Message ID | ZjpQTrCuL3dr3RK4@tucnak |
---|---|
State | New |
Headers | show |
Series | expansion: Use __trunchfbf2 calls rather than __extendhfbf2 [PR114907] | expand |
> Am 07.05.2024 um 18:02 schrieb Jakub Jelinek <jakub@redhat.com>: > > Hi! > > The HF and BF modes have the same size/precision and neither is > a subset nor superset of the other. > So, using either __extendhfbf2 or __trunchfbf2 is weird. > The expansion apparently emits __extendhfbf2, but on the libgcc side > we apparently have __trunchfbf2 implemented. > > I think it is easier to switch to using what is available rather than > adding new entrypoints to libgcc, even alias, because this is backportable. > > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? Ok - do we have any target patterns that need adjustments? Thanks, Richard > 2024-05-07 Jakub Jelinek <jakub@redhat.com> > > PR middle-end/114907 > * expr.cc (convert_mode_scalar): Use trunc_optab rather than > sext_optab for HF->BF conversions. > * optabs-libfuncs.cc (gen_trunc_conv_libfunc): Likewise. > > * gcc.dg/pr114907.c: New test. > > --- gcc/expr.cc.jj 2024-04-09 09:29:04.000000000 +0200 > +++ gcc/expr.cc 2024-05-06 13:21:33.933798494 +0200 > @@ -355,8 +355,16 @@ convert_mode_scalar (rtx to, rtx from, i > && REAL_MODE_FORMAT (from_mode) == &ieee_half_format)); > > if (GET_MODE_PRECISION (from_mode) == GET_MODE_PRECISION (to_mode)) > - /* Conversion between decimal float and binary float, same size. */ > - tab = DECIMAL_FLOAT_MODE_P (from_mode) ? trunc_optab : sext_optab; > + { > + if (REAL_MODE_FORMAT (to_mode) == &arm_bfloat_half_format > + && REAL_MODE_FORMAT (from_mode) == &ieee_half_format) > + /* libgcc implements just __trunchfbf2, not __extendhfbf2. */ > + tab = trunc_optab; > + else > + /* Conversion between decimal float and binary float, same > + size. */ > + tab = DECIMAL_FLOAT_MODE_P (from_mode) ? trunc_optab : sext_optab; > + } > else if (GET_MODE_PRECISION (from_mode) < GET_MODE_PRECISION (to_mode)) > tab = sext_optab; > else > --- gcc/optabs-libfuncs.cc.jj 2024-01-03 11:51:31.739728303 +0100 > +++ gcc/optabs-libfuncs.cc 2024-05-06 15:50:21.611027802 +0200 > @@ -589,7 +589,9 @@ gen_trunc_conv_libfunc (convert_optab ta > if (GET_MODE_CLASS (float_tmode) != GET_MODE_CLASS (float_fmode)) > gen_interclass_conv_libfunc (tab, opname, float_tmode, float_fmode); > > - if (GET_MODE_PRECISION (float_fmode) <= GET_MODE_PRECISION (float_tmode)) > + if (GET_MODE_PRECISION (float_fmode) <= GET_MODE_PRECISION (float_tmode) > + && (REAL_MODE_FORMAT (float_tmode) != &arm_bfloat_half_format > + || REAL_MODE_FORMAT (float_fmode) != &ieee_half_format)) > return; > > if (GET_MODE_CLASS (float_tmode) == GET_MODE_CLASS (float_fmode)) > --- gcc/testsuite/gcc.dg/pr114907.c.jj 2024-05-06 15:59:08.734958523 +0200 > +++ gcc/testsuite/gcc.dg/pr114907.c 2024-05-06 16:02:38.914139829 +0200 > @@ -0,0 +1,27 @@ > +/* PR middle-end/114907 */ > +/* { dg-do run } */ > +/* { dg-options "" } */ > +/* { dg-add-options float16 } */ > +/* { dg-require-effective-target float16_runtime } */ > +/* { dg-add-options bfloat16 } */ > +/* { dg-require-effective-target bfloat16_runtime } */ > + > +__attribute__((noipa)) _Float16 > +foo (__bf16 x) > +{ > + return (_Float16) x; > +} > + > +__attribute__((noipa)) __bf16 > +bar (_Float16 x) > +{ > + return (__bf16) x; > +} > + > +int > +main () > +{ > + if (foo (11.125bf16) != 11.125f16 > + || bar (11.125f16) != 11.125bf16) > + __builtin_abort (); > +} > > Jakub >
