[llvm] 3996867 - [X86] haddsub-2.ll - remove duplicate scalar tests (#207183)
via llvm-commits
llvm-commits at lists.llvm.org
Thu Jul 2 07:04:00 PDT 2026
Author: Simon Pilgrim
Date: 2026-07-02T14:03:55Z
New Revision: 3996867e69047fc5ec8aa95bfe06eb153696bdd2
URL: https://github.com/llvm/llvm-project/commit/3996867e69047fc5ec8aa95bfe06eb153696bdd2
DIFF: https://github.com/llvm/llvm-project/commit/3996867e69047fc5ec8aa95bfe06eb153696bdd2.diff
LOG: [X86] haddsub-2.ll - remove duplicate scalar tests (#207183)
These vectorisation tests exist in PhaseOrdering/X86/hadd.ll and
PhaseOrdering/X86/hsub.ll and the result shuffle tests exist in
CodeGen/X86/haddsub.ll and CodeGen/X86/phaddsub.ll
Help towards #143000
Added:
Modified:
llvm/test/CodeGen/X86/haddsub-2.ll
Removed:
################################################################################
diff --git a/llvm/test/CodeGen/X86/haddsub-2.ll b/llvm/test/CodeGen/X86/haddsub-2.ll
index bca446fa8fb56..c060b6b708302 100644
--- a/llvm/test/CodeGen/X86/haddsub-2.ll
+++ b/llvm/test/CodeGen/X86/haddsub-2.ll
@@ -1,816 +1,8 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+sse2,+sse3 | FileCheck %s --check-prefix=SSE --check-prefix=SSE3
-; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+sse2,+sse3,+ssse3 | FileCheck %s --check-prefix=SSE --check-prefix=SSSE3
-; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx | FileCheck %s --check-prefix=AVX --check-prefix=AVX1
-; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx2 | FileCheck %s --check-prefix=AVX --check-prefix=AVX2
-
-define <4 x float> @hadd_ps_test1(<4 x float> %A, <4 x float> %B) {
-; SSE-LABEL: hadd_ps_test1:
-; SSE: # %bb.0:
-; SSE-NEXT: haddps %xmm1, %xmm0
-; SSE-NEXT: retq
-;
-; AVX-LABEL: hadd_ps_test1:
-; AVX: # %bb.0:
-; AVX-NEXT: vhaddps %xmm1, %xmm0, %xmm0
-; AVX-NEXT: retq
- %vecext = extractelement <4 x float> %A, i32 0
- %vecext1 = extractelement <4 x float> %A, i32 1
- %add = fadd float %vecext, %vecext1
- %vecinit = insertelement <4 x float> undef, float %add, i32 0
- %vecext2 = extractelement <4 x float> %A, i32 2
- %vecext3 = extractelement <4 x float> %A, i32 3
- %add4 = fadd float %vecext2, %vecext3
- %vecinit5 = insertelement <4 x float> %vecinit, float %add4, i32 1
- %vecext6 = extractelement <4 x float> %B, i32 0
- %vecext7 = extractelement <4 x float> %B, i32 1
- %add8 = fadd float %vecext6, %vecext7
- %vecinit9 = insertelement <4 x float> %vecinit5, float %add8, i32 2
- %vecext10 = extractelement <4 x float> %B, i32 2
- %vecext11 = extractelement <4 x float> %B, i32 3
- %add12 = fadd float %vecext10, %vecext11
- %vecinit13 = insertelement <4 x float> %vecinit9, float %add12, i32 3
- ret <4 x float> %vecinit13
-}
-
-define <4 x float> @hadd_ps_test2(<4 x float> %A, <4 x float> %B) {
-; SSE-LABEL: hadd_ps_test2:
-; SSE: # %bb.0:
-; SSE-NEXT: haddps %xmm1, %xmm0
-; SSE-NEXT: retq
-;
-; AVX-LABEL: hadd_ps_test2:
-; AVX: # %bb.0:
-; AVX-NEXT: vhaddps %xmm1, %xmm0, %xmm0
-; AVX-NEXT: retq
- %vecext = extractelement <4 x float> %A, i32 2
- %vecext1 = extractelement <4 x float> %A, i32 3
- %add = fadd float %vecext, %vecext1
- %vecinit = insertelement <4 x float> undef, float %add, i32 1
- %vecext2 = extractelement <4 x float> %A, i32 0
- %vecext3 = extractelement <4 x float> %A, i32 1
- %add4 = fadd float %vecext2, %vecext3
- %vecinit5 = insertelement <4 x float> %vecinit, float %add4, i32 0
- %vecext6 = extractelement <4 x float> %B, i32 2
- %vecext7 = extractelement <4 x float> %B, i32 3
- %add8 = fadd float %vecext6, %vecext7
- %vecinit9 = insertelement <4 x float> %vecinit5, float %add8, i32 3
- %vecext10 = extractelement <4 x float> %B, i32 0
- %vecext11 = extractelement <4 x float> %B, i32 1
- %add12 = fadd float %vecext10, %vecext11
- %vecinit13 = insertelement <4 x float> %vecinit9, float %add12, i32 2
- ret <4 x float> %vecinit13
-}
-
-define <4 x float> @hsub_ps_test1(<4 x float> %A, <4 x float> %B) {
-; SSE-LABEL: hsub_ps_test1:
-; SSE: # %bb.0:
-; SSE-NEXT: hsubps %xmm1, %xmm0
-; SSE-NEXT: retq
-;
-; AVX-LABEL: hsub_ps_test1:
-; AVX: # %bb.0:
-; AVX-NEXT: vhsubps %xmm1, %xmm0, %xmm0
-; AVX-NEXT: retq
- %vecext = extractelement <4 x float> %A, i32 0
- %vecext1 = extractelement <4 x float> %A, i32 1
- %sub = fsub float %vecext, %vecext1
- %vecinit = insertelement <4 x float> undef, float %sub, i32 0
- %vecext2 = extractelement <4 x float> %A, i32 2
- %vecext3 = extractelement <4 x float> %A, i32 3
- %sub4 = fsub float %vecext2, %vecext3
- %vecinit5 = insertelement <4 x float> %vecinit, float %sub4, i32 1
- %vecext6 = extractelement <4 x float> %B, i32 0
- %vecext7 = extractelement <4 x float> %B, i32 1
- %sub8 = fsub float %vecext6, %vecext7
- %vecinit9 = insertelement <4 x float> %vecinit5, float %sub8, i32 2
- %vecext10 = extractelement <4 x float> %B, i32 2
- %vecext11 = extractelement <4 x float> %B, i32 3
- %sub12 = fsub float %vecext10, %vecext11
- %vecinit13 = insertelement <4 x float> %vecinit9, float %sub12, i32 3
- ret <4 x float> %vecinit13
-}
-
-define <4 x float> @hsub_ps_test2(<4 x float> %A, <4 x float> %B) {
-; SSE-LABEL: hsub_ps_test2:
-; SSE: # %bb.0:
-; SSE-NEXT: hsubps %xmm1, %xmm0
-; SSE-NEXT: retq
-;
-; AVX-LABEL: hsub_ps_test2:
-; AVX: # %bb.0:
-; AVX-NEXT: vhsubps %xmm1, %xmm0, %xmm0
-; AVX-NEXT: retq
- %vecext = extractelement <4 x float> %A, i32 2
- %vecext1 = extractelement <4 x float> %A, i32 3
- %sub = fsub float %vecext, %vecext1
- %vecinit = insertelement <4 x float> undef, float %sub, i32 1
- %vecext2 = extractelement <4 x float> %A, i32 0
- %vecext3 = extractelement <4 x float> %A, i32 1
- %sub4 = fsub float %vecext2, %vecext3
- %vecinit5 = insertelement <4 x float> %vecinit, float %sub4, i32 0
- %vecext6 = extractelement <4 x float> %B, i32 2
- %vecext7 = extractelement <4 x float> %B, i32 3
- %sub8 = fsub float %vecext6, %vecext7
- %vecinit9 = insertelement <4 x float> %vecinit5, float %sub8, i32 3
- %vecext10 = extractelement <4 x float> %B, i32 0
- %vecext11 = extractelement <4 x float> %B, i32 1
- %sub12 = fsub float %vecext10, %vecext11
- %vecinit13 = insertelement <4 x float> %vecinit9, float %sub12, i32 2
- ret <4 x float> %vecinit13
-}
-
-define <4 x i32> @phadd_d_test1(<4 x i32> %A, <4 x i32> %B) {
-; SSE3-LABEL: phadd_d_test1:
-; SSE3: # %bb.0:
-; SSE3-NEXT: movd %xmm0, %eax
-; SSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,1,1]
-; SSE3-NEXT: movd %xmm2, %ecx
-; SSE3-NEXT: addl %eax, %ecx
-; SSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]
-; SSE3-NEXT: movd %xmm2, %eax
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
-; SSE3-NEXT: movd %xmm0, %edx
-; SSE3-NEXT: addl %eax, %edx
-; SSE3-NEXT: movd %xmm1, %eax
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE3-NEXT: movd %xmm0, %esi
-; SSE3-NEXT: addl %eax, %esi
