[llvm] [X86] vector-reduce-add-*.ll - add 32-bit test coverage (PR #197152)
Simon Pilgrim via llvm-commits
llvm-commits at lists.llvm.org
Tue May 12 04:13:06 PDT 2026
https://github.com/RKSimon created https://github.com/llvm/llvm-project/pull/197152
None
>From a6b4041b116c7e3d160035761183333bf36f996e Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev at redking.me.uk>
Date: Tue, 12 May 2026 12:11:47 +0100
Subject: [PATCH] [X86] vector-reduce-add-*.ll - add 32-bit test coverage
---
.../CodeGen/X86/vector-reduce-add-mask.ll | 2538 +++++++++++------
.../CodeGen/X86/vector-reduce-add-sext.ll | 2387 ++++++++++------
.../CodeGen/X86/vector-reduce-add-zext.ll | 653 +++--
llvm/test/CodeGen/X86/vector-reduce-add.ll | 1713 +++++++----
4 files changed, 4834 insertions(+), 2457 deletions(-)
diff --git a/llvm/test/CodeGen/X86/vector-reduce-add-mask.ll b/llvm/test/CodeGen/X86/vector-reduce-add-mask.ll
index 3d85d5587a45f..43bc6f5ef8429 100644
--- a/llvm/test/CodeGen/X86/vector-reduce-add-mask.ll
+++ b/llvm/test/CodeGen/X86/vector-reduce-add-mask.ll
@@ -1,48 +1,84 @@
; 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=SSE2
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefixes=SSE41
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-SLOW
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx,+fast-hops | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-FAST
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,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=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,AVX1,AVX1-SLOW,X86-AVX1,X86-AVX1-SLOW
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-SLOW,X64-AVX1,X64-AVX1-SLOW
+; RUN: llc < %s -mtriple=i686-- -mattr=+avx,+fast-hops | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-FAST,X86-AVX1,X86-AVX1-FAST
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx,+fast-hops | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-FAST,X64-AVX1,X64-AVX1-FAST
+; RUN: llc < %s -mtriple=i686-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2,X86-AVX2
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2,X64-AVX2
+; 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
;
; vXi64
;
-define i64 @test_v2i64_v2i32(<2 x i64> %a0) {
-; SSE2-LABEL: test_v2i64_v2i32:
-; SSE2: # %bb.0:
-; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT: paddq %xmm0, %xmm1
-; SSE2-NEXT: movq %xmm1, %rax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v2i64_v2i32:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: paddq %xmm0, %xmm1
-; SSE41-NEXT: movq %xmm1, %rax
-; SSE41-NEXT: retq
-;
-; AVX1-LABEL: test_v2i64_v2i32:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vmovq %xmm0, %rax
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: test_v2i64_v2i32:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovq %xmm0, %rax
-; AVX2-NEXT: retq
+define i64 @test_v2i64_v2i32(<2 x i64> %a0) nounwind {
+; X86-SSE2-LABEL: test_v2i64_v2i32:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE2-NEXT: paddq %xmm0, %xmm1
+; X86-SSE2-NEXT: movd %xmm1, %eax
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE2-NEXT: movd %xmm0, %edx
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE-LABEL: test_v2i64_v2i32:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE-NEXT: paddq %xmm0, %xmm1
+; X64-SSE-NEXT: movq %xmm1, %rax
+; X64-SSE-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v2i64_v2i32:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: paddq %xmm0, %xmm1
+; X86-SSE4-NEXT: movd %xmm1, %eax
+; X86-SSE4-NEXT: pextrd $1, %xmm1, %edx
+; X86-SSE4-NEXT: retl
+;
