[llvm] 00a14f8 - [X86] vector-reduce-mul.ll - add 32-bit test coverage to the integer multiply tests (#196030)
via llvm-commits
llvm-commits at lists.llvm.org
Wed May 6 02:53:22 PDT 2026
Author: Simon Pilgrim
Date: 2026-05-06T09:53:16Z
New Revision: 00a14f8d1612b63e6104045450304511b32f6ce0
URL: https://github.com/llvm/llvm-project/commit/00a14f8d1612b63e6104045450304511b32f6ce0
DIFF: https://github.com/llvm/llvm-project/commit/00a14f8d1612b63e6104045450304511b32f6ce0.diff
LOG: [X86] vector-reduce-mul.ll - add 32-bit test coverage to the integer multiply tests (#196030)
Added:
Modified:
llvm/test/CodeGen/X86/vector-reduce-mul.ll
Removed:
################################################################################
diff --git a/llvm/test/CodeGen/X86/vector-reduce-mul.ll b/llvm/test/CodeGen/X86/vector-reduce-mul.ll
index 5e3ef32ef7e4a..103c3329ede5c 100644
--- a/llvm/test/CodeGen/X86/vector-reduce-mul.ll
+++ b/llvm/test/CodeGen/X86/vector-reduce-mul.ll
@@ -1,47 +1,147 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefixes=SSE,SSE2
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefixes=SSE,SSE41
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1OR2,AVX1
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX1OR2,AVX2
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefixes=AVX,AVX512,AVX512BW
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw,+avx512vl | FileCheck %s --check-prefixes=AVX,AVX512,AVX512BWVL
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512dq | FileCheck %s --check-prefixes=AVX,AVX512,AVX512DQ
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512dq,+avx512vl | FileCheck %s --check-prefixes=AVX,AVX512,AVX512DQVL
+; RUN: llc < %s -mtriple=i686-- -mattr=+sse2 | FileCheck %s --check-prefixes=SSE,X86-SSE,SSE2,X86-SSE2
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse2 | FileCheck %s --check-prefixes=SSE,X64-SSE,SSE2,X64-SSE2
+; RUN: llc < %s -mtriple=i686-- -mattr=+sse4.1 | FileCheck %s --check-prefixes=SSE,X86-SSE,SSE4,X86-SSE4
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse4.1 | FileCheck %s --check-prefixes=SSE,X64-SSE,SSE4,X64-SSE4
+; RUN: llc < %s -mtriple=i686-- -mattr=+sse4.2 | FileCheck %s --check-prefixes=SSE,X86-SSE,SSE4,X86-SSE4
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse4.2 | FileCheck %s --check-prefixes=SSE,X64-SSE,SSE4,X64-SSE4
+; RUN: llc < %s -mtriple=i686-- -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1OR2,AVX1,X86-AVX1
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1OR2,AVX1,X64-AVX1
+; RUN: llc < %s -mtriple=i686-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX1OR2,AVX2,X86-AVX2
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX1OR2,AVX2,X64-AVX2
+; RUN: llc < %s -mtriple=x86_64- -mattr=+avx512f | FileCheck %s --check-prefixes=AVX,AVX512,AVX512F
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefixes=AVX,AVX512,AVX512BW
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f,+avx512bw,+avx512vl | FileCheck %s --check-prefixes=AVX,AVX512,AVX512BWVL
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f,+avx512dq | FileCheck %s --check-prefixes=AVX,AVX512,AVX512DQ
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f,+avx512dq,+avx512vl | FileCheck %s --check-prefixes=AVX,AVX512,AVX512DQVL
;
; vXi64
;
-define i64 @test_v2i64(<2 x i64> %a0) {
-; SSE-LABEL: test_v2i64:
-; SSE: # %bb.0:
-; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE-NEXT: movdqa %xmm0, %xmm2
-; SSE-NEXT: psrlq $32, %xmm2
-; SSE-NEXT: pmuludq %xmm1, %xmm2
-; SSE-NEXT: movdqa %xmm0, %xmm3
-; SSE-NEXT: psrldq {{.*#+}} xmm3 = xmm3[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; SSE-NEXT: pmuludq %xmm0, %xmm3
-; SSE-NEXT: paddq %xmm2, %xmm3
-; SSE-NEXT: psllq $32, %xmm3
-; SSE-NEXT: pmuludq %xmm1, %xmm0
-; SSE-NEXT: paddq %xmm3, %xmm0
-; SSE-NEXT: movq %xmm0, %rax
-; SSE-NEXT: retq
-;
-; AVX1OR2-LABEL: test_v2i64:
-; AVX1OR2: # %bb.0:
-; AVX1OR2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1OR2-NEXT: vpsrlq $32, %xmm0, %xmm2
-; AVX1OR2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
-; AVX1OR2-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; AVX1OR2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
-; AVX1OR2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
-; AVX1OR2-NEXT: vpsllq $32, %xmm2, %xmm2
-; AVX1OR2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
-; AVX1OR2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
-; AVX1OR2-NEXT: vmovq %xmm0, %rax
-; AVX1OR2-NEXT: retq
+define i64 @test_v2i64(<2 x i64> %a0) nounwind {
+; X86-SSE2-LABEL: test_v2i64:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm2
+; X86-SSE2-NEXT: psrlq $32, %xmm2
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm2
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm3
+; X86-SSE2-NEXT: psrldq {{.*#+}} xmm3 = xmm3[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE2-NEXT: pmuludq %xmm0, %xmm3
+; X86-SSE2-NEXT: paddq %xmm2, %xmm3
+; X86-SSE2-NEXT: psllq $32, %xmm3
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm0
+; X86-SSE2-NEXT: paddq %xmm3, %xmm0
+; X86-SSE2-NEXT: movd %xmm0, %eax
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; X86-SSE2-NEXT: movd %xmm0, %edx
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE-LABEL: test_v2i64:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE-NEXT: movdqa %xmm0, %xmm2
+; X64-SSE-NEXT: psrlq $32, %xmm2
+; X64-SSE-NEXT: pmuludq %xmm1, %xmm2
+; X64-SSE-NEXT: movdqa %xmm0, %xmm3
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm3 = xmm3[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pmuludq %xmm0, %xmm3
+; X64-SSE-NEXT: paddq %xmm2, %xmm3
+; X64-SSE-NEXT: psllq $32, %xmm3
+; X64-SSE-NEXT: pmuludq %xmm1, %xmm0
+; X64-SSE-NEXT: paddq %xmm3, %xmm0
+; X64-SSE-NEXT: movq %xmm0, %rax
+; X64-SSE-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v2i64:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm2
+; X86-SSE4-NEXT: psrlq $32, %xmm2
+; X86-SSE4-NEXT: pmuludq %xmm1, %xmm2
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm3
+; X86-SSE4-NEXT: psrldq {{.*#+}} xmm3 = xmm3[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE4-NEXT: pmuludq %xmm0, %xmm3
+; X86-SSE4-NEXT: paddq %xmm2, %xmm3
+; X86-SSE4-NEXT: psllq $32, %xmm3
+; X86-SSE4-NEXT: pmuludq %xmm1, %xmm0
+; X86-SSE4-NEXT: paddq %xmm3, %xmm0
+; X86-SSE4-NEXT: movd %xmm0, %eax
+; X86-SSE4-NEXT: pextrd $1, %xmm0, %edx
+; X86-SSE4-NEXT: retl
+;
+; X86-AVX1-LABEL: test_v2i64:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X86-AVX1-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X86-AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v2i64:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X64-AVX1-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X64-AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
+; X64-AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovq %xmm0, %rax
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v2i64:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X86-AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X86-AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X86-AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
+; X86-AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v2i64:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X64-AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X64-AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X64-AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
+; X64-AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: retq
+;
+; AVX512F-LABEL: test_v2i64:
+; AVX512F: # %bb.0:
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512F-NEXT: vpsrlq $32, %xmm0, %xmm2
+; AVX512F-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; AVX512F-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX512F-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; AVX512F-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; AVX512F-NEXT: vpsllq $32, %xmm2, %xmm2
+; AVX512F-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; AVX512F-NEXT: vmovq %xmm0, %rax
+; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: test_v2i64:
; AVX512BW: # %bb.0:
@@ -90,80 +190,209 @@ define i64 @test_v2i64(<2 x i64> %a0) {
ret i64 %1
}
-define i64 @test_v4i64(<4 x i64> %a0) {
-; SSE-LABEL: test_v4i64:
-; SSE: # %bb.0:
-; SSE-NEXT: movdqa %xmm0, %xmm2
-; SSE-NEXT: psrlq $32, %xmm2
-; SSE-NEXT: pmuludq %xmm1, %xmm2
-; SSE-NEXT: movdqa %xmm1, %xmm3
-; SSE-NEXT: psrlq $32, %xmm3
-; SSE-NEXT: pmuludq %xmm0, %xmm3
-; SSE-NEXT: paddq %xmm2, %xmm3
-; SSE-NEXT: psllq $32, %xmm3
-; SSE-NEXT: pmuludq %xmm1, %xmm0
-; SSE-NEXT: paddq %xmm3, %xmm0
-; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE-NEXT: movdqa %xmm0, %xmm2
-; SSE-NEXT: psrlq $32, %xmm2
-; SSE-NEXT: pmuludq %xmm1, %xmm2
-; SSE-NEXT: movdqa %xmm0, %xmm3
-; SSE-NEXT: psrldq {{.*#+}} xmm3 = xmm3[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; SSE-NEXT: pmuludq %xmm0, %xmm3
-; SSE-NEXT: paddq %xmm2, %xmm3
-; SSE-NEXT: psllq $32, %xmm3
-; SSE-NEXT: pmuludq %xmm1, %xmm0
-; SSE-NEXT: paddq %xmm3, %xmm0
-; SSE-NEXT: movq %xmm0, %rax
-; SSE-NEXT: retq
-;
-; AVX1-LABEL: test_v4i64:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
-; AVX1-NEXT: vpsrlq $32, %xmm0, %xmm2
-; AVX1-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
-; AVX1-NEXT: vpsrlq $32, %xmm1, %xmm3
-; AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
-; AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
-; AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
-; AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpsrlq $32, %xmm0, %xmm2
-; AVX1-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
-; AVX1-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
-; AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
-; AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
-; AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vmovq %xmm0, %rax
-; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
+define i64 @test_v4i64(<4 x i64> %a0) nounwind {
+; X86-SSE2-LABEL: test_v4i64:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm2
+; X86-SSE2-NEXT: psrlq $32, %xmm2
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm2
+; X86-SSE2-NEXT: movdqa %xmm1, %xmm3
+; X86-SSE2-NEXT: psrlq $32, %xmm3
+; X86-SSE2-NEXT: pmuludq %xmm0, %xmm3
+; X86-SSE2-NEXT: paddq %xmm2, %xmm3
+; X86-SSE2-NEXT: psllq $32, %xmm3
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm0
+; X86-SSE2-NEXT: paddq %xmm3, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm2
+; X86-SSE2-NEXT: psrlq $32, %xmm2
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm2
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm3
+; X86-SSE2-NEXT: psrldq {{.*#+}} xmm3 = xmm3[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE2-NEXT: pmuludq %xmm0, %xmm3
+; X86-SSE2-NEXT: paddq %xmm2, %xmm3
+; X86-SSE2-NEXT: psllq $32, %xmm3
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm0
+; X86-SSE2-NEXT: paddq %xmm3, %xmm0
+; X86-SSE2-NEXT: movd %xmm0, %eax
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; X86-SSE2-NEXT: movd %xmm0, %edx
+; X86-SSE2-NEXT: retl
;
-; AVX2-LABEL: test_v4i64:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
-; AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
-; AVX2-NEXT: vpsrlq $32, %xmm1, %xmm3
-; AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
-; AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
-; AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
-; AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
-; AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
-; AVX2-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
-; AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
-; AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
-; AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
-; AVX2-NEXT: vmovq %xmm0, %rax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+; X64-SSE-LABEL: test_v4i64:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: movdqa %xmm0, %xmm2
+; X64-SSE-NEXT: psrlq $32, %xmm2
+; X64-SSE-NEXT: pmuludq %xmm1, %xmm2
+; X64-SSE-NEXT: movdqa %xmm1, %xmm3
+; X64-SSE-NEXT: psrlq $32, %xmm3
+; X64-SSE-NEXT: pmuludq %xmm0, %xmm3
+; X64-SSE-NEXT: paddq %xmm2, %xmm3
+; X64-SSE-NEXT: psllq $32, %xmm3
+; X64-SSE-NEXT: pmuludq %xmm1, %xmm0
+; X64-SSE-NEXT: paddq %xmm3, %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE-NEXT: movdqa %xmm0, %xmm2
+; X64-SSE-NEXT: psrlq $32, %xmm2
+; X64-SSE-NEXT: pmuludq %xmm1, %xmm2
+; X64-SSE-NEXT: movdqa %xmm0, %xmm3
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm3 = xmm3[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pmuludq %xmm0, %xmm3
+; X64-SSE-NEXT: paddq %xmm2, %xmm3
+; X64-SSE-NEXT: psllq $32, %xmm3
+; X64-SSE-NEXT: pmuludq %xmm1, %xmm0
+; X64-SSE-NEXT: paddq %xmm3, %xmm0
+; X64-SSE-NEXT: movq %xmm0, %rax
+; X64-SSE-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v4i64:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm2
+; X86-SSE4-NEXT: psrlq $32, %xmm2
+; X86-SSE4-NEXT: pmuludq %xmm1, %xmm2
+; X86-SSE4-NEXT: movdqa %xmm1, %xmm3
+; X86-SSE4-NEXT: psrlq $32, %xmm3
+; X86-SSE4-NEXT: pmuludq %xmm0, %xmm3
+; X86-SSE4-NEXT: paddq %xmm2, %xmm3
+; X86-SSE4-NEXT: psllq $32, %xmm3
+; X86-SSE4-NEXT: pmuludq %xmm1, %xmm0
+; X86-SSE4-NEXT: paddq %xmm3, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm2
+; X86-SSE4-NEXT: psrlq $32, %xmm2
+; X86-SSE4-NEXT: pmuludq %xmm1, %xmm2
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm3
+; X86-SSE4-NEXT: psrldq {{.*#+}} xmm3 = xmm3[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE4-NEXT: pmuludq %xmm0, %xmm3
+; X86-SSE4-NEXT: paddq %xmm2, %xmm3
+; X86-SSE4-NEXT: psllq $32, %xmm3
+; X86-SSE4-NEXT: pmuludq %xmm1, %xmm0
+; X86-SSE4-NEXT: paddq %xmm3, %xmm0
+; X86-SSE4-NEXT: movd %xmm0, %eax
+; X86-SSE4-NEXT: pextrd $1, %xmm0, %edx
+; X86-SSE4-NEXT: retl
+;
+; X86-AVX1-LABEL: test_v4i64:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
+; X86-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X86-AVX1-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpsrlq $32, %xmm1, %xmm3
+; X86-AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X86-AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X86-AVX1-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X86-AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX1-NEXT: vzeroupper
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v4i64:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
+; X64-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X64-AVX1-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X64-AVX1-NEXT: vpsrlq $32, %xmm1, %xmm3
+; X64-AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X64-AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
+; X64-AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X64-AVX1-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X64-AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
+; X64-AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovq %xmm0, %rax
+; X64-AVX1-NEXT: vzeroupper
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v4i64:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X86-AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X86-AVX2-NEXT: vpsrlq $32, %xmm1, %xmm3
+; X86-AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X86-AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X86-AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
+; X86-AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X86-AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X86-AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X86-AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
+; X86-AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX2-NEXT: vzeroupper
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v4i64:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X64-AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X64-AVX2-NEXT: vpsrlq $32, %xmm1, %xmm3
+; X64-AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X64-AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X64-AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
+; X64-AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X64-AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X64-AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X64-AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
+; X64-AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
+;
+; AVX512F-LABEL: test_v4i64:
+; AVX512F: # %bb.0:
+; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
+; AVX512F-NEXT: vpsrlq $32, %xmm0, %xmm2
+; AVX512F-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; AVX512F-NEXT: vpsrlq $32, %xmm1, %xmm3
+; AVX512F-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; AVX512F-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; AVX512F-NEXT: vpsllq $32, %xmm2, %xmm2
+; AVX512F-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512F-NEXT: vpsrlq $32, %xmm0, %xmm2
+; AVX512F-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; AVX512F-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX512F-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; AVX512F-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; AVX512F-NEXT: vpsllq $32, %xmm2, %xmm2
+; AVX512F-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; AVX512F-NEXT: vmovq %xmm0, %rax
+; AVX512F-NEXT: vzeroupper
+; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: test_v4i64:
; AVX512BW: # %bb.0:
@@ -237,125 +466,342 @@ define i64 @test_v4i64(<4 x i64> %a0) {
ret i64 %1
}
-define i64 @test_v8i64(<8 x i64> %a0) {
-; SSE-LABEL: test_v8i64:
-; SSE: # %bb.0:
-; SSE-NEXT: movdqa %xmm1, %xmm4
-; SSE-NEXT: psrlq $32, %xmm4
-; SSE-NEXT: pmuludq %xmm3, %xmm4
-; SSE-NEXT: movdqa %xmm3, %xmm5
-; SSE-NEXT: psrlq $32, %xmm5
-; SSE-NEXT: pmuludq %xmm1, %xmm5
-; SSE-NEXT: paddq %xmm4, %xmm5
-; SSE-NEXT: psllq $32, %xmm5
-; SSE-NEXT: pmuludq %xmm3, %xmm1
-; SSE-NEXT: paddq %xmm5, %xmm1
-; SSE-NEXT: movdqa %xmm0, %xmm3
-; SSE-NEXT: psrlq $32, %xmm3
-; SSE-NEXT: pmuludq %xmm2, %xmm3
-; SSE-NEXT: movdqa %xmm2, %xmm4
-; SSE-NEXT: psrlq $32, %xmm4
-; SSE-NEXT: pmuludq %xmm0, %xmm4
-; SSE-NEXT: paddq %xmm3, %xmm4
-; SSE-NEXT: psllq $32, %xmm4
-; SSE-NEXT: pmuludq %xmm2, %xmm0
-; SSE-NEXT: paddq %xmm4, %xmm0
-; SSE-NEXT: movdqa %xmm0, %xmm2
-; SSE-NEXT: psrlq $32, %xmm2
-; SSE-NEXT: pmuludq %xmm1, %xmm2
-; SSE-NEXT: movdqa %xmm1, %xmm3
-; SSE-NEXT: psrlq $32, %xmm3
-; SSE-NEXT: pmuludq %xmm0, %xmm3
-; SSE-NEXT: paddq %xmm2, %xmm3
-; SSE-NEXT: psllq $32, %xmm3
-; SSE-NEXT: pmuludq %xmm1, %xmm0
-; SSE-NEXT: paddq %xmm3, %xmm0
-; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE-NEXT: movdqa %xmm0, %xmm2
-; SSE-NEXT: psrlq $32, %xmm2
-; SSE-NEXT: pmuludq %xmm1, %xmm2
-; SSE-NEXT: movdqa %xmm0, %xmm3
-; SSE-NEXT: psrldq {{.*#+}} xmm3 = xmm3[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; SSE-NEXT: pmuludq %xmm0, %xmm3
-; SSE-NEXT: paddq %xmm2, %xmm3
-; SSE-NEXT: psllq $32, %xmm3
-; SSE-NEXT: pmuludq %xmm1, %xmm0
-; SSE-NEXT: paddq %xmm3, %xmm0
-; SSE-NEXT: movq %xmm0, %rax
-; SSE-NEXT: retq
-;
-; AVX1-LABEL: test_v8i64:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3
-; AVX1-NEXT: vpsrlq $32, %xmm3, %xmm4
-; AVX1-NEXT: vpmuludq %xmm2, %xmm4, %xmm4
-; AVX1-NEXT: vpsrlq $32, %xmm2, %xmm5
-; AVX1-NEXT: vpmuludq %xmm5, %xmm3, %xmm5
-; AVX1-NEXT: vpaddq %xmm4, %xmm5, %xmm4
-; AVX1-NEXT: vpsllq $32, %xmm4, %xmm4
-; AVX1-NEXT: vpmuludq %xmm2, %xmm3, %xmm2
-; AVX1-NEXT: vpaddq %xmm4, %xmm2, %xmm2
-; AVX1-NEXT: vpsrlq $32, %xmm0, %xmm3
-; AVX1-NEXT: vpmuludq %xmm1, %xmm3, %xmm3
-; AVX1-NEXT: vpsrlq $32, %xmm1, %xmm4
-; AVX1-NEXT: vpmuludq %xmm4, %xmm0, %xmm4
-; AVX1-NEXT: vpaddq %xmm3, %xmm4, %xmm3
-; AVX1-NEXT: vpsllq $32, %xmm3, %xmm3
-; AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm3, %xmm0, %xmm0
-; AVX1-NEXT: vpsrlq $32, %xmm0, %xmm1
-; AVX1-NEXT: vpmuludq %xmm2, %xmm1, %xmm1
-; AVX1-NEXT: vpsrlq $32, %xmm2, %xmm3
-; AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
-; AVX1-NEXT: vpaddq %xmm1, %xmm3, %xmm1
-; AVX1-NEXT: vpsllq $32, %xmm1, %xmm1
-; AVX1-NEXT: vpmuludq %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpsrlq $32, %xmm0, %xmm2
-; AVX1-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
-; AVX1-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
-; AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
-; AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
-; AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vmovq %xmm0, %rax
-; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
+define i64 @test_v8i64(<8 x i64> %a0) nounwind {
+; X86-SSE2-LABEL: test_v8i64:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: pushl %ebp
+; X86-SSE2-NEXT: movl %esp, %ebp
+; X86-SSE2-NEXT: andl $-16, %esp
+; X86-SSE2-NEXT: subl $16, %esp
+; X86-SSE2-NEXT: movdqa 8(%ebp), %xmm3
+; X86-SSE2-NEXT: movdqa %xmm1, %xmm4
+; X86-SSE2-NEXT: psrlq $32, %xmm4
+; X86-SSE2-NEXT: pmuludq %xmm3, %xmm4
+; X86-SSE2-NEXT: movdqa %xmm3, %xmm5
+; X86-SSE2-NEXT: psrlq $32, %xmm5
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm5
+; X86-SSE2-NEXT: paddq %xmm4, %xmm5
+; X86-SSE2-NEXT: psllq $32, %xmm5
+; X86-SSE2-NEXT: pmuludq %xmm3, %xmm1