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE3-NEXT: movd %xmm0, %eax
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[3,3,3,3]
-; SSE3-NEXT: movd %xmm0, %edi
-; SSE3-NEXT: addl %eax, %edi
-; SSE3-NEXT: movd %edi, %xmm0
-; SSE3-NEXT: movd %esi, %xmm1
-; SSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE3-NEXT: movd %edx, %xmm2
-; SSE3-NEXT: movd %ecx, %xmm0
-; SSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
-; SSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE3-NEXT: retq
-;
-; SSSE3-LABEL: phadd_d_test1:
-; SSSE3: # %bb.0:
-; SSSE3-NEXT: phaddd %xmm1, %xmm0
-; SSSE3-NEXT: retq
-;
-; AVX-LABEL: phadd_d_test1:
-; AVX: # %bb.0:
-; AVX-NEXT: vphaddd %xmm1, %xmm0, %xmm0
-; AVX-NEXT: retq
- %vecext = extractelement <4 x i32> %A, i32 0
- %vecext1 = extractelement <4 x i32> %A, i32 1
- %add = add i32 %vecext, %vecext1
- %vecinit = insertelement <4 x i32> undef, i32 %add, i32 0
- %vecext2 = extractelement <4 x i32> %A, i32 2
- %vecext3 = extractelement <4 x i32> %A, i32 3
- %add4 = add i32 %vecext2, %vecext3
- %vecinit5 = insertelement <4 x i32> %vecinit, i32 %add4, i32 1
- %vecext6 = extractelement <4 x i32> %B, i32 0
- %vecext7 = extractelement <4 x i32> %B, i32 1
- %add8 = add i32 %vecext6, %vecext7
- %vecinit9 = insertelement <4 x i32> %vecinit5, i32 %add8, i32 2
- %vecext10 = extractelement <4 x i32> %B, i32 2
- %vecext11 = extractelement <4 x i32> %B, i32 3
- %add12 = add i32 %vecext10, %vecext11
- %vecinit13 = insertelement <4 x i32> %vecinit9, i32 %add12, i32 3
- ret <4 x i32> %vecinit13
-}
-
-define <4 x i32> @phadd_d_test2(<4 x i32> %A, <4 x i32> %B) {
-; SSE3-LABEL: phadd_d_test2:
-; SSE3: # %bb.0:
-; SSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]
-; SSE3-NEXT: movd %xmm2, %eax
-; SSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm0[3,3,3,3]
-; SSE3-NEXT: movd %xmm2, %ecx
-; SSE3-NEXT: addl %eax, %ecx
-; SSE3-NEXT: movd %xmm0, %eax
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
-; SSE3-NEXT: movd %xmm0, %edx
-; SSE3-NEXT: addl %eax, %edx
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[3,3,3,3]
-; SSE3-NEXT: movd %xmm0, %eax
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE3-NEXT: movd %xmm0, %esi
-; SSE3-NEXT: addl %eax, %esi
-; SSE3-NEXT: movd %esi, %xmm0
-; SSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm1[1,1,1,1]
-; SSE3-NEXT: movd %xmm2, %eax
-; SSE3-NEXT: movd %xmm1, %esi
-; SSE3-NEXT: addl %eax, %esi
-; SSE3-NEXT: movd %esi, %xmm1
-; SSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE3-NEXT: movd %ecx, %xmm2
-; SSE3-NEXT: movd %edx, %xmm0
-; SSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
-; SSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE3-NEXT: retq
-;
-; SSSE3-LABEL: phadd_d_test2:
-; SSSE3: # %bb.0:
-; SSSE3-NEXT: phaddd %xmm1, %xmm0
-; SSSE3-NEXT: retq
-;
-; AVX-LABEL: phadd_d_test2:
-; AVX: # %bb.0:
-; AVX-NEXT: vphaddd %xmm1, %xmm0, %xmm0
-; AVX-NEXT: retq
- %vecext = extractelement <4 x i32> %A, i32 2
- %vecext1 = extractelement <4 x i32> %A, i32 3
- %add = add i32 %vecext, %vecext1
- %vecinit = insertelement <4 x i32> undef, i32 %add, i32 1
- %vecext2 = extractelement <4 x i32> %A, i32 0
- %vecext3 = extractelement <4 x i32> %A, i32 1
- %add4 = add i32 %vecext2, %vecext3
- %vecinit5 = insertelement <4 x i32> %vecinit, i32 %add4, i32 0
- %vecext6 = extractelement <4 x i32> %B, i32 3
- %vecext7 = extractelement <4 x i32> %B, i32 2
- %add8 = add i32 %vecext6, %vecext7
- %vecinit9 = insertelement <4 x i32> %vecinit5, i32 %add8, i32 3
- %vecext10 = extractelement <4 x i32> %B, i32 1
- %vecext11 = extractelement <4 x i32> %B, i32 0
- %add12 = add i32 %vecext10, %vecext11
- %vecinit13 = insertelement <4 x i32> %vecinit9, i32 %add12, i32 2
- ret <4 x i32> %vecinit13
-}
-
-define <4 x i32> @phsub_d_test1(<4 x i32> %A, <4 x i32> %B) {
-; SSE3-LABEL: phsub_d_test1:
-; SSE3: # %bb.0:
-; SSE3-NEXT: movd %xmm0, %eax
-; SSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,1,1]
-; SSE3-NEXT: movd %xmm2, %ecx
-; SSE3-NEXT: subl %ecx, %eax
-; SSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]
-; SSE3-NEXT: movd %xmm2, %ecx
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
-; SSE3-NEXT: movd %xmm0, %edx
-; SSE3-NEXT: subl %edx, %ecx
-; SSE3-NEXT: movd %xmm1, %edx
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE3-NEXT: movd %xmm0, %esi
-; SSE3-NEXT: subl %esi, %edx
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE3-NEXT: movd %xmm0, %esi
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[3,3,3,3]
-; SSE3-NEXT: movd %xmm0, %edi
-; SSE3-NEXT: subl %edi, %esi
-; SSE3-NEXT: movd %esi, %xmm0
-; SSE3-NEXT: movd %edx, %xmm1
-; SSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE3-NEXT: movd %ecx, %xmm2
-; SSE3-NEXT: movd %eax, %xmm0
-; SSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
-; SSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE3-NEXT: retq
-;
-; SSSE3-LABEL: phsub_d_test1:
-; SSSE3: # %bb.0:
-; SSSE3-NEXT: phsubd %xmm1, %xmm0
-; SSSE3-NEXT: retq
-;
-; AVX-LABEL: phsub_d_test1:
-; AVX: # %bb.0:
-; AVX-NEXT: vphsubd %xmm1, %xmm0, %xmm0
-; AVX-NEXT: retq
- %vecext = extractelement <4 x i32> %A, i32 0
- %vecext1 = extractelement <4 x i32> %A, i32 1
- %sub = sub i32 %vecext, %vecext1
- %vecinit = insertelement <4 x i32> undef, i32 %sub, i32 0
- %vecext2 = extractelement <4 x i32> %A, i32 2
- %vecext3 = extractelement <4 x i32> %A, i32 3
- %sub4 = sub i32 %vecext2, %vecext3
- %vecinit5 = insertelement <4 x i32> %vecinit, i32 %sub4, i32 1
- %vecext6 = extractelement <4 x i32> %B, i32 0
- %vecext7 = extractelement <4 x i32> %B, i32 1
- %sub8 = sub i32 %vecext6, %vecext7
- %vecinit9 = insertelement <4 x i32> %vecinit5, i32 %sub8, i32 2
- %vecext10 = extractelement <4 x i32> %B, i32 2
- %vecext11 = extractelement <4 x i32> %B, i32 3
- %sub12 = sub i32 %vecext10, %vecext11
- %vecinit13 = insertelement <4 x i32> %vecinit9, i32 %sub12, i32 3
- ret <4 x i32> %vecinit13
-}
-
-define <4 x i32> @phsub_d_test2(<4 x i32> %A, <4 x i32> %B) {
-; SSE3-LABEL: phsub_d_test2:
-; SSE3: # %bb.0:
-; SSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]
-; SSE3-NEXT: movd %xmm2, %eax
-; SSE3-NEXT: pshufd {{.*#+}} xmm2 = xmm0[3,3,3,3]
-; SSE3-NEXT: movd %xmm2, %ecx
-; SSE3-NEXT: subl %ecx, %eax
-; SSE3-NEXT: movd %xmm0, %ecx
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
-; SSE3-NEXT: movd %xmm0, %edx
-; SSE3-NEXT: subl %edx, %ecx
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE3-NEXT: movd %xmm0, %edx
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[3,3,3,3]
-; SSE3-NEXT: movd %xmm0, %esi
-; SSE3-NEXT: subl %esi, %edx
-; SSE3-NEXT: movd %edx, %xmm0
-; SSE3-NEXT: movd %xmm1, %edx
-; SSE3-NEXT: pshufd {{.*#+}} xmm1 = xmm1[1,1,1,1]
-; SSE3-NEXT: movd %xmm1, %esi
-; SSE3-NEXT: subl %esi, %edx
