| Message ID | 58F9C541.2090900@foss.arm.com |
|---|---|
| State | New |
| Headers | show |
Ping. https://gcc.gnu.org/ml/gcc-patches/2017-04/msg00934.html Thanks, Kyrill On 21/04/17 09:39, Kyrill Tkachov wrote: > Hi all, > > Consider the code: > typedef long long v2di __attribute__ ((vector_size (16))); > > void > store_laned (v2di x, long long *y) > { > y[0] = x[1]; > y[3] = x[0]; > } > > AArch64 GCC will generate: > store_laned: > umov x1, v0.d[0] > st1 {v0.d}[1], [x0] > str x1, [x0, 24] > ret > > It moves the zero lane into a core register and does a scalar store when instead it could have used a scalar FP store > that supports the required addressing mode: > store_laned: > st1 {v0.d}[1], [x0] > str d0, [x0, 24] > ret > > Combine already tries to match this pattern: > > Trying 10 -> 11: > Failed to match this instruction: > (set (mem:DI (plus:DI (reg/v/f:DI 76 [ y ]) > (const_int 24 [0x18])) [1 MEM[(long long int *)y_4(D) + 24B]+0 S8 A64]) > (vec_select:DI (reg/v:V2DI 75 [ x ]) > (parallel [ > (const_int 0 [0]) > ]))) > > but we don't match it in the backend. It's not hard to add it, so this patch does that for all the relevant vector modes. > With this patch we generate the second sequence above and in SPEC2006 eliminate some address computation instructions > because we use the more expressive STR instead of ST1 or we eliminate such moves to the integer registers because we > can just do the store of the D-reg. > > Bootstrapped and tested on aarch64-none-linux-gnu. > > Ok for trunk? > > Thanks, > Kyrill > > 2017-04-21 Kyrylo Tkachov <kyrylo.tkachov@arm.com> > > * config/aarch64/aarch64-simd.md (aarch64_store_lane0<mode>): > New pattern. > > 2017-04-21 Kyrylo Tkachov <kyrylo.tkachov@arm.com> > > * gcc.target/aarch64/store_lane0_str_1.c: New test. >
Ping. Thanks, Kyrill On 11/05/17 11:15, Kyrill Tkachov wrote: > Ping. > > https://gcc.gnu.org/ml/gcc-patches/2017-04/msg00934.html > > Thanks, > Kyrill > > On 21/04/17 09:39, Kyrill Tkachov wrote: >> Hi all, >> >> Consider the code: >> typedef long long v2di __attribute__ ((vector_size (16))); >> >> void >> store_laned (v2di x, long long *y) >> { >> y[0] = x[1]; >> y[3] = x[0]; >> } >> >> AArch64 GCC will generate: >> store_laned: >> umov x1, v0.d[0] >> st1 {v0.d}[1], [x0] >> str x1, [x0, 24] >> ret >> >> It moves the zero lane into a core register and does a scalar store when instead it could have used a scalar FP store >> that supports the required addressing mode: >> store_laned: >> st1 {v0.d}[1], [x0] >> str d0, [x0, 24] >> ret >> >> Combine already tries to match this pattern: >> >> Trying 10 -> 11: >> Failed to match this instruction: >> (set (mem:DI (plus:DI (reg/v/f:DI 76 [ y ]) >> (const_int 24 [0x18])) [1 MEM[(long long int *)y_4(D) + 24B]+0 S8 A64]) >> (vec_select:DI (reg/v:V2DI 75 [ x ]) >> (parallel [ >> (const_int 0 [0]) >> ]))) >> >> but we don't match it in the backend. It's not hard to add it, so this patch does that for all the relevant vector modes. >> With this patch we generate the second sequence above and in SPEC2006 eliminate some address computation instructions >> because we use the more expressive STR instead of ST1 or we eliminate such moves to the integer registers because we >> can just do the store of the D-reg. >> >> Bootstrapped and tested on aarch64-none-linux-gnu. >> >> Ok for trunk? >> >> Thanks, >> Kyrill >> >> 2017-04-21 Kyrylo Tkachov <kyrylo.tkachov@arm.com> >> >> * config/aarch64/aarch64-simd.md (aarch64_store_lane0<mode>): >> New pattern. >> >> 2017-04-21 Kyrylo Tkachov <kyrylo.tkachov@arm.com> >> >> * gcc.target/aarch64/store_lane0_str_1.c: New test. >> >
