This is an archive of the discontinued LLVM Phabricator instance.

Differential D147017

[X86] Add a fneg test for fma with a splat constant vector
ClosedPublic

Authored by e-kud on Mar 27 2023, 7:03 PM.

Details

Reviewers
pengfei
Commits
rG0bf6055e4994: [X86] Add a fneg test for fma with a splat constant vector

Diff Detail

Event Timeline

e-kud created this revision.Mar 27 2023, 7:03 PM
Herald added a project: Restricted Project. · View Herald TranscriptMar 27 2023, 7:03 PM
Herald added a subscriber: pengfei. · View Herald Transcript
e-kud requested review of this revision.Mar 27 2023, 7:03 PM
Herald added a project: Restricted Project. · View Herald TranscriptMar 27 2023, 7:03 PM
Herald added a subscriber: llvm-commits. · View Herald Transcript
e-kud added a child revision: D146494: [X86] Combine constant vector inputs for FMA.Mar 27 2023, 7:05 PM
Harbormaster completed remote builds in B222148: Diff 508867.Mar 27 2023, 7:28 PM
e-kud updated this revision to Diff 509880.Mar 30 2023, 7:30 PM

New tests

Harbormaster completed remote builds in B222900: Diff 509880.Mar 30 2023, 7:47 PM
e-kud updated this revision to Diff 510165.Mar 31 2023, 8:34 PM

Updated tests with undefs.
Use not only splat vector constants.

Harbormaster completed remote builds in B223105: Diff 510165.Mar 31 2023, 9:16 PM
e-kud updated this revision to Diff 511829.Apr 7 2023, 4:56 PM

Rebased

Harbormaster completed remote builds in B224315: Diff 511829.Apr 7 2023, 5:42 PM
e-kud updated this revision to Diff 521696.May 12 2023, 9:54 AM

Rebased

Harbormaster completed remote builds in B231633: Diff 521696.May 12 2023, 10:44 AM
pengfei accepted this revision.May 16 2023, 11:34 PM
This revision is now accepted and ready to land.May 16 2023, 11:34 PM
Closed by commit rG0bf6055e4994: [X86] Add a fneg test for fma with a splat constant vector (authored by e-kud, committed by pengfei). · Explain WhyMay 17 2023, 12:21 AM
This revision was automatically updated to reflect the committed changes.
pengfei added a commit: rG0bf6055e4994: [X86] Add a fneg test for fma with a splat constant vector.