-; SSE3-NEXT: movd %edx, %xmm1
-; SSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE3-NEXT: movd %eax, %xmm2
-; SSE3-NEXT: movd %ecx, %xmm0
-; SSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
-; SSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE3-NEXT: retq
-;
-; SSSE3-LABEL: phsub_d_test2:
-; SSSE3: # %bb.0:
-; SSSE3-NEXT: phsubd %xmm1, %xmm0
-; SSSE3-NEXT: retq
-;
-; AVX-LABEL: phsub_d_test2:
-; AVX: # %bb.0:
-; AVX-NEXT: vphsubd %xmm1, %xmm0, %xmm0
-; AVX-NEXT: retq
- %vecext = extractelement <4 x i32> %A, i32 2
- %vecext1 = extractelement <4 x i32> %A, i32 3
- %sub = sub i32 %vecext, %vecext1
- %vecinit = insertelement <4 x i32> undef, i32 %sub, i32 1
- %vecext2 = extractelement <4 x i32> %A, i32 0
- %vecext3 = extractelement <4 x i32> %A, i32 1
- %sub4 = sub i32 %vecext2, %vecext3
- %vecinit5 = insertelement <4 x i32> %vecinit, i32 %sub4, i32 0
- %vecext6 = extractelement <4 x i32> %B, i32 2
- %vecext7 = extractelement <4 x i32> %B, i32 3
- %sub8 = sub i32 %vecext6, %vecext7
- %vecinit9 = insertelement <4 x i32> %vecinit5, i32 %sub8, i32 3
- %vecext10 = extractelement <4 x i32> %B, i32 0
- %vecext11 = extractelement <4 x i32> %B, i32 1
- %sub12 = sub i32 %vecext10, %vecext11
- %vecinit13 = insertelement <4 x i32> %vecinit9, i32 %sub12, i32 2
- ret <4 x i32> %vecinit13
-}
-
-define <2 x double> @hadd_pd_test1(<2 x double> %A, <2 x double> %B) {
-; SSE-LABEL: hadd_pd_test1:
-; SSE: # %bb.0:
-; SSE-NEXT: haddpd %xmm1, %xmm0
-; SSE-NEXT: retq
-;
-; AVX-LABEL: hadd_pd_test1:
-; AVX: # %bb.0:
-; AVX-NEXT: vhaddpd %xmm1, %xmm0, %xmm0
-; AVX-NEXT: retq
- %vecext = extractelement <2 x double> %A, i32 0
- %vecext1 = extractelement <2 x double> %A, i32 1
- %add = fadd double %vecext, %vecext1
- %vecinit = insertelement <2 x double> undef, double %add, i32 0
- %vecext2 = extractelement <2 x double> %B, i32 0
- %vecext3 = extractelement <2 x double> %B, i32 1
- %add2 = fadd double %vecext2, %vecext3
- %vecinit2 = insertelement <2 x double> %vecinit, double %add2, i32 1
- ret <2 x double> %vecinit2
-}
-
-define <2 x double> @hadd_pd_test2(<2 x double> %A, <2 x double> %B) {
-; SSE-LABEL: hadd_pd_test2:
-; SSE: # %bb.0:
-; SSE-NEXT: haddpd %xmm1, %xmm0
-; SSE-NEXT: retq
-;
-; AVX-LABEL: hadd_pd_test2:
-; AVX: # %bb.0:
-; AVX-NEXT: vhaddpd %xmm1, %xmm0, %xmm0
-; AVX-NEXT: retq
- %vecext = extractelement <2 x double> %A, i32 1
- %vecext1 = extractelement <2 x double> %A, i32 0
- %add = fadd double %vecext, %vecext1
- %vecinit = insertelement <2 x double> undef, double %add, i32 0
- %vecext2 = extractelement <2 x double> %B, i32 1
- %vecext3 = extractelement <2 x double> %B, i32 0
- %add2 = fadd double %vecext2, %vecext3
- %vecinit2 = insertelement <2 x double> %vecinit, double %add2, i32 1
- ret <2 x double> %vecinit2
-}
-
-define <2 x double> @hsub_pd_test1(<2 x double> %A, <2 x double> %B) {
-; SSE-LABEL: hsub_pd_test1:
-; SSE: # %bb.0:
-; SSE-NEXT: hsubpd %xmm1, %xmm0
-; SSE-NEXT: retq
-;
-; AVX-LABEL: hsub_pd_test1:
-; AVX: # %bb.0:
-; AVX-NEXT: vhsubpd %xmm1, %xmm0, %xmm0
-; AVX-NEXT: retq
- %vecext = extractelement <2 x double> %A, i32 0
- %vecext1 = extractelement <2 x double> %A, i32 1
- %sub = fsub double %vecext, %vecext1
- %vecinit = insertelement <2 x double> undef, double %sub, i32 0
- %vecext2 = extractelement <2 x double> %B, i32 0
- %vecext3 = extractelement <2 x double> %B, i32 1
- %sub2 = fsub double %vecext2, %vecext3
- %vecinit2 = insertelement <2 x double> %vecinit, double %sub2, i32 1
- ret <2 x double> %vecinit2
-}
-
-define <2 x double> @hsub_pd_test2(<2 x double> %A, <2 x double> %B) {
-; SSE-LABEL: hsub_pd_test2:
-; SSE: # %bb.0:
-; SSE-NEXT: hsubpd %xmm1, %xmm0
-; SSE-NEXT: retq
-;
-; AVX-LABEL: hsub_pd_test2:
-; AVX: # %bb.0:
-; AVX-NEXT: vhsubpd %xmm1, %xmm0, %xmm0
-; AVX-NEXT: retq
- %vecext = extractelement <2 x double> %B, i32 0
- %vecext1 = extractelement <2 x double> %B, i32 1
- %sub = fsub double %vecext, %vecext1
- %vecinit = insertelement <2 x double> undef, double %sub, i32 1
- %vecext2 = extractelement <2 x double> %A, i32 0
- %vecext3 = extractelement <2 x double> %A, i32 1
- %sub2 = fsub double %vecext2, %vecext3
- %vecinit2 = insertelement <2 x double> %vecinit, double %sub2, i32 0
- ret <2 x double> %vecinit2
-}
-
-define <4 x double> @avx_vhadd_pd_test(<4 x double> %A, <4 x double> %B) {
-; SSE-LABEL: avx_vhadd_pd_test:
-; SSE: # %bb.0:
-; SSE-NEXT: haddpd %xmm1, %xmm0
-; SSE-NEXT: haddpd %xmm3, %xmm2
-; SSE-NEXT: movapd %xmm2, %xmm1
-; SSE-NEXT: retq
-;
-; AVX1-LABEL: avx_vhadd_pd_test:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm0[2,3],ymm1[2,3]
-; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1-NEXT: vhaddpd %ymm2, %ymm0, %ymm0
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: avx_vhadd_pd_test:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vhaddpd %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
-; AVX2-NEXT: retq
- %vecext = extractelement <4 x double> %A, i32 0
- %vecext1 = extractelement <4 x double> %A, i32 1
- %add = fadd double %vecext, %vecext1
- %vecinit = insertelement <4 x double> undef, double %add, i32 0
- %vecext2 = extractelement <4 x double> %A, i32 2
- %vecext3 = extractelement <4 x double> %A, i32 3
- %add4 = fadd double %vecext2, %vecext3
- %vecinit5 = insertelement <4 x double> %vecinit, double %add4, i32 1
- %vecext6 = extractelement <4 x double> %B, i32 0
- %vecext7 = extractelement <4 x double> %B, i32 1
- %add8 = fadd double %vecext6, %vecext7
- %vecinit9 = insertelement <4 x double> %vecinit5, double %add8, i32 2
- %vecext10 = extractelement <4 x double> %B, i32 2
- %vecext11 = extractelement <4 x double> %B, i32 3
- %add12 = fadd double %vecext10, %vecext11
- %vecinit13 = insertelement <4 x double> %vecinit9, double %add12, i32 3
- ret <4 x double> %vecinit13
-}
-
-define <4 x double> @avx_vhsub_pd_test(<4 x double> %A, <4 x double> %B) {
-; SSE-LABEL: avx_vhsub_pd_test:
-; SSE: # %bb.0:
-; SSE-NEXT: hsubpd %xmm1, %xmm0
-; SSE-NEXT: hsubpd %xmm3, %xmm2
-; SSE-NEXT: movapd %xmm2, %xmm1
-; SSE-NEXT: retq
-;
-; AVX1-LABEL: avx_vhsub_pd_test:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm0[2,3],ymm1[2,3]
-; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1-NEXT: vhsubpd %ymm2, %ymm0, %ymm0
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: avx_vhsub_pd_test:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vhsubpd %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vpermpd {{.*#+}} ymm0 = ymm0[0,2,1,3]
-; AVX2-NEXT: retq
- %vecext = extractelement <4 x double> %A, i32 0
- %vecext1 = extractelement <4 x double> %A, i32 1
- %sub = fsub double %vecext, %vecext1
- %vecinit = insertelement <4 x double> undef, double %sub, i32 0
- %vecext2 = extractelement <4 x double> %A, i32 2
- %vecext3 = extractelement <4 x double> %A, i32 3
