000, 0x3e700000, 0x3e900000, 0x3eb00000, > > > > > > + 0x3ed00000, 0x3ef00000, 0x3f100000, 0x3f300000) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sC_hi, > > > > > > + 0x3f500000, 0x3f700000, 0x3f900000, 0x3fb00000, > > > > > > + 0x3fd00000, 0x3ff00000, 0x40100000, 0x40300000, > > > > > > + 0x40500000, 0x40700000, 0x40900000, 0x40b00000, > > > > > > + 0x40d00000, 0x40f00000, 0x41100000, 0x00000000) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP7_lo, > > > > > > + 0xbc0e2f66, 0x460bda12, 0x43d638ef, 0xc3e11c3e, > > > > > > + 0xc2baa4e9, 0xc249da2d, 0xc1859b82, 0x40dd5b57, > > > > > > + 0x40494640, 0x40c730a8, 0xbf0f160e, 0x3e30e76f, > > > > > > + 0xbea81387, 0xbdb26a1c, 0xbd351e57, 0xbb4c01a0) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP7_hi, > > > > > > + 0x3c1d7bfb, 0x3c722cd1, 0x3c973f1c, 0x3c33a31b, > > > > > > + 0x3b862ef4, 0x3a27b3d0, 0xba3b5907, 0xba0efc22, > > > > > > + 0xb97f9f0f, 0xb8c8af50, 0xb7bdddfb, 0xb64f2950, > > > > > > + 0xb4e085b1, 0xb3731dfa, 0xb15a1f04, 0x00000000) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP6_lo, > > > > > > + 0x3e0910e9, 0x43761143, 0x4165ecdc, 0xc190f756, > > > > > > + 0xc08c097d, 0xc02ba813, 0xbf7f6bda, 0x3f2b1dc0, > > > > > > + 0x3ece105d, 0x3f426a94, 0xbadb0dc4, 0x3da43b17, > > > > > > + 0xbd51ab88, 0xbcaea23d, 0xbd3b6d8d, 0xbd6caaad) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP6_hi, > > > > > > + 0xbd795bed, 0xbd5fddda, 0xbd038f3b, 0xbc1cad63, > > > > > > + 0x3abb4766, 0x3b95f10b, 0x3b825873, 0x3afaea66, > > > > > > + 0x3a49f878, 0x39996bf3, 0x388f3e6c, 0x371bb0e3, > > > > > > + 0x35a8a5e6, 0x34369b17, 0x322487b0, 0x00000000) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP5_lo, > > > > > > + 0xb76dd6b9, 0xbe1c276d, 0x3c1dcf2f, 0x3dc1a78d, > > > > > > + 0x3d96f985, 0x3da2b61b, 0x3dc13397, 0x3dd2f670, > > > > > > + 0x3df48a0a, 0x3e06c5a8, 0x3e1a3aba, 0x3e27c405, > > > > > > + 0x3e2e78d0, 0x3e2c3e44, 0x3e1d3097, 0x3df4a8f4) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP5_hi, > > > > > > + 0x3da38508, 0x3d31416a, 0x3b562657, 0xbcaeeac9, > > > > > > + 0xbcce9419, 0xbcaaeac4, 0xbc49e7d0, 0xbba71ddd, > > > > > > + 0xbb003b0e, 0xba3f9a05, 0xb92c08a7, 0xb7ba9232, > > > > > > + 0xb64a0b0f, 0xb4dac169, 0xb2ab78ac, 0x00000000) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP4_lo, > > > > > > + 0xbeaaaaa5, 0xbeab0612, 0xbea7f01f, 0xbea4e120, > > > > > > + 0xbea387b7, 0xbea15962, 0xbe9d57f7, 0xbe976b5a, > > > > > > + 0xbe90230d, 0xbe880dff, 0xbe7479b3, 0xbe4c3d88, > > > > > > + 0xbe212482, 0xbdeb8cba, 0xbd5e78ad, 0x3c6b5e6e) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP4_hi, > > > > > > + 0x3d839143, 0x3dc21ee1, 0x3de347af, 0x3dcbec96, > > > > > > + 0x3d99ef2d, 0x3d542ea1, 0x3cdde701, 0x3c2cca67, > > > > > > + 0x3b81cb27, 0x3ac073a1, 0x39ac3032, 0x383a94d9, > > > > > > + 0x36ca081d, 0x355abd4c, 0x332b3cb6, 0x00000000) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP3_lo, > > > > > > + 0xb0343c7b, 0xbd6ee69d, 0xbd8f0da7, 0xbdae477d, > > > > > > + 0xbdcd2a1f, 0xbdeba80d, 0xbe0c443b, 0xbe293cf3, > > > > > > + 0xbe44f282, 0xbe5f3651, 0xbe81c7c0, 0xbe96d7ca, > > > > > > + 0xbea7fb8e, 0xbeb50e9e, 0xbec12efe, 