Revision Contents

PathSize
llvm/
test/
CodeGen/
X86/
86 lines
83 lines

Diff 510165

llvm/test/CodeGen/X86/avx2-fma-fneg-combine.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx2,+fma | FileCheck %s --check-prefix=X32 ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx2,+fma | FileCheck %s --check-prefix=X32
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2,+fma | FileCheck %s --check-prefix=X64 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2,+fma | FileCheck %s --check-prefix=X64
declare <8 x float> @llvm.fma.v8f32(<8 x float>, <8 x float>, <8 x float>) declare <8 x float> @llvm.fma.v8f32(<8 x float>, <8 x float>, <8 x float>)
declare <4 x float> @llvm.fma.v4f32(<4 x float>, <4 x float>, <4 x float>) declare <4 x float> @llvm.fma.v4f32(<4 x float>, <4 x float>, <4 x float>)
declare float @llvm.fma.f32(float, float, float) declare float @llvm.fma.f32(float, float, float)
declare <2 x double> @llvm.fma.v2f64(<2 x double>, <2 x double>, <2 x double>) declare <2 x double> @llvm.fma.v2f64(<2 x double>, <2 x double>, <2 x double>)
declare <4 x double> @llvm.fma.v4f64(<4 x double>, <4 x double>, <4 x double>)
; This test checks combinations of FNEG and FMA intrinsics ; This test checks combinations of FNEG and FMA intrinsics
define <8 x float> @test1(<8 x float> %a, <8 x float> %b, <8 x float> %c) { define <8 x float> @test1(<8 x float> %a, <8 x float> %b, <8 x float> %c) {
; X32-LABEL: test1: ; X32-LABEL: test1:
; X32: # %bb.0: ; X32: # %bb.0:
; X32-NEXT: vfmsub213ps {{.*#+}} ymm0 = (ymm1 * ymm0) - ymm2 ; X32-NEXT: vfmsub213ps {{.*#+}} ymm0 = (ymm1 * ymm0) - ymm2
; X32-NEXT: retl ; X32-NEXT: retl
▲ Show 20 Lines • Show All 126 Lines • ▼ Show 20 Lines
; X64-NEXT: vfnmadd213ps {{.*#+}} ymm0 = -(ymm1 * ymm0) + ymm2 ; X64-NEXT: vfnmadd213ps {{.*#+}} ymm0 = -(ymm1 * ymm0) + ymm2
; X64-NEXT: retq ; X64-NEXT: retq
%t0 = fsub float -0.0, %a %t0 = fsub float -0.0, %a
%t1 = insertelement <8 x float> undef, float %t0, i32 0 %t1 = insertelement <8 x float> undef, float %t0, i32 0
%t2 = shufflevector <8 x float> %t1, <8 x float> undef, <8 x i32> zeroinitializer %t2 = shufflevector <8 x float> %t1, <8 x float> undef, <8 x i32> zeroinitializer
%t3 = tail call nsz <8 x float> @llvm.fma.v8f32(<8 x float> %t2, <8 x float> %b, <8 x float> %c) %t3 = tail call nsz <8 x float> @llvm.fma.v8f32(<8 x float> %t2, <8 x float> %b, <8 x float> %c)
ret <8 x float> %t3 ret <8 x float> %t3
} }
define <4 x double> @test9(<4 x double> %a) {
; X32-LABEL: test9:
; X32: # %bb.0:
; X32-NEXT: vbroadcastsd {{.*#+}} ymm1 = [-5.0E-1,-5.0E-1,-5.0E-1,-5.0E-1]
; X32-NEXT: vbroadcastsd {{.*#+}} ymm2 = [5.0E-1,5.0E-1,5.0E-1,5.0E-1]
; X32-NEXT: vfmadd213pd {{.*#+}} ymm0 = (ymm2 * ymm0) + ymm1
; X32-NEXT: retl
;
; X64-LABEL: test9:
; X64: # %bb.0:
; X64-NEXT: vbroadcastsd {{.*#+}} ymm1 = [-5.0E-1,-5.0E-1,-5.0E-1,-5.0E-1]
; X64-NEXT: vbroadcastsd {{.*#+}} ymm2 = [5.0E-1,5.0E-1,5.0E-1,5.0E-1]
; X64-NEXT: vfmadd213pd {{.*#+}} ymm0 = (ymm2 * ymm0) + ymm1
; X64-NEXT: retq
%t = tail call <4 x double> @llvm.fma.v4f64(<4 x double> %a, <4 x double> <double 5.000000e-01, double 5.000000e-01, double 5.000000e-01, double 5.000000e-01>, <4 x double> <double -5.000000e-01, double -5.000000e-01, double -5.000000e-01, double -5.000000e-01>)
ret <4 x double> %t
}