+; X86-AVX1-LABEL: test_v2i64_v2i32:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v2i64_v2i32:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovq %xmm0, %rax
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v2i64_v2i32:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v2i64_v2i32:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: retq
;
; AVX512BW-LABEL: test_v2i64_v2i32:
; AVX512BW: # %bb.0:
@@ -64,48 +100,85 @@ define i64 @test_v2i64_v2i32(<2 x i64> %a0) {
ret i64 %2
}
-define i64 @test_v4i64_v4i16(<4 x i64> %a0) {
-; SSE2-LABEL: test_v4i64_v4i16:
-; SSE2: # %bb.0:
-; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE2-NEXT: paddq %xmm1, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT: paddq %xmm0, %xmm1
-; SSE2-NEXT: movq %xmm1, %rax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v4i64_v4i16:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE41-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE41-NEXT: paddq %xmm1, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: paddq %xmm0, %xmm1
-; SSE41-NEXT: movq %xmm1, %rax
-; SSE41-NEXT: retq
-;
-; AVX1-LABEL: test_v4i64_v4i16:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vmovq %xmm0, %rax
-; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: test_v4i64_v4i16:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovq %xmm0, %rax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+define i64 @test_v4i64_v4i16(<4 x i64> %a0) nounwind {
+; X86-SSE2-LABEL: test_v4i64_v4i16:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1
+; X86-SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
+; X86-SSE2-NEXT: paddq %xmm1, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE2-NEXT: paddq %xmm0, %xmm1
+; X86-SSE2-NEXT: movd %xmm1, %eax
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE2-NEXT: movd %xmm0, %edx
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE-LABEL: test_v4i64_v4i16:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; X64-SSE-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; X64-SSE-NEXT: paddq %xmm1, %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE-NEXT: paddq %xmm0, %xmm1
+; X64-SSE-NEXT: movq %xmm1, %rax
+; X64-SSE-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v4i64_v4i16:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1
+; X86-SSE4-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
+; X86-SSE4-NEXT: paddq %xmm1, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: paddq %xmm0, %xmm1
+; X86-SSE4-NEXT: movd %xmm1, %eax
+; X86-SSE4-NEXT: pextrd $1, %xmm1, %edx
+; X86-SSE4-NEXT: retl
+;
+; X86-AVX1-LABEL: test_v4i64_v4i16:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}, %ymm0, %ymm0
+; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddq %xmm1, %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_v4i16:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
+; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovq %xmm0, %rax
+; X64-AVX1-NEXT: vzeroupper
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v4i64_v4i16:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddq %xmm1, %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_v4i16:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
;
; AVX512BW-LABEL: test_v4i64_v4i16:
; AVX512BW: # %bb.0:
@@ -133,64 +206,127 @@ define i64 @test_v4i64_v4i16(<4 x i64> %a0) {
ret i64 %2
}
-define i64 @test_v8i64_v8i8(<8 x i64> %a0) {
-; SSE2-LABEL: test_v8i64_v8i8:
-; SSE2: # %bb.0:
-; SSE2-NEXT: psrlq $60, %xmm2
-; SSE2-NEXT: psrlq $60, %xmm0
-; SSE2-NEXT: paddq %xmm2, %xmm0
-; SSE2-NEXT: psrlq $60, %xmm3
-; SSE2-NEXT: psrlq $60, %xmm1
-; SSE2-NEXT: paddq %xmm3, %xmm1