0xbec4be92) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP3_hi, > > > > > > + 0xbebce070, 0xbead510e, 0xbe8ef7d6, 0xbe4b8704, > > > > > > + 0xbe083237, 0xbdaf7449, 0xbd2e1ec4, 0xbc83bf06, > > > > > > + 0xbbc3e0b5, 0xbb10aadc, 0xba0157db, 0xb88c18f2, > > > > > > + 0xb717b096, 0xb5a43bae, 0xb383012c, 0x00000000) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP2_lo, > > > > > > + 0x3f800000, 0x3f7f1f84, 0x3f7ebd11, 0x3f7e1e5f, > > > > > > + 0x3f7d609f, 0x3f7c842d, 0x3f7b00e5, 0x3f789580, > > > > > > + 0x3f75b8ad, 0x3f726fd9, 0x3f6cc59b, 0x3f63fb92, > > > > > > + 0x3f59ff97, 0x3f4f11d7, 0x3f3d7573, 0x3f24f360) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP2_hi, > > > > > > + 0x3f0cbfe7, 0x3eec1a69, 0x3eb0a801, 0x3e6753a2, > > > > > > + 0x3e132f1a, 0x3db7e7d3, 0x3d320845, 0x3c84d3d4, > > > > > > + 0x3bc477b7, 0x3b10d3da, 0x3a01601e, 0x388c1a3b, > > > > > > + 0x3717b0da, 0x35a43bce, 0x338306c6, 0x00000000) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP0_lo, > > > > > > + 0x00000000, 0x3d6fb9c9, 0x3d8fc35f, 0x3daf9169, > > > > > > + 0x3dcf49ab, 0x3deee849, 0x3e0f0ee8, 0x3e2e4984, > > > > > > + 0x3e4d2f8e, 0x3e6bb32e, 0x3e8c51cd, 0x3ea96163, > > > > > > + 0x3ec543f1, 0x3edfd735, 0x3f028438, 0x3f18abf0) > > > > > > + > > > > > > + float_block (LOCAL_DATA_NAME, _sP0_hi, > > > > > > + 0x3f2bc480, 0x3f3bec1c, 0x3f4f2e5b, 0x3f613c53, > > > > > > + 0x3f6ce37d, 0x3f743c4f, 0x3f7a5feb, 0x3f7dea85, > > > > > > + 0x3f7f3b3d, 0x3f7fb78c, 0x3f7fefd4, 0x3f7ffdd0, > > > > > > + 0x3f7fffb4, 0x3f7ffff6, 0x3f7fffff, 0x3f800000) > > > > > > + > > > > > > + .type LOCAL_DATA_NAME, @object > > > > > > + .size LOCAL_DATA_NAME, .-LOCAL_DATA_NAME > > > > > > -- > > > > > > 2.34.1 > > > > > > > > > > > > > > > > The data movement makes the assembler codes much harder to follow. > > > > > Sunil, what do you think of this patch series? > > > > > > > > What do you mean? The change on in how we define rodata or the movement > > > > to multiple files or something else? > > > > > > The glibc way to support data files for assembly codes is to define > > > data in C and use *.sym to generate offsets for assembly files, like > > > > I see. Although to be fair the entire SVML codebase bucks that trend. > > It is because libmvec codes were generated by ICC and processed > by scripts. > > > Seems like a more dramatic trend to move all the offsets to C. > > Since you are adding data by hand, you should do it in C. Since the plan is to integrate this piece-meal (function by function), think it's easier to integrate into a system that matches the rest of the unyet changed files. Once all of the SVML functions have been updated it will be simple enough to script the change from ASM -> C. Thoughts? > > > > > > > sysdeps/x86/cpu-features-offsets.sym:XSAVE_STATE_SIZE_OFFSET > > > offsetof (struct cpu_features, xsave_state_size) > > > sysdeps/x86_64/dl-trampoline.h: sub > > > _rtld_local_ro+RTLD_GLOBAL_RO_DL_X86_CPU_FEATURES_OFFSET+XSAVE_STATE_SIZE_OFFSET(%rip), > > > %RSP_LP > > > sysdeps/x86_64/dl-trampoline.h: sub > > > _dl_x86_cpu_features+XSAVE_STATE_SIZE_OFFSET(%rip), %RSP_LP > > > > > > -- > > > H.J. > > > > -- > H.J.
diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S b/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S index d74fc7731d..765e9ed7f7 100644 --- a/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S +++ b/sysdeps/x86_64/fpu/multiarch/svml_s_tanhf16_core_avx512.S @@ -70,94 +70,99 @@ * */ -/* Offsets for data table __svml_stanh_data_internal_avx512. Ordered - by use in the function. On cold-starts this might help the - prefetcher. Possibly a better idea is to interleave start/end so - that the prefetcher is less likely to detect a stream and pull - irrelivant lines into cache. */ -/* Offsets for data table __svml_stanh_data_internal. 4 bytes each. - */ + +#define LOCAL_DATA_NAME __svml_stanh_data_internal +#define LOCAL_DATA_NAME_UNALIGNED __svml_stanh_data_internal_unaligned +#include "svml_s_common_evex512_rodata_offsets.h" + +/* Offsets for data table __svml_stanh_data_internal_unaligned. + 4 bytes each. */ #define _iExpMantMask_UISA 0 #define _iMinIdxOfsMask_UISA 4 #define _iMaxIdxMask_UISA 8 #define _iExpMask 12 -/* Offsets for data table __svml_stanh_data_internal_al64. 64 bytes - each. */ -#define _sC_lo 0 -#define _sC_hi 64 -#define _sP7_lo 128 -#define _sP7_hi 192 -#define _sSignMask 256 -#define _sP6_lo 320 -#define _sP6_hi 384 -#define _sP5_lo 448 -#define _sP5_hi 512 -#define _sP4_lo 576 -#define _sP4_hi 640 -#define _sP3_lo 704 -#define _sP3_hi 768 -#define _sP2_lo 832 -#define _sP2_hi 896 -#define _sP0_lo 960 -#define _sP0_hi 1024 +/* Offsets for data table __svml_stanh_data_internal. Ordered + by use in the function. On cold-starts this might help the + prefetcher. Possibly a better idea is to interleave start/end so + that the prefetcher is less likely to detect a stream and pull + irrelivant lines into cache. */ + +/* Offsets for data table __svml_stanh_data_internal. + 64 bytes each. */ +#define _sC_lo 0 +#define _sC_hi 64 +#define _sP7_lo 128 +#define _sP7_hi 192 +#define _sP6_lo 256 +#define _sP6_hi 320 +#define _sP5_lo 384 +#define _sP5_hi 448 +#define _sP4_lo 512 +#define _sP4_hi 576 +#define _sP3_lo 640 +#define _sP3_hi 704 +#define _sP2_lo 768 +#define _sP2_hi 832 +#define _sP0_lo 896 +#define _sP0_hi 960 + #include <sysdep.h> -#define TANHF_DATA(x) ((x)+__svml_stanh_data_internal_al64) -#define TANHF_DATA_UNALIGNED(x) ((x)+__svml_stanh_data_internal) .section .text.evex512, "ax", @progbits ENTRY(_ZGVeN16v_tanhf_skx) - /* Here huge arguments, INF and NaNs are filtered out to callout. */ - vpandd TANHF_DATA_UNALIGNED(_iExpMantMask_UISA)(%rip){1to16}, %zmm0, %zmm1 - vpsubd TANHF_DATA_UNALIGNED(_iMinIdxOfsMask_UISA)(%rip){1to16}, %zmm1, %zmm2 + /* Here huge arguments, INF and NaNs are filtered out to + callout. */ + vpandd LOCAL_DATA_UNALIGNED(_iExpMantMask_UISA)(%rip){1to16}, %zmm0, %zmm1 + vpsubd LOCAL_DATA_UNALIGNED(_iMinIdxOfsMask_UISA)(%rip){1to16}, %zmm1, %zmm2 /* Selection arguments between [0, 0x03e00000] into zmm3. */ vpxord %zmm3, %zmm3, %zmm3 vpmaxsd %zmm3, %zmm2, %zmm3 - vpminsd TANHF_DATA_UNALIGNED(_iMaxIdxMask_UISA)(%rip){1to16}, %zmm3, %zmm3 + vpminsd LOCAL_DATA_UNALIGNED(_iMaxIdxMask_UISA)(%rip){1to16}, %zmm3, %zmm3 /* Setup permute indices in zmm3. */ vpsrld $21, %zmm3, %zmm3 /* Store if there are any special cases in k1. */ - vpcmpd $6, TANHF_DATA_UNALIGNED(_iExpMask)(%rip){1to16}, %zmm1, %k1 + vpcmpd $6, LOCAL_DATA_UNALIGNED(_iExpMask)(%rip){1to16}, %zmm1, %k1 - vmovaps TANHF_DATA(_sC_lo)(%rip), %zmm5 - vpermt2ps TANHF_DATA(_sC_hi)(%rip), %zmm3, %zmm5 + vmovaps LOCAL_DATA(_sC_lo)(%rip), %zmm5 + vpermt2ps LOCAL_DATA(_sC_hi)(%rip), %zmm3, %zmm5 - vmovaps TANHF_DATA(_sP7_lo)(%rip), %zmm2 - vpermt2ps TANHF_DATA(_sP7_hi)(%rip), %zmm3, %zmm2 + vmovaps LOCAL_DATA(_sP7_lo)(%rip), %zmm2 + vpermt2ps LOCAL_DATA(_sP7_hi)(%rip), %zmm3, %zmm2 /* Store absolute values of inputs in zmm1. */ - vmovaps TANHF_DATA(_sSignMask)(%rip), %zmm4 + vmovaps COMMON_DATA(_SignMask)(%rip), %zmm4 vandnps %zmm0, %zmm4, %zmm1 vsubps {rn-sae}, %zmm5, %zmm1, %zmm1 - vmovaps TANHF_DATA(_sP6_lo)(%rip), %zmm5 - vpermt2ps TANHF_DATA(_sP6_hi)(%rip), %zmm3, %zmm5 + vmovaps LOCAL_DATA(_sP6_lo)(%rip), %zmm5 + vpermt2ps LOCAL_DATA(_sP6_hi)(%rip), %zmm3, %zmm5 - vmovaps TANHF_DATA(_sP5_lo)(%rip), %zmm6 - vpermt2ps TANHF_DATA(_sP5_hi)(%rip), %zmm3, %zmm6 + vmovaps LOCAL_DATA(_sP5_lo)(%rip), %zmm6 + vpermt2ps LOCAL_DATA(_sP5_hi)(%rip), %zmm3, %zmm6 vfmadd213ps {rn-sae}, %zmm5, %zmm1, %zmm2 vfmadd213ps {rn-sae}, %zmm6, %zmm1, %zmm2 - vmovaps TANHF_DATA(_sP4_lo)(%rip), %zmm7 - vpermt2ps TANHF_DATA(_sP4_hi)(%rip), %zmm3, %zmm7 + vmovaps LOCAL_DATA(_sP4_lo)(%rip), %zmm7 + vpermt2ps LOCAL_DATA(_sP4_hi)(%rip), %zmm3, %zmm7 - vmovaps TANHF_DATA(_sP3_lo)(%rip), %zmm8 - vpermt2ps TANHF_DATA(_sP3_hi)(%rip), %zmm3, %zmm8 + vmovaps LOCAL_DATA(_sP3_lo)(%rip), %zmm8 + vpermt2ps LOCAL_DATA(_sP3_hi)(%rip), %zmm3, %zmm8 vfmadd213ps {rn-sae}, %zmm7, %zmm1, %zmm2 vfmadd213ps {rn-sae}, %zmm8, %zmm1, %zmm2 - vmovaps TANHF_DATA(_sP2_lo)(%rip), %zmm9 - vpermt2ps TANHF_DATA(_sP2_hi)(%rip), %zmm3, %zmm9 + vmovaps LOCAL_DATA(_sP2_lo)(%rip), %zmm9 + vpermt2ps LOCAL_DATA(_sP2_hi)(%rip), %zmm3, %zmm9 - vmovaps TANHF_DATA(_sP0_lo)(%rip), %zmm10 - vpermt2ps TANHF_DATA(_sP0_hi)(%rip), %zmm3, %zmm10 + vmovaps LOCAL_DATA(_sP0_lo)(%rip), %zmm10 + vpermt2ps LOCAL_DATA(_sP0_hi)(%rip), %zmm3, %zmm10 vfmadd213ps {rn-sae}, %zmm9, %zmm1, %zmm2 vfmadd213ps {rn-sae}, %zmm10, %zmm1, %zmm2 @@ -167,7 +172,7 @@ ENTRY(_ZGVeN16v_tanhf_skx) /* Go to special inputs processing branch. */ jne L(SPECIAL_VALUES_BRANCH) - # LOE rbx r12 r13 r14 r15 zmm0 zmm2 zmm4 + /* Wait until after branch of write over zmm0. */ vpternlogd $0xec, %zmm4, %zmm2, %zmm0 @@ -176,24 +181,24 @@ ENTRY(_ZGVeN16v_tanhf_skx) /* Cold case. edx has 1s where there was a special value that needs to be handled by a tanhf call. Optimize for code size - more so than speed here. */ + more so than speed here. */ L(SPECIAL_VALUES_BRANCH): - # LOE rbx rdx r12 r13 r14 r15 zmm0 zmm2 zmm4 - /* Use r13 to save/restore the stack. This allows us to use rbp as - callee save register saving code size. */ + + /* Use r13 to save/restore the stack. This allows us to use rbp + as callee save register saving code size. */ pushq %r13 - cfi_adjust_cfa_offset(8) - cfi_offset(r13, -16) - /* Need to callee save registers to preserve state across tanhf calls. - */ + cfi_adjust_cfa_offset (8) + cfi_offset (r13, -16) + /* Need to callee save registers to preserve state across tanhf + calls. */ pushq %rbx - cfi_adjust_cfa_offset(8) - cfi_offset(rbx, -24) + cfi_adjust_cfa_offset (8) + cfi_offset (rbx, -24) pushq %rbp - cfi_adjust_cfa_offset(8) - cfi_offset(rbp, -32) + cfi_adjust_cfa_offset (8) + cfi_offset (rbp, -32) movq %rsp, %r13 - cfi_def_cfa_register(r13) + cfi_def_cfa_register (r13) /* Align stack and make room for 2x zmm vectors. */ andq $-64, %rsp @@ -207,16 +212,17 @@ L(SPECIAL_VALUES_BRANCH): vzeroupper - /* edx has 1s where there was a special value that needs to be handled - by a tanhf call. */ + /* edx has 1s where there was a special value that needs to be + handled by a tanhf call. */ movl %edx, %ebx L(SPECIAL_VALUES_LOOP): - # LOE rbx rbp r12 r13 r14 r15 - /* use rbp as index for special value that is saved across calls to - tanhf. We technically don't need a callee save register here as offset - to rsp is always [0, 56] so we can restore rsp by realigning to 64. - Essentially the tradeoff is 1 extra save/restore vs 2 extra instructions - in the loop. Realigning also costs more code size. */ + + /* use rbp as index for special value that is saved across calls + to tanhf. We technically don't need a callee save register + here as offset to rsp is always [0, 56] so we can restore + rsp by realigning to 64. Essentially the tradeoff is 1 extra + save/restore vs 2 extra instructions in the loop. Realigning + also costs more code size. */ xorl %ebp, %ebp tzcntl %ebx, %ebp @@ -224,203 +230,141 @@ L(SPECIAL_VALUES_LOOP): vmovss 64(%rsp, %rbp, 4), %xmm0 call tanhf@PLT - /* No good way to avoid the store-forwarding fault this will cause on - return. `lfence` avoids the SF fault but at greater cost as it - serialized stack/callee save restoration. */ + /* No good way to avoid the store-forwarding fault this will + cause on return. `lfence` avoids the SF fault but at greater + cost as it serialized stack/callee save restoration. */ vmovss %xmm0, (%rsp, %rbp, 4) - blsrl %ebx, %ebx + blsrl %ebx, %ebx jnz L(SPECIAL_VALUES_LOOP) - # LOE r12 r13 r14 r15 + /* All results have been written to (%rsp). */ vmovaps (%rsp), %zmm0 /* Restore rsp. */ movq %r13, %rsp - cfi_def_cfa_register(rsp) + cfi_def_cfa_register (rsp) /* Restore callee save registers. */ popq %rbp - cfi_adjust_cfa_offset(-8) - cfi_restore(rbp) + cfi_adjust_cfa_offset (-8) + cfi_restore (rbp) popq %rbx - cfi_adjust_cfa_offset(-8) - cfi_restore(rbp) + cfi_adjust_cfa_offset (-8) + cfi_restore (rbp) popq %r13 - cfi_adjust_cfa_offset(-8) - cfi_restore(r13) + cfi_adjust_cfa_offset (-8) + cfi_restore (r13) ret END(_ZGVeN16v_tanhf_skx) - .section .rodata, "a" + .section .rodata.evex512, "a" .align 16 -#ifdef __svml_stanh_data_internal_typedef -typedef unsigned int VUINT32; -typedef struct - { - __declspec(align(4)) VUINT32 _iExpMantMask_UISA[1][1]; - __declspec(align(4)) VUINT32 _iMinIdxOfsMask_UISA[1][1]; - __declspec(align(4)) VUINT32 _iMaxIdxMask_UISA[1][1]; - __declspec(align(4)) VUINT32 _iExpMask[1][1]; - __declspec(align(64)) VUINT32 _sC_lo[16][1]; - __declspec(align(64)) VUINT32 _sC_hi[16][1]; - __declspec(align(64)) VUINT32 _sP7_lo[16][1]; - __declspec(align(64)) VUINT32 _sP7_hi[16][1]; - __declspec(align(64)) VUINT32 _sSignMask[16][1]; - __declspec(align(64)) VUINT32 _sP6_lo[16][1]; - __declspec(align(64)) VUINT32 _sP6_hi[16][1]; - __declspec(align(64)) VUINT32 _sP5_lo[16][1]; - __declspec(align(64)) VUINT32 _sP5_hi[16][1]; - __declspec(align(64)) VUINT32 _sP4_lo[16][1]; - __declspec(align(64)) VUINT32 _sP4_hi[16][1]; - __declspec(align(64)) VUINT32 _sP3_lo[16][1]; - __declspec(align(64)) VUINT32 _sP3_hi[16][1]; - __declspec(align(64)) VUINT32 _sP2_lo[16][1]; - __declspec(align(64)) VUINT32 _sP2_hi[16][1]; - __declspec(align(64)) VUINT32 _sP0_lo[16][1]; - __declspec(align(64)) VUINT32 _sP0_hi[16][1]; -} __svml_stanh_data_internal; -#endif - -__svml_stanh_data_internal: - .align 4 - /* _iExpMantMask_UISA */ - .long 0x7fe00000 - - .align 4 - /* _iMinIdxOfsMask_UISA */ - .long 0x3d400000 - - .align 4 - /* _iMaxIdxMask_UISA */ - .long 0x03e00000 - - .align 4 - /* _iExpMask */ - .long 0x7f000000 - - .align 64 -__svml_stanh_data_internal_al64: - .align 64 - /* _sC_lo */ - .long 0x00000000, 0x3d700000, 0x3d900000, 0x3db00000 - .long 0x3dd00000, 0x3df00000, 0x3e100000, 0x3e300000 - .long 0x3e500000, 0x3e700000, 0x3e900000, 0x3eb00000 - .long 0x3ed00000, 0x3ef00000, 0x3f100000, 0x3f300000 - - .align 64 - /* _sC_hi */ - .long 0x3f500000, 0x3f700000, 0x3f900000, 0x3fb00000 - .long 0x3fd00000, 0x3ff00000, 0x40100000, 0x40300000 - .long 0x40500000, 0x40700000, 0x40900000, 0x40b00000 - .long 0x40d00000, 0x40f00000, 0x41100000, 0x00000000 - - .align 64 - /* _sP7_lo */ - .long 0xbc0e2f66, 0x460bda12, 0x43d638ef, 0xc3e11c3e - .long 0xc2baa4e9, 0xc249da2d, 0xc1859b82, 0x40dd5b57 - .long 0x40494640, 0x40c730a8, 0xbf0f160e, 0x3e30e76f - .long 0xbea81387, 0xbdb26a1c, 0xbd351e57, 0xbb4c01a0 - - .align 64 - /* _sP7_hi */ - .long 0x3c1d7bfb, 0x3c722cd1, 0x3c973f1c, 0x3c33a31b - .long 0x3b862ef4, 0x3a27b3d0, 0xba3b5907, 0xba0efc22 - .long 0xb97f9f0f, 0xb8c8af50, 0xb7bdddfb, 0xb64f2950 - .long 0xb4e085b1, 0xb3731dfa, 0xb15a1f04, 0x00000000 - .align 64 - /* _sSignMask */ - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 - .long 0x80000000, 0x80000000, 0x80000000, 0x80000000 - - .align 64 - /* _sP6_lo */ - .long 0x3e0910e9, 0x43761143, 0x4165ecdc, 0xc190f756 - .long 0xc08c097d, 0xc02ba813, 0xbf7f6bda, 0x3f2b1dc0 - .long 0x3ece105d, 0x3f426a94, 0xbadb0dc4, 0x3da43b17 - .long 0xbd51ab88, 0xbcaea23d, 0xbd3b6d8d, 0xbd6caaad - - .align 64 - /* _sP6_hi */ - .long 0xbd795bed, 0xbd5fddda, 0xbd038f3b, 0xbc1cad63 - .long 0x3abb4766, 0x3b95f10b, 0x3b825873, 0x3afaea66 - .long 0x3a49f878, 0x39996bf3, 0x388f3e6c, 0x371bb0e3 - .long 0x35a8a5e6, 0x34369b17, 0x322487b0, 0x00000000 - - .align 64 - /* _sP5_lo */ - .long 0xb76dd6b9, 0xbe1c276d, 0x3c1dcf2f, 0x3dc1a78d - .long 0x3d96f985, 0x3da2b61b, 0x3dc13397, 0x3dd2f670 - .long 0x3df48a0a, 0x3e06c5a8, 0x3e1a3aba, 0x3e27c405 - .long 0x3e2e78d0, 0x3e2c3e44, 0x3e1d3097, 0x3df4a8f4 - - .align 64 - /* _sP5_hi */ - .long 0x3da38508, 0x3d31416a, 0x3b562657, 0xbcaeeac9 - .long 0xbcce9419, 0xbcaaeac4, 0xbc49e7d0, 0xbba71ddd - .long 0xbb003b0e, 0xba3f9a05, 0xb92c08a7, 0xb7ba9232 - .long 0xb64a0b0f, 0xb4dac169, 0xb2ab78ac, 0x00000000 - - .align 64 - /* _sP4_lo */ - .long 0xbeaaaaa5, 0xbeab0612, 0xbea7f01f, 0xbea4e120 - .long 0xbea387b7, 0xbea15962, 0xbe9d57f7, 0xbe976b5a - .long 0xbe90230d, 0xbe880dff, 0xbe7479b3, 0xbe4c3d88 - .long 0xbe212482, 0xbdeb8cba, 0xbd5e78ad, 0x3c6b5e6e - - .align 64 - /* _sP4_hi */ - .long 0x3d839143, 0x3dc21ee1, 0x3de347af, 0x3dcbec96 - .long 0x3d99ef2d, 0x3d542ea1, 0x3cdde701, 0x3c2cca67 - .long 0x3b81cb27, 0x3ac073a1, 0x39ac3032, 0x383a94d9 - .long 0x36ca081d, 0x355abd4c, 0x332b3cb6, 0x00000000 - - .align 64 - /* _sP3_lo */ - .long 0xb0343c7b, 0xbd6ee69d, 0xbd8f0da7, 0xbdae477d - .long 0xbdcd2a1f, 0xbdeba80d, 0xbe0c443b, 0xbe293cf3 - .long 0xbe44f282, 0xbe5f3651, 0xbe81c7c0, 0xbe96d7ca - .long 0xbea7fb8e, 0xbeb50e9e, 0xbec12efe, 0xbec4be92 - - .align 64 - /* _sP3_hi */ - .long 0xbebce070, 0xbead510e, 0xbe8ef7d6, 0xbe4b8704 - .long 0xbe083237, 0xbdaf7449, 0xbd2e1ec4, 0xbc83bf06 - .long 0xbbc3e0b5, 0xbb10aadc, 0xba0157db, 0xb88c18f2 - .long 0xb717b096, 0xb5a43bae, 0xb383012c, 0x00000000 - - .align 64 - /* _sP2_lo */ - .long 0x3f800000, 0x3f7f1f84, 0x3f7ebd11, 0x3f7e1e5f - .long 0x3f7d609f, 0x3f7c842d, 0x3f7b00e5, 0x3f789580 - .long 0x3f75b8ad, 0x3f726fd9, 0x3f6cc59b, 0x3f63fb92 - .long 0x3f59ff97, 0x3f4f11d7, 0x3f3d7573, 0x3f24f360 - - .align 64 - /* _sP2_hi */ - .long 0x3f0cbfe7, 0x3eec1a69, 0x3eb0a801, 0x3e6753a2 - .long 0x3e132f1a, 0x3db7e7d3, 0x3d320845, 0x3c84d3d4 - .long 0x3bc477b7, 0x3b10d3da, 0x3a01601e, 0x388c1a3b - .long 0x3717b0da, 0x35a43bce, 0x338306c6, 0x00000000 - - .align 64 - /* _sP0_lo */ - .long 0x00000000, 0x3d6fb9c9, 0x3d8fc35f, 0x3daf9169 - .long 0x3dcf49ab, 0x3deee849, 0x3e0f0ee8, 0x3e2e4984 - .long 0x3e4d2f8e, 0x3e6bb32e, 0x3e8c51cd, 0x3ea96163 - .long 0x3ec543f1, 0x3edfd735, 0x3f028438, 0x3f18abf0 - - .align 64 - /* _sP0_hi */ - .long 0x3f2bc480, 0x3f3bec1c, 0x3f4f2e5b, 0x3f613c53 - .long 0x3f6ce37d, 0x3f743c4f, 0x3f7a5feb, 0x3f7dea85 - .long 0x3f7f3b3d, 0x3f7fb78c, 0x3f7fefd4, 0x3f7ffdd0 - .long 0x3f7fffb4, 0x3f7ffff6, 0x3f7fffff, 0x3f800000 +LOCAL_DATA_NAME_UNALIGNED: + float_block (LOCAL_DATA_NAME_UNALIGNED, _iExpMantMask_UISA, 0x7fe00000) + float_block (LOCAL_DATA_NAME_UNALIGNED, _iMinIdxOfsMask_UISA, 0x3d400000) + float_block (LOCAL_DATA_NAME_UNALIGNED, _iMaxIdxMask_UISA, 0x03e00000) + float_block (LOCAL_DATA_NAME_UNALIGNED, _iExpMask, 0x7f000000) + .type LOCAL_DATA_NAME_UNALIGNED, @object + .size LOCAL_DATA_NAME_UNALIGNED, .-LOCAL_DATA_NAME_UNALIGNED .align 64 - .type __svml_stanh_data_internal_al64, @object - .size __svml_stanh_data_internal_al64, .-__svml_stanh_data_internal_al64 - .type __svml_stanh_data_internal, @object - .size __svml_stanh_data_internal, .-__svml_stanh_data_internal +LOCAL_DATA_NAME: + float_block (LOCAL_DATA_NAME, _sC_lo, + 0x00000000, 0x3d700000, 0x3d900000, 0x3db00000, + 0x3dd00000, 0x3df00000, 0x3e100000, 0x3e300000, + 0x3e500000, 0x3e700000, 0x3e900000, 0x3eb00000, + 0x3ed00000, 0x3ef00000, 0x3f100000, 0x3f300000) + + float_block (LOCAL_DATA_NAME, _sC_hi, + 0x3f500000, 0x3f700000, 0x3f900000, 0x3fb00000, + 0x3fd00000, 0x3ff00000, 0x40100000, 0x40300000, + 0x40500000, 0x40700000, 0x40900000, 0x40b00000, + 0x40d00000, 0x40f00000, 0x41100000, 0x00000000) + + float_block (LOCAL_DATA_NAME, _sP7_lo, + 0xbc0e2f66, 0x460bda12, 0x43d638ef, 0xc3e11c3e, + 0xc2baa4e9, 0xc249da2d, 0xc1859b82, 0x40dd5b57, + 0x40494640, 0x40c730a8, 0xbf0f160e, 0x3e30e76f, + 0xbea81387, 0xbdb26a1c, 0xbd351e57, 0xbb4c01a0) + + float_block (LOCAL_DATA_NAME, _sP7_hi, + 