define <4 x double> @test10(<4 x double> %a, <4 x double> %b) {
; X32-LABEL: test10:
; X32: # %bb.0:
; X32-NEXT: vmovapd {{.*#+}} ymm2 = <-9.5E+0,u,-5.5E+0,-2.5E+0>
; X32-NEXT: vfmadd213pd {{.*#+}} ymm2 = (ymm0 * ymm2) + ymm1
; X32-NEXT: vfmadd231pd {{.*#+}} ymm1 = (ymm0 * mem) + ymm1
; X32-NEXT: vaddpd %ymm1, %ymm2, %ymm0
; X32-NEXT: retl
;
; X64-LABEL: test10:
; X64: # %bb.0:
; X64-NEXT: vmovapd {{.*#+}} ymm2 = <-9.5E+0,u,-5.5E+0,-2.5E+0>
; X64-NEXT: vfmadd213pd {{.*#+}} ymm2 = (ymm0 * ymm2) + ymm1
; X64-NEXT: vfmadd231pd {{.*#+}} ymm1 = (ymm0 * mem) + ymm1
; X64-NEXT: vaddpd %ymm1, %ymm2, %ymm0
; X64-NEXT: retq
%t0 = tail call <4 x double> @llvm.fma.v4f64(<4 x double> %a, <4 x double> <double -95.00000e-01, double undef, double -55.00000e-01, double -25.00000e-01>, <4 x double> %b)
%t1 = tail call <4 x double> @llvm.fma.v4f64(<4 x double> %a, <4 x double> <double 95.00000e-01, double undef, double 55.00000e-01, double 25.00000e-01>, <4 x double> %b)
%t2 = fadd <4 x double> %t0, %t1
ret <4 x double> %t2
}
define <4 x double> @test11(<4 x double> %a) {
; X32-LABEL: test11:
; X32: # %bb.0:
; X32-NEXT: vbroadcastf128 {{.*#+}} ymm1 = [5.0E-1,2.5E+0,5.0E-1,2.5E+0]
; X32-NEXT: # ymm1 = mem[0,1,0,1]
; X32-NEXT: vaddpd %ymm1, %ymm0, %ymm0
; X32-NEXT: vfmadd213pd {{.*#+}} ymm0 = (ymm1 * ymm0) + mem
; X32-NEXT: retl
;
; X64-LABEL: test11:
; X64: # %bb.0:
; X64-NEXT: vbroadcastf128 {{.*#+}} ymm1 = [5.0E-1,2.5E+0,5.0E-1,2.5E+0]
; X64-NEXT: # ymm1 = mem[0,1,0,1]
; X64-NEXT: vaddpd %ymm1, %ymm0, %ymm0
; X64-NEXT: vfmadd213pd {{.*#+}} ymm0 = (ymm1 * ymm0) + mem
; X64-NEXT: retq
%t0 = fadd <4 x double> %a, <double 5.000000e-01, double 25.00000e-01, double 5.000000e-01, double 25.00000e-01>
%t1 = tail call <4 x double> @llvm.fma.v4f64(<4 x double> %t0, <4 x double> <double 5.000000e-01, double 25.00000e-01, double 5.000000e-01, double 25.00000e-01>, <4 x double> <double -5.000000e-01, double -25.00000e-01, double -5.000000e-01, double -25.00000e-01>)
ret <4 x double> %t1
}
define <4 x double> @test12(<4 x double> %a, <4 x double> %b) {
; X32-LABEL: test12:
; X32: # %bb.0:
; X32-NEXT: vmovapd {{.*#+}} ymm2 = [7.5E+0,2.5E+0,5.5E+0,9.5E+0]
; X32-NEXT: vfmadd213pd {{.*#+}} ymm2 = (ymm0 * ymm2) + mem
; X32-NEXT: vmovapd {{.*#+}} ymm0 = <u,2.5E+0,5.5E+0,9.5E+0>
; X32-NEXT: vfmadd213pd {{.*#+}} ymm0 = (ymm1 * ymm0) + mem
; X32-NEXT: vaddpd %ymm0, %ymm2, %ymm0
; X32-NEXT: retl
;
; X64-LABEL: test12:
; X64: # %bb.0:
; X64-NEXT: vmovapd {{.*#+}} ymm2 = [7.5E+0,2.5E+0,5.5E+0,9.5E+0]
; X64-NEXT: vfmadd213pd {{.*#+}} ymm2 = (ymm0 * ymm2) + mem
; X64-NEXT: vmovapd {{.*#+}} ymm0 = <u,2.5E+0,5.5E+0,9.5E+0>
; X64-NEXT: vfmadd213pd {{.*#+}} ymm0 = (ymm1 * ymm0) + mem
; X64-NEXT: vaddpd %ymm0, %ymm2, %ymm0
; X64-NEXT: retq
%t0 = tail call <4 x double> @llvm.fma.v4f64(<4 x double> %a, <4 x double> <double 75.00000e-01, double 25.00000e-01, double 55.00000e-01, double 95.00000e-01>, <4 x double> <double -75.00000e-01, double undef, double -55.00000e-01, double -95.00000e-01>)
%t1 = tail call <4 x double> @llvm.fma.v4f64(<4 x double> %b, <4 x double> <double undef, double 25.00000e-01, double 55.00000e-01, double 95.00000e-01>, <4 x double> <double -75.00000e-01, double -25.00000e-01, double -55.00000e-01, double -95.00000e-01>)
%t2 = fadd <4 x double> %t0, %t1
ret <4 x double> %t2
}