+; X86-SSE2-NEXT: paddq %xmm5, %xmm1
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm3
+; X86-SSE2-NEXT: psrlq $32, %xmm3
+; X86-SSE2-NEXT: pmuludq %xmm2, %xmm3
+; X86-SSE2-NEXT: movdqa %xmm2, %xmm4
+; X86-SSE2-NEXT: psrlq $32, %xmm4
+; X86-SSE2-NEXT: pmuludq %xmm0, %xmm4
+; X86-SSE2-NEXT: paddq %xmm3, %xmm4
+; X86-SSE2-NEXT: psllq $32, %xmm4
+; X86-SSE2-NEXT: pmuludq %xmm2, %xmm0
+; X86-SSE2-NEXT: paddq %xmm4, %xmm0
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm2
+; X86-SSE2-NEXT: psrlq $32, %xmm2
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm2
+; X86-SSE2-NEXT: movdqa %xmm1, %xmm3
+; X86-SSE2-NEXT: psrlq $32, %xmm3
+; X86-SSE2-NEXT: pmuludq %xmm0, %xmm3
+; X86-SSE2-NEXT: paddq %xmm2, %xmm3
+; X86-SSE2-NEXT: psllq $32, %xmm3
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm0
+; X86-SSE2-NEXT: paddq %xmm3, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm2
+; X86-SSE2-NEXT: psrlq $32, %xmm2
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm2
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm3
+; X86-SSE2-NEXT: psrldq {{.*#+}} xmm3 = xmm3[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE2-NEXT: pmuludq %xmm0, %xmm3
+; X86-SSE2-NEXT: paddq %xmm2, %xmm3
+; X86-SSE2-NEXT: psllq $32, %xmm3
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm0
+; X86-SSE2-NEXT: paddq %xmm3, %xmm0
+; X86-SSE2-NEXT: movd %xmm0, %eax
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; X86-SSE2-NEXT: movd %xmm0, %edx
+; X86-SSE2-NEXT: movl %ebp, %esp
+; X86-SSE2-NEXT: popl %ebp
+; X86-SSE2-NEXT: retl
;
-; AVX2-LABEL: test_v8i64:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpsrlq $32, %ymm0, %ymm2
-; AVX2-NEXT: vpmuludq %ymm1, %ymm2, %ymm2
-; AVX2-NEXT: vpsrlq $32, %ymm1, %ymm3
-; AVX2-NEXT: vpmuludq %ymm3, %ymm0, %ymm3
-; AVX2-NEXT: vpaddq %ymm2, %ymm3, %ymm2
-; AVX2-NEXT: vpsllq $32, %ymm2, %ymm2
-; AVX2-NEXT: vpmuludq %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vpaddq %ymm2, %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
-; AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
-; AVX2-NEXT: vpsrlq $32, %xmm1, %xmm3
-; AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
-; AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
-; AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
-; AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
-; AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
-; AVX2-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
-; AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
-; AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
-; AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
-; AVX2-NEXT: vmovq %xmm0, %rax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+; X64-SSE-LABEL: test_v8i64:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: movdqa %xmm1, %xmm4
+; X64-SSE-NEXT: psrlq $32, %xmm4
+; X64-SSE-NEXT: pmuludq %xmm3, %xmm4
+; X64-SSE-NEXT: movdqa %xmm3, %xmm5
+; X64-SSE-NEXT: psrlq $32, %xmm5
+; X64-SSE-NEXT: pmuludq %xmm1, %xmm5
+; X64-SSE-NEXT: paddq %xmm4, %xmm5
+; X64-SSE-NEXT: psllq $32, %xmm5
+; X64-SSE-NEXT: pmuludq %xmm3, %xmm1
+; X64-SSE-NEXT: paddq %xmm5, %xmm1
+; X64-SSE-NEXT: movdqa %xmm0, %xmm3
+; X64-SSE-NEXT: psrlq $32, %xmm3
+; X64-SSE-NEXT: pmuludq %xmm2, %xmm3
+; X64-SSE-NEXT: movdqa %xmm2, %xmm4
+; X64-SSE-NEXT: psrlq $32, %xmm4
+; X64-SSE-NEXT: pmuludq %xmm0, %xmm4
+; X64-SSE-NEXT: paddq %xmm3, %xmm4
+; X64-SSE-NEXT: psllq $32, %xmm4
+; X64-SSE-NEXT: pmuludq %xmm2, %xmm0
+; X64-SSE-NEXT: paddq %xmm4, %xmm0
+; X64-SSE-NEXT: movdqa %xmm0, %xmm2
+; X64-SSE-NEXT: psrlq $32, %xmm2
+; X64-SSE-NEXT: pmuludq %xmm1, %xmm2
+; X64-SSE-NEXT: movdqa %xmm1, %xmm3
+; X64-SSE-NEXT: psrlq $32, %xmm3
+; X64-SSE-NEXT: pmuludq %xmm0, %xmm3
+; X64-SSE-NEXT: paddq %xmm2, %xmm3
+; X64-SSE-NEXT: psllq $32, %xmm3
+; X64-SSE-NEXT: pmuludq %xmm1, %xmm0
+; X64-SSE-NEXT: paddq %xmm3, %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE-NEXT: movdqa %xmm0, %xmm2
+; X64-SSE-NEXT: psrlq $32, %xmm2
+; X64-SSE-NEXT: pmuludq %xmm1, %xmm2
+; X64-SSE-NEXT: movdqa %xmm0, %xmm3
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm3 = xmm3[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pmuludq %xmm0, %xmm3
+; X64-SSE-NEXT: paddq %xmm2, %xmm3
+; X64-SSE-NEXT: psllq $32, %xmm3
+; X64-SSE-NEXT: pmuludq %xmm1, %xmm0
+; X64-SSE-NEXT: paddq %xmm3, %xmm0
+; X64-SSE-NEXT: movq %xmm0, %rax
+; X64-SSE-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v8i64:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pushl %ebp
+; X86-SSE4-NEXT: movl %esp, %ebp
+; X86-SSE4-NEXT: andl $-16, %esp
+; X86-SSE4-NEXT: subl $16, %esp
+; X86-SSE4-NEXT: movdqa 8(%ebp), %xmm3
+; X86-SSE4-NEXT: movdqa %xmm1, %xmm4
+; X86-SSE4-NEXT: psrlq $32, %xmm4
+; X86-SSE4-NEXT: pmuludq %xmm3, %xmm4
+; X86-SSE4-NEXT: movdqa %xmm3, %xmm5
+; X86-SSE4-NEXT: psrlq $32, %xmm5
+; X86-SSE4-NEXT: pmuludq %xmm1, %xmm5
+; X86-SSE4-NEXT: paddq %xmm4, %xmm5
+; X86-SSE4-NEXT: psllq $32, %xmm5
+; X86-SSE4-NEXT: pmuludq %xmm3, %xmm1
+; X86-SSE4-NEXT: paddq %xmm5, %xmm1
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm3
+; X86-SSE4-NEXT: psrlq $32, %xmm3
+; X86-SSE4-NEXT: pmuludq %xmm2, %xmm3
+; X86-SSE4-NEXT: movdqa %xmm2, %xmm4
+; X86-SSE4-NEXT: psrlq $32, %xmm4
+; X86-SSE4-NEXT: pmuludq %xmm0, %xmm4
+; X86-SSE4-NEXT: paddq %xmm3, %xmm4
+; X86-SSE4-NEXT: psllq $32, %xmm4
+; X86-SSE4-NEXT: pmuludq %xmm2, %xmm0
+; X86-SSE4-NEXT: paddq %xmm4, %xmm0
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm2
+; X86-SSE4-NEXT: psrlq $32, %xmm2
+; X86-SSE4-NEXT: pmuludq %xmm1, %xmm2
+; X86-SSE4-NEXT: movdqa %xmm1, %xmm3
+; X86-SSE4-NEXT: psrlq $32, %xmm3
+; X86-SSE4-NEXT: pmuludq %xmm0, %xmm3
+; X86-SSE4-NEXT: paddq %xmm2, %xmm3
+; X86-SSE4-NEXT: psllq $32, %xmm3
+; X86-SSE4-NEXT: pmuludq %xmm1, %xmm0
+; X86-SSE4-NEXT: paddq %xmm3, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm2
+; X86-SSE4-NEXT: psrlq $32, %xmm2
+; X86-SSE4-NEXT: pmuludq %xmm1, %xmm2
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm3
+; X86-SSE4-NEXT: psrldq {{.*#+}} xmm3 = xmm3[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE4-NEXT: pmuludq %xmm0, %xmm3
+; X86-SSE4-NEXT: paddq %xmm2, %xmm3
+; X86-SSE4-NEXT: psllq $32, %xmm3
+; X86-SSE4-NEXT: pmuludq %xmm1, %xmm0
+; X86-SSE4-NEXT: paddq %xmm3, %xmm0
+; X86-SSE4-NEXT: movd %xmm0, %eax
+; X86-SSE4-NEXT: pextrd $1, %xmm0, %edx
+; X86-SSE4-NEXT: movl %ebp, %esp
+; X86-SSE4-NEXT: popl %ebp
+; X86-SSE4-NEXT: retl
+;
+; X86-AVX1-LABEL: test_v8i64:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
+; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3
+; X86-AVX1-NEXT: vpsrlq $32, %xmm3, %xmm4
+; X86-AVX1-NEXT: vpmuludq %xmm2, %xmm4, %xmm4
+; X86-AVX1-NEXT: vpsrlq $32, %xmm2, %xmm5
+; X86-AVX1-NEXT: vpmuludq %xmm5, %xmm3, %xmm5
+; X86-AVX1-NEXT: vpaddq %xmm4, %xmm5, %xmm4
+; X86-AVX1-NEXT: vpsllq $32, %xmm4, %xmm4
+; X86-AVX1-NEXT: vpmuludq %xmm2, %xmm3, %xmm2
+; X86-AVX1-NEXT: vpaddq %xmm4, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm3
+; X86-AVX1-NEXT: vpmuludq %xmm1, %xmm3, %xmm3
+; X86-AVX1-NEXT: vpsrlq $32, %xmm1, %xmm4
+; X86-AVX1-NEXT: vpmuludq %xmm4, %xmm0, %xmm4
+; X86-AVX1-NEXT: vpaddq %xmm3, %xmm4, %xmm3
+; X86-AVX1-NEXT: vpsllq $32, %xmm3, %xmm3
+; X86-AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm3, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm1
+; X86-AVX1-NEXT: vpmuludq %xmm2, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpsrlq $32, %xmm2, %xmm3
+; X86-AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm3, %xmm1
+; X86-AVX1-NEXT: vpsllq $32, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpmuludq %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X86-AVX1-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X86-AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX1-NEXT: vzeroupper
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v8i64:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
+; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3
+; X64-AVX1-NEXT: vpsrlq $32, %xmm3, %xmm4
+; X64-AVX1-NEXT: vpmuludq %xmm2, %xmm4, %xmm4
+; X64-AVX1-NEXT: vpsrlq $32, %xmm2, %xmm5
+; X64-AVX1-NEXT: vpmuludq %xmm5, %xmm3, %xmm5
+; X64-AVX1-NEXT: vpaddq %xmm4, %xmm5, %xmm4
+; X64-AVX1-NEXT: vpsllq $32, %xmm4, %xmm4
+; X64-AVX1-NEXT: vpmuludq %xmm2, %xmm3, %xmm2
+; X64-AVX1-NEXT: vpaddq %xmm4, %xmm2, %xmm2
+; X64-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm3
+; X64-AVX1-NEXT: vpmuludq %xmm1, %xmm3, %xmm3
+; X64-AVX1-NEXT: vpsrlq $32, %xmm1, %xmm4
+; X64-AVX1-NEXT: vpmuludq %xmm4, %xmm0, %xmm4
+; X64-AVX1-NEXT: vpaddq %xmm3, %xmm4, %xmm3
+; X64-AVX1-NEXT: vpsllq $32, %xmm3, %xmm3
+; X64-AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm3, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm1
+; X64-AVX1-NEXT: vpmuludq %xmm2, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpsrlq $32, %xmm2, %xmm3
+; X64-AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm3, %xmm1
+; X64-AVX1-NEXT: vpsllq $32, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpmuludq %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X64-AVX1-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X64-AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
+; X64-AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovq %xmm0, %rax
+; X64-AVX1-NEXT: vzeroupper
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v8i64:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpsrlq $32, %ymm0, %ymm2
+; X86-AVX2-NEXT: vpmuludq %ymm1, %ymm2, %ymm2
+; X86-AVX2-NEXT: vpsrlq $32, %ymm1, %ymm3
+; X86-AVX2-NEXT: vpmuludq %ymm3, %ymm0, %ymm3
+; X86-AVX2-NEXT: vpaddq %ymm2, %ymm3, %ymm2
+; X86-AVX2-NEXT: vpsllq $32, %ymm2, %ymm2
+; X86-AVX2-NEXT: vpmuludq %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpaddq %ymm2, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X86-AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X86-AVX2-NEXT: vpsrlq $32, %xmm1, %xmm3
+; X86-AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X86-AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X86-AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
+; X86-AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X86-AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X86-AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X86-AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
+; X86-AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX2-NEXT: vzeroupper
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v8i64:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpsrlq $32, %ymm0, %ymm2
+; X64-AVX2-NEXT: vpmuludq %ymm1, %ymm2, %ymm2
+; X64-AVX2-NEXT: vpsrlq $32, %ymm1, %ymm3
+; X64-AVX2-NEXT: vpmuludq %ymm3, %ymm0, %ymm3
+; X64-AVX2-NEXT: vpaddq %ymm2, %ymm3, %ymm2
+; X64-AVX2-NEXT: vpsllq $32, %ymm2, %ymm2
+; X64-AVX2-NEXT: vpmuludq %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpaddq %ymm2, %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X64-AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X64-AVX2-NEXT: vpsrlq $32, %xmm1, %xmm3
+; X64-AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X64-AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X64-AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
+; X64-AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X64-AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X64-AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X64-AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
+; X64-AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
+;
+; AVX512F-LABEL: test_v8i64:
+; AVX512F: # %bb.0:
+; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT: vpsrlq $32, %ymm0, %ymm2
+; AVX512F-NEXT: vpmuludq %ymm1, %ymm2, %ymm2
+; AVX512F-NEXT: vpsrlq $32, %ymm1, %ymm3
+; AVX512F-NEXT: vpmuludq %ymm3, %ymm0, %ymm3
+; AVX512F-NEXT: vpaddq %ymm2, %ymm3, %ymm2
+; AVX512F-NEXT: vpsllq $32, %ymm2, %ymm2
+; AVX512F-NEXT: vpmuludq %ymm1, %ymm0, %ymm0
+; AVX512F-NEXT: vpaddq %ymm2, %ymm0, %ymm0
+; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
+; AVX512F-NEXT: vpsrlq $32, %xmm0, %xmm2
+; AVX512F-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; AVX512F-NEXT: vpsrlq $32, %xmm1, %xmm3
+; AVX512F-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; AVX512F-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; AVX512F-NEXT: vpsllq $32, %xmm2, %xmm2
+; AVX512F-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512F-NEXT: vpsrlq $32, %xmm0, %xmm2
+; AVX512F-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; AVX512F-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX512F-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; AVX512F-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; AVX512F-NEXT: vpsllq $32, %xmm2, %xmm2
+; AVX512F-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; AVX512F-NEXT: vmovq %xmm0, %rax
+; AVX512F-NEXT: vzeroupper
+; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: test_v8i64:
; AVX512BW: # %bb.0:
@@ -450,215 +896,592 @@ define i64 @test_v8i64(<8 x i64> %a0) {
ret i64 %1
}
-define i64 @test_v16i64(<16 x i64> %a0) {
-; SSE-LABEL: test_v16i64:
-; SSE: # %bb.0:
-; SSE-NEXT: movdqa %xmm2, %xmm8
-; SSE-NEXT: psrlq $32, %xmm8
-; SSE-NEXT: pmuludq %xmm6, %xmm8
-; SSE-NEXT: movdqa %xmm6, %xmm9
-; SSE-NEXT: psrlq $32, %xmm9
-; SSE-NEXT: pmuludq %xmm2, %xmm9
-; SSE-NEXT: paddq %xmm8, %xmm9
-; SSE-NEXT: psllq $32, %xmm9
-; SSE-NEXT: pmuludq %xmm6, %xmm2
-; SSE-NEXT: paddq %xmm9, %xmm2
-; SSE-NEXT: movdqa %xmm0, %xmm6
-; SSE-NEXT: psrlq $32, %xmm6
-; SSE-NEXT: pmuludq %xmm4, %xmm6
-; SSE-NEXT: movdqa %xmm4, %xmm8
-; SSE-NEXT: psrlq $32, %xmm8
-; SSE-NEXT: pmuludq %xmm0, %xmm8
-; SSE-NEXT: paddq %xmm6, %xmm8
-; SSE-NEXT: psllq $32, %xmm8
-; SSE-NEXT: pmuludq %xmm4, %xmm0
-; SSE-NEXT: paddq %xmm8, %xmm0
-; SSE-NEXT: movdqa %xmm3, %xmm4
-; SSE-NEXT: psrlq $32, %xmm4
-; SSE-NEXT: pmuludq %xmm7, %xmm4
-; SSE-NEXT: movdqa %xmm7, %xmm6
-; SSE-NEXT: psrlq $32, %xmm6
-; SSE-NEXT: pmuludq %xmm3, %xmm6
-; SSE-NEXT: paddq %xmm4, %xmm6
-; SSE-NEXT: psllq $32, %xmm6
-; SSE-NEXT: pmuludq %xmm7, %xmm3
-; SSE-NEXT: paddq %xmm6, %xmm3
-; SSE-NEXT: movdqa %xmm1, %xmm4
-; SSE-NEXT: psrlq $32, %xmm4
-; SSE-NEXT: pmuludq %xmm5, %xmm4
-; SSE-NEXT: movdqa %xmm5, %xmm6
-; SSE-NEXT: psrlq $32, %xmm6
-; SSE-NEXT: pmuludq %xmm1, %xmm6
-; SSE-NEXT: paddq %xmm4, %xmm6
-; SSE-NEXT: psllq $32, %xmm6
-; SSE-NEXT: pmuludq %xmm5, %xmm1
-; SSE-NEXT: paddq %xmm6, %xmm1
-; SSE-NEXT: movdqa %xmm1, %xmm4
-; SSE-NEXT: psrlq $32, %xmm4
-; SSE-NEXT: pmuludq %xmm3, %xmm4
-; SSE-NEXT: movdqa %xmm3, %xmm5
-; SSE-NEXT: psrlq $32, %xmm5
-; SSE-NEXT: pmuludq %xmm1, %xmm5
-; SSE-NEXT: paddq %xmm4, %xmm5
-; SSE-NEXT: psllq $32, %xmm5
-; SSE-NEXT: pmuludq %xmm3, %xmm1
-; SSE-NEXT: paddq %xmm5, %xmm1
-; SSE-NEXT: movdqa %xmm0, %xmm3
-; SSE-NEXT: psrlq $32, %xmm3
-; SSE-NEXT: pmuludq %xmm2, %xmm3
-; SSE-NEXT: movdqa %xmm2, %xmm4
-; SSE-NEXT: psrlq $32, %xmm4
-; SSE-NEXT: pmuludq %xmm0, %xmm4
-; SSE-NEXT: paddq %xmm3, %xmm4
-; SSE-NEXT: psllq $32, %xmm4
-; SSE-NEXT: pmuludq %xmm2, %xmm0
-; SSE-NEXT: paddq %xmm4, %xmm0
-; SSE-NEXT: movdqa %xmm0, %xmm2
-; SSE-NEXT: psrlq $32, %xmm2
-; SSE-NEXT: pmuludq %xmm1, %xmm2
-; SSE-NEXT: movdqa %xmm1, %xmm3
-; SSE-NEXT: psrlq $32, %xmm3
-; SSE-NEXT: pmuludq %xmm0, %xmm3
-; SSE-NEXT: paddq %xmm2, %xmm3
-; SSE-NEXT: psllq $32, %xmm3
-; SSE-NEXT: pmuludq %xmm1, %xmm0
-; SSE-NEXT: paddq %xmm3, %xmm0
-; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE-NEXT: movdqa %xmm0, %xmm2
-; SSE-NEXT: psrlq $32, %xmm2
-; SSE-NEXT: pmuludq %xmm1, %xmm2
-; SSE-NEXT: movdqa %xmm0, %xmm3
-; SSE-NEXT: psrldq {{.*#+}} xmm3 = xmm3[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; SSE-NEXT: pmuludq %xmm0, %xmm3
-; SSE-NEXT: paddq %xmm2, %xmm3
-; SSE-NEXT: psllq $32, %xmm3
-; SSE-NEXT: pmuludq %xmm1, %xmm0
-; SSE-NEXT: paddq %xmm3, %xmm0
-; SSE-NEXT: movq %xmm0, %rax
-; SSE-NEXT: retq
-;
-; AVX1-LABEL: test_v16i64:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vpsrlq $32, %xmm1, %xmm4
-; AVX1-NEXT: vpmuludq %xmm3, %xmm4, %xmm4
-; AVX1-NEXT: vpsrlq $32, %xmm3, %xmm5
-; AVX1-NEXT: vpmuludq %xmm5, %xmm1, %xmm5
-; AVX1-NEXT: vpaddq %xmm4, %xmm5, %xmm4
-; AVX1-NEXT: vpsllq $32, %xmm4, %xmm4
-; AVX1-NEXT: vpmuludq %xmm3, %xmm1, %xmm5
-; AVX1-NEXT: vpaddq %xmm4, %xmm5, %xmm4
-; AVX1-NEXT: vpsrlq $32, %xmm0, %xmm5
-; AVX1-NEXT: vpmuludq %xmm2, %xmm5, %xmm5
-; AVX1-NEXT: vpsrlq $32, %xmm2, %xmm6
-; AVX1-NEXT: vpmuludq %xmm6, %xmm0, %xmm6
-; AVX1-NEXT: vpaddq %xmm5, %xmm6, %xmm5
-; AVX1-NEXT: vpsllq $32, %xmm5, %xmm5
-; AVX1-NEXT: vpmuludq %xmm2, %xmm0, %xmm6
-; AVX1-NEXT: vpaddq %xmm5, %xmm6, %xmm5
-; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm3
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT: vpsrlq $32, %xmm1, %xmm6
-; AVX1-NEXT: vpmuludq %xmm3, %xmm6, %xmm6
-; AVX1-NEXT: vpsrlq $32, %xmm3, %xmm7
-; AVX1-NEXT: vpmuludq %xmm7, %xmm1, %xmm7
-; AVX1-NEXT: vpaddq %xmm6, %xmm7, %xmm6
-; AVX1-NEXT: vpsllq $32, %xmm6, %xmm6
-; AVX1-NEXT: vpmuludq %xmm3, %xmm1, %xmm1
-; AVX1-NEXT: vpaddq %xmm6, %xmm1, %xmm1
-; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT: vpsrlq $32, %xmm0, %xmm3
-; AVX1-NEXT: vpmuludq %xmm2, %xmm3, %xmm3
-; AVX1-NEXT: vpsrlq $32, %xmm2, %xmm6
-; AVX1-NEXT: vpmuludq %xmm6, %xmm0, %xmm6
-; AVX1-NEXT: vpaddq %xmm3, %xmm6, %xmm3
-; AVX1-NEXT: vpsllq $32, %xmm3, %xmm3
-; AVX1-NEXT: vpmuludq %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm3, %xmm0, %xmm0
-; AVX1-NEXT: vpsrlq $32, %xmm0, %xmm2
-; AVX1-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
-; AVX1-NEXT: vpsrlq $32, %xmm1, %xmm3
-; AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
-; AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
-; AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
-; AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vpsrlq $32, %xmm5, %xmm1
-; AVX1-NEXT: vpmuludq %xmm4, %xmm1, %xmm1
-; AVX1-NEXT: vpsrlq $32, %xmm4, %xmm2
-; AVX1-NEXT: vpmuludq %xmm2, %xmm5, %xmm2
-; AVX1-NEXT: vpaddq %xmm1, %xmm2, %xmm1
-; AVX1-NEXT: vpsllq $32, %xmm1, %xmm1
-; AVX1-NEXT: vpmuludq %xmm4, %xmm5, %xmm2
-; AVX1-NEXT: vpaddq %xmm1, %xmm2, %xmm1
-; AVX1-NEXT: vpsrlq $32, %xmm1, %xmm2
-; AVX1-NEXT: vpmuludq %xmm0, %xmm2, %xmm2
-; AVX1-NEXT: vpsrlq $32, %xmm0, %xmm3
-; AVX1-NEXT: vpmuludq %xmm3, %xmm1, %xmm3
-; AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
-; AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
-; AVX1-NEXT: vpmuludq %xmm0, %xmm1, %xmm0
-; AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpsrlq $32, %xmm0, %xmm2
-; AVX1-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
-; AVX1-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
-; AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
-; AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
-; AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vmovq %xmm0, %rax
-; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
+define i64 @test_v16i64(<16 x i64> %a0) nounwind {
+; X86-SSE2-LABEL: test_v16i64:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: pushl %ebp
+; X86-SSE2-NEXT: movl %esp, %ebp
+; X86-SSE2-NEXT: andl $-16, %esp
+; X86-SSE2-NEXT: subl $16, %esp
+; X86-SSE2-NEXT: movdqa 72(%ebp), %xmm4
+; X86-SSE2-NEXT: movdqa 56(%ebp), %xmm3
+; X86-SSE2-NEXT: movdqa %xmm2, %xmm5
+; X86-SSE2-NEXT: psrlq $32, %xmm5
+; X86-SSE2-NEXT: pmuludq %xmm3, %xmm5
+; X86-SSE2-NEXT: movdqa %xmm3, %xmm6
+; X86-SSE2-NEXT: psrlq $32, %xmm6
+; X86-SSE2-NEXT: pmuludq %xmm2, %xmm6
+; X86-SSE2-NEXT: paddq %xmm5, %xmm6
+; X86-SSE2-NEXT: movdqa 24(%ebp), %xmm5
+; X86-SSE2-NEXT: psllq $32, %xmm6
+; X86-SSE2-NEXT: pmuludq %xmm3, %xmm2
+; X86-SSE2-NEXT: paddq %xmm6, %xmm2
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm3
+; X86-SSE2-NEXT: psrlq $32, %xmm3
+; X86-SSE2-NEXT: pmuludq %xmm5, %xmm3
+; X86-SSE2-NEXT: movdqa %xmm5, %xmm6
+; X86-SSE2-NEXT: psrlq $32, %xmm6
+; X86-SSE2-NEXT: pmuludq %xmm0, %xmm6
+; X86-SSE2-NEXT: paddq %xmm3, %xmm6
+; X86-SSE2-NEXT: movdqa 8(%ebp), %xmm3
+; X86-SSE2-NEXT: psllq $32, %xmm6
+; X86-SSE2-NEXT: pmuludq %xmm5, %xmm0
+; X86-SSE2-NEXT: paddq %xmm6, %xmm0
+; X86-SSE2-NEXT: movdqa %xmm3, %xmm5
+; X86-SSE2-NEXT: psrlq $32, %xmm5
+; X86-SSE2-NEXT: pmuludq %xmm4, %xmm5
+; X86-SSE2-NEXT: movdqa %xmm4, %xmm6
+; X86-SSE2-NEXT: psrlq $32, %xmm6
+; X86-SSE2-NEXT: pmuludq %xmm3, %xmm6
+; X86-SSE2-NEXT: paddq %xmm5, %xmm6
+; X86-SSE2-NEXT: movdqa 40(%ebp), %xmm5
+; X86-SSE2-NEXT: psllq $32, %xmm6
+; X86-SSE2-NEXT: pmuludq %xmm4, %xmm3
+; X86-SSE2-NEXT: paddq %xmm6, %xmm3
+; X86-SSE2-NEXT: movdqa %xmm1, %xmm4
+; X86-SSE2-NEXT: psrlq $32, %xmm4
+; X86-SSE2-NEXT: pmuludq %xmm5, %xmm4
+; X86-SSE2-NEXT: movdqa %xmm5, %xmm6
+; X86-SSE2-NEXT: psrlq $32, %xmm6
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm6
+; X86-SSE2-NEXT: paddq %xmm4, %xmm6
+; X86-SSE2-NEXT: psllq $32, %xmm6
+; X86-SSE2-NEXT: pmuludq %xmm5, %xmm1
+; X86-SSE2-NEXT: paddq %xmm6, %xmm1
+; X86-SSE2-NEXT: movdqa %xmm1, %xmm4
+; X86-SSE2-NEXT: psrlq $32, %xmm4
+; X86-SSE2-NEXT: pmuludq %xmm3, %xmm4
+; X86-SSE2-NEXT: movdqa %xmm3, %xmm5
+; X86-SSE2-NEXT: psrlq $32, %xmm5
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm5
+; X86-SSE2-NEXT: paddq %xmm4, %xmm5
+; X86-SSE2-NEXT: psllq $32, %xmm5
+; X86-SSE2-NEXT: pmuludq %xmm3, %xmm1