- %sub4 = fsub double %vecext2, %vecext3
- %vecinit5 = insertelement <4 x double> %vecinit, double %sub4, i32 1
- %vecext6 = extractelement <4 x double> %B, i32 0
- %vecext7 = extractelement <4 x double> %B, i32 1
- %sub8 = fsub double %vecext6, %vecext7
- %vecinit9 = insertelement <4 x double> %vecinit5, double %sub8, i32 2
- %vecext10 = extractelement <4 x double> %B, i32 2
- %vecext11 = extractelement <4 x double> %B, i32 3
- %sub12 = fsub double %vecext10, %vecext11
- %vecinit13 = insertelement <4 x double> %vecinit9, double %sub12, i32 3
- ret <4 x double> %vecinit13
-}
-
-define <8 x i32> @avx2_vphadd_d_test(<8 x i32> %A, <8 x i32> %B) {
-; SSE3-LABEL: avx2_vphadd_d_test:
-; SSE3: # %bb.0:
-; SSE3-NEXT: movd %xmm0, %ecx
-; SSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm0[1,1,1,1]
-; SSE3-NEXT: movd %xmm4, %eax
-; SSE3-NEXT: addl %ecx, %eax
-; SSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm0[2,3,2,3]
-; SSE3-NEXT: movd %xmm4, %edx
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
-; SSE3-NEXT: movd %xmm0, %ecx
-; SSE3-NEXT: addl %edx, %ecx
-; SSE3-NEXT: movd %xmm1, %edx
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE3-NEXT: movd %xmm0, %esi
-; SSE3-NEXT: addl %edx, %esi
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE3-NEXT: movd %xmm0, %edx
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[3,3,3,3]
-; SSE3-NEXT: movd %xmm0, %edi
-; SSE3-NEXT: addl %edx, %edi
-; SSE3-NEXT: movd %xmm2, %r8d
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]
-; SSE3-NEXT: movd %xmm0, %edx
-; SSE3-NEXT: addl %r8d, %edx
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
-; SSE3-NEXT: movd %xmm0, %r8d
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm2[3,3,3,3]
-; SSE3-NEXT: movd %xmm0, %r9d
-; SSE3-NEXT: addl %r8d, %r9d
-; SSE3-NEXT: movd %xmm3, %r8d
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,1,1]
-; SSE3-NEXT: movd %xmm0, %r10d
-; SSE3-NEXT: addl %r8d, %r10d
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
-; SSE3-NEXT: movd %xmm0, %r8d
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm3[3,3,3,3]
-; SSE3-NEXT: movd %xmm0, %r11d
-; SSE3-NEXT: addl %r8d, %r11d
-; SSE3-NEXT: movd %edi, %xmm0
-; SSE3-NEXT: movd %esi, %xmm1
-; SSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE3-NEXT: movd %ecx, %xmm2
-; SSE3-NEXT: movd %eax, %xmm0
-; SSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
-; SSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE3-NEXT: movd %r11d, %xmm1
-; SSE3-NEXT: movd %r10d, %xmm2
-; SSE3-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
-; SSE3-NEXT: movd %r9d, %xmm3
-; SSE3-NEXT: movd %edx, %xmm1
-; SSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]
-; SSE3-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
-; SSE3-NEXT: retq
-;
-; SSSE3-LABEL: avx2_vphadd_d_test:
-; SSSE3: # %bb.0:
-; SSSE3-NEXT: phaddd %xmm1, %xmm0
-; SSSE3-NEXT: phaddd %xmm3, %xmm2
-; SSSE3-NEXT: movdqa %xmm2, %xmm1
-; SSSE3-NEXT: retq
-;
-; AVX1-LABEL: avx2_vphadd_d_test:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
-; AVX1-NEXT: vphaddd %xmm2, %xmm1, %xmm1
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2
-; AVX1-NEXT: vphaddd %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: avx2_vphadd_d_test:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vphaddd %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
-; AVX2-NEXT: retq
- %vecext = extractelement <8 x i32> %A, i32 0
- %vecext1 = extractelement <8 x i32> %A, i32 1
- %add = add i32 %vecext, %vecext1
- %vecinit = insertelement <8 x i32> undef, i32 %add, i32 0
- %vecext2 = extractelement <8 x i32> %A, i32 2
- %vecext3 = extractelement <8 x i32> %A, i32 3
- %add4 = add i32 %vecext2, %vecext3
- %vecinit5 = insertelement <8 x i32> %vecinit, i32 %add4, i32 1
- %vecext6 = extractelement <8 x i32> %A, i32 4
- %vecext7 = extractelement <8 x i32> %A, i32 5
- %add8 = add i32 %vecext6, %vecext7
- %vecinit9 = insertelement <8 x i32> %vecinit5, i32 %add8, i32 2
- %vecext10 = extractelement <8 x i32> %A, i32 6
- %vecext11 = extractelement <8 x i32> %A, i32 7
- %add12 = add i32 %vecext10, %vecext11
- %vecinit13 = insertelement <8 x i32> %vecinit9, i32 %add12, i32 3
- %vecext14 = extractelement <8 x i32> %B, i32 0
- %vecext15 = extractelement <8 x i32> %B, i32 1
- %add16 = add i32 %vecext14, %vecext15
- %vecinit17 = insertelement <8 x i32> %vecinit13, i32 %add16, i32 4
- %vecext18 = extractelement <8 x i32> %B, i32 2
- %vecext19 = extractelement <8 x i32> %B, i32 3
- %add20 = add i32 %vecext18, %vecext19
- %vecinit21 = insertelement <8 x i32> %vecinit17, i32 %add20, i32 5
- %vecext22 = extractelement <8 x i32> %B, i32 4
- %vecext23 = extractelement <8 x i32> %B, i32 5
- %add24 = add i32 %vecext22, %vecext23
- %vecinit25 = insertelement <8 x i32> %vecinit21, i32 %add24, i32 6
- %vecext26 = extractelement <8 x i32> %B, i32 6
- %vecext27 = extractelement <8 x i32> %B, i32 7
- %add28 = add i32 %vecext26, %vecext27
- %vecinit29 = insertelement <8 x i32> %vecinit25, i32 %add28, i32 7
- ret <8 x i32> %vecinit29
-}
-
-define <16 x i16> @avx2_vphadd_w_test(<16 x i16> %a, <16 x i16> %b) nounwind {
-; SSE3-LABEL: avx2_vphadd_w_test:
-; SSE3: # %bb.0:
-; SSE3-NEXT: pushq %rbp
-; SSE3-NEXT: pushq %r15
-; SSE3-NEXT: pushq %r14
-; SSE3-NEXT: pushq %r13
-; SSE3-NEXT: pushq %r12
-; SSE3-NEXT: pushq %rbx
-; SSE3-NEXT: movd %xmm0, %ecx
-; SSE3-NEXT: pextrw $1, %xmm0, %eax
-; SSE3-NEXT: addl %ecx, %eax
-; SSE3-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE3-NEXT: pextrw $2, %xmm0, %edx
-; SSE3-NEXT: pextrw $3, %xmm0, %eax
-; SSE3-NEXT: addl %edx, %eax
-; SSE3-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE3-NEXT: pextrw $4, %xmm0, %edx
-; SSE3-NEXT: pextrw $5, %xmm0, %esi
-; SSE3-NEXT: addl %edx, %esi
-; SSE3-NEXT: pextrw $6, %xmm0, %edx
-; SSE3-NEXT: pextrw $7, %xmm0, %r8d
-; SSE3-NEXT: addl %edx, %r8d
-; SSE3-NEXT: movd %xmm1, %edx
-; SSE3-NEXT: pextrw $1, %xmm1, %r10d
-; SSE3-NEXT: addl %edx, %r10d
-; SSE3-NEXT: pextrw $2, %xmm1, %edx
-; SSE3-NEXT: pextrw $3, %xmm1, %ebx
-; SSE3-NEXT: addl %edx, %ebx
-; SSE3-NEXT: pextrw $4, %xmm1, %edx
-; SSE3-NEXT: pextrw $5, %xmm1, %r14d
-; SSE3-NEXT: addl %edx, %r14d
-; SSE3-NEXT: pextrw $6, %xmm1, %edx
-; SSE3-NEXT: pextrw $7, %xmm1, %r12d
-; SSE3-NEXT: addl %edx, %r12d
-; SSE3-NEXT: movd %xmm2, %edi
-; SSE3-NEXT: pextrw $1, %xmm2, %edx
-; SSE3-NEXT: addl %edi, %edx
-; SSE3-NEXT: pextrw $2, %xmm2, %r9d
-; SSE3-NEXT: pextrw $3, %xmm2, %edi
-; SSE3-NEXT: addl %r9d, %edi
-; SSE3-NEXT: pextrw $4, %xmm2, %r11d
-; SSE3-NEXT: pextrw $5, %xmm2, %r9d
-; SSE3-NEXT: addl %r11d, %r9d
-; SSE3-NEXT: pextrw $6, %xmm2, %ebp
-; SSE3-NEXT: pextrw $7, %xmm2, %r11d
-; SSE3-NEXT: addl %ebp, %r11d
-; SSE3-NEXT: movd %xmm3, %r15d
-; SSE3-NEXT: pextrw $1, %xmm3, %ebp