llvm/test/CodeGen/X86/fma-fneg-combine-2.ll

Show First 20 Lines • Show All 120 Lines • ▼ Show 20 Lines
; FMA4-NEXT: vfnmsubss {{.*#+}} xmm0 = -(xmm0 * mem) - xmm1 ; FMA4-NEXT: vfnmsubss {{.*#+}} xmm0 = -(xmm0 * mem) - xmm1
; FMA4-NEXT: retq ; FMA4-NEXT: retq
%m = fmul float %x, 42.0 %m = fmul float %x, 42.0
%fma = call nsz float @llvm.fma.f32(float %x, float -42.0, float %m) %fma = call nsz float @llvm.fma.f32(float %x, float -42.0, float %m)
%nfma = fneg float %fma %nfma = fneg float %fma
ret float %nfma ret float %nfma
} }
define <4 x double> @negated_constant_v4f64(<4 x double> %a) {
; FMA3-LABEL: negated_constant_v4f64:
; FMA3: # %bb.0:
; FMA3-NEXT: vmovapd {{.*#+}} ymm1 = [5.0E-1,2.5E-1,1.25E-1,6.25E-2]
; FMA3-NEXT: vfmadd213pd {{.*#+}} ymm0 = (ymm1 * ymm0) + mem
; FMA3-NEXT: retq
;
; FMA4-LABEL: negated_constant_v4f64:
; FMA4: # %bb.0:
; FMA4-NEXT: vmovapd {{.*#+}} ymm1 = [5.0E-1,2.5E-1,1.25E-1,6.25E-2]
; FMA4-NEXT: vfmaddpd {{.*#+}} ymm0 = (ymm0 * ymm1) + mem
; FMA4-NEXT: retq
%t = tail call <4 x double> @llvm.fma.v4f64(<4 x double> %a, <4 x double> <double 5.000000e-01, double 2.5000000e-01, double 1.25000000e-01, double 0.62500000e-01>, <4 x double> <double -5.000000e-01, double -2.5000000e-01, double -1.25000000e-01, double -0.62500000e-01>)
ret <4 x double> %t
}
define <4 x double> @negated_constant_v4f64_2fmas(<4 x double> %a, <4 x double> %b) {
; FMA3-LABEL: negated_constant_v4f64_2fmas:
; FMA3: # %bb.0:
; FMA3-NEXT: vmovapd {{.*#+}} ymm2 = <-5.0E-1,u,-2.5E+0,-4.5E+0>
; FMA3-NEXT: vfmadd213pd {{.*#+}} ymm2 = (ymm0 * ymm2) + ymm1
; FMA3-NEXT: vfmadd231pd {{.*#+}} ymm1 = (ymm0 * mem) + ymm1
; FMA3-NEXT: vaddpd %ymm1, %ymm2, %ymm0
; FMA3-NEXT: retq
;
; FMA4-LABEL: negated_constant_v4f64_2fmas:
; FMA4: # %bb.0:
; FMA4-NEXT: vfmaddpd {{.*#+}} ymm2 = (ymm0 * mem) + ymm1
; FMA4-NEXT: vfmaddpd {{.*#+}} ymm0 = (ymm0 * mem) + ymm1
; FMA4-NEXT: vaddpd %ymm0, %ymm2, %ymm0
; FMA4-NEXT: retq
%t0 = tail call <4 x double> @llvm.fma.v4f64(<4 x double> %a, <4 x double> <double -5.000000e-01, double undef, double -25.000000e-01, double -45.000000e-01>, <4 x double> %b)
%t1 = tail call <4 x double> @llvm.fma.v4f64(<4 x double> %a, <4 x double> <double 5.000000e-01, double undef, double 25.000000e-01, double 45.000000e-01>, <4 x double> %b)
%t2 = fadd <4 x double> %t0, %t1
ret <4 x double> %t2
}
define <4 x double> @negated_constant_v4f64_fadd(<4 x double> %a) {
; FMA3-LABEL: negated_constant_v4f64_fadd:
; FMA3: # %bb.0:
; FMA3-NEXT: vbroadcastf128 {{.*#+}} ymm1 = [1.5E+0,1.25E-1,1.5E+0,1.25E-1]
; FMA3-NEXT: # ymm1 = mem[0,1,0,1]