+; X86-SSE2-NEXT: paddq %xmm5, %xmm1
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm3
+; X86-SSE2-NEXT: psrlq $32, %xmm3
+; X86-SSE2-NEXT: pmuludq %xmm2, %xmm3
+; X86-SSE2-NEXT: movdqa %xmm2, %xmm4
+; X86-SSE2-NEXT: psrlq $32, %xmm4
+; X86-SSE2-NEXT: pmuludq %xmm0, %xmm4
+; X86-SSE2-NEXT: paddq %xmm3, %xmm4
+; X86-SSE2-NEXT: psllq $32, %xmm4
+; X86-SSE2-NEXT: pmuludq %xmm2, %xmm0
+; X86-SSE2-NEXT: paddq %xmm4, %xmm0
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm2
+; X86-SSE2-NEXT: psrlq $32, %xmm2
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm2
+; X86-SSE2-NEXT: movdqa %xmm1, %xmm3
+; X86-SSE2-NEXT: psrlq $32, %xmm3
+; X86-SSE2-NEXT: pmuludq %xmm0, %xmm3
+; X86-SSE2-NEXT: paddq %xmm2, %xmm3
+; X86-SSE2-NEXT: psllq $32, %xmm3
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm0
+; X86-SSE2-NEXT: paddq %xmm3, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm2
+; X86-SSE2-NEXT: psrlq $32, %xmm2
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm2
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm3
+; X86-SSE2-NEXT: psrldq {{.*#+}} xmm3 = xmm3[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE2-NEXT: pmuludq %xmm0, %xmm3
+; X86-SSE2-NEXT: paddq %xmm2, %xmm3
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm0
+; X86-SSE2-NEXT: psllq $32, %xmm3
+; X86-SSE2-NEXT: paddq %xmm3, %xmm0
+; X86-SSE2-NEXT: movd %xmm0, %eax
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; X86-SSE2-NEXT: movd %xmm0, %edx
+; X86-SSE2-NEXT: movl %ebp, %esp
+; X86-SSE2-NEXT: popl %ebp
+; X86-SSE2-NEXT: retl
;
-; AVX2-LABEL: test_v16i64:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpsrlq $32, %ymm1, %ymm4
-; AVX2-NEXT: vpmuludq %ymm3, %ymm4, %ymm4
-; AVX2-NEXT: vpsrlq $32, %ymm3, %ymm5
-; AVX2-NEXT: vpmuludq %ymm5, %ymm1, %ymm5
-; AVX2-NEXT: vpaddq %ymm4, %ymm5, %ymm4
-; AVX2-NEXT: vpsllq $32, %ymm4, %ymm4
-; AVX2-NEXT: vpmuludq %ymm3, %ymm1, %ymm1
-; AVX2-NEXT: vpaddq %ymm4, %ymm1, %ymm1
-; AVX2-NEXT: vpsrlq $32, %ymm0, %ymm3
-; AVX2-NEXT: vpmuludq %ymm2, %ymm3, %ymm3
-; AVX2-NEXT: vpsrlq $32, %ymm2, %ymm4
-; AVX2-NEXT: vpmuludq %ymm4, %ymm0, %ymm4
-; AVX2-NEXT: vpaddq %ymm3, %ymm4, %ymm3
-; AVX2-NEXT: vpsllq $32, %ymm3, %ymm3
-; AVX2-NEXT: vpmuludq %ymm2, %ymm0, %ymm0
-; AVX2-NEXT: vpaddq %ymm3, %ymm0, %ymm0
-; AVX2-NEXT: vpsrlq $32, %ymm0, %ymm2
-; AVX2-NEXT: vpmuludq %ymm1, %ymm2, %ymm2
-; AVX2-NEXT: vpsrlq $32, %ymm1, %ymm3
-; AVX2-NEXT: vpmuludq %ymm3, %ymm0, %ymm3
-; AVX2-NEXT: vpaddq %ymm2, %ymm3, %ymm2
-; AVX2-NEXT: vpsllq $32, %ymm2, %ymm2
-; AVX2-NEXT: vpmuludq %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vpaddq %ymm2, %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
-; AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
-; AVX2-NEXT: vpsrlq $32, %xmm1, %xmm3
-; AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
-; AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
-; AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
-; AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
-; AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
-; AVX2-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
-; AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
-; AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
-; AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
-; AVX2-NEXT: vmovq %xmm0, %rax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+; X64-SSE-LABEL: test_v16i64:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: movdqa %xmm2, %xmm8
+; X64-SSE-NEXT: psrlq $32, %xmm8
+; X64-SSE-NEXT: pmuludq %xmm6, %xmm8
+; X64-SSE-NEXT: movdqa %xmm6, %xmm9
+; X64-SSE-NEXT: psrlq $32, %xmm9
+; X64-SSE-NEXT: pmuludq %xmm2, %xmm9
+; X64-SSE-NEXT: paddq %xmm8, %xmm9
+; X64-SSE-NEXT: psllq $32, %xmm9
+; X64-SSE-NEXT: pmuludq %xmm6, %xmm2
+; X64-SSE-NEXT: paddq %xmm9, %xmm2
+; X64-SSE-NEXT: movdqa %xmm0, %xmm6
+; X64-SSE-NEXT: psrlq $32, %xmm6
+; X64-SSE-NEXT: pmuludq %xmm4, %xmm6
+; X64-SSE-NEXT: movdqa %xmm4, %xmm8
+; X64-SSE-NEXT: psrlq $32, %xmm8
+; X64-SSE-NEXT: pmuludq %xmm0, %xmm8
+; X64-SSE-NEXT: paddq %xmm6, %xmm8
+; X64-SSE-NEXT: psllq $32, %xmm8
+; X64-SSE-NEXT: pmuludq %xmm4, %xmm0
+; X64-SSE-NEXT: paddq %xmm8, %xmm0
+; X64-SSE-NEXT: movdqa %xmm3, %xmm4
+; X64-SSE-NEXT: psrlq $32, %xmm4
+; X64-SSE-NEXT: pmuludq %xmm7, %xmm4
+; X64-SSE-NEXT: movdqa %xmm7, %xmm6
+; X64-SSE-NEXT: psrlq $32, %xmm6
+; X64-SSE-NEXT: pmuludq %xmm3, %xmm6
+; X64-SSE-NEXT: paddq %xmm4, %xmm6
+; X64-SSE-NEXT: psllq $32, %xmm6
+; X64-SSE-NEXT: pmuludq %xmm7, %xmm3
+; X64-SSE-NEXT: paddq %xmm6, %xmm3
+; X64-SSE-NEXT: movdqa %xmm1, %xmm4
+; X64-SSE-NEXT: psrlq $32, %xmm4
+; X64-SSE-NEXT: pmuludq %xmm5, %xmm4
+; X64-SSE-NEXT: movdqa %xmm5, %xmm6
+; X64-SSE-NEXT: psrlq $32, %xmm6
+; X64-SSE-NEXT: pmuludq %xmm1, %xmm6
+; X64-SSE-NEXT: paddq %xmm4, %xmm6
+; X64-SSE-NEXT: psllq $32, %xmm6
+; X64-SSE-NEXT: pmuludq %xmm5, %xmm1
+; X64-SSE-NEXT: paddq %xmm6, %xmm1
+; X64-SSE-NEXT: movdqa %xmm1, %xmm4
+; X64-SSE-NEXT: psrlq $32, %xmm4
+; X64-SSE-NEXT: pmuludq %xmm3, %xmm4
+; X64-SSE-NEXT: movdqa %xmm3, %xmm5
+; X64-SSE-NEXT: psrlq $32, %xmm5
+; X64-SSE-NEXT: pmuludq %xmm1, %xmm5
+; X64-SSE-NEXT: paddq %xmm4, %xmm5
+; X64-SSE-NEXT: psllq $32, %xmm5
+; X64-SSE-NEXT: pmuludq %xmm3, %xmm1
+; X64-SSE-NEXT: paddq %xmm5, %xmm1
+; X64-SSE-NEXT: movdqa %xmm0, %xmm3
+; X64-SSE-NEXT: psrlq $32, %xmm3
+; X64-SSE-NEXT: pmuludq %xmm2, %xmm3
+; X64-SSE-NEXT: movdqa %xmm2, %xmm4
+; X64-SSE-NEXT: psrlq $32, %xmm4
+; X64-SSE-NEXT: pmuludq %xmm0, %xmm4
+; X64-SSE-NEXT: paddq %xmm3, %xmm4
+; X64-SSE-NEXT: psllq $32, %xmm4
+; X64-SSE-NEXT: pmuludq %xmm2, %xmm0
+; X64-SSE-NEXT: paddq %xmm4, %xmm0
+; X64-SSE-NEXT: movdqa %xmm0, %xmm2
+; X64-SSE-NEXT: psrlq $32, %xmm2
+; X64-SSE-NEXT: pmuludq %xmm1, %xmm2
+; X64-SSE-NEXT: movdqa %xmm1, %xmm3
+; X64-SSE-NEXT: psrlq $32, %xmm3
+; X64-SSE-NEXT: pmuludq %xmm0, %xmm3
+; X64-SSE-NEXT: paddq %xmm2, %xmm3
+; X64-SSE-NEXT: psllq $32, %xmm3
+; X64-SSE-NEXT: pmuludq %xmm1, %xmm0
+; X64-SSE-NEXT: paddq %xmm3, %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE-NEXT: movdqa %xmm0, %xmm2
+; X64-SSE-NEXT: psrlq $32, %xmm2
+; X64-SSE-NEXT: pmuludq %xmm1, %xmm2
+; X64-SSE-NEXT: movdqa %xmm0, %xmm3
+; X64-SSE-NEXT: psrldq {{.*#+}} xmm3 = xmm3[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE-NEXT: pmuludq %xmm0, %xmm3
+; X64-SSE-NEXT: paddq %xmm2, %xmm3
+; X64-SSE-NEXT: psllq $32, %xmm3
+; X64-SSE-NEXT: pmuludq %xmm1, %xmm0
+; X64-SSE-NEXT: paddq %xmm3, %xmm0
+; X64-SSE-NEXT: movq %xmm0, %rax
+; X64-SSE-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v16i64:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pushl %ebp
+; X86-SSE4-NEXT: movl %esp, %ebp
+; X86-SSE4-NEXT: andl $-16, %esp
+; X86-SSE4-NEXT: subl $16, %esp
+; X86-SSE4-NEXT: movdqa 72(%ebp), %xmm4
+; X86-SSE4-NEXT: movdqa 56(%ebp), %xmm3
+; X86-SSE4-NEXT: movdqa %xmm2, %xmm5
+; X86-SSE4-NEXT: psrlq $32, %xmm5
+; X86-SSE4-NEXT: pmuludq %xmm3, %xmm5
+; X86-SSE4-NEXT: movdqa %xmm3, %xmm6
+; X86-SSE4-NEXT: psrlq $32, %xmm6
+; X86-SSE4-NEXT: pmuludq %xmm2, %xmm6
+; X86-SSE4-NEXT: paddq %xmm5, %xmm6
+; X86-SSE4-NEXT: movdqa 24(%ebp), %xmm5
+; X86-SSE4-NEXT: psllq $32, %xmm6
+; X86-SSE4-NEXT: pmuludq %xmm3, %xmm2
+; X86-SSE4-NEXT: paddq %xmm6, %xmm2
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm3
+; X86-SSE4-NEXT: psrlq $32, %xmm3
+; X86-SSE4-NEXT: pmuludq %xmm5, %xmm3
+; X86-SSE4-NEXT: movdqa %xmm5, %xmm6
+; X86-SSE4-NEXT: psrlq $32, %xmm6
+; X86-SSE4-NEXT: pmuludq %xmm0, %xmm6
+; X86-SSE4-NEXT: paddq %xmm3, %xmm6
+; X86-SSE4-NEXT: movdqa 8(%ebp), %xmm3
+; X86-SSE4-NEXT: psllq $32, %xmm6
+; X86-SSE4-NEXT: pmuludq %xmm5, %xmm0
+; X86-SSE4-NEXT: paddq %xmm6, %xmm0
+; X86-SSE4-NEXT: movdqa %xmm3, %xmm5
+; X86-SSE4-NEXT: psrlq $32, %xmm5
+; X86-SSE4-NEXT: pmuludq %xmm4, %xmm5
+; X86-SSE4-NEXT: movdqa %xmm4, %xmm6
+; X86-SSE4-NEXT: psrlq $32, %xmm6
+; X86-SSE4-NEXT: pmuludq %xmm3, %xmm6
+; X86-SSE4-NEXT: paddq %xmm5, %xmm6
+; X86-SSE4-NEXT: movdqa 40(%ebp), %xmm5
+; X86-SSE4-NEXT: psllq $32, %xmm6
+; X86-SSE4-NEXT: pmuludq %xmm4, %xmm3
+; X86-SSE4-NEXT: paddq %xmm6, %xmm3
+; X86-SSE4-NEXT: movdqa %xmm1, %xmm4
+; X86-SSE4-NEXT: psrlq $32, %xmm4
+; X86-SSE4-NEXT: pmuludq %xmm5, %xmm4
+; X86-SSE4-NEXT: movdqa %xmm5, %xmm6
+; X86-SSE4-NEXT: psrlq $32, %xmm6
+; X86-SSE4-NEXT: pmuludq %xmm1, %xmm6
+; X86-SSE4-NEXT: paddq %xmm4, %xmm6
+; X86-SSE4-NEXT: psllq $32, %xmm6
+; X86-SSE4-NEXT: pmuludq %xmm5, %xmm1
+; X86-SSE4-NEXT: paddq %xmm6, %xmm1
+; X86-SSE4-NEXT: movdqa %xmm1, %xmm4
+; X86-SSE4-NEXT: psrlq $32, %xmm4
+; X86-SSE4-NEXT: pmuludq %xmm3, %xmm4
+; X86-SSE4-NEXT: movdqa %xmm3, %xmm5
+; X86-SSE4-NEXT: psrlq $32, %xmm5
+; X86-SSE4-NEXT: pmuludq %xmm1, %xmm5
+; X86-SSE4-NEXT: paddq %xmm4, %xmm5
+; X86-SSE4-NEXT: psllq $32, %xmm5
+; X86-SSE4-NEXT: pmuludq %xmm3, %xmm1
+; X86-SSE4-NEXT: paddq %xmm5, %xmm1
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm3
+; X86-SSE4-NEXT: psrlq $32, %xmm3
+; X86-SSE4-NEXT: pmuludq %xmm2, %xmm3
+; X86-SSE4-NEXT: movdqa %xmm2, %xmm4
+; X86-SSE4-NEXT: psrlq $32, %xmm4
+; X86-SSE4-NEXT: pmuludq %xmm0, %xmm4
+; X86-SSE4-NEXT: paddq %xmm3, %xmm4
+; X86-SSE4-NEXT: psllq $32, %xmm4
+; X86-SSE4-NEXT: pmuludq %xmm2, %xmm0
+; X86-SSE4-NEXT: paddq %xmm4, %xmm0
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm2
+; X86-SSE4-NEXT: psrlq $32, %xmm2
+; X86-SSE4-NEXT: pmuludq %xmm1, %xmm2
+; X86-SSE4-NEXT: movdqa %xmm1, %xmm3
+; X86-SSE4-NEXT: psrlq $32, %xmm3
+; X86-SSE4-NEXT: pmuludq %xmm0, %xmm3
+; X86-SSE4-NEXT: paddq %xmm2, %xmm3
+; X86-SSE4-NEXT: psllq $32, %xmm3
+; X86-SSE4-NEXT: pmuludq %xmm1, %xmm0
+; X86-SSE4-NEXT: paddq %xmm3, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm2
+; X86-SSE4-NEXT: psrlq $32, %xmm2
+; X86-SSE4-NEXT: pmuludq %xmm1, %xmm2
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm3
+; X86-SSE4-NEXT: psrldq {{.*#+}} xmm3 = xmm3[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE4-NEXT: pmuludq %xmm0, %xmm3
+; X86-SSE4-NEXT: paddq %xmm2, %xmm3
+; X86-SSE4-NEXT: pmuludq %xmm1, %xmm0
+; X86-SSE4-NEXT: psllq $32, %xmm3
+; X86-SSE4-NEXT: paddq %xmm3, %xmm0
+; X86-SSE4-NEXT: movd %xmm0, %eax
+; X86-SSE4-NEXT: pextrd $1, %xmm0, %edx
+; X86-SSE4-NEXT: movl %ebp, %esp
+; X86-SSE4-NEXT: popl %ebp
+; X86-SSE4-NEXT: retl
+;
+; X86-AVX1-LABEL: test_v16i64:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: pushl %ebp
+; X86-AVX1-NEXT: movl %esp, %ebp
+; X86-AVX1-NEXT: andl $-32, %esp
+; X86-AVX1-NEXT: subl $32, %esp
+; X86-AVX1-NEXT: vmovdqa 8(%ebp), %xmm3
+; X86-AVX1-NEXT: vmovdqa 24(%ebp), %xmm5
+; X86-AVX1-NEXT: vpsrlq $32, %xmm3, %xmm4
+; X86-AVX1-NEXT: vpmuludq %xmm4, %xmm1, %xmm4
+; X86-AVX1-NEXT: vpsrlq $32, %xmm1, %xmm6
+; X86-AVX1-NEXT: vpmuludq %xmm3, %xmm6, %xmm6
+; X86-AVX1-NEXT: vpaddq %xmm6, %xmm4, %xmm4
+; X86-AVX1-NEXT: vpsllq $32, %xmm4, %xmm4
+; X86-AVX1-NEXT: vpmuludq %xmm3, %xmm1, %xmm3
+; X86-AVX1-NEXT: vpaddq %xmm4, %xmm3, %xmm3
+; X86-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm4
+; X86-AVX1-NEXT: vpmuludq %xmm2, %xmm4, %xmm4
+; X86-AVX1-NEXT: vpsrlq $32, %xmm2, %xmm6
+; X86-AVX1-NEXT: vpmuludq %xmm6, %xmm0, %xmm6
+; X86-AVX1-NEXT: vpaddq %xmm4, %xmm6, %xmm4
+; X86-AVX1-NEXT: vpsllq $32, %xmm4, %xmm4
+; X86-AVX1-NEXT: vpmuludq %xmm2, %xmm0, %xmm6
+; X86-AVX1-NEXT: vpaddq %xmm4, %xmm6, %xmm4
+; X86-AVX1-NEXT: vpsrlq $32, %xmm5, %xmm6
+; X86-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X86-AVX1-NEXT: vpmuludq %xmm6, %xmm1, %xmm6
+; X86-AVX1-NEXT: vpsrlq $32, %xmm1, %xmm7
+; X86-AVX1-NEXT: vpmuludq %xmm5, %xmm7, %xmm7
+; X86-AVX1-NEXT: vpaddq %xmm7, %xmm6, %xmm6
+; X86-AVX1-NEXT: vpsllq $32, %xmm6, %xmm6
+; X86-AVX1-NEXT: vpmuludq %xmm5, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpaddq %xmm6, %xmm1, %xmm1
+; X86-AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X86-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpmuludq %xmm2, %xmm5, %xmm5
+; X86-AVX1-NEXT: vpsrlq $32, %xmm2, %xmm6
+; X86-AVX1-NEXT: vpmuludq %xmm6, %xmm0, %xmm6
+; X86-AVX1-NEXT: vpaddq %xmm5, %xmm6, %xmm5
+; X86-AVX1-NEXT: vpsllq $32, %xmm5, %xmm5
+; X86-AVX1-NEXT: vpmuludq %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm5, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X86-AVX1-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpsrlq $32, %xmm1, %xmm5
+; X86-AVX1-NEXT: vpmuludq %xmm5, %xmm0, %xmm5
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm5, %xmm2
+; X86-AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpsrlq $32, %xmm4, %xmm1
+; X86-AVX1-NEXT: vpmuludq %xmm3, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpsrlq $32, %xmm3, %xmm2
+; X86-AVX1-NEXT: vpmuludq %xmm2, %xmm4, %xmm2
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm2, %xmm1
+; X86-AVX1-NEXT: vpsllq $32, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpmuludq %xmm3, %xmm4, %xmm2
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm2, %xmm1
+; X86-AVX1-NEXT: vpsrlq $32, %xmm1, %xmm2
+; X86-AVX1-NEXT: vpmuludq %xmm0, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm3
+; X86-AVX1-NEXT: vpmuludq %xmm3, %xmm1, %xmm3
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X86-AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpmuludq %xmm0, %xmm1, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X86-AVX1-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X86-AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX1-NEXT: movl %ebp, %esp
+; X86-AVX1-NEXT: popl %ebp
+; X86-AVX1-NEXT: vzeroupper
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v16i64:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vpsrlq $32, %xmm1, %xmm4
+; X64-AVX1-NEXT: vpmuludq %xmm3, %xmm4, %xmm4
+; X64-AVX1-NEXT: vpsrlq $32, %xmm3, %xmm5
+; X64-AVX1-NEXT: vpmuludq %xmm5, %xmm1, %xmm5
+; X64-AVX1-NEXT: vpaddq %xmm4, %xmm5, %xmm4
+; X64-AVX1-NEXT: vpsllq $32, %xmm4, %xmm4
+; X64-AVX1-NEXT: vpmuludq %xmm3, %xmm1, %xmm5
+; X64-AVX1-NEXT: vpaddq %xmm4, %xmm5, %xmm4
+; X64-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm5
+; X64-AVX1-NEXT: vpmuludq %xmm2, %xmm5, %xmm5
+; X64-AVX1-NEXT: vpsrlq $32, %xmm2, %xmm6
+; X64-AVX1-NEXT: vpmuludq %xmm6, %xmm0, %xmm6
+; X64-AVX1-NEXT: vpaddq %xmm5, %xmm6, %xmm5
+; X64-AVX1-NEXT: vpsllq $32, %xmm5, %xmm5
+; X64-AVX1-NEXT: vpmuludq %xmm2, %xmm0, %xmm6
+; X64-AVX1-NEXT: vpaddq %xmm5, %xmm6, %xmm5
+; X64-AVX1-NEXT: vextractf128 $1, %ymm3, %xmm3
+; X64-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X64-AVX1-NEXT: vpsrlq $32, %xmm1, %xmm6
+; X64-AVX1-NEXT: vpmuludq %xmm3, %xmm6, %xmm6
+; X64-AVX1-NEXT: vpsrlq $32, %xmm3, %xmm7
+; X64-AVX1-NEXT: vpmuludq %xmm7, %xmm1, %xmm7
+; X64-AVX1-NEXT: vpaddq %xmm6, %xmm7, %xmm6
+; X64-AVX1-NEXT: vpsllq $32, %xmm6, %xmm6
+; X64-AVX1-NEXT: vpmuludq %xmm3, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpaddq %xmm6, %xmm1, %xmm1
+; X64-AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X64-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm3
+; X64-AVX1-NEXT: vpmuludq %xmm2, %xmm3, %xmm3
+; X64-AVX1-NEXT: vpsrlq $32, %xmm2, %xmm6
+; X64-AVX1-NEXT: vpmuludq %xmm6, %xmm0, %xmm6
+; X64-AVX1-NEXT: vpaddq %xmm3, %xmm6, %xmm3
+; X64-AVX1-NEXT: vpsllq $32, %xmm3, %xmm3
+; X64-AVX1-NEXT: vpmuludq %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm3, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X64-AVX1-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X64-AVX1-NEXT: vpsrlq $32, %xmm1, %xmm3
+; X64-AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X64-AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
+; X64-AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpsrlq $32, %xmm5, %xmm1
+; X64-AVX1-NEXT: vpmuludq %xmm4, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpsrlq $32, %xmm4, %xmm2
+; X64-AVX1-NEXT: vpmuludq %xmm2, %xmm5, %xmm2
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm2, %xmm1
+; X64-AVX1-NEXT: vpsllq $32, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpmuludq %xmm4, %xmm5, %xmm2
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm2, %xmm1
+; X64-AVX1-NEXT: vpsrlq $32, %xmm1, %xmm2
+; X64-AVX1-NEXT: vpmuludq %xmm0, %xmm2, %xmm2
+; X64-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm3
+; X64-AVX1-NEXT: vpmuludq %xmm3, %xmm1, %xmm3
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X64-AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
+; X64-AVX1-NEXT: vpmuludq %xmm0, %xmm1, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X64-AVX1-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X64-AVX1-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX1-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X64-AVX1-NEXT: vpsllq $32, %xmm2, %xmm2
+; X64-AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovq %xmm0, %rax
+; X64-AVX1-NEXT: vzeroupper
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v16i64:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: pushl %ebp
+; X86-AVX2-NEXT: movl %esp, %ebp
+; X86-AVX2-NEXT: andl $-32, %esp
+; X86-AVX2-NEXT: subl $32, %esp
+; X86-AVX2-NEXT: vmovdqa 8(%ebp), %ymm3
+; X86-AVX2-NEXT: vpsrlq $32, %ymm1, %ymm4
+; X86-AVX2-NEXT: vpmuludq %ymm3, %ymm4, %ymm4
+; X86-AVX2-NEXT: vpsrlq $32, %ymm3, %ymm5
+; X86-AVX2-NEXT: vpmuludq %ymm5, %ymm1, %ymm5
+; X86-AVX2-NEXT: vpaddq %ymm4, %ymm5, %ymm4
+; X86-AVX2-NEXT: vpsllq $32, %ymm4, %ymm4
+; X86-AVX2-NEXT: vpmuludq %ymm3, %ymm1, %ymm1
+; X86-AVX2-NEXT: vpaddq %ymm4, %ymm1, %ymm1
+; X86-AVX2-NEXT: vpsrlq $32, %ymm0, %ymm3
+; X86-AVX2-NEXT: vpmuludq %ymm2, %ymm3, %ymm3
+; X86-AVX2-NEXT: vpsrlq $32, %ymm2, %ymm4
+; X86-AVX2-NEXT: vpmuludq %ymm4, %ymm0, %ymm4
+; X86-AVX2-NEXT: vpaddq %ymm3, %ymm4, %ymm3
+; X86-AVX2-NEXT: vpsllq $32, %ymm3, %ymm3
+; X86-AVX2-NEXT: vpmuludq %ymm2, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpaddq %ymm3, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpsrlq $32, %ymm0, %ymm2
+; X86-AVX2-NEXT: vpmuludq %ymm1, %ymm2, %ymm2
+; X86-AVX2-NEXT: vpsrlq $32, %ymm1, %ymm3
+; X86-AVX2-NEXT: vpmuludq %ymm3, %ymm0, %ymm3
+; X86-AVX2-NEXT: vpaddq %ymm2, %ymm3, %ymm2
+; X86-AVX2-NEXT: vpsllq $32, %ymm2, %ymm2
+; X86-AVX2-NEXT: vpmuludq %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpaddq %ymm2, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X86-AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X86-AVX2-NEXT: vpsrlq $32, %xmm1, %xmm3
+; X86-AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X86-AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X86-AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
+; X86-AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X86-AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X86-AVX2-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X86-AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X86-AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
+; X86-AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX2-NEXT: movl %ebp, %esp
+; X86-AVX2-NEXT: popl %ebp
+; X86-AVX2-NEXT: vzeroupper
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v16i64:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpsrlq $32, %ymm1, %ymm4
+; X64-AVX2-NEXT: vpmuludq %ymm3, %ymm4, %ymm4
+; X64-AVX2-NEXT: vpsrlq $32, %ymm3, %ymm5
+; X64-AVX2-NEXT: vpmuludq %ymm5, %ymm1, %ymm5
+; X64-AVX2-NEXT: vpaddq %ymm4, %ymm5, %ymm4
+; X64-AVX2-NEXT: vpsllq $32, %ymm4, %ymm4
+; X64-AVX2-NEXT: vpmuludq %ymm3, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpaddq %ymm4, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpsrlq $32, %ymm0, %ymm3
+; X64-AVX2-NEXT: vpmuludq %ymm2, %ymm3, %ymm3
+; X64-AVX2-NEXT: vpsrlq $32, %ymm2, %ymm4
+; X64-AVX2-NEXT: vpmuludq %ymm4, %ymm0, %ymm4
+; X64-AVX2-NEXT: vpaddq %ymm3, %ymm4, %ymm3
+; X64-AVX2-NEXT: vpsllq $32, %ymm3, %ymm3
+; X64-AVX2-NEXT: vpmuludq %ymm2, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpaddq %ymm3, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpsrlq $32, %ymm0, %ymm2
+; X64-AVX2-NEXT: vpmuludq %ymm1, %ymm2, %ymm2
+; X64-AVX2-NEXT: vpsrlq $32, %ymm1, %ymm3
+; X64-AVX2-NEXT: vpmuludq %ymm3, %ymm0, %ymm3
+; X64-AVX2-NEXT: vpaddq %ymm2, %ymm3, %ymm2
+; X64-AVX2-NEXT: vpsllq $32, %ymm2, %ymm2
+; X64-AVX2-NEXT: vpmuludq %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpaddq %ymm2, %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X64-AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X64-AVX2-NEXT: vpsrlq $32, %xmm1, %xmm3
+; X64-AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X64-AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X64-AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
+; X64-AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpsrlq $32, %xmm0, %xmm2
+; X64-AVX2-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; X64-AVX2-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-AVX2-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; X64-AVX2-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X64-AVX2-NEXT: vpsllq $32, %xmm2, %xmm2
+; X64-AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
+;
+; AVX512F-LABEL: test_v16i64:
+; AVX512F: # %bb.0:
+; AVX512F-NEXT: vpsrlq $32, %zmm0, %zmm2
+; AVX512F-NEXT: vpmuludq %zmm1, %zmm2, %zmm2
+; AVX512F-NEXT: vpsrlq $32, %zmm1, %zmm3
+; AVX512F-NEXT: vpmuludq %zmm3, %zmm0, %zmm3
+; AVX512F-NEXT: vpaddq %zmm2, %zmm3, %zmm2
+; AVX512F-NEXT: vpsllq $32, %zmm2, %zmm2