-; SSE3-NEXT: addl %r15d, %ebp
-; SSE3-NEXT: pextrw $2, %xmm3, %r13d
-; SSE3-NEXT: pextrw $3, %xmm3, %r15d
-; SSE3-NEXT: addl %r13d, %r15d
-; SSE3-NEXT: pextrw $4, %xmm3, %r13d
-; SSE3-NEXT: pextrw $5, %xmm3, %ecx
-; SSE3-NEXT: addl %r13d, %ecx
-; SSE3-NEXT: pextrw $6, %xmm3, %r13d
-; SSE3-NEXT: pextrw $7, %xmm3, %eax
-; SSE3-NEXT: addl %r13d, %eax
-; SSE3-NEXT: movd %r12d, %xmm4
-; SSE3-NEXT: movd %r14d, %xmm2
-; SSE3-NEXT: movd %ebx, %xmm5
-; SSE3-NEXT: movd %r10d, %xmm3
-; SSE3-NEXT: movd %r8d, %xmm6
-; SSE3-NEXT: movd %esi, %xmm7
-; SSE3-NEXT: movd {{[-0-9]+}}(%r{{[sb]}}p), %xmm8 # 4-byte Folded Reload
-; SSE3-NEXT: # xmm8 = mem[0],zero,zero,zero
-; SSE3-NEXT: movd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Folded Reload
-; SSE3-NEXT: # xmm0 = mem[0],zero,zero,zero
-; SSE3-NEXT: movd %eax, %xmm9
-; SSE3-NEXT: movd %ecx, %xmm10
-; SSE3-NEXT: movd %r15d, %xmm11
-; SSE3-NEXT: movd %ebp, %xmm12
-; SSE3-NEXT: movd %r11d, %xmm13
-; SSE3-NEXT: movd %r9d, %xmm14
-; SSE3-NEXT: movd %edi, %xmm15
-; SSE3-NEXT: movd %edx, %xmm1
-; SSE3-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
-; SSE3-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
-; SSE3-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
-; SSE3-NEXT: punpcklwd {{.*#+}} xmm7 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
-; SSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3]
-; SSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1]
-; SSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm3[0]
-; SSE3-NEXT: punpcklwd {{.*#+}} xmm10 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
-; SSE3-NEXT: punpcklwd {{.*#+}} xmm12 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
-; SSE3-NEXT: punpckldq {{.*#+}} xmm12 = xmm12[0],xmm10[0],xmm12[1],xmm10[1]
-; SSE3-NEXT: punpcklwd {{.*#+}} xmm14 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
-; SSE3-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm15[0],xmm1[1],xmm15[1],xmm1[2],xmm15[2],xmm1[3],xmm15[3]
-; SSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm14[0],xmm1[1],xmm14[1]
-; SSE3-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm12[0]
-; SSE3-NEXT: popq %rbx
-; SSE3-NEXT: popq %r12
-; SSE3-NEXT: popq %r13
-; SSE3-NEXT: popq %r14
-; SSE3-NEXT: popq %r15
-; SSE3-NEXT: popq %rbp
-; SSE3-NEXT: retq
-;
-; SSSE3-LABEL: avx2_vphadd_w_test:
-; SSSE3: # %bb.0:
-; SSSE3-NEXT: phaddw %xmm1, %xmm0
-; SSSE3-NEXT: phaddw %xmm3, %xmm2
-; SSSE3-NEXT: movdqa %xmm2, %xmm1
-; SSSE3-NEXT: retq
-;
-; AVX1-LABEL: avx2_vphadd_w_test:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
-; AVX1-NEXT: vphaddw %xmm2, %xmm1, %xmm1
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2
-; AVX1-NEXT: vphaddw %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: avx2_vphadd_w_test:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vphaddw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
-; AVX2-NEXT: retq
- %vecext = extractelement <16 x i16> %a, i32 0
- %vecext1 = extractelement <16 x i16> %a, i32 1
- %add = add i16 %vecext, %vecext1
- %vecinit = insertelement <16 x i16> undef, i16 %add, i32 0
- %vecext4 = extractelement <16 x i16> %a, i32 2
- %vecext6 = extractelement <16 x i16> %a, i32 3
- %add8 = add i16 %vecext4, %vecext6
- %vecinit10 = insertelement <16 x i16> %vecinit, i16 %add8, i32 1
- %vecext11 = extractelement <16 x i16> %a, i32 4
- %vecext13 = extractelement <16 x i16> %a, i32 5
- %add15 = add i16 %vecext11, %vecext13
- %vecinit17 = insertelement <16 x i16> %vecinit10, i16 %add15, i32 2
- %vecext18 = extractelement <16 x i16> %a, i32 6
- %vecext20 = extractelement <16 x i16> %a, i32 7
- %add22 = add i16 %vecext18, %vecext20
- %vecinit24 = insertelement <16 x i16> %vecinit17, i16 %add22, i32 3
- %vecext25 = extractelement <16 x i16> %a, i32 8
- %vecext27 = extractelement <16 x i16> %a, i32 9
- %add29 = add i16 %vecext25, %vecext27
- %vecinit31 = insertelement <16 x i16> %vecinit24, i16 %add29, i32 4
- %vecext32 = extractelement <16 x i16> %a, i32 10
- %vecext34 = extractelement <16 x i16> %a, i32 11
- %add36 = add i16 %vecext32, %vecext34
- %vecinit38 = insertelement <16 x i16> %vecinit31, i16 %add36, i32 5
- %vecext39 = extractelement <16 x i16> %a, i32 12
- %vecext41 = extractelement <16 x i16> %a, i32 13
- %add43 = add i16 %vecext39, %vecext41
- %vecinit45 = insertelement <16 x i16> %vecinit38, i16 %add43, i32 6
- %vecext46 = extractelement <16 x i16> %a, i32 14
- %vecext48 = extractelement <16 x i16> %a, i32 15
- %add50 = add i16 %vecext46, %vecext48
- %vecinit52 = insertelement <16 x i16> %vecinit45, i16 %add50, i32 7
- %vecext53 = extractelement <16 x i16> %b, i32 0
- %vecext55 = extractelement <16 x i16> %b, i32 1
- %add57 = add i16 %vecext53, %vecext55
- %vecinit59 = insertelement <16 x i16> %vecinit52, i16 %add57, i32 8
- %vecext60 = extractelement <16 x i16> %b, i32 2
- %vecext62 = extractelement <16 x i16> %b, i32 3
- %add64 = add i16 %vecext60, %vecext62
- %vecinit66 = insertelement <16 x i16> %vecinit59, i16 %add64, i32 9
- %vecext67 = extractelement <16 x i16> %b, i32 4
- %vecext69 = extractelement <16 x i16> %b, i32 5
- %add71 = add i16 %vecext67, %vecext69
- %vecinit73 = insertelement <16 x i16> %vecinit66, i16 %add71, i32 10
- %vecext74 = extractelement <16 x i16> %b, i32 6
- %vecext76 = extractelement <16 x i16> %b, i32 7
- %add78 = add i16 %vecext74, %vecext76
- %vecinit80 = insertelement <16 x i16> %vecinit73, i16 %add78, i32 11
- %vecext81 = extractelement <16 x i16> %b, i32 8
- %vecext83 = extractelement <16 x i16> %b, i32 9
- %add85 = add i16 %vecext81, %vecext83
- %vecinit87 = insertelement <16 x i16> %vecinit80, i16 %add85, i32 12
- %vecext88 = extractelement <16 x i16> %b, i32 10
- %vecext90 = extractelement <16 x i16> %b, i32 11
- %add92 = add i16 %vecext88, %vecext90
- %vecinit94 = insertelement <16 x i16> %vecinit87, i16 %add92, i32 13
- %vecext95 = extractelement <16 x i16> %b, i32 12
- %vecext97 = extractelement <16 x i16> %b, i32 13
- %add99 = add i16 %vecext95, %vecext97
- %vecinit101 = insertelement <16 x i16> %vecinit94, i16 %add99, i32 14
- %vecext102 = extractelement <16 x i16> %b, i32 14
- %vecext104 = extractelement <16 x i16> %b, i32 15
- %add106 = add i16 %vecext102, %vecext104
- %vecinit108 = insertelement <16 x i16> %vecinit101, i16 %add106, i32 15
- ret <16 x i16> %vecinit108
-}
+; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+sse2,+sse3 | FileCheck %s --check-prefix=SSE
+; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+sse2,+sse3,+ssse3 | FileCheck %s --check-prefix=SSE
+; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx | FileCheck %s --check-prefix=AVX
+; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx2 | FileCheck %s --check-prefix=AVX
; Verify that we don't select horizontal subs in the following functions.