; FMA3-NEXT: vaddpd %ymm1, %ymm0, %ymm0
; FMA3-NEXT: vfmadd213pd {{.*#+}} ymm0 = (ymm1 * ymm0) + mem
; FMA3-NEXT: retq
;
; FMA4-LABEL: negated_constant_v4f64_fadd:
; FMA4: # %bb.0:
; FMA4-NEXT: vbroadcastf128 {{.*#+}} ymm1 = [1.5E+0,1.25E-1,1.5E+0,1.25E-1]
; FMA4-NEXT: # ymm1 = mem[0,1,0,1]
; FMA4-NEXT: vaddpd %ymm1, %ymm0, %ymm0
; FMA4-NEXT: vfmaddpd {{.*#+}} ymm0 = (ymm0 * ymm1) + mem
; FMA4-NEXT: retq
%t0 = fadd <4 x double> %a, <double 15.000000e-01, double 1.25000000e-01, double 15.000000e-01, double 1.25000000e-01>
%t1 = tail call <4 x double> @llvm.fma.v4f64(<4 x double> %t0, <4 x double> <double 15.000000e-01, double 1.25000000e-01, double 15.000000e-01, double 1.25000000e-01>, <4 x double> <double -15.000000e-01, double -1.25000000e-01, double -15.000000e-01, double -1.25000000e-01>)
ret <4 x double> %t1
}
define <4 x double> @negated_constant_v4f64_2fma_undefs(<4 x double> %a, <4 x double> %b) {
; FMA3-LABEL: negated_constant_v4f64_2fma_undefs:
; FMA3: # %bb.0:
; FMA3-NEXT: vmovapd {{.*#+}} ymm2 = [5.0E-1,5.0E-1,5.0E-1,5.0E-1]
; FMA3-NEXT: vfmadd213pd {{.*#+}} ymm2 = (ymm0 * ymm2) + mem
; FMA3-NEXT: vmovapd {{.*#+}} ymm0 = <u,5.0E-1,5.0E-1,5.0E-1>
; FMA3-NEXT: vfmadd213pd {{.*#+}} ymm0 = (ymm1 * ymm0) + mem
; FMA3-NEXT: vaddpd %ymm0, %ymm2, %ymm0
; FMA3-NEXT: retq
;
; FMA4-LABEL: negated_constant_v4f64_2fma_undefs:
; FMA4: # %bb.0:
; FMA4-NEXT: vmovapd {{.*#+}} ymm2 = [5.0E-1,5.0E-1,5.0E-1,5.0E-1]
; FMA4-NEXT: vfmaddpd {{.*#+}} ymm0 = (ymm0 * ymm2) + mem
; FMA4-NEXT: vmovapd {{.*#+}} ymm2 = <u,5.0E-1,5.0E-1,5.0E-1>
; FMA4-NEXT: vfmaddpd {{.*#+}} ymm1 = (ymm1 * ymm2) + mem
; FMA4-NEXT: vaddpd %ymm1, %ymm0, %ymm0
; FMA4-NEXT: retq
%t0 = tail call <4 x double> @llvm.fma.v4f64(<4 x double> %a, <4 x double> <double 5.000000e-01, double 5.000000e-01, double 5.000000e-01, double 5.000000e-01>, <4 x double> <double -5.000000e-01, double undef, double -5.000000e-01, double -5.000000e-01>)
%t1 = tail call <4 x double> @llvm.fma.v4f64(<4 x double> %b, <4 x double> <double undef, double 5.000000e-01, double 5.000000e-01, double 5.000000e-01>, <4 x double> <double -5.000000e-01, double -5.000000e-01, double -5.000000e-01, double -5.000000e-01>)
%t2 = fadd <4 x double> %t0, %t1
ret <4 x double> %t2
}
declare float @llvm.fma.f32(float, float, float) declare float @llvm.fma.f32(float, float, float)
declare <4 x float> @llvm.fma.v4f32(<4 x float>, <4 x float>, <4 x float>) declare <4 x float> @llvm.fma.v4f32(<4 x float>, <4 x float>, <4 x float>)
declare <4 x double> @llvm.fma.v4f64(<4 x double>, <4 x double>, <4 x double>)