+; AVX512F-NEXT: vpmuludq %zmm1, %zmm0, %zmm0
+; AVX512F-NEXT: vpaddq %zmm2, %zmm0, %zmm0
+; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT: vpsrlq $32, %ymm0, %ymm2
+; AVX512F-NEXT: vpmuludq %ymm1, %ymm2, %ymm2
+; AVX512F-NEXT: vpsrlq $32, %ymm1, %ymm3
+; AVX512F-NEXT: vpmuludq %ymm3, %ymm0, %ymm3
+; AVX512F-NEXT: vpaddq %ymm2, %ymm3, %ymm2
+; AVX512F-NEXT: vpsllq $32, %ymm2, %ymm2
+; AVX512F-NEXT: vpmuludq %ymm1, %ymm0, %ymm0
+; AVX512F-NEXT: vpaddq %ymm2, %ymm0, %ymm0
+; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
+; AVX512F-NEXT: vpsrlq $32, %xmm0, %xmm2
+; AVX512F-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; AVX512F-NEXT: vpsrlq $32, %xmm1, %xmm3
+; AVX512F-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; AVX512F-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; AVX512F-NEXT: vpsllq $32, %xmm2, %xmm2
+; AVX512F-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512F-NEXT: vpsrlq $32, %xmm0, %xmm2
+; AVX512F-NEXT: vpmuludq %xmm1, %xmm2, %xmm2
+; AVX512F-NEXT: vpsrldq {{.*#+}} xmm3 = xmm0[12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; AVX512F-NEXT: vpmuludq %xmm3, %xmm0, %xmm3
+; AVX512F-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; AVX512F-NEXT: vpsllq $32, %xmm2, %xmm2
+; AVX512F-NEXT: vpmuludq %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; AVX512F-NEXT: vmovq %xmm0, %rax
+; AVX512F-NEXT: vzeroupper
+; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: test_v16i64:
; AVX512BW: # %bb.0:
@@ -775,32 +1598,32 @@ define i64 @test_v16i64(<16 x i64> %a0) {
; vXi32
;
-define i32 @test_v2i32(<2 x i32> %a0) {
+define i32 @test_v2i32(<2 x i32> %a0) nounwind {
; SSE2-LABEL: test_v2i32:
; SSE2: # %bb.0:
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; SSE2-NEXT: pmuludq %xmm0, %xmm1
; SSE2-NEXT: movd %xmm1, %eax
-; SSE2-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
;
-; SSE41-LABEL: test_v2i32:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE41-NEXT: pmulld %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: retq
+; SSE4-LABEL: test_v2i32:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; SSE4-NEXT: pmulld %xmm0, %xmm1
+; SSE4-NEXT: movd %xmm1, %eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX-LABEL: test_v2i32:
; AVX: # %bb.0:
; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; AVX-NEXT: vpmulld %xmm1, %xmm0, %xmm0
; AVX-NEXT: vmovd %xmm0, %eax
-; AVX-NEXT: retq
+; AVX-NEXT: ret{{[l|q]}}
%1 = call i32 @llvm.vector.reduce.mul.v2i32(<2 x i32> %a0)
ret i32 %1
}
-define i32 @test_v4i32(<4 x i32> %a0) {
+define i32 @test_v4i32(<4 x i32> %a0) nounwind {
; SSE2-LABEL: test_v4i32:
; SSE2: # %bb.0:
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
@@ -812,16 +1635,16 @@ define i32 @test_v4i32(<4 x i32> %a0) {
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm3[0,0,0,0]
; SSE2-NEXT: pmuludq %xmm1, %xmm0
; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
;
-; SSE41-LABEL: test_v4i32:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: pmulld %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: pmulld %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: retq
+; SSE4-LABEL: test_v4i32:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; SSE4-NEXT: pmulld %xmm0, %xmm1
+; SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; SSE4-NEXT: pmulld %xmm1, %xmm0
+; SSE4-NEXT: movd %xmm0, %eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX-LABEL: test_v4i32:
; AVX: # %bb.0:
@@ -830,12 +1653,12 @@ define i32 @test_v4i32(<4 x i32> %a0) {
; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; AVX-NEXT: vpmulld %xmm1, %xmm0, %xmm0
; AVX-NEXT: vmovd %xmm0, %eax
-; AVX-NEXT: retq
+; AVX-NEXT: ret{{[l|q]}}
%1 = call i32 @llvm.vector.reduce.mul.v4i32(<4 x i32> %a0)
ret i32 %1
}
-define i32 @test_v8i32(<8 x i32> %a0) {
+define i32 @test_v8i32(<8 x i32> %a0) nounwind {
; SSE2-LABEL: test_v8i32:
; SSE2: # %bb.0:
; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,3,3]
@@ -854,17 +1677,17 @@ define i32 @test_v8i32(<8 x i32> %a0) {
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,0,0,0]
; SSE2-NEXT: pmuludq %xmm2, %xmm0
; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v8i32:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmulld %xmm1, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: pmulld %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: pmulld %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
+;
+; SSE4-LABEL: test_v8i32:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pmulld %xmm1, %xmm0
+; SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; SSE4-NEXT: pmulld %xmm0, %xmm1
+; SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; SSE4-NEXT: pmulld %xmm1, %xmm0
+; SSE4-NEXT: movd %xmm0, %eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX1-LABEL: test_v8i32:
; AVX1: # %bb.0:
@@ -876,7 +1699,7 @@ define i32 @test_v8i32(<8 x i32> %a0) {
; AVX1-NEXT: vpmulld %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vmovd %xmm0, %eax
; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
+; AVX1-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v8i32:
; AVX2: # %bb.0:
@@ -888,7 +1711,7 @@ define i32 @test_v8i32(<8 x i32> %a0) {
; AVX2-NEXT: vpmulld %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v8i32:
; AVX512: # %bb.0:
@@ -905,52 +1728,111 @@ define i32 @test_v8i32(<8 x i32> %a0) {
ret i32 %1
}
-define i32 @test_v16i32(<16 x i32> %a0) {
-; SSE2-LABEL: test_v16i32:
-; SSE2: # %bb.0:
-; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm1[1,1,3,3]
-; SSE2-NEXT: pmuludq %xmm3, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
-; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
-; SSE2-NEXT: pmuludq %xmm4, %xmm3
-; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,2,2,3]
-; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1]
-; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3]
-; SSE2-NEXT: pmuludq %xmm2, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
-; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
-; SSE2-NEXT: pmuludq %xmm4, %xmm2
-; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,2,2,3]
-; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]
-; SSE2-NEXT: pmuludq %xmm1, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm3[0,0,2,2]
-; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,2,2]
-; SSE2-NEXT: pmuludq %xmm1, %xmm2
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,2,2,3]
-; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT: pmuludq %xmm0, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,2,2,2]
-; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,2,2]
-; SSE2-NEXT: pmuludq %xmm0, %xmm2
-; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,0,0,0]
-; SSE2-NEXT: pmuludq %xmm1, %xmm0
-; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v16i32:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmulld %xmm3, %xmm1
-; SSE41-NEXT: pmulld %xmm2, %xmm0
-; SSE41-NEXT: pmulld %xmm1, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: pmulld %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: pmulld %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: retq
+define i32 @test_v16i32(<16 x i32> %a0) nounwind {
+; X86-SSE2-LABEL: test_v16i32:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: pushl %ebp
+; X86-SSE2-NEXT: movl %esp, %ebp
+; X86-SSE2-NEXT: andl $-16, %esp
+; X86-SSE2-NEXT: subl $16, %esp
+; X86-SSE2-NEXT: movdqa 8(%ebp), %xmm4
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm1[1,1,3,3]
+; X86-SSE2-NEXT: pmuludq %xmm4, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,2,2,3]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm4[1,1,3,3]
+; X86-SSE2-NEXT: pmuludq %xmm5, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm1[0,2,2,3]
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3]
+; X86-SSE2-NEXT: pmuludq %xmm2, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; X86-SSE2-NEXT: pmuludq %xmm4, %xmm2
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,2,2,3]
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]
+; X86-SSE2-NEXT: pmuludq %xmm3, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,0,2,2]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,2,2]
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm2
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,2,2,3]
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE2-NEXT: pmuludq %xmm0, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,2,2,2]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,2,2]
+; X86-SSE2-NEXT: pmuludq %xmm0, %xmm2
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,0,0,0]
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm0
+; X86-SSE2-NEXT: movd %xmm0, %eax
+; X86-SSE2-NEXT: movl %ebp, %esp
+; X86-SSE2-NEXT: popl %ebp
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE2-LABEL: test_v16i32:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm1[1,1,3,3]
+; X64-SSE2-NEXT: pmuludq %xmm3, %xmm1
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; X64-SSE2-NEXT: pmuludq %xmm4, %xmm3
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[0,2,2,3]
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm0[1,1,3,3]
+; X64-SSE2-NEXT: pmuludq %xmm2, %xmm0
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; X64-SSE2-NEXT: pmuludq %xmm4, %xmm2
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm2[0,2,2,3]
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1]
+; X64-SSE2-NEXT: pmuludq %xmm1, %xmm0
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm3[0,0,2,2]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,2,2]
+; X64-SSE2-NEXT: pmuludq %xmm1, %xmm2
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[0,2,2,3]
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE2-NEXT: pmuludq %xmm0, %xmm1
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,2,2,2]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,2,2]
+; X64-SSE2-NEXT: pmuludq %xmm0, %xmm2
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[0,0,0,0]
+; X64-SSE2-NEXT: pmuludq %xmm1, %xmm0
+; X64-SSE2-NEXT: movd %xmm0, %eax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v16i32:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pushl %ebp
+; X86-SSE4-NEXT: movl %esp, %ebp
+; X86-SSE4-NEXT: andl $-16, %esp
+; X86-SSE4-NEXT: subl $16, %esp
+; X86-SSE4-NEXT: pmulld %xmm2, %xmm0
+; X86-SSE4-NEXT: pmulld 8(%ebp), %xmm1
+; X86-SSE4-NEXT: pmulld %xmm0, %xmm1
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE4-NEXT: pmulld %xmm1, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE4-NEXT: pmulld %xmm0, %xmm1
+; X86-SSE4-NEXT: movd %xmm1, %eax
+; X86-SSE4-NEXT: movl %ebp, %esp
+; X86-SSE4-NEXT: popl %ebp
+; X86-SSE4-NEXT: retl
+;
+; X64-SSE4-LABEL: test_v16i32:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: pmulld %xmm3, %xmm1
+; X64-SSE4-NEXT: pmulld %xmm2, %xmm0
+; X64-SSE4-NEXT: pmulld %xmm1, %xmm0
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE4-NEXT: pmulld %xmm0, %xmm1
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE4-NEXT: pmulld %xmm1, %xmm0
+; X64-SSE4-NEXT: movd %xmm0, %eax
+; X64-SSE4-NEXT: retq
;
; AVX1-LABEL: test_v16i32:
; AVX1: # %bb.0:
@@ -965,7 +1847,7 @@ define i32 @test_v16i32(<16 x i32> %a0) {
; AVX1-NEXT: vpmulld %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vmovd %xmm0, %eax
; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
+; AVX1-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v16i32:
; AVX2: # %bb.0:
@@ -978,7 +1860,7 @@ define i32 @test_v16i32(<16 x i32> %a0) {
; AVX2-NEXT: vpmulld %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v16i32:
; AVX512: # %bb.0:
@@ -997,120 +1879,263 @@ define i32 @test_v16i32(<16 x i32> %a0) {
ret i32 %1
}
-define i32 @test_v32i32(<32 x i32> %a0) {
-; SSE2-LABEL: test_v32i32:
-; SSE2: # %bb.0:
-; SSE2-NEXT: pshufd {{.*#+}} xmm8 = xmm2[1,1,3,3]
-; SSE2-NEXT: pmuludq %xmm6, %xmm2
-; SSE2-NEXT: pshufd {{.*#+}} xmm9 = xmm2[0,2,2,3]
-; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm6[1,1,3,3]
-; SSE2-NEXT: pmuludq %xmm8, %xmm2
-; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm2[0,2,2,3]
-; SSE2-NEXT: punpckldq {{.*#+}} xmm9 = xmm9[0],xmm6[0],xmm9[1],xmm6[1]
-; SSE2-NEXT: pshufd {{.*#+}} xmm8 = xmm0[1,1,3,3]
-; SSE2-NEXT: pmuludq %xmm4, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm0[0,2,2,3]
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
-; SSE2-NEXT: pmuludq %xmm8, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm0[0,2,2,3]
-; SSE2-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm4[0],xmm6[1],xmm4[1]
-; SSE2-NEXT: pmuludq %xmm9, %xmm6
-; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
-; SSE2-NEXT: pmuludq %xmm7, %xmm3
-; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,2,2,3]
-; SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
-; SSE2-NEXT: pmuludq %xmm4, %xmm7
-; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm7[0,2,2,3]
-; SSE2-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]
-; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm1[1,1,3,3]
-; SSE2-NEXT: pmuludq %xmm5, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
-; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
-; SSE2-NEXT: pmuludq %xmm4, %xmm5
-; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm5[0,2,2,3]
-; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1]
-; SSE2-NEXT: pmuludq %xmm3, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
-; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm7[0,0,2,2]
-; SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm5[0,0,2,2]
-; SSE2-NEXT: pmuludq %xmm3, %xmm4
-; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm4[0,2,2,3]
-; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]
-; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm6[0,2,2,3]
-; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,2,2]
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
-; SSE2-NEXT: pmuludq %xmm2, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm0[0,2,2,3]
-; SSE2-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
-; SSE2-NEXT: pmuludq %xmm1, %xmm3
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm3[0,2,2,3]
-; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm4[0,0,2,2]
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
-; SSE2-NEXT: pmuludq %xmm2, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm0[0,2,2,3]
-; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
-; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm1[2,3,2,3]
-; SSE2-NEXT: pmuludq %xmm1, %xmm2
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,2,2,2]
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
-; SSE2-NEXT: pmuludq %xmm1, %xmm0
-; SSE2-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
-; SSE2-NEXT: pmuludq %xmm2, %xmm0
-; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v32i32:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmulld %xmm6, %xmm2
-; SSE41-NEXT: pmulld %xmm4, %xmm0
-; SSE41-NEXT: pmulld %xmm2, %xmm0
-; SSE41-NEXT: pmulld %xmm7, %xmm3
-; SSE41-NEXT: pmulld %xmm5, %xmm1
-; SSE41-NEXT: pmulld %xmm3, %xmm1
-; SSE41-NEXT: pmulld %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE41-NEXT: pmulld %xmm1, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE41-NEXT: pmulld %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: retq
-;
-; AVX1-LABEL: test_v32i32:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vpmulld %xmm3, %xmm1, %xmm4
-; AVX1-NEXT: vpmulld %xmm2, %xmm0, %xmm5
-; AVX1-NEXT: vpmulld %xmm4, %xmm5, %xmm4
-; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm3
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT: vpmulld %xmm3, %xmm1, %xmm1
-; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT: vpmulld %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vpmulld %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpmulld %xmm0, %xmm4, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpmulld %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX1-NEXT: vpmulld %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vmovd %xmm0, %eax
-; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
+define i32 @test_v32i32(<32 x i32> %a0) nounwind {
+; X86-SSE2-LABEL: test_v32i32:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: pushl %ebp
+; X86-SSE2-NEXT: movl %esp, %ebp
+; X86-SSE2-NEXT: andl $-16, %esp
+; X86-SSE2-NEXT: subl $16, %esp
+; X86-SSE2-NEXT: movdqa 72(%ebp), %xmm5
+; X86-SSE2-NEXT: movdqa 24(%ebp), %xmm7
+; X86-SSE2-NEXT: movdqa 56(%ebp), %xmm4
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm2[1,1,3,3]
+; X86-SSE2-NEXT: pmuludq %xmm4, %xmm2
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm2[0,2,2,3]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm4[1,1,3,3]
+; X86-SSE2-NEXT: pmuludq %xmm3, %xmm2
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm2[0,2,2,3]
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm3[0],xmm6[1],xmm3[1]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,3,3]
+; X86-SSE2-NEXT: pmuludq %xmm7, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm0[0,2,2,3]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm7[1,1,3,3]
+; X86-SSE2-NEXT: pmuludq %xmm3, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm0[0,2,2,3]
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]
+; X86-SSE2-NEXT: movdqa 8(%ebp), %xmm3
+; X86-SSE2-NEXT: pmuludq %xmm6, %xmm4
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm3[1,1,3,3]
+; X86-SSE2-NEXT: pmuludq %xmm5, %xmm3
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm3[0,2,2,3]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; X86-SSE2-NEXT: pmuludq %xmm7, %xmm5
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm5[0,2,2,3]
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm3[0],xmm6[1],xmm3[1]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm1[1,1,3,3]
+; X86-SSE2-NEXT: movdqa 40(%ebp), %xmm7
+; X86-SSE2-NEXT: pmuludq %xmm7, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
+; X86-SSE2-NEXT: pmuludq %xmm3, %xmm7
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm7[0,2,2,3]
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]
+; X86-SSE2-NEXT: pmuludq %xmm6, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm1[0,2,2,3]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm5[0,0,2,2]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm7[0,0,2,2]
+; X86-SSE2-NEXT: pmuludq %xmm5, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm1[0,2,2,3]
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm5[0],xmm3[1],xmm5[1]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,2,2,3]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,2,2]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; X86-SSE2-NEXT: pmuludq %xmm2, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm0[0,2,2,3]
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
+; X86-SSE2-NEXT: pmuludq %xmm3, %xmm4
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm4[0,2,2,3]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,0,2,2]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[0,2,2,3]
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm2[2,3,2,3]
+; X86-SSE2-NEXT: pmuludq %xmm2, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm0[2,2,2,2]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; X86-SSE2-NEXT: pmuludq %xmm2, %xmm0
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
+; X86-SSE2-NEXT: pmuludq %xmm1, %xmm0
+; X86-SSE2-NEXT: movd %xmm0, %eax
+; X86-SSE2-NEXT: movl %ebp, %esp
+; X86-SSE2-NEXT: popl %ebp
+; X86-SSE2-NEXT: retl
;
-; AVX2-LABEL: test_v32i32:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpmulld %ymm3, %ymm1, %ymm1
-; AVX2-NEXT: vpmulld %ymm2, %ymm0, %ymm0
-; AVX2-NEXT: vpmulld %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpmulld %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpmulld %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX2-NEXT: vpmulld %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+; X64-SSE2-LABEL: test_v32i32:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm8 = xmm2[1,1,3,3]
+; X64-SSE2-NEXT: pmuludq %xmm6, %xmm2
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm9 = xmm2[0,2,2,3]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm6[1,1,3,3]
+; X64-SSE2-NEXT: pmuludq %xmm8, %xmm2
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm2[0,2,2,3]
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm9 = xmm9[0],xmm6[0],xmm9[1],xmm6[1]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm8 = xmm0[1,1,3,3]
+; X64-SSE2-NEXT: pmuludq %xmm4, %xmm0
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm6 = xmm0[0,2,2,3]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm4[1,1,3,3]
+; X64-SSE2-NEXT: pmuludq %xmm8, %xmm0
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm0[0,2,2,3]
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm4[0],xmm6[1],xmm4[1]
+; X64-SSE2-NEXT: pmuludq %xmm9, %xmm6
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm3[1,1,3,3]
+; X64-SSE2-NEXT: pmuludq %xmm7, %xmm3
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm3[0,2,2,3]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm7 = xmm7[1,1,3,3]
+; X64-SSE2-NEXT: pmuludq %xmm4, %xmm7
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm7[0,2,2,3]