On Tue, May 07, 2024 at 08:57:00PM +0200, Richard Biener wrote: > > > > Am 07.05.2024 um 18:02 schrieb Jakub Jelinek <jakub@redhat.com>: > > > > Hi! > > > > The HF and BF modes have the same size/precision and neither is > > a subset nor superset of the other. > > So, using either __extendhfbf2 or __trunchfbf2 is weird. > > The expansion apparently emits __extendhfbf2, but on the libgcc side > > we apparently have __trunchfbf2 implemented. > > > > I think it is easier to switch to using what is available rather than > > adding new entrypoints to libgcc, even alias, because this is backportable. > > > > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? > > Ok - do we have any target patterns that need adjustments? I don't think so. BFmode is i386/aarch64/arm/riscv backend only from what I can see, I've done make mddump for all of them and none of the tmp-mddump.md files show any matches for hfbf (nor bfhf). Jakub
--- gcc/expr.cc.jj 2024-04-09 09:29:04.000000000 +0200 +++ gcc/expr.cc 2024-05-06 13:21:33.933798494 +0200 @@ -355,8 +355,16 @@ convert_mode_scalar (rtx to, rtx from, i && REAL_MODE_FORMAT (from_mode) == &ieee_half_format)); if (GET_MODE_PRECISION (from_mode) == GET_MODE_PRECISION (to_mode)) - /* Conversion between decimal float and binary float, same size. */ - tab = DECIMAL_FLOAT_MODE_P (from_mode) ? trunc_optab : sext_optab; + { + if (REAL_MODE_FORMAT (to_mode) == &arm_bfloat_half_format + && REAL_MODE_FORMAT (from_mode) == &ieee_half_format) + /* libgcc implements just __trunchfbf2, not __extendhfbf2. */ + tab = trunc_optab; + else + /* Conversion between decimal float and binary float, same + size. */ + tab = DECIMAL_FLOAT_MODE_P (from_mode) ? trunc_optab : sext_optab; + } else if (GET_MODE_PRECISION (from_mode) < GET_MODE_PRECISION (to_mode)) tab = sext_optab; else --- gcc/optabs-libfuncs.cc.jj 2024-01-03 11:51:31.739728303 +0100 +++ gcc/optabs-libfuncs.cc 2024-05-06 15:50:21.611027802 +0200 @@ -589,7 +589,9 @@ gen_trunc_conv_libfunc (convert_optab ta if (GET_MODE_CLASS (float_tmode) != GET_MODE_CLASS (float_fmode)) gen_interclass_conv_libfunc (tab, opname, float_tmode, float_fmode); - if (GET_MODE_PRECISION (float_fmode) <= GET_MODE_PRECISION (float_tmode)) + if (GET_MODE_PRECISION (float_fmode) <= GET_MODE_PRECISION (float_tmode) + && (REAL_MODE_FORMAT (float_tmode) != &arm_bfloat_half_format + || REAL_MODE_FORMAT (float_fmode) != &ieee_half_format)) return; if (GET_MODE_CLASS (float_tmode) == GET_MODE_CLASS (float_fmode)) --- gcc/testsuite/gcc.dg/pr114907.c.jj 2024-05-06 15:59:08.734958523 +0200 +++ gcc/testsuite/gcc.dg/pr114907.c 2024-05-06 16:02:38.914139829 +0200 @@ -0,0 +1,27 @@ +/* PR middle-end/114907 */ +/* { dg-do run } */ +/* { dg-options "" } */ +/* { dg-add-options float16 } */ +/* { dg-require-effective-target float16_runtime } */ +/* { dg-add-options bfloat16 } */ +/* { dg-require-effective-target bfloat16_runtime } */ + +__attribute__((noipa)) _Float16 +foo (__bf16 x) +{ + return (_Float16) x; +} + +__attribute__((noipa)) __bf16 +bar (_Float16 x) +{ + return (__bf16) x; +} + +int +main () +{ + if (foo (11.125bf16) != 11.125f16 + || bar (11.125f16) != 11.125bf16) + __builtin_abort (); +}