@@ -970,454 +162,3 @@ define <2 x double> @not_a_hsub_3(<2 x double> %A, <2 x double> %B) {
%vecinit2 = insertelement <2 x double> %vecinit, double %sub2, i32 0
ret <2 x double> %vecinit2
}
-
-; Test AVX horizontal add/sub of packed single/double precision
-; floating point values from 256-bit vectors.
-
-define <8 x float> @avx_vhadd_ps(<8 x float> %a, <8 x float> %b) {
-; SSE-LABEL: avx_vhadd_ps:
-; SSE: # %bb.0:
-; SSE-NEXT: haddps %xmm2, %xmm0
-; SSE-NEXT: haddps %xmm3, %xmm1
-; SSE-NEXT: retq
-;
-; AVX-LABEL: avx_vhadd_ps:
-; AVX: # %bb.0:
-; AVX-NEXT: vhaddps %ymm1, %ymm0, %ymm0
-; AVX-NEXT: retq
- %vecext = extractelement <8 x float> %a, i32 0
- %vecext1 = extractelement <8 x float> %a, i32 1
- %add = fadd float %vecext, %vecext1
- %vecinit = insertelement <8 x float> undef, float %add, i32 0
- %vecext2 = extractelement <8 x float> %a, i32 2
- %vecext3 = extractelement <8 x float> %a, i32 3
- %add4 = fadd float %vecext2, %vecext3
- %vecinit5 = insertelement <8 x float> %vecinit, float %add4, i32 1
- %vecext6 = extractelement <8 x float> %b, i32 0
- %vecext7 = extractelement <8 x float> %b, i32 1
- %add8 = fadd float %vecext6, %vecext7
- %vecinit9 = insertelement <8 x float> %vecinit5, float %add8, i32 2
- %vecext10 = extractelement <8 x float> %b, i32 2
- %vecext11 = extractelement <8 x float> %b, i32 3
- %add12 = fadd float %vecext10, %vecext11
- %vecinit13 = insertelement <8 x float> %vecinit9, float %add12, i32 3
- %vecext14 = extractelement <8 x float> %a, i32 4
- %vecext15 = extractelement <8 x float> %a, i32 5
- %add16 = fadd float %vecext14, %vecext15
- %vecinit17 = insertelement <8 x float> %vecinit13, float %add16, i32 4
- %vecext18 = extractelement <8 x float> %a, i32 6
- %vecext19 = extractelement <8 x float> %a, i32 7
- %add20 = fadd float %vecext18, %vecext19
- %vecinit21 = insertelement <8 x float> %vecinit17, float %add20, i32 5
- %vecext22 = extractelement <8 x float> %b, i32 4
- %vecext23 = extractelement <8 x float> %b, i32 5
- %add24 = fadd float %vecext22, %vecext23
- %vecinit25 = insertelement <8 x float> %vecinit21, float %add24, i32 6
- %vecext26 = extractelement <8 x float> %b, i32 6
- %vecext27 = extractelement <8 x float> %b, i32 7
- %add28 = fadd float %vecext26, %vecext27
- %vecinit29 = insertelement <8 x float> %vecinit25, float %add28, i32 7
- ret <8 x float> %vecinit29
-}
-
-define <8 x float> @avx_vhsub_ps(<8 x float> %a, <8 x float> %b) {
-; SSE-LABEL: avx_vhsub_ps:
-; SSE: # %bb.0:
-; SSE-NEXT: hsubps %xmm2, %xmm0
-; SSE-NEXT: hsubps %xmm3, %xmm1
-; SSE-NEXT: retq
-;
-; AVX-LABEL: avx_vhsub_ps:
-; AVX: # %bb.0:
-; AVX-NEXT: vhsubps %ymm1, %ymm0, %ymm0
-; AVX-NEXT: retq
- %vecext = extractelement <8 x float> %a, i32 0
- %vecext1 = extractelement <8 x float> %a, i32 1
- %sub = fsub float %vecext, %vecext1
- %vecinit = insertelement <8 x float> undef, float %sub, i32 0
- %vecext2 = extractelement <8 x float> %a, i32 2
- %vecext3 = extractelement <8 x float> %a, i32 3
- %sub4 = fsub float %vecext2, %vecext3
- %vecinit5 = insertelement <8 x float> %vecinit, float %sub4, i32 1
- %vecext6 = extractelement <8 x float> %b, i32 0
- %vecext7 = extractelement <8 x float> %b, i32 1
- %sub8 = fsub float %vecext6, %vecext7
- %vecinit9 = insertelement <8 x float> %vecinit5, float %sub8, i32 2
- %vecext10 = extractelement <8 x float> %b, i32 2
- %vecext11 = extractelement <8 x float> %b, i32 3
- %sub12 = fsub float %vecext10, %vecext11
- %vecinit13 = insertelement <8 x float> %vecinit9, float %sub12, i32 3
- %vecext14 = extractelement <8 x float> %a, i32 4
- %vecext15 = extractelement <8 x float> %a, i32 5
- %sub16 = fsub float %vecext14, %vecext15
- %vecinit17 = insertelement <8 x float> %vecinit13, float %sub16, i32 4
- %vecext18 = extractelement <8 x float> %a, i32 6
- %vecext19 = extractelement <8 x float> %a, i32 7
- %sub20 = fsub float %vecext18, %vecext19
- %vecinit21 = insertelement <8 x float> %vecinit17, float %sub20, i32 5
- %vecext22 = extractelement <8 x float> %b, i32 4
- %vecext23 = extractelement <8 x float> %b, i32 5
- %sub24 = fsub float %vecext22, %vecext23
- %vecinit25 = insertelement <8 x float> %vecinit21, float %sub24, i32 6
- %vecext26 = extractelement <8 x float> %b, i32 6
- %vecext27 = extractelement <8 x float> %b, i32 7
- %sub28 = fsub float %vecext26, %vecext27
- %vecinit29 = insertelement <8 x float> %vecinit25, float %sub28, i32 7
- ret <8 x float> %vecinit29
-}
-
-define <4 x double> @avx_hadd_pd(<4 x double> %a, <4 x double> %b) {
-; SSE-LABEL: avx_hadd_pd:
-; SSE: # %bb.0:
-; SSE-NEXT: haddpd %xmm2, %xmm0
-; SSE-NEXT: haddpd %xmm3, %xmm1
-; SSE-NEXT: retq
-;
-; AVX-LABEL: avx_hadd_pd:
-; AVX: # %bb.0:
-; AVX-NEXT: vhaddpd %ymm1, %ymm0, %ymm0
-; AVX-NEXT: retq
- %vecext = extractelement <4 x double> %a, i32 0
- %vecext1 = extractelement <4 x double> %a, i32 1
- %add = fadd double %vecext, %vecext1
- %vecinit = insertelement <4 x double> undef, double %add, i32 0
- %vecext2 = extractelement <4 x double> %b, i32 0
- %vecext3 = extractelement <4 x double> %b, i32 1
- %add4 = fadd double %vecext2, %vecext3
- %vecinit5 = insertelement <4 x double> %vecinit, double %add4, i32 1
- %vecext6 = extractelement <4 x double> %a, i32 2
- %vecext7 = extractelement <4 x double> %a, i32 3
- %add8 = fadd double %vecext6, %vecext7
- %vecinit9 = insertelement <4 x double> %vecinit5, double %add8, i32 2
- %vecext10 = extractelement <4 x double> %b, i32 2
- %vecext11 = extractelement <4 x double> %b, i32 3
- %add12 = fadd double %vecext10, %vecext11
- %vecinit13 = insertelement <4 x double> %vecinit9, double %add12, i32 3
- ret <4 x double> %vecinit13
-}
-
-define <4 x double> @avx_hsub_pd(<4 x double> %a, <4 x double> %b) {
-; SSE-LABEL: avx_hsub_pd:
-; SSE: # %bb.0:
-; SSE-NEXT: hsubpd %xmm2, %xmm0
-; SSE-NEXT: hsubpd %xmm3, %xmm1
-; SSE-NEXT: retq
-;
-; AVX-LABEL: avx_hsub_pd:
-; AVX: # %bb.0:
-; AVX-NEXT: vhsubpd %ymm1, %ymm0, %ymm0
-; AVX-NEXT: retq
- %vecext = extractelement <4 x double> %a, i32 0
- %vecext1 = extractelement <4 x double> %a, i32 1
- %sub = fsub double %vecext, %vecext1
- %vecinit = insertelement <4 x double> undef, double %sub, i32 0
- %vecext2 = extractelement <4 x double> %b, i32 0
- %vecext3 = extractelement <4 x double> %b, i32 1
- %sub4 = fsub double %vecext2, %vecext3
- %vecinit5 = insertelement <4 x double> %vecinit, double %sub4, i32 1
- %vecext6 = extractelement <4 x double> %a, i32 2
- %vecext7 = extractelement <4 x double> %a, i32 3
- %sub8 = fsub double %vecext6, %vecext7
- %vecinit9 = insertelement <4 x double> %vecinit5, double %sub8, i32 2
- %vecext10 = extractelement <4 x double> %b, i32 2
- %vecext11 = extractelement <4 x double> %b, i32 3
- %sub12 = fsub double %vecext10, %vecext11
- %vecinit13 = insertelement <4 x double> %vecinit9, double %sub12, i32 3
- ret <4 x double> %vecinit13
-}
-
-; Test AVX2 horizontal add of packed integer values from 256-bit vectors.