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm1[1,1,3,3]
+; X64-SSE2-NEXT: pmuludq %xmm5, %xmm1
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm5[1,1,3,3]
+; X64-SSE2-NEXT: pmuludq %xmm4, %xmm5
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm5[0,2,2,3]
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1]
+; X64-SSE2-NEXT: pmuludq %xmm3, %xmm1
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm7[0,0,2,2]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm4 = xmm5[0,0,2,2]
+; X64-SSE2-NEXT: pmuludq %xmm3, %xmm4
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm4[0,2,2,3]
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm6[0,2,2,3]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,2,2]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; X64-SSE2-NEXT: pmuludq %xmm2, %xmm0
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm0[0,2,2,3]
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
+; X64-SSE2-NEXT: pmuludq %xmm1, %xmm3
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm3[0,2,2,3]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm4[0,0,2,2]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; X64-SSE2-NEXT: pmuludq %xmm2, %xmm0
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm0[0,2,2,3]
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm1[2,3,2,3]
+; X64-SSE2-NEXT: pmuludq %xmm1, %xmm2
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,2,2,2]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,2,2]
+; X64-SSE2-NEXT: pmuludq %xmm1, %xmm0
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
+; X64-SSE2-NEXT: pmuludq %xmm2, %xmm0
+; X64-SSE2-NEXT: movd %xmm0, %eax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v32i32:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pushl %ebp
+; X86-SSE4-NEXT: movl %esp, %ebp
+; X86-SSE4-NEXT: andl $-16, %esp
+; X86-SSE4-NEXT: subl $16, %esp
+; X86-SSE4-NEXT: movdqa 8(%ebp), %xmm3
+; X86-SSE4-NEXT: pmulld 56(%ebp), %xmm2
+; X86-SSE4-NEXT: pmulld 24(%ebp), %xmm0
+; X86-SSE4-NEXT: pmulld %xmm2, %xmm0
+; X86-SSE4-NEXT: pmulld 72(%ebp), %xmm3
+; X86-SSE4-NEXT: pmulld 40(%ebp), %xmm1
+; X86-SSE4-NEXT: pmulld %xmm3, %xmm1
+; X86-SSE4-NEXT: pmulld %xmm0, %xmm1
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE4-NEXT: pmulld %xmm1, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE4-NEXT: pmulld %xmm0, %xmm1
+; X86-SSE4-NEXT: movd %xmm1, %eax
+; X86-SSE4-NEXT: movl %ebp, %esp
+; X86-SSE4-NEXT: popl %ebp
+; X86-SSE4-NEXT: retl
+;
+; X64-SSE4-LABEL: test_v32i32:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: pmulld %xmm6, %xmm2
+; X64-SSE4-NEXT: pmulld %xmm4, %xmm0
+; X64-SSE4-NEXT: pmulld %xmm2, %xmm0
+; X64-SSE4-NEXT: pmulld %xmm7, %xmm3
+; X64-SSE4-NEXT: pmulld %xmm5, %xmm1
+; X64-SSE4-NEXT: pmulld %xmm3, %xmm1
+; X64-SSE4-NEXT: pmulld %xmm0, %xmm1
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X64-SSE4-NEXT: pmulld %xmm1, %xmm0
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-SSE4-NEXT: pmulld %xmm0, %xmm1
+; X64-SSE4-NEXT: movd %xmm1, %eax
+; X64-SSE4-NEXT: retq
+;
+; X86-AVX1-LABEL: test_v32i32:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: pushl %ebp
+; X86-AVX1-NEXT: movl %esp, %ebp
+; X86-AVX1-NEXT: andl $-32, %esp
+; X86-AVX1-NEXT: subl $32, %esp
+; X86-AVX1-NEXT: vpmulld %xmm2, %xmm0, %xmm3
+; X86-AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X86-AVX1-NEXT: vpmulld %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpmulld 8(%ebp), %xmm1, %xmm2
+; X86-AVX1-NEXT: vpmulld %xmm2, %xmm3, %xmm2
+; X86-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X86-AVX1-NEXT: vpmulld 24(%ebp), %xmm1, %xmm1
+; X86-AVX1-NEXT: vpmulld %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpmulld %xmm0, %xmm2, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpmulld %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX1-NEXT: vpmulld %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: movl %ebp, %esp
+; X86-AVX1-NEXT: popl %ebp
+; X86-AVX1-NEXT: vzeroupper
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v32i32:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vpmulld %xmm3, %xmm1, %xmm4
+; X64-AVX1-NEXT: vpmulld %xmm2, %xmm0, %xmm5
+; X64-AVX1-NEXT: vpmulld %xmm4, %xmm5, %xmm4
+; X64-AVX1-NEXT: vextractf128 $1, %ymm3, %xmm3
+; X64-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X64-AVX1-NEXT: vpmulld %xmm3, %xmm1, %xmm1
+; X64-AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X64-AVX1-NEXT: vpmulld %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpmulld %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpmulld %xmm0, %xmm4, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpmulld %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX1-NEXT: vpmulld %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-NEXT: vzeroupper
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v32i32:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: pushl %ebp
+; X86-AVX2-NEXT: movl %esp, %ebp
+; X86-AVX2-NEXT: andl $-32, %esp
+; X86-AVX2-NEXT: subl $32, %esp
+; X86-AVX2-NEXT: vpmulld %ymm2, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpmulld 8(%ebp), %ymm1, %ymm1
+; X86-AVX2-NEXT: vpmulld %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpmulld %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpmulld %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX2-NEXT: vpmulld %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: movl %ebp, %esp
+; X86-AVX2-NEXT: popl %ebp
+; X86-AVX2-NEXT: vzeroupper
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v32i32:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpmulld %ymm3, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpmulld %ymm2, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpmulld %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpmulld %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpmulld %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX2-NEXT: vpmulld %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovd %xmm0, %eax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
;
; AVX512-LABEL: test_v32i32:
; AVX512: # %bb.0:
@@ -1134,7 +2159,7 @@ define i32 @test_v32i32(<32 x i32> %a0) {
; vXi16
;
-define i16 @test_v2i16(<2 x i16> %a0) {
+define i16 @test_v2i16(<2 x i16> %a0) nounwind {
; SSE-LABEL: test_v2i16:
; SSE: # %bb.0:
; SSE-NEXT: movdqa %xmm0, %xmm1
@@ -1142,7 +2167,7 @@ define i16 @test_v2i16(<2 x i16> %a0) {
; SSE-NEXT: pmullw %xmm0, %xmm1
; SSE-NEXT: movd %xmm1, %eax
; SSE-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE-NEXT: retq
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX-LABEL: test_v2i16:
; AVX: # %bb.0:
@@ -1150,12 +2175,12 @@ define i16 @test_v2i16(<2 x i16> %a0) {
; AVX-NEXT: vpmullw %xmm1, %xmm0, %xmm0
; AVX-NEXT: vmovd %xmm0, %eax
; AVX-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX-NEXT: retq
+; AVX-NEXT: ret{{[l|q]}}
%1 = call i16 @llvm.vector.reduce.mul.v2i16(<2 x i16> %a0)
ret i16 %1
}
-define i16 @test_v4i16(<4 x i16> %a0) {
+define i16 @test_v4i16(<4 x i16> %a0) nounwind {
; SSE-LABEL: test_v4i16:
; SSE: # %bb.0:
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
@@ -1165,7 +2190,7 @@ define i16 @test_v4i16(<4 x i16> %a0) {
; SSE-NEXT: pmullw %xmm1, %xmm0
; SSE-NEXT: movd %xmm0, %eax
; SSE-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE-NEXT: retq
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX-LABEL: test_v4i16:
; AVX: # %bb.0:
@@ -1175,12 +2200,12 @@ define i16 @test_v4i16(<4 x i16> %a0) {
; AVX-NEXT: vpmullw %xmm1, %xmm0, %xmm0
; AVX-NEXT: vmovd %xmm0, %eax
; AVX-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX-NEXT: retq
+; AVX-NEXT: ret{{[l|q]}}
%1 = call i16 @llvm.vector.reduce.mul.v4i16(<4 x i16> %a0)
ret i16 %1
}
-define i16 @test_v8i16(<8 x i16> %a0) {
+define i16 @test_v8i16(<8 x i16> %a0) nounwind {
; SSE-LABEL: test_v8i16:
; SSE: # %bb.0:
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
@@ -1192,7 +2217,7 @@ define i16 @test_v8i16(<8 x i16> %a0) {
; SSE-NEXT: pmullw %xmm0, %xmm1
; SSE-NEXT: movd %xmm1, %eax
; SSE-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE-NEXT: retq
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX-LABEL: test_v8i16:
; AVX: # %bb.0:
@@ -1204,12 +2229,12 @@ define i16 @test_v8i16(<8 x i16> %a0) {
; AVX-NEXT: vpmullw %xmm1, %xmm0, %xmm0
; AVX-NEXT: vmovd %xmm0, %eax
; AVX-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX-NEXT: retq
+; AVX-NEXT: ret{{[l|q]}}
%1 = call i16 @llvm.vector.reduce.mul.v8i16(<8 x i16> %a0)
ret i16 %1
}
-define i16 @test_v16i16(<16 x i16> %a0) {
+define i16 @test_v16i16(<16 x i16> %a0) nounwind {
; SSE-LABEL: test_v16i16:
; SSE: # %bb.0:
; SSE-NEXT: pmullw %xmm1, %xmm0
@@ -1222,7 +2247,7 @@ define i16 @test_v16i16(<16 x i16> %a0) {
; SSE-NEXT: pmullw %xmm0, %xmm1
; SSE-NEXT: movd %xmm1, %eax
; SSE-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE-NEXT: retq
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX1-LABEL: test_v16i16:
; AVX1: # %bb.0:
@@ -1237,7 +2262,7 @@ define i16 @test_v16i16(<16 x i16> %a0) {
; AVX1-NEXT: vmovd %xmm0, %eax
; AVX1-NEXT: # kill: def $ax killed $ax killed $eax
; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
+; AVX1-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v16i16:
; AVX2: # %bb.0:
@@ -1252,7 +2277,7 @@ define i16 @test_v16i16(<16 x i16> %a0) {
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v16i16:
; AVX512: # %bb.0:
@@ -1272,22 +2297,44 @@ define i16 @test_v16i16(<16 x i16> %a0) {
ret i16 %1
}
-define i16 @test_v32i16(<32 x i16> %a0) {
-; SSE-LABEL: test_v32i16:
-; SSE: # %bb.0:
-; SSE-NEXT: pmullw %xmm3, %xmm1
-; SSE-NEXT: pmullw %xmm2, %xmm0
-; SSE-NEXT: pmullw %xmm1, %xmm0
-; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE-NEXT: pmullw %xmm0, %xmm1
-; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE-NEXT: pmullw %xmm1, %xmm0
-; SSE-NEXT: movdqa %xmm0, %xmm1
-; SSE-NEXT: psrld $16, %xmm1
-; SSE-NEXT: pmullw %xmm0, %xmm1
-; SSE-NEXT: movd %xmm1, %eax
-; SSE-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE-NEXT: retq
+define i16 @test_v32i16(<32 x i16> %a0) nounwind {
+; X86-SSE-LABEL: test_v32i16:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: pushl %ebp
+; X86-SSE-NEXT: movl %esp, %ebp
+; X86-SSE-NEXT: andl $-16, %esp
+; X86-SSE-NEXT: subl $16, %esp
+; X86-SSE-NEXT: pmullw %xmm2, %xmm0
+; X86-SSE-NEXT: pmullw 8(%ebp), %xmm1
+; X86-SSE-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE-NEXT: movdqa %xmm1, %xmm0
+; X86-SSE-NEXT: psrld $16, %xmm0
+; X86-SSE-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE-NEXT: movd %xmm0, %eax
+; X86-SSE-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-SSE-NEXT: movl %ebp, %esp
+; X86-SSE-NEXT: popl %ebp
+; X86-SSE-NEXT: retl
+;
+; X64-SSE-LABEL: test_v32i16:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: pmullw %xmm3, %xmm1
+; X64-SSE-NEXT: pmullw %xmm2, %xmm0
+; X64-SSE-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE-NEXT: movdqa %xmm0, %xmm1
+; X64-SSE-NEXT: psrld $16, %xmm1
+; X64-SSE-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE-NEXT: movd %xmm1, %eax
+; X64-SSE-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-SSE-NEXT: retq
;
; AVX1-LABEL: test_v32i16:
; AVX1: # %bb.0:
@@ -1305,7 +2352,7 @@ define i16 @test_v32i16(<32 x i16> %a0) {
; AVX1-NEXT: vmovd %xmm0, %eax
; AVX1-NEXT: # kill: def $ax killed $ax killed $eax
; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
+; AVX1-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v32i16:
; AVX2: # %bb.0:
@@ -1321,7 +2368,24 @@ define i16 @test_v32i16(<32 x i16> %a0) {
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
+;
+; AVX512F-LABEL: test_v32i16:
+; AVX512F: # %bb.0:
+; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT: vpmullw %ymm1, %ymm0, %ymm0
+; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
+; AVX512F-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512F-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; AVX512F-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpsrld $16, %xmm0, %xmm1
+; AVX512F-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vmovd %xmm0, %eax
+; AVX512F-NEXT: # kill: def $ax killed $ax killed $eax
+; AVX512F-NEXT: vzeroupper
+; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: test_v32i16:
; AVX512BW: # %bb.0:
@@ -1394,68 +2458,168 @@ define i16 @test_v32i16(<32 x i16> %a0) {
ret i16 %1
}
-define i16 @test_v64i16(<64 x i16> %a0) {
-; SSE-LABEL: test_v64i16:
-; SSE: # %bb.0:
-; SSE-NEXT: pmullw %xmm6, %xmm2
-; SSE-NEXT: pmullw %xmm4, %xmm0
-; SSE-NEXT: pmullw %xmm2, %xmm0
-; SSE-NEXT: pmullw %xmm7, %xmm3
-; SSE-NEXT: pmullw %xmm5, %xmm1
-; SSE-NEXT: pmullw %xmm3, %xmm1
-; SSE-NEXT: pmullw %xmm0, %xmm1
-; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE-NEXT: pmullw %xmm1, %xmm0
-; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE-NEXT: pmullw %xmm0, %xmm1
-; SSE-NEXT: movdqa %xmm1, %xmm0
-; SSE-NEXT: psrld $16, %xmm0
-; SSE-NEXT: pmullw %xmm1, %xmm0
-; SSE-NEXT: movd %xmm0, %eax
-; SSE-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE-NEXT: retq
+define i16 @test_v64i16(<64 x i16> %a0) nounwind {
+; X86-SSE-LABEL: test_v64i16:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: pushl %ebp
+; X86-SSE-NEXT: movl %esp, %ebp
+; X86-SSE-NEXT: andl $-16, %esp
+; X86-SSE-NEXT: subl $16, %esp
+; X86-SSE-NEXT: movdqa 8(%ebp), %xmm3
+; X86-SSE-NEXT: pmullw 56(%ebp), %xmm2
+; X86-SSE-NEXT: pmullw 24(%ebp), %xmm0
+; X86-SSE-NEXT: pmullw %xmm2, %xmm0
+; X86-SSE-NEXT: pmullw 72(%ebp), %xmm3
+; X86-SSE-NEXT: pmullw 40(%ebp), %xmm1
+; X86-SSE-NEXT: pmullw %xmm3, %xmm1
+; X86-SSE-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE-NEXT: movdqa %xmm1, %xmm0
+; X86-SSE-NEXT: psrld $16, %xmm0
+; X86-SSE-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE-NEXT: movd %xmm0, %eax
+; X86-SSE-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-SSE-NEXT: movl %ebp, %esp
+; X86-SSE-NEXT: popl %ebp
+; X86-SSE-NEXT: retl
;
-; AVX1-LABEL: test_v64i16:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vpmullw %xmm3, %xmm1, %xmm4
-; AVX1-NEXT: vpmullw %xmm2, %xmm0, %xmm5
-; AVX1-NEXT: vpmullw %xmm4, %xmm5, %xmm4
-; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm3
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT: vpmullw %xmm3, %xmm1, %xmm1
-; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT: vpmullw %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpmullw %xmm0, %xmm4, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpsrld $16, %xmm0, %xmm1
-; AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vmovd %xmm0, %eax
-; AVX1-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
+; X64-SSE-LABEL: test_v64i16:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: pmullw %xmm6, %xmm2
+; X64-SSE-NEXT: pmullw %xmm4, %xmm0
+; X64-SSE-NEXT: pmullw %xmm2, %xmm0
+; X64-SSE-NEXT: pmullw %xmm7, %xmm3
+; X64-SSE-NEXT: pmullw %xmm5, %xmm1
+; X64-SSE-NEXT: pmullw %xmm3, %xmm1
+; X64-SSE-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X64-SSE-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-SSE-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE-NEXT: movdqa %xmm1, %xmm0
+; X64-SSE-NEXT: psrld $16, %xmm0
+; X64-SSE-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE-NEXT: movd %xmm0, %eax
+; X64-SSE-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-SSE-NEXT: retq
;
-; AVX2-LABEL: test_v64i16:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpmullw %ymm3, %ymm1, %ymm1
-; AVX2-NEXT: vpmullw %ymm2, %ymm0, %ymm0
-; AVX2-NEXT: vpmullw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
-; AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+; X86-AVX1-LABEL: test_v64i16:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: pushl %ebp
+; X86-AVX1-NEXT: movl %esp, %ebp
+; X86-AVX1-NEXT: andl $-32, %esp
+; X86-AVX1-NEXT: subl $32, %esp
+; X86-AVX1-NEXT: vpmullw %xmm2, %xmm0, %xmm3
+; X86-AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X86-AVX1-NEXT: vpmullw %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpmullw 8(%ebp), %xmm1, %xmm2
+; X86-AVX1-NEXT: vpmullw %xmm2, %xmm3, %xmm2
+; X86-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X86-AVX1-NEXT: vpmullw 24(%ebp), %xmm1, %xmm1
+; X86-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpmullw %xmm0, %xmm2, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpsrld $16, %xmm0, %xmm1
+; X86-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-AVX1-NEXT: movl %ebp, %esp
+; X86-AVX1-NEXT: popl %ebp
+; X86-AVX1-NEXT: vzeroupper
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v64i16:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vpmullw %xmm3, %xmm1, %xmm4
+; X64-AVX1-NEXT: vpmullw %xmm2, %xmm0, %xmm5
+; X64-AVX1-NEXT: vpmullw %xmm4, %xmm5, %xmm4
+; X64-AVX1-NEXT: vextractf128 $1, %ymm3, %xmm3
+; X64-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X64-AVX1-NEXT: vpmullw %xmm3, %xmm1, %xmm1
+; X64-AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X64-AVX1-NEXT: vpmullw %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpmullw %xmm0, %xmm4, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpsrld $16, %xmm0, %xmm1
+; X64-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-AVX1-NEXT: vzeroupper
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v64i16:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: pushl %ebp
+; X86-AVX2-NEXT: movl %esp, %ebp
+; X86-AVX2-NEXT: andl $-32, %esp
+; X86-AVX2-NEXT: subl $32, %esp
+; X86-AVX2-NEXT: vpmullw %ymm2, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpmullw 8(%ebp), %ymm1, %ymm1
+; X86-AVX2-NEXT: vpmullw %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
+; X86-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-AVX2-NEXT: movl %ebp, %esp
+; X86-AVX2-NEXT: popl %ebp
+; X86-AVX2-NEXT: vzeroupper
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v64i16:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpmullw %ymm3, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpmullw %ymm2, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpmullw %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
+; X64-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovd %xmm0, %eax
+; X64-AVX2-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
+;
+; AVX512F-LABEL: test_v64i16:
+; AVX512F: # %bb.0:
+; AVX512F-NEXT: vextracti64x4 $1, %zmm1, %ymm2
+; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm3
+; AVX512F-NEXT: vpmullw %ymm2, %ymm3, %ymm2
+; AVX512F-NEXT: vpmullw %ymm1, %ymm0, %ymm0
+; AVX512F-NEXT: vpmullw %ymm2, %ymm0, %ymm0
+; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
+; AVX512F-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512F-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; AVX512F-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpsrld $16, %xmm0, %xmm1
+; AVX512F-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vmovd %xmm0, %eax
+; AVX512F-NEXT: # kill: def $ax killed $ax killed $eax
+; AVX512F-NEXT: vzeroupper
+; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: test_v64i16:
; AVX512BW: # %bb.0:
@@ -1540,7 +2704,7 @@ define i16 @test_v64i16(<64 x i16> %a0) {
; vXi8
;
-define i8 @test_v2i8(<2 x i8> %a0) {
+define i8 @test_v2i8(<2 x i8> %a0) nounwind {
; SSE-LABEL: test_v2i8:
; SSE: # %bb.0:
; SSE-NEXT: movdqa %xmm0, %xmm1
@@ -1548,7 +2712,7 @@ define i8 @test_v2i8(<2 x i8> %a0) {
; SSE-NEXT: pmullw %xmm0, %xmm1
; SSE-NEXT: movd %xmm1, %eax
; SSE-NEXT: # kill: def $al killed $al killed $eax
-; SSE-NEXT: retq
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX-LABEL: test_v2i8:
; AVX: # %bb.0:
@@ -1556,12 +2720,12 @@ define i8 @test_v2i8(<2 x i8> %a0) {
; AVX-NEXT: vpmullw %xmm1, %xmm0, %xmm0
; AVX-NEXT: vmovd %xmm0, %eax
; AVX-NEXT: # kill: def $al killed $al killed $eax
-; AVX-NEXT: retq
+; AVX-NEXT: ret{{[l|q]}}
%1 = call i8 @llvm.vector.reduce.mul.v2i8(<2 x i8> %a0)
ret i8 %1
}
-define i8 @test_v4i8(<4 x i8> %a0) {
+define i8 @test_v4i8(<4 x i8> %a0) nounwind {
; SSE2-LABEL: test_v4i8:
; SSE2: # %bb.0:
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
@@ -1572,19 +2736,19 @@ define i8 @test_v4i8(<4 x i8> %a0) {
; SSE2-NEXT: pmullw %xmm1, %xmm0
; SSE2-NEXT: movd %xmm0, %eax
; SSE2-NEXT: # kill: def $al killed $al killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v4i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE41-NEXT: pmullw %xmm0, %xmm1
-; SSE41-NEXT: movdqa %xmm1, %xmm0
-; SSE41-NEXT: psrld $16, %xmm0
-; SSE41-NEXT: pmullw %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: # kill: def $al killed $al killed $eax
-; SSE41-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
+;
+; SSE4-LABEL: test_v4i8:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; SSE4-NEXT: pmullw %xmm0, %xmm1
+; SSE4-NEXT: movdqa %xmm1, %xmm0
+; SSE4-NEXT: psrld $16, %xmm0
+; SSE4-NEXT: pmullw %xmm1, %xmm0
+; SSE4-NEXT: movd %xmm0, %eax
+; SSE4-NEXT: # kill: def $al killed $al killed $eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX-LABEL: test_v4i8:
; AVX: # %bb.0:
@@ -1595,12 +2759,12 @@ define i8 @test_v4i8(<4 x i8> %a0) {
; AVX-NEXT: vpmullw %xmm1, %xmm0, %xmm0
; AVX-NEXT: vmovd %xmm0, %eax
; AVX-NEXT: # kill: def $al killed $al killed $eax
-; AVX-NEXT: retq
+; AVX-NEXT: ret{{[l|q]}}
%1 = call i8 @llvm.vector.reduce.mul.v4i8(<4 x i8> %a0)
