diff options
| author | Richard Henderson <rth@redhat.com> | 2010-10-18 19:12:00 -0700 |
|---|---|---|
| committer | Richard Henderson <rth@gcc.gnu.org> | 2010-10-18 19:12:00 -0700 |
| commit | 89509419968e2be6c7880a97968f07fab6b0e3b2 (patch) | |
| tree | 23aa0fcfd49c69dcded07a79443419e6decfd644 /gcc/ChangeLog | |
| parent | a11930ba8d19cc26848dc06d4a2571642af735b9 (diff) | |
| download | gcc-89509419968e2be6c7880a97968f07fab6b0e3b2.zip gcc-89509419968e2be6c7880a97968f07fab6b0e3b2.tar.gz gcc-89509419968e2be6c7880a97968f07fab6b0e3b2.tar.bz2 | |
Simplify FMA4 patterns with FMA rtx code.
Also fix incorrect rtl generation for scalar instructions.
From-SVN: r165676
Diffstat (limited to 'gcc/ChangeLog')
| -rw-r--r-- | gcc/ChangeLog | 43 |
1 files changed, 43 insertions, 0 deletions
diff --git a/gcc/ChangeLog b/gcc/ChangeLog index f9eea06..8f1595d 100644 --- a/gcc/ChangeLog +++ b/gcc/ChangeLog @@ -1,3 +1,46 @@ +2010-10-18 Richard Henderson <rth@redhat.com> + + * config/i386/i386.c (IX86_BUILTIN_VFMSUBSS, IX86_BUILTIN_VFMSUBSD, + IX86_BUILTIN_VFMSUBPS, IX86_BUILTIN_VFMSUBPD, + IX86_BUILTIN_VFMSUBADDPS, IX86_BUILTIN_VFMSUBADDPD, + IX86_BUILTIN_VFNMADDSS, IX86_BUILTIN_VFNMADDSD, + IX86_BUILTIN_VFNMADDPS, IX86_BUILTIN_VFNMADDPD, + IX86_BUILTIN_VFNMSUBSS, IX86_BUILTIN_VFNMSUBSD, + IX86_BUILTIN_VFNMSUBPS, IX86_BUILTIN_VFNMSUBPD, + IX86_BUILTIN_VFMSUBADDPS256, IX86_BUILTIN_VFMSUBADDPD256, + IX86_BUILTIN_VFNMADDPS256, IX86_BUILTIN_VFNMADDPD256, + IX86_BUILTIN_VFNMSUBPS256, IX86_BUILTIN_VFNMSUBPD256): Remove. + (bdesc_multi_arg): Remove the corresponding builtins. + * config/i386/i386.md (UNSPEC_FMA4_INTRINSIC): Remove. + (UNSPEC_FMA4_FMSUBADD): Remove. + (UNSPEC_FMADDSUB): Rename from UNSPEC_FMA4_FMADDSUB. + * config/i386/sse.md (FMA4MODEF4): Remove. + (FMAMODE): Add. + (fma<mode>4): New expander. + (*fma4i_fmadd_<mode>): Macroize from fma4i_fmadd<mode>4 patterns, + and use FMA rtx code instead of UNSPEC_FMA4_INTRINSIC. + (*fma4i_fmsub_<mode>): Similarly. + (*fma4i_fnmadd_<mode>): Similarly. + (*fma4i_fnmsub_<mode>): Similarly. + (fma4i_vmfmadd_<mode>): Scalar patterns zero-extend, not merge + with the first operand. + (fma4i_fmaddsub_<mode>): Represent with UNSPEC_FMADDSUB instead + of explicit arithmetic. Macroize with AVXMODEF2P. + (*fma4i_fmsubadd_<mode>): Represent with UNSPEC_FMADDSUB + NEG. + (xop_frcz<mode>2): Macroize with FMAMODE. + (xop_vmfrcz<mode>2): Scalar patterns zero-extend, not merge with + the first operand. + * config/i386/fma4intrin.h (_mm_msub_ps): Use vfmadd intrinsic with + extra negations. + (_mm_msub_pd, _mm_msub_ss, _mm_msub_sd): Likewise. + (_mm_nmacc_ps, _mm_nmacc_pd, _mm_nmacc_ss, _mm_nmacc_sd): Likewise. + (_mm_nmsub_ps, _mm_nmsub_pd, _mm_nmsub_ss, _mm_nmsub_sd): Likewise. + (_mm256_msub_ps, _mm256_msub_pd): Likewise. + (_mm256_nmacc_ps, _mm256_nmacc_pd): Likewise. + (_mm256_nmsub_ps, _mm256_nmsub_pd): Likewise. + (_mm_msubadd_ps): Use vfmaddsub intrinsic with extra negation. + (_mm_msubadd_pd, _mm256_msubadd_ps, _mm256_msubadd_pd): Likewise. + 2010-10-18 Bernd Schmidt <bernds@codesourcery.com> PR rtl-optimization/45966 |