-
-define <8 x i32> @avx2_hadd_d(<8 x i32> %a, <8 x i32> %b) {
-; SSE3-LABEL: avx2_hadd_d:
-; SSE3: # %bb.0:
-; SSE3-NEXT: movd %xmm0, %ecx
-; SSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm0[1,1,1,1]
-; SSE3-NEXT: movd %xmm4, %eax
-; SSE3-NEXT: addl %ecx, %eax
-; SSE3-NEXT: pshufd {{.*#+}} xmm4 = xmm0[2,3,2,3]
-; SSE3-NEXT: movd %xmm4, %edx
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm0[3,3,3,3]
-; SSE3-NEXT: movd %xmm0, %ecx
-; SSE3-NEXT: addl %edx, %ecx
-; SSE3-NEXT: movd %xmm2, %edx
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm2[1,1,1,1]
-; SSE3-NEXT: movd %xmm0, %esi
-; SSE3-NEXT: addl %edx, %esi
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
-; SSE3-NEXT: movd %xmm0, %edx
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm2[3,3,3,3]
-; SSE3-NEXT: movd %xmm0, %edi
-; SSE3-NEXT: addl %edx, %edi
-; SSE3-NEXT: movd %xmm1, %r8d
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE3-NEXT: movd %xmm0, %edx
-; SSE3-NEXT: addl %r8d, %edx
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE3-NEXT: movd %xmm0, %r8d
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm1[3,3,3,3]
-; SSE3-NEXT: movd %xmm0, %r9d
-; SSE3-NEXT: addl %r8d, %r9d
-; SSE3-NEXT: movd %xmm3, %r8d
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm3[1,1,1,1]
-; SSE3-NEXT: movd %xmm0, %r10d
-; SSE3-NEXT: addl %r8d, %r10d
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
-; SSE3-NEXT: movd %xmm0, %r8d
-; SSE3-NEXT: pshufd {{.*#+}} xmm0 = xmm3[3,3,3,3]
-; SSE3-NEXT: movd %xmm0, %r11d
-; SSE3-NEXT: addl %r8d, %r11d
-; SSE3-NEXT: movd %edi, %xmm0
-; SSE3-NEXT: movd %esi, %xmm1
-; SSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; SSE3-NEXT: movd %ecx, %xmm2
-; SSE3-NEXT: movd %eax, %xmm0
-; SSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
-; SSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; SSE3-NEXT: movd %r11d, %xmm1
-; SSE3-NEXT: movd %r10d, %xmm2
-; SSE3-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
-; SSE3-NEXT: movd %r9d, %xmm3
-; SSE3-NEXT: movd %edx, %xmm1
-; SSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]
-; SSE3-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm2[0]
-; SSE3-NEXT: retq
-;
-; SSSE3-LABEL: avx2_hadd_d:
-; SSSE3: # %bb.0:
-; SSSE3-NEXT: phaddd %xmm2, %xmm0
-; SSSE3-NEXT: phaddd %xmm3, %xmm1
-; SSSE3-NEXT: retq
-;
-; AVX1-LABEL: avx2_hadd_d:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3
-; AVX1-NEXT: vphaddd %xmm2, %xmm3, %xmm2
-; AVX1-NEXT: vphaddd %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: avx2_hadd_d:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vphaddd %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: retq
- %vecext = extractelement <8 x i32> %a, i32 0
- %vecext1 = extractelement <8 x i32> %a, i32 1
- %add = add i32 %vecext, %vecext1
- %vecinit = insertelement <8 x i32> undef, i32 %add, i32 0
- %vecext2 = extractelement <8 x i32> %a, i32 2
- %vecext3 = extractelement <8 x i32> %a, i32 3
- %add4 = add i32 %vecext2, %vecext3
- %vecinit5 = insertelement <8 x i32> %vecinit, i32 %add4, i32 1
- %vecext6 = extractelement <8 x i32> %b, i32 0
- %vecext7 = extractelement <8 x i32> %b, i32 1
- %add8 = add i32 %vecext6, %vecext7
- %vecinit9 = insertelement <8 x i32> %vecinit5, i32 %add8, i32 2
- %vecext10 = extractelement <8 x i32> %b, i32 2
- %vecext11 = extractelement <8 x i32> %b, i32 3
- %add12 = add i32 %vecext10, %vecext11
- %vecinit13 = insertelement <8 x i32> %vecinit9, i32 %add12, i32 3
- %vecext14 = extractelement <8 x i32> %a, i32 4
- %vecext15 = extractelement <8 x i32> %a, i32 5
- %add16 = add i32 %vecext14, %vecext15
- %vecinit17 = insertelement <8 x i32> %vecinit13, i32 %add16, i32 4
- %vecext18 = extractelement <8 x i32> %a, i32 6
- %vecext19 = extractelement <8 x i32> %a, i32 7
- %add20 = add i32 %vecext18, %vecext19
- %vecinit21 = insertelement <8 x i32> %vecinit17, i32 %add20, i32 5
- %vecext22 = extractelement <8 x i32> %b, i32 4
- %vecext23 = extractelement <8 x i32> %b, i32 5
- %add24 = add i32 %vecext22, %vecext23
- %vecinit25 = insertelement <8 x i32> %vecinit21, i32 %add24, i32 6
- %vecext26 = extractelement <8 x i32> %b, i32 6
- %vecext27 = extractelement <8 x i32> %b, i32 7
- %add28 = add i32 %vecext26, %vecext27
- %vecinit29 = insertelement <8 x i32> %vecinit25, i32 %add28, i32 7
- ret <8 x i32> %vecinit29
-}
-
-define <16 x i16> @avx2_hadd_w(<16 x i16> %a, <16 x i16> %b) nounwind {
-; SSE3-LABEL: avx2_hadd_w:
-; SSE3: # %bb.0:
-; SSE3-NEXT: pushq %rbp
-; SSE3-NEXT: pushq %r15
-; SSE3-NEXT: pushq %r14
-; SSE3-NEXT: pushq %r13
-; SSE3-NEXT: pushq %r12
-; SSE3-NEXT: pushq %rbx
-; SSE3-NEXT: movd %xmm0, %eax
-; SSE3-NEXT: pextrw $1, %xmm0, %edx
-; SSE3-NEXT: addl %eax, %edx
-; SSE3-NEXT: pextrw $2, %xmm0, %eax
-; SSE3-NEXT: pextrw $3, %xmm0, %esi
-; SSE3-NEXT: addl %eax, %esi
-; SSE3-NEXT: pextrw $4, %xmm0, %eax
-; SSE3-NEXT: pextrw $5, %xmm0, %r9d
-; SSE3-NEXT: addl %eax, %r9d
-; SSE3-NEXT: pextrw $6, %xmm0, %eax
-; SSE3-NEXT: pextrw $7, %xmm0, %r10d
-; SSE3-NEXT: addl %eax, %r10d
-; SSE3-NEXT: movd %xmm1, %ecx
-; SSE3-NEXT: pextrw $1, %xmm1, %eax
-; SSE3-NEXT: addl %ecx, %eax
-; SSE3-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE3-NEXT: pextrw $2, %xmm1, %edi
-; SSE3-NEXT: pextrw $3, %xmm1, %eax
-; SSE3-NEXT: addl %edi, %eax
-; SSE3-NEXT: movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; SSE3-NEXT: pextrw $4, %xmm1, %r8d
-; SSE3-NEXT: pextrw $5, %xmm1, %edi
-; SSE3-NEXT: addl %r8d, %edi
-; SSE3-NEXT: pextrw $6, %xmm1, %r11d
-; SSE3-NEXT: pextrw $7, %xmm1, %r8d
-; SSE3-NEXT: addl %r11d, %r8d
-; SSE3-NEXT: movd %xmm2, %r11d
-; SSE3-NEXT: pextrw $1, %xmm2, %ebp
-; SSE3-NEXT: addl %r11d, %ebp
-; SSE3-NEXT: pextrw $2, %xmm2, %r11d
-; SSE3-NEXT: pextrw $3, %xmm2, %r14d
-; SSE3-NEXT: addl %r11d, %r14d
-; SSE3-NEXT: pextrw $4, %xmm2, %r11d
-; SSE3-NEXT: pextrw $5, %xmm2, %r15d
-; SSE3-NEXT: addl %r11d, %r15d
-; SSE3-NEXT: pextrw $6, %xmm2, %r11d
-; SSE3-NEXT: pextrw $7, %xmm2, %r12d
-; SSE3-NEXT: addl %r11d, %r12d
-; SSE3-NEXT: movd %xmm3, %ebx
-; SSE3-NEXT: pextrw $1, %xmm3, %r11d
-; SSE3-NEXT: addl %ebx, %r11d
-; SSE3-NEXT: pextrw $2, %xmm3, %r13d
-; SSE3-NEXT: pextrw $3, %xmm3, %ebx
-; SSE3-NEXT: addl %r13d, %ebx
-; SSE3-NEXT: pextrw $4, %xmm3, %r13d