ret i8 %1
}
-define i8 @test_v8i8(<8 x i8> %a0) {
+define i8 @test_v8i8(<8 x i8> %a0) nounwind {
; SSE2-LABEL: test_v8i8:
; SSE2: # %bb.0:
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
@@ -1613,21 +2777,21 @@ define i8 @test_v8i8(<8 x i8> %a0) {
; SSE2-NEXT: pmullw %xmm0, %xmm1
; SSE2-NEXT: movd %xmm1, %eax
; SSE2-NEXT: # kill: def $al killed $al killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v8i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: pmullw %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: pmullw %xmm1, %xmm0
-; SSE41-NEXT: movdqa %xmm0, %xmm1
-; SSE41-NEXT: psrld $16, %xmm1
-; SSE41-NEXT: pmullw %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: # kill: def $al killed $al killed $eax
-; SSE41-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
+;
+; SSE4-LABEL: test_v8i8:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; SSE4-NEXT: pmullw %xmm0, %xmm1
+; SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; SSE4-NEXT: pmullw %xmm1, %xmm0
+; SSE4-NEXT: movdqa %xmm0, %xmm1
+; SSE4-NEXT: psrld $16, %xmm1
+; SSE4-NEXT: pmullw %xmm0, %xmm1
+; SSE4-NEXT: movd %xmm1, %eax
+; SSE4-NEXT: # kill: def $al killed $al killed $eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX-LABEL: test_v8i8:
; AVX: # %bb.0:
@@ -1640,12 +2804,12 @@ define i8 @test_v8i8(<8 x i8> %a0) {
; AVX-NEXT: vpmullw %xmm1, %xmm0, %xmm0
; AVX-NEXT: vmovd %xmm0, %eax
; AVX-NEXT: # kill: def $al killed $al killed $eax
-; AVX-NEXT: retq
+; AVX-NEXT: ret{{[l|q]}}
%1 = call i8 @llvm.vector.reduce.mul.v8i8(<8 x i8> %a0)
ret i8 %1
}
-define i8 @test_v16i8(<16 x i8> %a0) {
+define i8 @test_v16i8(<16 x i8> %a0) nounwind {
; SSE2-LABEL: test_v16i8:
; SSE2: # %bb.0:
; SSE2-NEXT: movdqa %xmm0, %xmm1
@@ -1661,23 +2825,23 @@ define i8 @test_v16i8(<16 x i8> %a0) {
; SSE2-NEXT: pmullw %xmm0, %xmm1
; SSE2-NEXT: movd %xmm1, %eax
; SSE2-NEXT: # kill: def $al killed $al killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v16i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; SSE41-NEXT: punpckhbw {{.*#+}} xmm0 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE41-NEXT: pmullw %xmm1, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: pmullw %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: pmullw %xmm1, %xmm0
-; SSE41-NEXT: movdqa %xmm0, %xmm1
-; SSE41-NEXT: psrld $16, %xmm1
-; SSE41-NEXT: pmullw %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: # kill: def $al killed $al killed $eax
-; SSE41-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
+;
+; SSE4-LABEL: test_v16i8:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pmovzxbw {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; SSE4-NEXT: punpckhbw {{.*#+}} xmm0 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; SSE4-NEXT: pmullw %xmm1, %xmm0
+; SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; SSE4-NEXT: pmullw %xmm0, %xmm1
+; SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; SSE4-NEXT: pmullw %xmm1, %xmm0
+; SSE4-NEXT: movdqa %xmm0, %xmm1
+; SSE4-NEXT: psrld $16, %xmm1
+; SSE4-NEXT: pmullw %xmm0, %xmm1
+; SSE4-NEXT: movd %xmm1, %eax
+; SSE4-NEXT: # kill: def $al killed $al killed $eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX-LABEL: test_v16i8:
; AVX: # %bb.0:
@@ -1692,12 +2856,12 @@ define i8 @test_v16i8(<16 x i8> %a0) {
; AVX-NEXT: vpmullw %xmm1, %xmm0, %xmm0
; AVX-NEXT: vmovd %xmm0, %eax
; AVX-NEXT: # kill: def $al killed $al killed $eax
-; AVX-NEXT: retq
+; AVX-NEXT: ret{{[l|q]}}
%1 = call i8 @llvm.vector.reduce.mul.v16i8(<16 x i8> %a0)
ret i8 %1
}
-define i8 @test_v32i8(<32 x i8> %a0) {
+define i8 @test_v32i8(<32 x i8> %a0) nounwind {
; SSE2-LABEL: test_v32i8:
; SSE2: # %bb.0:
; SSE2-NEXT: movdqa %xmm1, %xmm2
@@ -1718,27 +2882,27 @@ define i8 @test_v32i8(<32 x i8> %a0) {
; SSE2-NEXT: pmullw %xmm0, %xmm1
; SSE2-NEXT: movd %xmm1, %eax
; SSE2-NEXT: # kill: def $al killed $al killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v32i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm2 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
-; SSE41-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE41-NEXT: pmullw %xmm2, %xmm1
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; SSE41-NEXT: punpckhbw {{.*#+}} xmm0 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE41-NEXT: pmullw %xmm2, %xmm0
-; SSE41-NEXT: pmullw %xmm1, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: pmullw %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: pmullw %xmm1, %xmm0
-; SSE41-NEXT: movdqa %xmm0, %xmm1
-; SSE41-NEXT: psrld $16, %xmm1
-; SSE41-NEXT: pmullw %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: # kill: def $al killed $al killed $eax
-; SSE41-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
+;
+; SSE4-LABEL: test_v32i8:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pmovzxbw {{.*#+}} xmm2 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
+; SSE4-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; SSE4-NEXT: pmullw %xmm2, %xmm1
+; SSE4-NEXT: pmovzxbw {{.*#+}} xmm2 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; SSE4-NEXT: punpckhbw {{.*#+}} xmm0 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; SSE4-NEXT: pmullw %xmm2, %xmm0
+; SSE4-NEXT: pmullw %xmm1, %xmm0
+; SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; SSE4-NEXT: pmullw %xmm0, %xmm1
+; SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; SSE4-NEXT: pmullw %xmm1, %xmm0
+; SSE4-NEXT: movdqa %xmm0, %xmm1
+; SSE4-NEXT: psrld $16, %xmm1
+; SSE4-NEXT: pmullw %xmm0, %xmm1
+; SSE4-NEXT: movd %xmm1, %eax
+; SSE4-NEXT: # kill: def $al killed $al killed $eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX1-LABEL: test_v32i8:
; AVX1: # %bb.0:
@@ -1759,7 +2923,7 @@ define i8 @test_v32i8(<32 x i8> %a0) {
; AVX1-NEXT: vmovd %xmm0, %eax
; AVX1-NEXT: # kill: def $al killed $al killed $eax
; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
+; AVX1-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v32i8:
; AVX2: # %bb.0:
@@ -1777,7 +2941,7 @@ define i8 @test_v32i8(<32 x i8> %a0) {
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: # kill: def $al killed $al killed $eax
; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v32i8:
; AVX512: # %bb.0:
@@ -1800,66 +2964,140 @@ define i8 @test_v32i8(<32 x i8> %a0) {
ret i8 %1
}
-define i8 @test_v64i8(<64 x i8> %a0) {
-; SSE2-LABEL: test_v64i8:
-; SSE2: # %bb.0:
-; SSE2-NEXT: movdqa %xmm3, %xmm4
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: pmullw %xmm4, %xmm3
-; SSE2-NEXT: movdqa %xmm1, %xmm4
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: pmullw %xmm4, %xmm1
-; SSE2-NEXT: pmullw %xmm3, %xmm1
-; SSE2-NEXT: movdqa %xmm2, %xmm3
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: pmullw %xmm3, %xmm2
-; SSE2-NEXT: movdqa %xmm0, %xmm3
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: pmullw %xmm3, %xmm0
-; SSE2-NEXT: pmullw %xmm2, %xmm0
-; SSE2-NEXT: pmullw %xmm1, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT: pmullw %xmm0, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE2-NEXT: pmullw %xmm1, %xmm0
-; SSE2-NEXT: movdqa %xmm0, %xmm1
-; SSE2-NEXT: psrld $16, %xmm1
-; SSE2-NEXT: pmullw %xmm0, %xmm1
-; SSE2-NEXT: movd %xmm1, %eax
-; SSE2-NEXT: # kill: def $al killed $al killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v64i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm4 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero,xmm3[4],zero,xmm3[5],zero,xmm3[6],zero,xmm3[7],zero
-; SSE41-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE41-NEXT: pmullw %xmm4, %xmm3
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm4 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
-; SSE41-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE41-NEXT: pmullw %xmm4, %xmm1
-; SSE41-NEXT: pmullw %xmm3, %xmm1
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm3 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero,xmm2[4],zero,xmm2[5],zero,xmm2[6],zero,xmm2[7],zero
-; SSE41-NEXT: punpckhbw {{.*#+}} xmm2 = xmm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE41-NEXT: pmullw %xmm3, %xmm2
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm3 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; SSE41-NEXT: punpckhbw {{.*#+}} xmm0 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE41-NEXT: pmullw %xmm3, %xmm0
-; SSE41-NEXT: pmullw %xmm2, %xmm0
-; SSE41-NEXT: pmullw %xmm1, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: pmullw %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: pmullw %xmm1, %xmm0
-; SSE41-NEXT: movdqa %xmm0, %xmm1
-; SSE41-NEXT: psrld $16, %xmm1
-; SSE41-NEXT: pmullw %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: # kill: def $al killed $al killed $eax
-; SSE41-NEXT: retq
+define i8 @test_v64i8(<64 x i8> %a0) nounwind {
+; X86-SSE2-LABEL: test_v64i8:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: pushl %ebp
+; X86-SSE2-NEXT: movl %esp, %ebp
+; X86-SSE2-NEXT: andl $-16, %esp
+; X86-SSE2-NEXT: subl $16, %esp
+; X86-SSE2-NEXT: movdqa 8(%ebp), %xmm3
+; X86-SSE2-NEXT: movdqa %xmm3, %xmm4
+; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X86-SSE2-NEXT: pmullw %xmm4, %xmm3
+; X86-SSE2-NEXT: movdqa %xmm1, %xmm4
+; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X86-SSE2-NEXT: pmullw %xmm4, %xmm1
+; X86-SSE2-NEXT: pmullw %xmm3, %xmm1
+; X86-SSE2-NEXT: movdqa %xmm2, %xmm3
+; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X86-SSE2-NEXT: pmullw %xmm3, %xmm2
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm3
+; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X86-SSE2-NEXT: pmullw %xmm3, %xmm0
+; X86-SSE2-NEXT: pmullw %xmm2, %xmm0
+; X86-SSE2-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE2-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE2-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm1
+; X86-SSE2-NEXT: psrld $16, %xmm1
+; X86-SSE2-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE2-NEXT: movd %xmm1, %eax
+; X86-SSE2-NEXT: # kill: def $al killed $al killed $eax
+; X86-SSE2-NEXT: movl %ebp, %esp
+; X86-SSE2-NEXT: popl %ebp
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE2-LABEL: test_v64i8:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: movdqa %xmm3, %xmm4
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X64-SSE2-NEXT: pmullw %xmm4, %xmm3
+; X64-SSE2-NEXT: movdqa %xmm1, %xmm4
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X64-SSE2-NEXT: pmullw %xmm4, %xmm1
+; X64-SSE2-NEXT: pmullw %xmm3, %xmm1
+; X64-SSE2-NEXT: movdqa %xmm2, %xmm3
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X64-SSE2-NEXT: pmullw %xmm3, %xmm2
+; X64-SSE2-NEXT: movdqa %xmm0, %xmm3
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X64-SSE2-NEXT: pmullw %xmm3, %xmm0
+; X64-SSE2-NEXT: pmullw %xmm2, %xmm0
+; X64-SSE2-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE2-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE2-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE2-NEXT: movdqa %xmm0, %xmm1
+; X64-SSE2-NEXT: psrld $16, %xmm1
+; X64-SSE2-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE2-NEXT: movd %xmm1, %eax
+; X64-SSE2-NEXT: # kill: def $al killed $al killed $eax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v64i8:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pushl %ebp
+; X86-SSE4-NEXT: movl %esp, %ebp
+; X86-SSE4-NEXT: andl $-16, %esp
+; X86-SSE4-NEXT: subl $16, %esp
+; X86-SSE4-NEXT: movdqa 8(%ebp), %xmm3
+; X86-SSE4-NEXT: pmovzxbw {{.*#+}} xmm4 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero,xmm3[4],zero,xmm3[5],zero,xmm3[6],zero,xmm3[7],zero
+; X86-SSE4-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE4-NEXT: pmullw %xmm4, %xmm3
+; X86-SSE4-NEXT: pmovzxbw {{.*#+}} xmm4 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
+; X86-SSE4-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE4-NEXT: pmullw %xmm4, %xmm1
+; X86-SSE4-NEXT: pmullw %xmm3, %xmm1
+; X86-SSE4-NEXT: pmovzxbw {{.*#+}} xmm3 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero,xmm2[4],zero,xmm2[5],zero,xmm2[6],zero,xmm2[7],zero
+; X86-SSE4-NEXT: punpckhbw {{.*#+}} xmm2 = xmm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE4-NEXT: pmullw %xmm3, %xmm2
+; X86-SSE4-NEXT: pmovzxbw {{.*#+}} xmm3 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; X86-SSE4-NEXT: punpckhbw {{.*#+}} xmm0 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE4-NEXT: pmullw %xmm3, %xmm0
+; X86-SSE4-NEXT: pmullw %xmm2, %xmm0
+; X86-SSE4-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE4-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm1
+; X86-SSE4-NEXT: psrld $16, %xmm1
+; X86-SSE4-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE4-NEXT: movd %xmm1, %eax
+; X86-SSE4-NEXT: # kill: def $al killed $al killed $eax
+; X86-SSE4-NEXT: movl %ebp, %esp
+; X86-SSE4-NEXT: popl %ebp
+; X86-SSE4-NEXT: retl
+;
+; X64-SSE4-LABEL: test_v64i8:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: pmovzxbw {{.*#+}} xmm4 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero,xmm3[4],zero,xmm3[5],zero,xmm3[6],zero,xmm3[7],zero
+; X64-SSE4-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE4-NEXT: pmullw %xmm4, %xmm3
+; X64-SSE4-NEXT: pmovzxbw {{.*#+}} xmm4 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
+; X64-SSE4-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE4-NEXT: pmullw %xmm4, %xmm1
+; X64-SSE4-NEXT: pmullw %xmm3, %xmm1
+; X64-SSE4-NEXT: pmovzxbw {{.*#+}} xmm3 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero,xmm2[4],zero,xmm2[5],zero,xmm2[6],zero,xmm2[7],zero
+; X64-SSE4-NEXT: punpckhbw {{.*#+}} xmm2 = xmm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE4-NEXT: pmullw %xmm3, %xmm2
+; X64-SSE4-NEXT: pmovzxbw {{.*#+}} xmm3 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; X64-SSE4-NEXT: punpckhbw {{.*#+}} xmm0 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE4-NEXT: pmullw %xmm3, %xmm0
+; X64-SSE4-NEXT: pmullw %xmm2, %xmm0
+; X64-SSE4-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE4-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE4-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE4-NEXT: movdqa %xmm0, %xmm1
+; X64-SSE4-NEXT: psrld $16, %xmm1
+; X64-SSE4-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE4-NEXT: movd %xmm1, %eax
+; X64-SSE4-NEXT: # kill: def $al killed $al killed $eax
+; X64-SSE4-NEXT: retq
;
; AVX1-LABEL: test_v64i8:
; AVX1: # %bb.0:
@@ -1889,7 +3127,7 @@ define i8 @test_v64i8(<64 x i8> %a0) {
; AVX1-NEXT: vmovd %xmm0, %eax
; AVX1-NEXT: # kill: def $al killed $al killed $eax
; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
+; AVX1-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v64i8:
; AVX2: # %bb.0:
@@ -1911,7 +3149,30 @@ define i8 @test_v64i8(<64 x i8> %a0) {
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: # kill: def $al killed $al killed $eax
; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
+;
+; AVX512F-LABEL: test_v64i8:
+; AVX512F: # %bb.0:
+; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm1
+; AVX512F-NEXT: vpunpckhbw {{.*#+}} ymm2 = ymm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
+; AVX512F-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
+; AVX512F-NEXT: vpmullw %ymm2, %ymm1, %ymm1
+; AVX512F-NEXT: vpunpckhbw {{.*#+}} ymm2 = ymm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
+; AVX512F-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
+; AVX512F-NEXT: vpmullw %ymm2, %ymm0, %ymm0
+; AVX512F-NEXT: vpmullw %ymm1, %ymm0, %ymm0
+; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
+; AVX512F-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512F-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; AVX512F-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpsrld $16, %xmm0, %xmm1
+; AVX512F-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vmovd %xmm0, %eax
+; AVX512F-NEXT: # kill: def $al killed $al killed $eax
+; AVX512F-NEXT: vzeroupper
+; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: test_v64i8:
; AVX512BW: # %bb.0:
@@ -2002,180 +3263,422 @@ define i8 @test_v64i8(<64 x i8> %a0) {
ret i8 %1
}
-define i8 @test_v128i8(<128 x i8> %a0) {
-; SSE2-LABEL: test_v128i8:
-; SSE2: # %bb.0:
-; SSE2-NEXT: movdqa %xmm7, %xmm8
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm8 = xmm8[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm7 = xmm7[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: pmullw %xmm8, %xmm7
-; SSE2-NEXT: movdqa %xmm3, %xmm8
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm8 = xmm8[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: pmullw %xmm8, %xmm3
-; SSE2-NEXT: pmullw %xmm7, %xmm3
-; SSE2-NEXT: movdqa %xmm5, %xmm7
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm7 = xmm7[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm5 = xmm5[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: pmullw %xmm7, %xmm5
-; SSE2-NEXT: movdqa %xmm1, %xmm7
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm7 = xmm7[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: pmullw %xmm7, %xmm1
-; SSE2-NEXT: pmullw %xmm5, %xmm1
-; SSE2-NEXT: pmullw %xmm3, %xmm1
-; SSE2-NEXT: movdqa %xmm6, %xmm3
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: pmullw %xmm3, %xmm6
-; SSE2-NEXT: movdqa %xmm2, %xmm3
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: pmullw %xmm3, %xmm2
-; SSE2-NEXT: pmullw %xmm6, %xmm2
-; SSE2-NEXT: movdqa %xmm4, %xmm3
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: pmullw %xmm3, %xmm4
-; SSE2-NEXT: movdqa %xmm0, %xmm3
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: pmullw %xmm3, %xmm0
-; SSE2-NEXT: pmullw %xmm4, %xmm0
-; SSE2-NEXT: pmullw %xmm2, %xmm0
-; SSE2-NEXT: pmullw %xmm1, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT: pmullw %xmm0, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE2-NEXT: pmullw %xmm1, %xmm0
-; SSE2-NEXT: movdqa %xmm0, %xmm1
-; SSE2-NEXT: psrld $16, %xmm1
-; SSE2-NEXT: pmullw %xmm0, %xmm1
-; SSE2-NEXT: movd %xmm1, %eax
-; SSE2-NEXT: # kill: def $al killed $al killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v128i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm8 = xmm7[0],zero,xmm7[1],zero,xmm7[2],zero,xmm7[3],zero,xmm7[4],zero,xmm7[5],zero,xmm7[6],zero,xmm7[7],zero
-; SSE41-NEXT: punpckhbw {{.*#+}} xmm7 = xmm7[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE41-NEXT: pmullw %xmm8, %xmm7
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm8 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero,xmm3[4],zero,xmm3[5],zero,xmm3[6],zero,xmm3[7],zero
-; SSE41-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE41-NEXT: pmullw %xmm8, %xmm3
-; SSE41-NEXT: pmullw %xmm7, %xmm3
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm7 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero,xmm5[4],zero,xmm5[5],zero,xmm5[6],zero,xmm5[7],zero
-; SSE41-NEXT: punpckhbw {{.*#+}} xmm5 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE41-NEXT: pmullw %xmm7, %xmm5
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm7 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
-; SSE41-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE41-NEXT: pmullw %xmm7, %xmm1
-; SSE41-NEXT: pmullw %xmm5, %xmm1
-; SSE41-NEXT: pmullw %xmm3, %xmm1
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm3 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero,xmm6[4],zero,xmm6[5],zero,xmm6[6],zero,xmm6[7],zero
-; SSE41-NEXT: punpckhbw {{.*#+}} xmm6 = xmm6[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE41-NEXT: pmullw %xmm3, %xmm6
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm3 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero,xmm2[4],zero,xmm2[5],zero,xmm2[6],zero,xmm2[7],zero
-; SSE41-NEXT: punpckhbw {{.*#+}} xmm2 = xmm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE41-NEXT: pmullw %xmm3, %xmm2
-; SSE41-NEXT: pmullw %xmm6, %xmm2
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm3 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero,xmm4[4],zero,xmm4[5],zero,xmm4[6],zero,xmm4[7],zero
-; SSE41-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE41-NEXT: pmullw %xmm3, %xmm4
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm3 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; SSE41-NEXT: punpckhbw {{.*#+}} xmm0 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE41-NEXT: pmullw %xmm3, %xmm0
-; SSE41-NEXT: pmullw %xmm4, %xmm0
-; SSE41-NEXT: pmullw %xmm2, %xmm0
-; SSE41-NEXT: pmullw %xmm1, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: pmullw %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: pmullw %xmm1, %xmm0
-; SSE41-NEXT: movdqa %xmm0, %xmm1
-; SSE41-NEXT: psrld $16, %xmm1
-; SSE41-NEXT: pmullw %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: # kill: def $al killed $al killed $eax
-; SSE41-NEXT: retq
-;
-; AVX1-LABEL: test_v128i8:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm4
-; AVX1-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX1-NEXT: vpmovzxbw {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero,xmm4[4],zero,xmm4[5],zero,xmm4[6],zero,xmm4[7],zero
-; AVX1-NEXT: vpmullw %xmm5, %xmm4, %xmm4
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm5
-; AVX1-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX1-NEXT: vpmovzxbw {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero,xmm5[4],zero,xmm5[5],zero,xmm5[6],zero,xmm5[7],zero