On Fri, Apr 21, 2017 at 09:39:29AM +0100, Kyrill Tkachov wrote: > Hi all, > > Consider the code: > typedef long long v2di __attribute__ ((vector_size (16))); > void > store_laned (v2di x, long long *y) > { > y[0] = x[1]; > y[3] = x[0]; > } > > AArch64 GCC will generate: > store_laned: > umov x1, v0.d[0] > st1 {v0.d}[1], [x0] > str x1, [x0, 24] > ret > > It moves the zero lane into a core register and does a scalar store when instead it could have used a scalar FP store > that supports the required addressing mode: > store_laned: > st1 {v0.d}[1], [x0] > str d0, [x0, 24] > ret > > Combine already tries to match this pattern: > > Trying 10 -> 11: > Failed to match this instruction: > (set (mem:DI (plus:DI (reg/v/f:DI 76 [ y ]) > (const_int 24 [0x18])) [1 MEM[(long long int *)y_4(D) + 24B]+0 S8 A64]) > (vec_select:DI (reg/v:V2DI 75 [ x ]) > (parallel [ > (const_int 0 [0]) > ]))) > > but we don't match it in the backend. It's not hard to add it, so this patch does that for all the relevant vector modes. > With this patch we generate the second sequence above and in SPEC2006 eliminate some address computation instructions > because we use the more expressive STR instead of ST1 or we eliminate such moves to the integer registers because we > can just do the store of the D-reg. Good spot! > Bootstrapped and tested on aarch64-none-linux-gnu. > > Ok for trunk? OK. Thanks, James > 2017-04-21 Kyrylo Tkachov <kyrylo.tkachov@arm.com> > > * config/aarch64/aarch64-simd.md (aarch64_store_lane0<mode>): > New pattern. > > 2017-04-21 Kyrylo Tkachov <kyrylo.tkachov@arm.com> > > * gcc.target/aarch64/store_lane0_str_1.c: New test. >
diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md index f36665e27bb66e0f1fb42443ce7b506bd2bf6914..bf13f0753a856a13ae92ceeb44291df3dc379a13 100644 --- a/gcc/config/aarch64/aarch64-simd.md +++ b/gcc/config/aarch64/aarch64-simd.md @@ -153,6 +153,19 @@ (define_insn "*aarch64_simd_mov<mode>" (set_attr "length" "4,4,4,8,8,8,4")] ) +;; When storing lane zero we can use the normal STR and its more permissive +;; addressing modes. + +(define_insn "aarch64_store_lane0<mode>" + [(set (match_operand:<VEL> 0 "memory_operand" "=m") + (vec_select:<VEL> (match_operand:VALL_F16 1 "register_operand" "w") + (parallel [(match_operand 2 "const_int_operand" "n")])))] + "TARGET_SIMD + && ENDIAN_LANE_N (<MODE>mode, INTVAL (operands[2])) == 0" + "str\\t%<Vetype>1, %0" + [(set_attr "type" "neon_store1_1reg<q>")] +) + (define_insn "load_pair<mode>" [(set (match_operand:VD 0 "register_operand" "=w") (match_operand:VD 1 "aarch64_mem_pair_operand" "Ump")) diff --git a/gcc/testsuite/gcc.target/aarch64/store_lane0_str_1.c b/gcc/testsuite/gcc.target/aarch64/store_lane0_str_1.c new file mode 100644 index 0000000000000000000000000000000000000000..4464fec2c1f24c212be4fc6c94b509843fd0058e --- /dev/null +++ b/gcc/testsuite/gcc.target/aarch64/store_lane0_str_1.c @@ -0,0 +1,54 @@ +/* { dg-do compile } */ +/* { dg-options "-O2" } */ + +typedef int v2si __attribute__ ((vector_size (8))); +typedef float v2sf __attribute__ ((vector_size (8))); +typedef short v4hi __attribute__ ((vector_size (8))); +typedef __fp16 v4hf __attribute__ ((vector_size (8))); +typedef char v8qi __attribute__ ((vector_size (8))); + +typedef int v4si __attribute__ ((vector_size (16))); +typedef float v4sf __attribute__ ((vector_size (16))); +typedef short v8hi __attribute__ ((vector_size (16))); +typedef __fp16 v8hf __attribute__ ((vector_size (16))); +typedef char v16qi __attribute__ ((vector_size (16))); +typedef long long v2di __attribute__ ((vector_size (16))); +typedef double v2df __attribute__ ((vector_size (16))); + +#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ +#define LANE(N) (N - 1) +#else +#define LANE(N) 0 +#endif + +#define FUNC(T, E, N) \ +void \ +store_lane_##T (T x, E *y) \ +{ \ + y[0] = x[N - 1 - LANE (N)]; \ + y[3] = x[LANE (N)]; \ +} + +FUNC (v2si, int, 2) +FUNC (v2sf, float, 2) +FUNC (v4hi, short, 4) +FUNC (v4hf, __fp16, 4) +FUNC (v8qi, char, 8) + +FUNC (v4si, int, 4) +FUNC (v4sf, float, 4) +FUNC (v8hi, short, 8) +FUNC (v8hf, __fp16, 8) +FUNC (v16qi, char, 16) +FUNC (v2di, long long, 2) +FUNC (v2df, double, 2) + +/* When storing lane zero of a vector we can use the scalar STR instruction + that supports more addressing modes. */ + +/* { dg-final { scan-assembler-times "str\ts\[0-9\]+" 4 } } */ +/* { dg-final { scan-assembler-times "str\tb\[0-9\]+" 2 } } */ +/* { dg-final { scan-assembler-times "str\th\[0-9\]+" 4 } } */ +/* { dg-final { scan-assembler-times "str\td\[0-9\]+" 2 } } */ +/* { dg-final { scan-assembler-not "umov" } } */ +/* { dg-final { scan-assembler-not "dup" } } */
Hi all, Consider the code: typedef long long v2di __attribute__ ((vector_size (16))); void store_laned (v2di x, long long *y) { y[0] = x[1]; y[3] = x[0]; } AArch64 GCC will generate: store_laned: umov x1, v0.d[0] st1 {v0.d}[1], [x0] str x1, [x0, 24] ret It moves the zero lane into a core register and does a scalar store when instead it could have used a scalar FP store that supports the required addressing mode: store_laned: st1 {v0.d}[1], [x0] str d0, [x0, 24] ret Combine already tries to match this pattern: Trying 10 -> 11: Failed to match this instruction: (set (mem:DI (plus:DI (reg/v/f:DI 76 [ y ]) (const_int 24 [0x18])) [1 MEM[(long long int *)y_4(D) + 24B]+0 S8 A64]) (vec_select:DI (reg/v:V2DI 75 [ x ]) (parallel [ (const_int 0 [0]) ]))) but we don't match it in the backend. It's not hard to add it, so this patch does that for all the relevant vector modes. With this patch we generate the second sequence above and in SPEC2006 eliminate some address computation instructions because we use the more expressive STR instead of ST1 or we eliminate such moves to the integer registers because we can just do the store of the D-reg. Bootstrapped and tested on aarch64-none-linux-gnu. Ok for trunk? Thanks, Kyrill 2017-04-21 Kyrylo Tkachov <kyrylo.tkachov@arm.com> * config/aarch64/aarch64-simd.md (aarch64_store_lane0<mode>): New pattern. 2017-04-21 Kyrylo Tkachov <kyrylo.tkachov@arm.com> * gcc.target/aarch64/store_lane0_str_1.c: New test.