-; SSE3-NEXT: pextrw $5, %xmm3, %ecx
-; SSE3-NEXT: addl %r13d, %ecx
-; SSE3-NEXT: pextrw $6, %xmm3, %r13d
-; SSE3-NEXT: pextrw $7, %xmm3, %eax
-; SSE3-NEXT: addl %r13d, %eax
-; SSE3-NEXT: movd %r12d, %xmm4
-; SSE3-NEXT: movd %r15d, %xmm2
-; SSE3-NEXT: movd %r14d, %xmm5
-; SSE3-NEXT: movd %ebp, %xmm3
-; SSE3-NEXT: movd %r10d, %xmm6
-; SSE3-NEXT: movd %r9d, %xmm7
-; SSE3-NEXT: movd %esi, %xmm8
-; SSE3-NEXT: movd %edx, %xmm0
-; SSE3-NEXT: movd %eax, %xmm9
-; SSE3-NEXT: movd %ecx, %xmm10
-; SSE3-NEXT: movd %ebx, %xmm11
-; SSE3-NEXT: movd %r11d, %xmm12
-; SSE3-NEXT: movd %r8d, %xmm13
-; SSE3-NEXT: movd %edi, %xmm14
-; SSE3-NEXT: movd {{[-0-9]+}}(%r{{[sb]}}p), %xmm15 # 4-byte Folded Reload
-; SSE3-NEXT: # xmm15 = mem[0],zero,zero,zero
-; SSE3-NEXT: movd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 4-byte Folded Reload
-; SSE3-NEXT: # xmm1 = mem[0],zero,zero,zero
-; SSE3-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
-; SSE3-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1],xmm3[2],xmm5[2],xmm3[3],xmm5[3]
-; SSE3-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
-; SSE3-NEXT: punpcklwd {{.*#+}} xmm7 = xmm7[0],xmm6[0],xmm7[1],xmm6[1],xmm7[2],xmm6[2],xmm7[3],xmm6[3]
-; SSE3-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm8[0],xmm0[1],xmm8[1],xmm0[2],xmm8[2],xmm0[3],xmm8[3]
-; SSE3-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1]
-; SSE3-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm3[0]
-; SSE3-NEXT: punpcklwd {{.*#+}} xmm10 = xmm10[0],xmm9[0],xmm10[1],xmm9[1],xmm10[2],xmm9[2],xmm10[3],xmm9[3]
-; SSE3-NEXT: punpcklwd {{.*#+}} xmm12 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3]
-; SSE3-NEXT: punpckldq {{.*#+}} xmm12 = xmm12[0],xmm10[0],xmm12[1],xmm10[1]
-; SSE3-NEXT: punpcklwd {{.*#+}} xmm14 = xmm14[0],xmm13[0],xmm14[1],xmm13[1],xmm14[2],xmm13[2],xmm14[3],xmm13[3]
-; SSE3-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm15[0],xmm1[1],xmm15[1],xmm1[2],xmm15[2],xmm1[3],xmm15[3]
-; SSE3-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm14[0],xmm1[1],xmm14[1]
-; SSE3-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm12[0]
-; SSE3-NEXT: popq %rbx
-; SSE3-NEXT: popq %r12
-; SSE3-NEXT: popq %r13
-; SSE3-NEXT: popq %r14
-; SSE3-NEXT: popq %r15
-; SSE3-NEXT: popq %rbp
-; SSE3-NEXT: retq
-;
-; SSSE3-LABEL: avx2_hadd_w:
-; SSSE3: # %bb.0:
-; SSSE3-NEXT: phaddw %xmm2, %xmm0
-; SSSE3-NEXT: phaddw %xmm3, %xmm1
-; SSSE3-NEXT: retq
-;
-; AVX1-LABEL: avx2_hadd_w:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3
-; AVX1-NEXT: vphaddw %xmm2, %xmm3, %xmm2
-; AVX1-NEXT: vphaddw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: avx2_hadd_w:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vphaddw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: retq
- %vecext = extractelement <16 x i16> %a, i32 0
- %vecext1 = extractelement <16 x i16> %a, i32 1
- %add = add i16 %vecext, %vecext1
- %vecinit = insertelement <16 x i16> undef, i16 %add, i32 0
- %vecext4 = extractelement <16 x i16> %a, i32 2
- %vecext6 = extractelement <16 x i16> %a, i32 3
- %add8 = add i16 %vecext4, %vecext6
- %vecinit10 = insertelement <16 x i16> %vecinit, i16 %add8, i32 1
- %vecext11 = extractelement <16 x i16> %a, i32 4
- %vecext13 = extractelement <16 x i16> %a, i32 5
- %add15 = add i16 %vecext11, %vecext13
- %vecinit17 = insertelement <16 x i16> %vecinit10, i16 %add15, i32 2
- %vecext18 = extractelement <16 x i16> %a, i32 6
- %vecext20 = extractelement <16 x i16> %a, i32 7
- %add22 = add i16 %vecext18, %vecext20
- %vecinit24 = insertelement <16 x i16> %vecinit17, i16 %add22, i32 3
- %vecext25 = extractelement <16 x i16> %a, i32 8
- %vecext27 = extractelement <16 x i16> %a, i32 9
- %add29 = add i16 %vecext25, %vecext27
- %vecinit31 = insertelement <16 x i16> %vecinit24, i16 %add29, i32 8
- %vecext32 = extractelement <16 x i16> %a, i32 10
- %vecext34 = extractelement <16 x i16> %a, i32 11
- %add36 = add i16 %vecext32, %vecext34
- %vecinit38 = insertelement <16 x i16> %vecinit31, i16 %add36, i32 9
- %vecext39 = extractelement <16 x i16> %a, i32 12
- %vecext41 = extractelement <16 x i16> %a, i32 13
- %add43 = add i16 %vecext39, %vecext41
- %vecinit45 = insertelement <16 x i16> %vecinit38, i16 %add43, i32 10
- %vecext46 = extractelement <16 x i16> %a, i32 14
- %vecext48 = extractelement <16 x i16> %a, i32 15
- %add50 = add i16 %vecext46, %vecext48
- %vecinit52 = insertelement <16 x i16> %vecinit45, i16 %add50, i32 11
- %vecext53 = extractelement <16 x i16> %b, i32 0
- %vecext55 = extractelement <16 x i16> %b, i32 1
- %add57 = add i16 %vecext53, %vecext55
- %vecinit59 = insertelement <16 x i16> %vecinit52, i16 %add57, i32 4
- %vecext60 = extractelement <16 x i16> %b, i32 2
- %vecext62 = extractelement <16 x i16> %b, i32 3
- %add64 = add i16 %vecext60, %vecext62
- %vecinit66 = insertelement <16 x i16> %vecinit59, i16 %add64, i32 5
- %vecext67 = extractelement <16 x i16> %b, i32 4
- %vecext69 = extractelement <16 x i16> %b, i32 5
- %add71 = add i16 %vecext67, %vecext69
- %vecinit73 = insertelement <16 x i16> %vecinit66, i16 %add71, i32 6
- %vecext74 = extractelement <16 x i16> %b, i32 6
- %vecext76 = extractelement <16 x i16> %b, i32 7
- %add78 = add i16 %vecext74, %vecext76
- %vecinit80 = insertelement <16 x i16> %vecinit73, i16 %add78, i32 7
- %vecext81 = extractelement <16 x i16> %b, i32 8
- %vecext83 = extractelement <16 x i16> %b, i32 9
- %add85 = add i16 %vecext81, %vecext83
- %vecinit87 = insertelement <16 x i16> %vecinit80, i16 %add85, i32 12
- %vecext88 = extractelement <16 x i16> %b, i32 10
- %vecext90 = extractelement <16 x i16> %b, i32 11
- %add92 = add i16 %vecext88, %vecext90
- %vecinit94 = insertelement <16 x i16> %vecinit87, i16 %add92, i32 13
- %vecext95 = extractelement <16 x i16> %b, i32 12
- %vecext97 = extractelement <16 x i16> %b, i32 13
- %add99 = add i16 %vecext95, %vecext97
- %vecinit101 = insertelement <16 x i16> %vecinit94, i16 %add99, i32 14
- %vecext102 = extractelement <16 x i16> %b, i32 14
- %vecext104 = extractelement <16 x i16> %b, i32 15
- %add106 = add i16 %vecext102, %vecext104
- %vecinit108 = insertelement <16 x i16> %vecinit101, i16 %add106, i32 15
- ret <16 x i16> %vecinit108
-}
More information about the llvm-commits
mailing list