| Message ID | ydd7esnqlmi.fsf@CeBiTec.Uni-Bielefeld.DE |
|---|---|
| State | New |
| Headers | show |
| Series | [testsuite,i386] Guard against incomplete AVX512F support in Solaris as | expand |
On Fri, Jan 12, 2018 at 1:58 PM, Rainer Orth <ro@cebitec.uni-bielefeld.de> wrote: > Recent versions of the Solaris/x86 assembler do support AVX512F, but > there are a couple of bugs: it chokes on > > vaddsd {rn-sae}, %xmm2, %xmm0, %xmm0 > vaddsd {rn-sae}, %xmm0, %xmm1, %xmm5{%k1} > vaddsd {rn-sae}, %xmm2, %xmm0, %xmm0{%k1}{z} > > and similar for vaddss, vdivs[ds], vmuls[ds], vsubs[ds]; also on > > vgetmantsd $2, {sae}, %xmm1, %xmm0, %xmm3 > vgetmantss $2, {sae}, %xmm1, %xmm0, %xmm3 > > Since this creates quite a bit of testsuite noise, I extended the > avx512f effective-target check to also test for those. Doing so, I > noticed that the check_effective_target_avx512f proc is duplicated in > gcc.target/i386/i386.exp. For the moment I changed both, but they > should be consolidated in one place to avoid confusion. Please remove the one in i386.exp. > Bootstrapped without regressions on i386-pc-solaris2.11 and > x86_64-pc-linux-gnu. Ok for mainline and the gcc-6 and gcc-7 branches > after a soak time? > > > Rainer > > -- > ----------------------------------------------------------------------------- > Rainer Orth, Center for Biotechnology, Bielefeld University > > > 2017-09-28 Rainer Orth <ro@CeBiTec.Uni-Bielefeld.DE> > > * lib/target-supports.exp (check_effective_target_avx512f): Also > check for __builtin_ia32_addsd_round, > __builtin_ia32_getmantsd_round. > * gcc.target/i386/i386.exp (check_effective_target_avx512f): > Likewise. OK with the above mentioned removal. Thanks, Uros.
# HG changeset patch # Parent d01cdb0ecacbbf61de4cd12700ceb5a4eec05345 Guard against incomplete AVX512F support in Solaris as diff --git a/gcc/testsuite/gcc.target/i386/i386.exp b/gcc/testsuite/gcc.target/i386/i386.exp --- a/gcc/testsuite/gcc.target/i386/i386.exp +++ b/gcc/testsuite/gcc.target/i386/i386.exp @@ -250,6 +250,17 @@ proc check_effective_target_avx512f { } { return __builtin_ia32_pandq512_mask (__X, __Y, __X, -1); } + + typedef double __m128d __attribute__ ((__vector_size__ (16))); + __m128d _mm128_add (__m128d a) + { + return __builtin_ia32_addsd_round (a, a, 8); + } + + __m128d _mm128_getmant (__m128d a) + { + return __builtin_ia32_getmantsd_round (a, a, 0, 8); + } } "-mavx512f" ] } diff --git a/gcc/testsuite/lib/target-supports.exp b/gcc/testsuite/lib/target-supports.exp --- a/gcc/testsuite/lib/target-supports.exp +++ b/gcc/testsuite/lib/target-supports.exp @@ -7590,11 +7590,22 @@ proc check_prefer_avx128 { } { proc check_effective_target_avx512f { } { return [check_no_compiler_messages avx512f object { typedef double __m512d __attribute__ ((__vector_size__ (64))); + typedef double __m128d __attribute__ ((__vector_size__ (16))); __m512d _mm512_add (__m512d a) { return __builtin_ia32_addpd512_mask (a, a, a, 1, 4); } + + __m128d _mm128_add (__m128d a) + { + return __builtin_ia32_addsd_round (a, a, 8); + } + + __m128d _mm128_getmant (__m128d a) + { + return __builtin_ia32_getmantsd_round (a, a, 0, 8); + } } "-O2 -mavx512f" ] }
Recent versions of the Solaris/x86 assembler do support AVX512F, but there are a couple of bugs: it chokes on vaddsd {rn-sae}, %xmm2, %xmm0, %xmm0 vaddsd {rn-sae}, %xmm0, %xmm1, %xmm5{%k1} vaddsd {rn-sae}, %xmm2, %xmm0, %xmm0{%k1}{z} and similar for vaddss, vdivs[ds], vmuls[ds], vsubs[ds]; also on vgetmantsd $2, {sae}, %xmm1, %xmm0, %xmm3 vgetmantss $2, {sae}, %xmm1, %xmm0, %xmm3 Since this creates quite a bit of testsuite noise, I extended the avx512f effective-target check to also test for those. Doing so, I noticed that the check_effective_target_avx512f proc is duplicated in gcc.target/i386/i386.exp. For the moment I changed both, but they should be consolidated in one place to avoid confusion. Bootstrapped without regressions on i386-pc-solaris2.11 and x86_64-pc-linux-gnu. Ok for mainline and the gcc-6 and gcc-7 branches after a soak time? Rainer