-; AVX1-NEXT: vpmullw %xmm6, %xmm5, %xmm5
-; AVX1-NEXT: vpmullw %xmm4, %xmm5, %xmm4
-; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm5
-; AVX1-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX1-NEXT: vpmovzxbw {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero,xmm5[4],zero,xmm5[5],zero,xmm5[6],zero,xmm5[7],zero
-; AVX1-NEXT: vpmullw %xmm6, %xmm5, %xmm5
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm6
-; AVX1-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm6[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX1-NEXT: vpmovzxbw {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero,xmm6[4],zero,xmm6[5],zero,xmm6[6],zero,xmm6[7],zero
-; AVX1-NEXT: vpmullw %xmm7, %xmm6, %xmm6
-; AVX1-NEXT: vpmullw %xmm5, %xmm6, %xmm5
-; AVX1-NEXT: vpmullw %xmm4, %xmm5, %xmm4
-; AVX1-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX1-NEXT: vpmovzxbw {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero,xmm3[4],zero,xmm3[5],zero,xmm3[6],zero,xmm3[7],zero
-; AVX1-NEXT: vpmullw %xmm5, %xmm3, %xmm3
-; AVX1-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX1-NEXT: vpmovzxbw {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
-; AVX1-NEXT: vpmullw %xmm5, %xmm1, %xmm1
-; AVX1-NEXT: vpmullw %xmm3, %xmm1, %xmm1
-; AVX1-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX1-NEXT: vpmovzxbw {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero,xmm2[4],zero,xmm2[5],zero,xmm2[6],zero,xmm2[7],zero
-; AVX1-NEXT: vpmullw %xmm3, %xmm2, %xmm2
-; AVX1-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; AVX1-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; AVX1-NEXT: vpmullw %xmm3, %xmm0, %xmm0
-; AVX1-NEXT: vpmullw %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpmullw %xmm4, %xmm0, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpsrld $16, %xmm0, %xmm1
-; AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vmovd %xmm0, %eax
-; AVX1-NEXT: # kill: def $al killed $al killed $eax
-; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
+define i8 @test_v128i8(<128 x i8> %a0) nounwind {
+; X86-SSE2-LABEL: test_v128i8:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: pushl %ebp
+; X86-SSE2-NEXT: movl %esp, %ebp
+; X86-SSE2-NEXT: andl $-16, %esp
+; X86-SSE2-NEXT: subl $16, %esp
+; X86-SSE2-NEXT: movdqa 8(%ebp), %xmm3
+; X86-SSE2-NEXT: movdqa 72(%ebp), %xmm4
+; X86-SSE2-NEXT: movdqa %xmm4, %xmm5
+; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm5 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X86-SSE2-NEXT: pmullw %xmm5, %xmm4
+; X86-SSE2-NEXT: movdqa %xmm3, %xmm5
+; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm5 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X86-SSE2-NEXT: pmullw %xmm5, %xmm3
+; X86-SSE2-NEXT: movdqa 40(%ebp), %xmm5
+; X86-SSE2-NEXT: pmullw %xmm4, %xmm3
+; X86-SSE2-NEXT: movdqa %xmm5, %xmm4
+; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm5 = xmm5[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X86-SSE2-NEXT: pmullw %xmm4, %xmm5
+; X86-SSE2-NEXT: movdqa %xmm1, %xmm4
+; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X86-SSE2-NEXT: pmullw %xmm4, %xmm1
+; X86-SSE2-NEXT: pmullw %xmm5, %xmm1
+; X86-SSE2-NEXT: movdqa 56(%ebp), %xmm4
+; X86-SSE2-NEXT: pmullw %xmm3, %xmm1
+; X86-SSE2-NEXT: movdqa %xmm4, %xmm3
+; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X86-SSE2-NEXT: pmullw %xmm3, %xmm4
+; X86-SSE2-NEXT: movdqa %xmm2, %xmm3
+; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X86-SSE2-NEXT: pmullw %xmm3, %xmm2
+; X86-SSE2-NEXT: movdqa 24(%ebp), %xmm3
+; X86-SSE2-NEXT: pmullw %xmm4, %xmm2
+; X86-SSE2-NEXT: movdqa %xmm3, %xmm4
+; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X86-SSE2-NEXT: pmullw %xmm4, %xmm3
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm4
+; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X86-SSE2-NEXT: pmullw %xmm4, %xmm0
+; X86-SSE2-NEXT: pmullw %xmm3, %xmm0
+; X86-SSE2-NEXT: pmullw %xmm2, %xmm0
+; X86-SSE2-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE2-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE2-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm1
+; X86-SSE2-NEXT: psrld $16, %xmm1
+; X86-SSE2-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE2-NEXT: movd %xmm1, %eax
+; X86-SSE2-NEXT: # kill: def $al killed $al killed $eax
+; X86-SSE2-NEXT: movl %ebp, %esp
+; X86-SSE2-NEXT: popl %ebp
+; X86-SSE2-NEXT: retl
;
-; AVX2-LABEL: test_v128i8:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
-; AVX2-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm3[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
-; AVX2-NEXT: vpmullw %ymm4, %ymm3, %ymm3
-; AVX2-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
-; AVX2-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
-; AVX2-NEXT: vpmullw %ymm4, %ymm1, %ymm1
-; AVX2-NEXT: vpmullw %ymm3, %ymm1, %ymm1
-; AVX2-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
-; AVX2-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
-; AVX2-NEXT: vpmullw %ymm3, %ymm2, %ymm2
-; AVX2-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
-; AVX2-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
-; AVX2-NEXT: vpmullw %ymm3, %ymm0, %ymm0
-; AVX2-NEXT: vpmullw %ymm2, %ymm0, %ymm0
-; AVX2-NEXT: vpmullw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
-; AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: # kill: def $al killed $al killed $eax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+; X64-SSE2-LABEL: test_v128i8:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: movdqa %xmm7, %xmm8
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm8 = xmm8[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm7 = xmm7[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X64-SSE2-NEXT: pmullw %xmm8, %xmm7
+; X64-SSE2-NEXT: movdqa %xmm3, %xmm8
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm8 = xmm8[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X64-SSE2-NEXT: pmullw %xmm8, %xmm3
+; X64-SSE2-NEXT: pmullw %xmm7, %xmm3
+; X64-SSE2-NEXT: movdqa %xmm5, %xmm7
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm7 = xmm7[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm5 = xmm5[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X64-SSE2-NEXT: pmullw %xmm7, %xmm5
+; X64-SSE2-NEXT: movdqa %xmm1, %xmm7
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm7 = xmm7[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X64-SSE2-NEXT: pmullw %xmm7, %xmm1
+; X64-SSE2-NEXT: pmullw %xmm5, %xmm1
+; X64-SSE2-NEXT: pmullw %xmm3, %xmm1
+; X64-SSE2-NEXT: movdqa %xmm6, %xmm3
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X64-SSE2-NEXT: pmullw %xmm3, %xmm6
+; X64-SSE2-NEXT: movdqa %xmm2, %xmm3
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X64-SSE2-NEXT: pmullw %xmm3, %xmm2
+; X64-SSE2-NEXT: pmullw %xmm6, %xmm2
+; X64-SSE2-NEXT: movdqa %xmm4, %xmm3
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X64-SSE2-NEXT: pmullw %xmm3, %xmm4
+; X64-SSE2-NEXT: movdqa %xmm0, %xmm3
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X64-SSE2-NEXT: pmullw %xmm3, %xmm0
+; X64-SSE2-NEXT: pmullw %xmm4, %xmm0
+; X64-SSE2-NEXT: pmullw %xmm2, %xmm0
+; X64-SSE2-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE2-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE2-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE2-NEXT: movdqa %xmm0, %xmm1
+; X64-SSE2-NEXT: psrld $16, %xmm1
+; X64-SSE2-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE2-NEXT: movd %xmm1, %eax
+; X64-SSE2-NEXT: # kill: def $al killed $al killed $eax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v128i8:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pushl %ebp
+; X86-SSE4-NEXT: movl %esp, %ebp
+; X86-SSE4-NEXT: andl $-16, %esp
+; X86-SSE4-NEXT: subl $16, %esp
+; X86-SSE4-NEXT: movdqa 8(%ebp), %xmm3
+; X86-SSE4-NEXT: movdqa 72(%ebp), %xmm4
+; X86-SSE4-NEXT: pmovzxbw {{.*#+}} xmm5 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero,xmm4[4],zero,xmm4[5],zero,xmm4[6],zero,xmm4[7],zero
+; X86-SSE4-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE4-NEXT: pmullw %xmm5, %xmm4
+; X86-SSE4-NEXT: pmovzxbw {{.*#+}} xmm5 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero,xmm3[4],zero,xmm3[5],zero,xmm3[6],zero,xmm3[7],zero
+; X86-SSE4-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE4-NEXT: pmullw %xmm5, %xmm3
+; X86-SSE4-NEXT: movdqa 40(%ebp), %xmm5
+; X86-SSE4-NEXT: pmullw %xmm4, %xmm3
+; X86-SSE4-NEXT: pmovzxbw {{.*#+}} xmm4 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero,xmm5[4],zero,xmm5[5],zero,xmm5[6],zero,xmm5[7],zero
+; X86-SSE4-NEXT: punpckhbw {{.*#+}} xmm5 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE4-NEXT: pmullw %xmm4, %xmm5
+; X86-SSE4-NEXT: pmovzxbw {{.*#+}} xmm4 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
+; X86-SSE4-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE4-NEXT: pmullw %xmm4, %xmm1
+; X86-SSE4-NEXT: pmullw %xmm5, %xmm1
+; X86-SSE4-NEXT: movdqa 56(%ebp), %xmm4
+; X86-SSE4-NEXT: pmullw %xmm3, %xmm1
+; X86-SSE4-NEXT: pmovzxbw {{.*#+}} xmm3 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero,xmm4[4],zero,xmm4[5],zero,xmm4[6],zero,xmm4[7],zero
+; X86-SSE4-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE4-NEXT: pmullw %xmm3, %xmm4
+; X86-SSE4-NEXT: pmovzxbw {{.*#+}} xmm3 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero,xmm2[4],zero,xmm2[5],zero,xmm2[6],zero,xmm2[7],zero
+; X86-SSE4-NEXT: punpckhbw {{.*#+}} xmm2 = xmm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE4-NEXT: pmullw %xmm3, %xmm2
+; X86-SSE4-NEXT: movdqa 24(%ebp), %xmm3
+; X86-SSE4-NEXT: pmullw %xmm4, %xmm2
+; X86-SSE4-NEXT: pmovzxbw {{.*#+}} xmm4 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero,xmm3[4],zero,xmm3[5],zero,xmm3[6],zero,xmm3[7],zero
+; X86-SSE4-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE4-NEXT: pmullw %xmm4, %xmm3
+; X86-SSE4-NEXT: pmovzxbw {{.*#+}} xmm4 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; X86-SSE4-NEXT: punpckhbw {{.*#+}} xmm0 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE4-NEXT: pmullw %xmm4, %xmm0
+; X86-SSE4-NEXT: pmullw %xmm3, %xmm0
+; X86-SSE4-NEXT: pmullw %xmm2, %xmm0
+; X86-SSE4-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE4-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm1
+; X86-SSE4-NEXT: psrld $16, %xmm1
+; X86-SSE4-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE4-NEXT: movd %xmm1, %eax
+; X86-SSE4-NEXT: # kill: def $al killed $al killed $eax
+; X86-SSE4-NEXT: movl %ebp, %esp
+; X86-SSE4-NEXT: popl %ebp
+; X86-SSE4-NEXT: retl
+;
+; X64-SSE4-LABEL: test_v128i8:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: pmovzxbw {{.*#+}} xmm8 = xmm7[0],zero,xmm7[1],zero,xmm7[2],zero,xmm7[3],zero,xmm7[4],zero,xmm7[5],zero,xmm7[6],zero,xmm7[7],zero
+; X64-SSE4-NEXT: punpckhbw {{.*#+}} xmm7 = xmm7[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE4-NEXT: pmullw %xmm8, %xmm7
+; X64-SSE4-NEXT: pmovzxbw {{.*#+}} xmm8 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero,xmm3[4],zero,xmm3[5],zero,xmm3[6],zero,xmm3[7],zero
+; X64-SSE4-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE4-NEXT: pmullw %xmm8, %xmm3
+; X64-SSE4-NEXT: pmullw %xmm7, %xmm3
+; X64-SSE4-NEXT: pmovzxbw {{.*#+}} xmm7 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero,xmm5[4],zero,xmm5[5],zero,xmm5[6],zero,xmm5[7],zero
+; X64-SSE4-NEXT: punpckhbw {{.*#+}} xmm5 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE4-NEXT: pmullw %xmm7, %xmm5
+; X64-SSE4-NEXT: pmovzxbw {{.*#+}} xmm7 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
+; X64-SSE4-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE4-NEXT: pmullw %xmm7, %xmm1
+; X64-SSE4-NEXT: pmullw %xmm5, %xmm1
+; X64-SSE4-NEXT: pmullw %xmm3, %xmm1
+; X64-SSE4-NEXT: pmovzxbw {{.*#+}} xmm3 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero,xmm6[4],zero,xmm6[5],zero,xmm6[6],zero,xmm6[7],zero
+; X64-SSE4-NEXT: punpckhbw {{.*#+}} xmm6 = xmm6[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE4-NEXT: pmullw %xmm3, %xmm6
+; X64-SSE4-NEXT: pmovzxbw {{.*#+}} xmm3 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero,xmm2[4],zero,xmm2[5],zero,xmm2[6],zero,xmm2[7],zero
+; X64-SSE4-NEXT: punpckhbw {{.*#+}} xmm2 = xmm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE4-NEXT: pmullw %xmm3, %xmm2
+; X64-SSE4-NEXT: pmullw %xmm6, %xmm2
+; X64-SSE4-NEXT: pmovzxbw {{.*#+}} xmm3 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero,xmm4[4],zero,xmm4[5],zero,xmm4[6],zero,xmm4[7],zero
+; X64-SSE4-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE4-NEXT: pmullw %xmm3, %xmm4
+; X64-SSE4-NEXT: pmovzxbw {{.*#+}} xmm3 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; X64-SSE4-NEXT: punpckhbw {{.*#+}} xmm0 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE4-NEXT: pmullw %xmm3, %xmm0
+; X64-SSE4-NEXT: pmullw %xmm4, %xmm0
+; X64-SSE4-NEXT: pmullw %xmm2, %xmm0
+; X64-SSE4-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE4-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE4-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE4-NEXT: movdqa %xmm0, %xmm1
+; X64-SSE4-NEXT: psrld $16, %xmm1
+; X64-SSE4-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE4-NEXT: movd %xmm1, %eax
+; X64-SSE4-NEXT: # kill: def $al killed $al killed $eax
+; X64-SSE4-NEXT: retq
+;
+; X86-AVX1-LABEL: test_v128i8:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: pushl %ebp
+; X86-AVX1-NEXT: movl %esp, %ebp
+; X86-AVX1-NEXT: andl $-32, %esp
+; X86-AVX1-NEXT: subl $32, %esp
+; X86-AVX1-NEXT: vmovdqa 8(%ebp), %xmm3
+; X86-AVX1-NEXT: vmovdqa 24(%ebp), %xmm4
+; X86-AVX1-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero,xmm4[4],zero,xmm4[5],zero,xmm4[6],zero,xmm4[7],zero
+; X86-AVX1-NEXT: vpmullw %xmm5, %xmm4, %xmm4
+; X86-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm5
+; X86-AVX1-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero,xmm5[4],zero,xmm5[5],zero,xmm5[6],zero,xmm5[7],zero
+; X86-AVX1-NEXT: vpmullw %xmm6, %xmm5, %xmm5
+; X86-AVX1-NEXT: vpmullw %xmm4, %xmm5, %xmm4
+; X86-AVX1-NEXT: vextractf128 $1, %ymm2, %xmm5
+; X86-AVX1-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero,xmm5[4],zero,xmm5[5],zero,xmm5[6],zero,xmm5[7],zero
+; X86-AVX1-NEXT: vpmullw %xmm6, %xmm5, %xmm5
+; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm6
+; X86-AVX1-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm6[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero,xmm6[4],zero,xmm6[5],zero,xmm6[6],zero,xmm6[7],zero
+; X86-AVX1-NEXT: vpmullw %xmm7, %xmm6, %xmm6
+; X86-AVX1-NEXT: vpmullw %xmm5, %xmm6, %xmm5
+; X86-AVX1-NEXT: vpmullw %xmm4, %xmm5, %xmm4
+; X86-AVX1-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero,xmm3[4],zero,xmm3[5],zero,xmm3[6],zero,xmm3[7],zero
+; X86-AVX1-NEXT: vpmullw %xmm5, %xmm3, %xmm3
+; X86-AVX1-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
+; X86-AVX1-NEXT: vpmullw %xmm5, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpmullw %xmm3, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero,xmm2[4],zero,xmm2[5],zero,xmm2[6],zero,xmm2[7],zero
+; X86-AVX1-NEXT: vpmullw %xmm3, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; X86-AVX1-NEXT: vpmullw %xmm3, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpmullw %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpmullw %xmm4, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpsrld $16, %xmm0, %xmm1
+; X86-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: # kill: def $al killed $al killed $eax
+; X86-AVX1-NEXT: movl %ebp, %esp
+; X86-AVX1-NEXT: popl %ebp
+; X86-AVX1-NEXT: vzeroupper
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v128i8:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vextractf128 $1, %ymm3, %xmm4
+; X64-AVX1-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm4[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm4 = xmm4[0],zero,xmm4[1],zero,xmm4[2],zero,xmm4[3],zero,xmm4[4],zero,xmm4[5],zero,xmm4[6],zero,xmm4[7],zero
+; X64-AVX1-NEXT: vpmullw %xmm5, %xmm4, %xmm4
+; X64-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm5
+; X64-AVX1-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero,xmm5[4],zero,xmm5[5],zero,xmm5[6],zero,xmm5[7],zero
+; X64-AVX1-NEXT: vpmullw %xmm6, %xmm5, %xmm5
+; X64-AVX1-NEXT: vpmullw %xmm4, %xmm5, %xmm4
+; X64-AVX1-NEXT: vextractf128 $1, %ymm2, %xmm5
+; X64-AVX1-NEXT: vpunpckhbw {{.*#+}} xmm6 = xmm5[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm5 = xmm5[0],zero,xmm5[1],zero,xmm5[2],zero,xmm5[3],zero,xmm5[4],zero,xmm5[5],zero,xmm5[6],zero,xmm5[7],zero
+; X64-AVX1-NEXT: vpmullw %xmm6, %xmm5, %xmm5
+; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm6
+; X64-AVX1-NEXT: vpunpckhbw {{.*#+}} xmm7 = xmm6[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm6 = xmm6[0],zero,xmm6[1],zero,xmm6[2],zero,xmm6[3],zero,xmm6[4],zero,xmm6[5],zero,xmm6[6],zero,xmm6[7],zero
+; X64-AVX1-NEXT: vpmullw %xmm7, %xmm6, %xmm6
+; X64-AVX1-NEXT: vpmullw %xmm5, %xmm6, %xmm5
+; X64-AVX1-NEXT: vpmullw %xmm4, %xmm5, %xmm4
+; X64-AVX1-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm3 = xmm3[0],zero,xmm3[1],zero,xmm3[2],zero,xmm3[3],zero,xmm3[4],zero,xmm3[5],zero,xmm3[6],zero,xmm3[7],zero
+; X64-AVX1-NEXT: vpmullw %xmm5, %xmm3, %xmm3
+; X64-AVX1-NEXT: vpunpckhbw {{.*#+}} xmm5 = xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
+; X64-AVX1-NEXT: vpmullw %xmm5, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpmullw %xmm3, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm2 = xmm2[0],zero,xmm2[1],zero,xmm2[2],zero,xmm2[3],zero,xmm2[4],zero,xmm2[5],zero,xmm2[6],zero,xmm2[7],zero
+; X64-AVX1-NEXT: vpmullw %xmm3, %xmm2, %xmm2
+; X64-AVX1-NEXT: vpunpckhbw {{.*#+}} xmm3 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; X64-AVX1-NEXT: vpmullw %xmm3, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpmullw %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpmullw %xmm4, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpsrld $16, %xmm0, %xmm1
+; X64-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-NEXT: # kill: def $al killed $al killed $eax
+; X64-AVX1-NEXT: vzeroupper
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v128i8:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: pushl %ebp
+; X86-AVX2-NEXT: movl %esp, %ebp
+; X86-AVX2-NEXT: andl $-32, %esp
+; X86-AVX2-NEXT: subl $32, %esp
+; X86-AVX2-NEXT: vmovdqa 8(%ebp), %ymm3
+; X86-AVX2-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
+; X86-AVX2-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm3[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
+; X86-AVX2-NEXT: vpmullw %ymm4, %ymm3, %ymm3
+; X86-AVX2-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
+; X86-AVX2-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
+; X86-AVX2-NEXT: vpmullw %ymm4, %ymm1, %ymm1
+; X86-AVX2-NEXT: vpmullw %ymm3, %ymm1, %ymm1
+; X86-AVX2-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
+; X86-AVX2-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
+; X86-AVX2-NEXT: vpmullw %ymm3, %ymm2, %ymm2
+; X86-AVX2-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
+; X86-AVX2-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
+; X86-AVX2-NEXT: vpmullw %ymm3, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpmullw %ymm2, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpmullw %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
+; X86-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: # kill: def $al killed $al killed $eax
+; X86-AVX2-NEXT: movl %ebp, %esp
+; X86-AVX2-NEXT: popl %ebp
+; X86-AVX2-NEXT: vzeroupper
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v128i8:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
+; X64-AVX2-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm3[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
+; X64-AVX2-NEXT: vpmullw %ymm4, %ymm3, %ymm3
+; X64-AVX2-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
+; X64-AVX2-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
+; X64-AVX2-NEXT: vpmullw %ymm4, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpmullw %ymm3, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
+; X64-AVX2-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
+; X64-AVX2-NEXT: vpmullw %ymm3, %ymm2, %ymm2
+; X64-AVX2-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
+; X64-AVX2-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
+; X64-AVX2-NEXT: vpmullw %ymm3, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpmullw %ymm2, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpmullw %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
+; X64-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovd %xmm0, %eax
+; X64-AVX2-NEXT: # kill: def $al killed $al killed $eax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
+;
+; AVX512F-LABEL: test_v128i8:
+; AVX512F: # %bb.0:
+; AVX512F-NEXT: vextracti64x4 $1, %zmm1, %ymm2
+; AVX512F-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm2[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
+; AVX512F-NEXT: vpunpcklbw {{.*#+}} ymm2 = ymm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
+; AVX512F-NEXT: vpmullw %ymm3, %ymm2, %ymm2
+; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm3