0x3c1d7bfb, 0x3c722cd1, 0x3c973f1c, 0x3c33a31b, + 0x3b862ef4, 0x3a27b3d0, 0xba3b5907, 0xba0efc22, + 0xb97f9f0f, 0xb8c8af50, 0xb7bdddfb, 0xb64f2950, + 0xb4e085b1, 0xb3731dfa, 0xb15a1f04, 0x00000000) + + float_block (LOCAL_DATA_NAME, _sP6_lo, + 0x3e0910e9, 0x43761143, 0x4165ecdc, 0xc190f756, + 0xc08c097d, 0xc02ba813, 0xbf7f6bda, 0x3f2b1dc0, + 0x3ece105d, 0x3f426a94, 0xbadb0dc4, 0x3da43b17, + 0xbd51ab88, 0xbcaea23d, 0xbd3b6d8d, 0xbd6caaad) + + float_block (LOCAL_DATA_NAME, _sP6_hi, + 0xbd795bed, 0xbd5fddda, 0xbd038f3b, 0xbc1cad63, + 0x3abb4766, 0x3b95f10b, 0x3b825873, 0x3afaea66, + 0x3a49f878, 0x39996bf3, 0x388f3e6c, 0x371bb0e3, + 0x35a8a5e6, 0x34369b17, 0x322487b0, 0x00000000) + + float_block (LOCAL_DATA_NAME, _sP5_lo, + 0xb76dd6b9, 0xbe1c276d, 0x3c1dcf2f, 0x3dc1a78d, + 0x3d96f985, 0x3da2b61b, 0x3dc13397, 0x3dd2f670, + 0x3df48a0a, 0x3e06c5a8, 0x3e1a3aba, 0x3e27c405, + 0x3e2e78d0, 0x3e2c3e44, 0x3e1d3097, 0x3df4a8f4) + + float_block (LOCAL_DATA_NAME, _sP5_hi, + 0x3da38508, 0x3d31416a, 0x3b562657, 0xbcaeeac9, + 0xbcce9419, 0xbcaaeac4, 0xbc49e7d0, 0xbba71ddd, + 0xbb003b0e, 0xba3f9a05, 0xb92c08a7, 0xb7ba9232, + 0xb64a0b0f, 0xb4dac169, 0xb2ab78ac, 0x00000000) + + float_block (LOCAL_DATA_NAME, _sP4_lo, + 0xbeaaaaa5, 0xbeab0612, 0xbea7f01f, 0xbea4e120, + 0xbea387b7, 0xbea15962, 0xbe9d57f7, 0xbe976b5a, + 0xbe90230d, 0xbe880dff, 0xbe7479b3, 0xbe4c3d88, + 0xbe212482, 0xbdeb8cba, 0xbd5e78ad, 0x3c6b5e6e) + + float_block (LOCAL_DATA_NAME, _sP4_hi, + 0x3d839143, 0x3dc21ee1, 0x3de347af, 0x3dcbec96, + 0x3d99ef2d, 0x3d542ea1, 0x3cdde701, 0x3c2cca67, + 0x3b81cb27, 0x3ac073a1, 0x39ac3032, 0x383a94d9, + 0x36ca081d, 0x355abd4c, 0x332b3cb6, 0x00000000) + + float_block (LOCAL_DATA_NAME, _sP3_lo, + 0xb0343c7b, 0xbd6ee69d, 0xbd8f0da7, 0xbdae477d, + 0xbdcd2a1f, 0xbdeba80d, 0xbe0c443b, 0xbe293cf3, + 0xbe44f282, 0xbe5f3651, 0xbe81c7c0, 0xbe96d7ca, + 0xbea7fb8e, 0xbeb50e9e, 0xbec12efe, 0xbec4be92) + + float_block (LOCAL_DATA_NAME, _sP3_hi, + 0xbebce070, 0xbead510e, 0xbe8ef7d6, 0xbe4b8704, + 0xbe083237, 0xbdaf7449, 0xbd2e1ec4, 0xbc83bf06, + 0xbbc3e0b5, 0xbb10aadc, 0xba0157db, 0xb88c18f2, + 0xb717b096, 0xb5a43bae, 0xb383012c, 0x00000000) + + float_block (LOCAL_DATA_NAME, _sP2_lo, + 0x3f800000, 0x3f7f1f84, 0x3f7ebd11, 0x3f7e1e5f, + 0x3f7d609f, 0x3f7c842d, 0x3f7b00e5, 0x3f789580, + 0x3f75b8ad, 0x3f726fd9, 0x3f6cc59b, 0x3f63fb92, + 0x3f59ff97, 0x3f4f11d7, 0x3f3d7573, 0x3f24f360) + + float_block (LOCAL_DATA_NAME, _sP2_hi, + 0x3f0cbfe7, 0x3eec1a69, 0x3eb0a801, 0x3e6753a2, + 0x3e132f1a, 0x3db7e7d3, 0x3d320845, 0x3c84d3d4, + 0x3bc477b7, 0x3b10d3da, 0x3a01601e, 0x388c1a3b, + 0x3717b0da, 0x35a43bce, 0x338306c6, 0x00000000) + + float_block (LOCAL_DATA_NAME, _sP0_lo, + 0x00000000, 0x3d6fb9c9, 0x3d8fc35f, 0x3daf9169, + 0x3dcf49ab, 0x3deee849, 0x3e0f0ee8, 0x3e2e4984, + 0x3e4d2f8e, 0x3e6bb32e, 0x3e8c51cd, 0x3ea96163, + 0x3ec543f1, 0x3edfd735, 0x3f028438, 0x3f18abf0) + + float_block (LOCAL_DATA_NAME, _sP0_hi, + 0x3f2bc480, 0x3f3bec1c, 0x3f4f2e5b, 0x3f613c53, + 0x3f6ce37d, 0x3f743c4f, 0x3f7a5feb, 0x3f7dea85, + 0x3f7f3b3d, 0x3f7fb78c, 0x3f7fefd4, 0x3f7ffdd0, + 0x3f7fffb4, 0x3f7ffff6, 0x3f7fffff, 0x3f800000) + + .type LOCAL_DATA_NAME, @object + .size LOCAL_DATA_NAME, .-LOCAL_DATA_NAME