Avx512f #933
Avx512f #933
Conversation
merge from base
merge base
… cvtss_i32, cvt_roundss_u32, cvt_roundss_u64, cvtss_i64, cvtss_u32, cvtss_u64
… cvtsd_i32, cvt_roundsd_u32, cvt_roundsd_u64, cvtsd_i64, cvtsd_u32, cvtsd_u64
… cvt_roundsi64_sd, cvt_roundi64_sd, cvt_roundu32_ss, cvt_roundu64_ss, cvt_roundu64_sd
…i64, cvttsd_i32, cvtt_roundsd_u32, cvtt_roundsd_u64, cvttsd_i64, cvttsd_u32, cvttsd_u64, cvtt_roundss_si32, cvtt_roundss_i32, cvtt_roundss_si64, cvtt_roundss_i64, cvttss_i32, cvtt_roundss_u32, cvtt_roundss_u64, cvttss_i64, cvttss_u32, cvttss_u64;
|
r? @Amanieu (rust_highfive has picked a reviewer for you, use r? to override) |
|
It seems "mm_cvt_roundss_si64()" ..., i64,si64 cause i686-unknown-linux-gnu test error. |
| /// [Intel's documentation](https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=mm_mask_fmsub_ss&expand=2668) | ||
| #[inline] | ||
| #[target_feature(enable = "avx512f")] | ||
| #[cfg_attr(test, assert_instr(vfmadd213ss))] //should be vfmsub213ss |
There was a problem hiding this comment.
This seems wrong: it should be generating the vfmsub213ss instruction. Maybe you need to invoke the subtract intrinsic directly instead of negating extractc.
There was a problem hiding this comment.
I follow LLVM code, it uses llvm.fma.f32. I think vfmadd213ss and vfmsub213ss have the same latency, and vfmaddss comes first when generating.
__m128 test_mm_mask3_fmsub_ss(__m128 __W, __m128 __X, __m128 __Y, __mmask8 __U){
// CHECK-LABEL: @test_mm_mask3_fmsub_ss
// CHECK: [[NEG:%.+]] = fneg <4 x float> [[ORIGC:%.+]]
// CHECK: [[A:%.+]] = extractelement <4 x float> %{{.}}, i64 0
// CHECK-NEXT: [[B:%.+]] = extractelement <4 x float> %{{.}}, i64 0
// CHECK-NEXT: [[C:%.+]] = extractelement <4 x float> [[NEG]], i64 0
// CHECK-NEXT: [[FMA:%.+]] = call float @llvm.fma.f32(float [[A]], float [[B]], float [[C]])
// CHECK-NEXT: [[C2:%.+]] = extractelement <4 x float> [[ORIGC]], i64 0
// CHECK-NEXT: bitcast i8 %{{.}} to <8 x i1>
// CHECK-NEXT: extractelement <8 x i1> %{{.}}, i64 0
// CHECK-NEXT: [[SEL:%.+]] = select i1 %{{.*}}, float [[FMA]], float [[C2]]
// CHECK-NEXT: insertelement <4 x float> [[ORIGC]], float [[SEL]], i64 0
return _mm_mask3_fmsub_ss(__W, __X, __Y, __U);
}
There was a problem hiding this comment.
I think the problem is that you are using 0. - extractc instead of -extractc. The former is a subtraction while the latter is a negation. They have different behavior in edge cases like negative zero.
mask_fmsub: ss,sd; fmsub_round: ss,sd;
fnmadd: ss,sd; fnmsub: ss,sd; fnmadd_round: ss,sd; fnmsub_round: ss,sd;
fixupimm: ss,sd; fixupimm_round: ss,sd;
cvt_roundss_sd, cvtss_sd, cvtroundsd_ss, cvtsd_ss;
cvt_roundss_si32, cvt_roundss_i32, cvt_roundss_si64, cvt_roundss_i64, cvtss_i32, cvt_roundss_u32, cvt_roundss_u64, cvtss_i64, cvtss_u32, cvtss_u64;
cvt_roundsd_si32, cvt_roundsd_i32, cvt_roundsd_si64, cvt_roundsd_i64, cvtsd_i32, cvt_roundsd_u32, cvt_roundsd_u64, cvtsd_i64, cvtsd_u32, cvtsd_u64;
cvt_roundsi32_ss, cvt_roundi32_ss, cvt_roundsi64_ss, cvt_roundi64_ss, cvt_roundsi64_sd, cvt_roundi64_sd, cvt_roundu32_ss, cvt_roundu64_ss, cvt_roundu64_sd;
cvti32_ss; cvti32_sd; cvti64_ss; cvti64_sd;
cvtt_roundsd_si32, cvtt_roundsd_i32, cvtt_roundsd_si64, cvtt_roundsd_i64, cvttsd_i32, cvtt_roundsd_u32, cvtt_roundsd_u64, cvttsd_i64, cvttsd_u32, cvttsd_u64, cvtt_roundss_si32, cvtt_roundss_i32, cvtt_roundss_si64, cvtt_roundss_i64, cvttss_i32, cvtt_roundss_u32, cvtt_roundss_u64, cvttss_i64, cvttss_u32, cvttss_u64;
cvtu32_ss, cvtu32_sd, cvtu64_ss, cvtu64_sd;
mm_comi_ss; mm_comi_sd;