+; AVX512F-NEXT: vpunpckhbw {{.*#+}} ymm4 = ymm3[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
+; AVX512F-NEXT: vpunpcklbw {{.*#+}} ymm3 = ymm3[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
+; AVX512F-NEXT: vpmullw %ymm4, %ymm3, %ymm3
+; AVX512F-NEXT: vpmullw %ymm2, %ymm3, %ymm2
+; AVX512F-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
+; AVX512F-NEXT: vpunpcklbw {{.*#+}} ymm1 = ymm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
+; AVX512F-NEXT: vpmullw %ymm3, %ymm1, %ymm1
+; AVX512F-NEXT: vpunpckhbw {{.*#+}} ymm3 = ymm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15,24,24,25,25,26,26,27,27,28,28,29,29,30,30,31,31]
+; AVX512F-NEXT: vpunpcklbw {{.*#+}} ymm0 = ymm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,16,16,17,17,18,18,19,19,20,20,21,21,22,22,23,23]
+; AVX512F-NEXT: vpmullw %ymm3, %ymm0, %ymm0
+; AVX512F-NEXT: vpmullw %ymm1, %ymm0, %ymm0
+; AVX512F-NEXT: vpmullw %ymm2, %ymm0, %ymm0
+; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
+; AVX512F-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512F-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; AVX512F-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vpsrld $16, %xmm0, %xmm1
+; AVX512F-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512F-NEXT: vmovd %xmm0, %eax
+; AVX512F-NEXT: # kill: def $al killed $al killed $eax
+; AVX512F-NEXT: vzeroupper
+; AVX512F-NEXT: retq
;
; AVX512BW-LABEL: test_v128i8:
; AVX512BW: # %bb.0:
@@ -2296,163 +3799,467 @@ define i8 @test_v128i8(<128 x i8> %a0) {
; Legalization
;
-define i8 @illegal_v4i8(i8 %a0, ptr %a1) {
-; SSE2-LABEL: illegal_v4i8:
-; SSE2: # %bb.0:
-; SSE2-NEXT: movl %edi, %eax
-; SSE2-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE2-NEXT: pmullw %xmm0, %xmm1
-; SSE2-NEXT: movdqa %xmm1, %xmm0
-; SSE2-NEXT: psrld $16, %xmm0
-; SSE2-NEXT: pmullw %xmm1, %xmm0
-; SSE2-NEXT: movd %xmm0, %ecx
-; SSE2-NEXT: # kill: def $al killed $al killed $eax
-; SSE2-NEXT: mulb %cl
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: illegal_v4i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: movl %edi, %eax
-; SSE41-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE41-NEXT: pmullw %xmm0, %xmm1
-; SSE41-NEXT: movdqa %xmm1, %xmm0
-; SSE41-NEXT: psrld $16, %xmm0
-; SSE41-NEXT: pmullw %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %ecx
-; SSE41-NEXT: # kill: def $al killed $al killed $eax
-; SSE41-NEXT: mulb %cl
-; SSE41-NEXT: retq
-;
-; AVX-LABEL: illegal_v4i8:
-; AVX: # %bb.0:
-; AVX-NEXT: movl %edi, %eax
-; AVX-NEXT: vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vpsrld $16, %xmm0, %xmm1
-; AVX-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vmovd %xmm0, %ecx
-; AVX-NEXT: # kill: def $al killed $al killed $eax
-; AVX-NEXT: mulb %cl
-; AVX-NEXT: retq
+define i8 @illegal_v4i8(i8 %a0, ptr %a1) nounwind {
+; X86-SSE2-LABEL: illegal_v4i8:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
+; X86-SSE2-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-SSE2-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE2-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE2-NEXT: movdqa %xmm1, %xmm0
+; X86-SSE2-NEXT: psrld $16, %xmm0
+; X86-SSE2-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE2-NEXT: movd %xmm0, %ecx
+; X86-SSE2-NEXT: mulb %cl
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE2-LABEL: illegal_v4i8:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: movl %edi, %eax
+; X64-SSE2-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-SSE2-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE2-NEXT: movdqa %xmm1, %xmm0
+; X64-SSE2-NEXT: psrld $16, %xmm0
+; X64-SSE2-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE2-NEXT: movd %xmm0, %ecx
+; X64-SSE2-NEXT: # kill: def $al killed $al killed $eax
+; X64-SSE2-NEXT: mulb %cl
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: illegal_v4i8:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: movzbl {{[0-9]+}}(%esp), %eax
+; X86-SSE4-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-SSE4-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X86-SSE4-NEXT: pmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE4-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE4-NEXT: movdqa %xmm1, %xmm0
+; X86-SSE4-NEXT: psrld $16, %xmm0
+; X86-SSE4-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE4-NEXT: movd %xmm0, %ecx
+; X86-SSE4-NEXT: mulb %cl
+; X86-SSE4-NEXT: retl
+;
+; X64-SSE4-LABEL: illegal_v4i8:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: movl %edi, %eax
+; X64-SSE4-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X64-SSE4-NEXT: pmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-SSE4-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE4-NEXT: movdqa %xmm1, %xmm0
+; X64-SSE4-NEXT: psrld $16, %xmm0
+; X64-SSE4-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE4-NEXT: movd %xmm0, %ecx
+; X64-SSE4-NEXT: # kill: def $al killed $al killed $eax
+; X64-SSE4-NEXT: mulb %cl
+; X64-SSE4-NEXT: retq
+;
+; X86-AVX1-LABEL: illegal_v4i8:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: movzbl {{[0-9]+}}(%esp), %eax
+; X86-AVX1-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX1-NEXT: vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X86-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpsrld $16, %xmm0, %xmm1
+; X86-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %ecx
+; X86-AVX1-NEXT: mulb %cl
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: illegal_v4i8:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: movl %edi, %eax
+; X64-AVX1-NEXT: vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X64-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpsrld $16, %xmm0, %xmm1
+; X64-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovd %xmm0, %ecx
+; X64-AVX1-NEXT: # kill: def $al killed $al killed $eax
+; X64-AVX1-NEXT: mulb %cl
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: illegal_v4i8:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
+; X86-AVX2-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX2-NEXT: vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X86-AVX2-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
+; X86-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %ecx
+; X86-AVX2-NEXT: mulb %cl
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: illegal_v4i8:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: movl %edi, %eax
+; X64-AVX2-NEXT: vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X64-AVX2-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
+; X64-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovd %xmm0, %ecx
+; X64-AVX2-NEXT: # kill: def $al killed $al killed $eax
+; X64-AVX2-NEXT: mulb %cl
+; X64-AVX2-NEXT: retq
+;
+; AVX512-LABEL: illegal_v4i8:
+; AVX512: # %bb.0:
+; AVX512-NEXT: movl %edi, %eax
+; AVX512-NEXT: vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; AVX512-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; AVX512-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: vpsrld $16, %xmm0, %xmm1
+; AVX512-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: vmovd %xmm0, %ecx
+; AVX512-NEXT: # kill: def $al killed $al killed $eax
+; AVX512-NEXT: mulb %cl
+; AVX512-NEXT: retq
%ld = load <4 x i8>, ptr %a1, align 4
%rdx = call i8 @llvm.vector.reduce.mul.v4i8(<4 x i8> %ld)
%mul = mul i8 %a0, %rdx
ret i8 %mul
}
-define i8 @illegal_v8i8(i8 %a0, ptr %a1) {
-; SSE2-LABEL: illegal_v8i8:
-; SSE2: # %bb.0:
-; SSE2-NEXT: movl %edi, %eax
-; SSE2-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT: pmullw %xmm0, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE2-NEXT: pmullw %xmm1, %xmm0
-; SSE2-NEXT: movdqa %xmm0, %xmm1
-; SSE2-NEXT: psrld $16, %xmm1
-; SSE2-NEXT: pmullw %xmm0, %xmm1
-; SSE2-NEXT: movd %xmm1, %ecx
-; SSE2-NEXT: # kill: def $al killed $al killed $eax
-; SSE2-NEXT: mulb %cl
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: illegal_v8i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: movl %edi, %eax
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm0 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: pmullw %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: pmullw %xmm1, %xmm0
-; SSE41-NEXT: movdqa %xmm0, %xmm1
-; SSE41-NEXT: psrld $16, %xmm1
-; SSE41-NEXT: pmullw %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %ecx
-; SSE41-NEXT: # kill: def $al killed $al killed $eax
-; SSE41-NEXT: mulb %cl
-; SSE41-NEXT: retq
-;
-; AVX-LABEL: illegal_v8i8:
-; AVX: # %bb.0:
-; AVX-NEXT: movl %edi, %eax
-; AVX-NEXT: vpmovzxbw {{.*#+}} xmm0 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero
-; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vpsrld $16, %xmm0, %xmm1
-; AVX-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vmovd %xmm0, %ecx
-; AVX-NEXT: # kill: def $al killed $al killed $eax
-; AVX-NEXT: mulb %cl
-; AVX-NEXT: retq
+define i8 @illegal_v8i8(i8 %a0, ptr %a1) nounwind {
+; X86-SSE2-LABEL: illegal_v8i8:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
+; X86-SSE2-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-SSE2-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE2-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE2-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm1
+; X86-SSE2-NEXT: psrld $16, %xmm1
+; X86-SSE2-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE2-NEXT: movd %xmm1, %ecx
+; X86-SSE2-NEXT: mulb %cl
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE2-LABEL: illegal_v8i8:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: movl %edi, %eax
+; X64-SSE2-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE2-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE2-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE2-NEXT: movdqa %xmm0, %xmm1
+; X64-SSE2-NEXT: psrld $16, %xmm1
+; X64-SSE2-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE2-NEXT: movd %xmm1, %ecx
+; X64-SSE2-NEXT: # kill: def $al killed $al killed $eax
+; X64-SSE2-NEXT: mulb %cl
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: illegal_v8i8:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: movzbl {{[0-9]+}}(%esp), %eax
+; X86-SSE4-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-SSE4-NEXT: pmovzxbw {{.*#+}} xmm0 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE4-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm1
+; X86-SSE4-NEXT: psrld $16, %xmm1
+; X86-SSE4-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE4-NEXT: movd %xmm1, %ecx
+; X86-SSE4-NEXT: mulb %cl
+; X86-SSE4-NEXT: retl
+;
+; X64-SSE4-LABEL: illegal_v8i8:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: movl %edi, %eax
+; X64-SSE4-NEXT: pmovzxbw {{.*#+}} xmm0 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE4-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE4-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE4-NEXT: movdqa %xmm0, %xmm1
+; X64-SSE4-NEXT: psrld $16, %xmm1
+; X64-SSE4-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE4-NEXT: movd %xmm1, %ecx
+; X64-SSE4-NEXT: # kill: def $al killed $al killed $eax
+; X64-SSE4-NEXT: mulb %cl
+; X64-SSE4-NEXT: retq
+;
+; X86-AVX1-LABEL: illegal_v8i8:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: movzbl {{[0-9]+}}(%esp), %eax
+; X86-AVX1-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm0 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpsrld $16, %xmm0, %xmm1
+; X86-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %ecx
+; X86-AVX1-NEXT: mulb %cl
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: illegal_v8i8:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: movl %edi, %eax
+; X64-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm0 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpsrld $16, %xmm0, %xmm1
+; X64-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovd %xmm0, %ecx
+; X64-AVX1-NEXT: # kill: def $al killed $al killed $eax
+; X64-AVX1-NEXT: mulb %cl
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: illegal_v8i8:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: movzbl {{[0-9]+}}(%esp), %eax
+; X86-AVX2-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-AVX2-NEXT: vpmovzxbw {{.*#+}} xmm0 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
+; X86-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %ecx
+; X86-AVX2-NEXT: mulb %cl
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: illegal_v8i8:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: movl %edi, %eax
+; X64-AVX2-NEXT: vpmovzxbw {{.*#+}} xmm0 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
+; X64-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovd %xmm0, %ecx
+; X64-AVX2-NEXT: # kill: def $al killed $al killed $eax
+; X64-AVX2-NEXT: mulb %cl
+; X64-AVX2-NEXT: retq
+;
+; AVX512-LABEL: illegal_v8i8:
+; AVX512: # %bb.0:
+; AVX512-NEXT: movl %edi, %eax
+; AVX512-NEXT: vpmovzxbw {{.*#+}} xmm0 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero
+; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; AVX512-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: vpsrld $16, %xmm0, %xmm1
+; AVX512-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: vmovd %xmm0, %ecx
+; AVX512-NEXT: # kill: def $al killed $al killed $eax
+; AVX512-NEXT: mulb %cl
+; AVX512-NEXT: retq
%ld = load <8 x i8>, ptr %a1, align 4
%rdx = call i8 @llvm.vector.reduce.mul.v8i8(<8 x i8> %ld)
%mul = mul i8 %a0, %rdx
ret i8 %mul
}
-define i8 @PR51858(i128 %arg) {
-; SSE2-LABEL: PR51858:
-; SSE2: # %bb.0:
-; SSE2-NEXT: movq %rdi, %xmm0
-; SSE2-NEXT: movq %rsi, %xmm1
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: pmullw %xmm1, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT: pmullw %xmm0, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE2-NEXT: pmullw %xmm1, %xmm0
-; SSE2-NEXT: movdqa %xmm0, %xmm1
-; SSE2-NEXT: psrld $16, %xmm1
-; SSE2-NEXT: pmullw %xmm0, %xmm1
-; SSE2-NEXT: movd %xmm1, %eax
-; SSE2-NEXT: # kill: def $al killed $al killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: PR51858:
-; SSE41: # %bb.0:
-; SSE41-NEXT: movq %rdi, %xmm0
-; SSE41-NEXT: movq %rsi, %xmm1
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; SSE41-NEXT: pmullw %xmm1, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: pmullw %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: pmullw %xmm1, %xmm0
-; SSE41-NEXT: movdqa %xmm0, %xmm1
-; SSE41-NEXT: psrld $16, %xmm1
-; SSE41-NEXT: pmullw %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: # kill: def $al killed $al killed $eax
-; SSE41-NEXT: retq
-;
-; AVX-LABEL: PR51858:
-; AVX: # %bb.0:
-; AVX-NEXT: vmovq %rdi, %xmm0
-; AVX-NEXT: vmovq %rsi, %xmm1
-; AVX-NEXT: vpmovzxbw {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
-; AVX-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; AVX-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vpsrld $16, %xmm0, %xmm1
-; AVX-NEXT: vpmullw %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vmovd %xmm0, %eax
-; AVX-NEXT: # kill: def $al killed $al killed $eax
-; AVX-NEXT: retq
+define i8 @PR51858(i128 %arg) nounwind {
+; X86-SSE2-LABEL: PR51858:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: pushl %ebp
+; X86-SSE2-NEXT: movl %esp, %ebp
+; X86-SSE2-NEXT: andl $-16, %esp
+; X86-SSE2-NEXT: subl $16, %esp
+; X86-SSE2-NEXT: movdqu 8(%ebp), %xmm0
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm1
+; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X86-SSE2-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE2-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE2-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm1
+; X86-SSE2-NEXT: psrld $16, %xmm1
+; X86-SSE2-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE2-NEXT: movd %xmm1, %eax
+; X86-SSE2-NEXT: # kill: def $al killed $al killed $eax
+; X86-SSE2-NEXT: movl %ebp, %esp
+; X86-SSE2-NEXT: popl %ebp
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE2-LABEL: PR51858:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: movq %rdi, %xmm0
+; X64-SSE2-NEXT: movq %rsi, %xmm1
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
+; X64-SSE2-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE2-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE2-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE2-NEXT: movdqa %xmm0, %xmm1
+; X64-SSE2-NEXT: psrld $16, %xmm1
+; X64-SSE2-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE2-NEXT: movd %xmm1, %eax
+; X64-SSE2-NEXT: # kill: def $al killed $al killed $eax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: PR51858:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pushl %ebp
+; X86-SSE4-NEXT: movl %esp, %ebp
+; X86-SSE4-NEXT: andl $-16, %esp
+; X86-SSE4-NEXT: subl $16, %esp
+; X86-SSE4-NEXT: movdqu 8(%ebp), %xmm0
+; X86-SSE4-NEXT: pmovzxbw {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; X86-SSE4-NEXT: punpckhbw {{.*#+}} xmm0 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-SSE4-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE4-NEXT: pmullw %xmm1, %xmm0
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm1
+; X86-SSE4-NEXT: psrld $16, %xmm1
+; X86-SSE4-NEXT: pmullw %xmm0, %xmm1
+; X86-SSE4-NEXT: movd %xmm1, %eax
+; X86-SSE4-NEXT: # kill: def $al killed $al killed $eax
+; X86-SSE4-NEXT: movl %ebp, %esp
+; X86-SSE4-NEXT: popl %ebp
+; X86-SSE4-NEXT: retl
+;
+; X64-SSE4-LABEL: PR51858:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: movq %rdi, %xmm0
+; X64-SSE4-NEXT: movq %rsi, %xmm1
+; X64-SSE4-NEXT: pmovzxbw {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
+; X64-SSE4-NEXT: pmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; X64-SSE4-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE4-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE4-NEXT: pmullw %xmm1, %xmm0
+; X64-SSE4-NEXT: movdqa %xmm0, %xmm1
+; X64-SSE4-NEXT: psrld $16, %xmm1
+; X64-SSE4-NEXT: pmullw %xmm0, %xmm1
+; X64-SSE4-NEXT: movd %xmm1, %eax
+; X64-SSE4-NEXT: # kill: def $al killed $al killed $eax
+; X64-SSE4-NEXT: retq
+;
+; X86-AVX1-LABEL: PR51858:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: pushl %ebp
+; X86-AVX1-NEXT: movl %esp, %ebp
+; X86-AVX1-NEXT: andl $-16, %esp
+; X86-AVX1-NEXT: subl $16, %esp
+; X86-AVX1-NEXT: vmovdqu 8(%ebp), %xmm0
+; X86-AVX1-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; X86-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpsrld $16, %xmm0, %xmm1
+; X86-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: # kill: def $al killed $al killed $eax
+; X86-AVX1-NEXT: movl %ebp, %esp
+; X86-AVX1-NEXT: popl %ebp
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: PR51858:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vmovq %rdi, %xmm0
+; X64-AVX1-NEXT: vmovq %rsi, %xmm1
+; X64-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
+; X64-AVX1-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; X64-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpsrld $16, %xmm0, %xmm1
+; X64-AVX1-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-NEXT: # kill: def $al killed $al killed $eax
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: PR51858:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: pushl %ebp
+; X86-AVX2-NEXT: movl %esp, %ebp
+; X86-AVX2-NEXT: andl $-16, %esp
+; X86-AVX2-NEXT: subl $16, %esp
+; X86-AVX2-NEXT: vmovdqu 8(%ebp), %xmm0
+; X86-AVX2-NEXT: vpunpckhbw {{.*#+}} xmm1 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X86-AVX2-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; X86-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
+; X86-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: # kill: def $al killed $al killed $eax
+; X86-AVX2-NEXT: movl %ebp, %esp
+; X86-AVX2-NEXT: popl %ebp
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: PR51858:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vmovq %rdi, %xmm0
+; X64-AVX2-NEXT: vmovq %rsi, %xmm1
+; X64-AVX2-NEXT: vpmovzxbw {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
+; X64-AVX2-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; X64-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
+; X64-AVX2-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovd %xmm0, %eax
+; X64-AVX2-NEXT: # kill: def $al killed $al killed $eax
+; X64-AVX2-NEXT: retq
+;
+; AVX512-LABEL: PR51858:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vmovq %rdi, %xmm0
+; AVX512-NEXT: vmovq %rsi, %xmm1
+; AVX512-NEXT: vpmovzxbw {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero
+; AVX512-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; AVX512-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; AVX512-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: vpsrld $16, %xmm0, %xmm1
+; AVX512-NEXT: vpmullw %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: vmovd %xmm0, %eax
+; AVX512-NEXT: # kill: def $al killed $al killed $eax
+; AVX512-NEXT: retq
%vec = bitcast i128 %arg to <16 x i8>
%red = tail call i8 @llvm.vector.reduce.mul.v16i8(<16 x i8> %vec)
ret i8 %red
@@ -2483,3 +4290,5 @@ declare i8 @llvm.vector.reduce.mul.v16i8(<16 x i8>)
declare i8 @llvm.vector.reduce.mul.v32i8(<32 x i8>)
declare i8 @llvm.vector.reduce.mul.v64i8(<64 x i8>)
declare i8 @llvm.vector.reduce.mul.v128i8(<128 x i8>)
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; AVX1OR2: {{.*}}
More information about the llvm-commits
mailing list