-; SSE2-NEXT: paddq %xmm0, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE2-NEXT: paddq %xmm1, %xmm0
-; SSE2-NEXT: movq %xmm0, %rax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v8i64_v8i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: psrlq $60, %xmm2
-; SSE41-NEXT: psrlq $60, %xmm0
-; SSE41-NEXT: paddq %xmm2, %xmm0
-; SSE41-NEXT: psrlq $60, %xmm3
-; SSE41-NEXT: psrlq $60, %xmm1
-; SSE41-NEXT: paddq %xmm3, %xmm1
-; SSE41-NEXT: paddq %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE41-NEXT: paddq %xmm1, %xmm0
-; SSE41-NEXT: movq %xmm0, %rax
-; SSE41-NEXT: retq
-;
-; AVX1-LABEL: test_v8i64_v8i8:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vpsrlq $60, %xmm1, %xmm2
-; AVX1-NEXT: vpsrlq $60, %xmm0, %xmm3
-; AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT: vpsrlq $60, %xmm1, %xmm1
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT: vpsrlq $60, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm0, %xmm2, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vmovq %xmm0, %rax
-; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: test_v8i64_v8i8:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpsrlq $60, %ymm1, %ymm1
-; AVX2-NEXT: vpsrlq $60, %ymm0, %ymm0
-; AVX2-NEXT: vpaddq %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovq %xmm0, %rax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+define i64 @test_v8i64_v8i8(<8 x i64> %a0) nounwind {
+; X86-SSE2-LABEL: test_v8i64_v8i8:
+; 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: psrlq $60, %xmm2
+; X86-SSE2-NEXT: psrlq $60, %xmm0
+; X86-SSE2-NEXT: paddq %xmm2, %xmm0
+; X86-SSE2-NEXT: psrlq $60, %xmm3
+; X86-SSE2-NEXT: psrlq $60, %xmm1
+; X86-SSE2-NEXT: paddq %xmm3, %xmm1
+; X86-SSE2-NEXT: paddq %xmm0, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE2-NEXT: paddq %xmm1, %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
+;
+; X64-SSE-LABEL: test_v8i64_v8i8:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: psrlq $60, %xmm2
+; X64-SSE-NEXT: psrlq $60, %xmm0
+; X64-SSE-NEXT: paddq %xmm2, %xmm0
+; X64-SSE-NEXT: psrlq $60, %xmm3
+; X64-SSE-NEXT: psrlq $60, %xmm1
+; X64-SSE-NEXT: paddq %xmm3, %xmm1
+; X64-SSE-NEXT: paddq %xmm0, %xmm1
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X64-SSE-NEXT: paddq %xmm1, %xmm0
+; X64-SSE-NEXT: movq %xmm0, %rax
+; X64-SSE-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v8i64_v8i8:
+; 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: psrlq $60, %xmm2
+; X86-SSE4-NEXT: psrlq $60, %xmm0
+; X86-SSE4-NEXT: paddq %xmm2, %xmm0
+; X86-SSE4-NEXT: psrlq $60, %xmm3
+; X86-SSE4-NEXT: psrlq $60, %xmm1
+; X86-SSE4-NEXT: paddq %xmm3, %xmm1
+; X86-SSE4-NEXT: paddq %xmm0, %xmm1
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE4-NEXT: paddq %xmm1, %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_v8i8:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vpsrlq $60, %xmm1, %xmm2
+; X86-AVX1-NEXT: vpsrlq $60, %xmm0, %xmm3
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X86-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X86-AVX1-NEXT: vpsrlq $60, %xmm1, %xmm1
+; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X86-AVX1-NEXT: vpsrlq $60, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm0, %xmm2, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddq %xmm1, %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_v8i8:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vpsrlq $60, %xmm1, %xmm2
+; X64-AVX1-NEXT: vpsrlq $60, %xmm0, %xmm3
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X64-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X64-AVX1-NEXT: vpsrlq $60, %xmm1, %xmm1
+; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X64-AVX1-NEXT: vpsrlq $60, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm0, %xmm2, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovq %xmm0, %rax
+; X64-AVX1-NEXT: vzeroupper
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v8i64_v8i8:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpsrlq $60, %ymm1, %ymm1
+; X86-AVX2-NEXT: vpsrlq $60, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpaddq %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddq %xmm1, %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_v8i8:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpsrlq $60, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpsrlq $60, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpaddq %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
;
; AVX512-LABEL: test_v8i64_v8i8:
; AVX512: # %bb.0:
@@ -206,94 +342,219 @@ define i64 @test_v8i64_v8i8(<8 x i64> %a0) {
ret i64 %2
}
-define i64 @test_v16i64_v16i8(<16 x i64> %a0) {
-; SSE2-LABEL: test_v16i64_v16i8:
-; SSE2: # %bb.0:
-; SSE2-NEXT: movdqa {{.*#+}} xmm8 = [1,1]
-; SSE2-NEXT: pand %xmm8, %xmm5
-; SSE2-NEXT: pand %xmm8, %xmm1
-; SSE2-NEXT: paddq %xmm5, %xmm1
-; SSE2-NEXT: pand %xmm8, %xmm7
-; SSE2-NEXT: pand %xmm8, %xmm3
-; SSE2-NEXT: paddq %xmm7, %xmm3
-; SSE2-NEXT: paddq %xmm1, %xmm3
-; SSE2-NEXT: pand %xmm8, %xmm4
-; SSE2-NEXT: pand %xmm8, %xmm0
-; SSE2-NEXT: paddq %xmm4, %xmm0
-; SSE2-NEXT: pand %xmm8, %xmm6
-; SSE2-NEXT: pand %xmm8, %xmm2
-; SSE2-NEXT: paddq %xmm6, %xmm2
-; SSE2-NEXT: paddq %xmm0, %xmm2
-; SSE2-NEXT: paddq %xmm3, %xmm2
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
-; SSE2-NEXT: paddq %xmm2, %xmm0
-; SSE2-NEXT: movq %xmm0, %rax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v16i64_v16i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovsxbq {{.*#+}} xmm8 = [1,1]
-; SSE41-NEXT: pand %xmm8, %xmm5
-; SSE41-NEXT: pand %xmm8, %xmm1
-; SSE41-NEXT: paddq %xmm5, %xmm1
-; SSE41-NEXT: pand %xmm8, %xmm7
-; SSE41-NEXT: pand %xmm8, %xmm3
-; SSE41-NEXT: paddq %xmm7, %xmm3
-; SSE41-NEXT: paddq %xmm1, %xmm3
-; SSE41-NEXT: pand %xmm8, %xmm4
-; SSE41-NEXT: pand %xmm8, %xmm0
-; SSE41-NEXT: paddq %xmm4, %xmm0
-; SSE41-NEXT: pand %xmm8, %xmm6
-; SSE41-NEXT: pand %xmm8, %xmm2
-; SSE41-NEXT: paddq %xmm6, %xmm2
-; SSE41-NEXT: paddq %xmm0, %xmm2
-; SSE41-NEXT: paddq %xmm3, %xmm2
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
-; SSE41-NEXT: paddq %xmm2, %xmm0
-; SSE41-NEXT: movq %xmm0, %rax
-; SSE41-NEXT: retq
-;
-; AVX1-LABEL: test_v16i64_v16i8:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vbroadcastsd {{.*#+}} ymm4 = [1,1,1,1]
-; AVX1-NEXT: vandps %ymm4, %ymm0, %ymm0
-; AVX1-NEXT: vandps %ymm4, %ymm2, %ymm2
-; AVX1-NEXT: vandps %ymm4, %ymm1, %ymm1
-; AVX1-NEXT: vandps %ymm4, %ymm3, %ymm3
-; AVX1-NEXT: vpaddq %xmm3, %xmm1, %xmm4
-; AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm5
-; AVX1-NEXT: vpaddq %xmm4, %xmm5, %xmm4
-; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm3
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT: vpaddq %xmm3, %xmm1, %xmm1
-; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm0, %xmm4, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vmovq %xmm0, %rax
-; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: test_v16i64_v16i8:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm4 = [1,1,1,1]
-; AVX2-NEXT: vpand %ymm4, %ymm2, %ymm2
-; AVX2-NEXT: vpand %ymm4, %ymm0, %ymm0
-; AVX2-NEXT: vpaddq %ymm2, %ymm0, %ymm0
-; AVX2-NEXT: vpand %ymm4, %ymm3, %ymm2
-; AVX2-NEXT: vpand %ymm4, %ymm1, %ymm1
-; AVX2-NEXT: vpaddq %ymm2, %ymm1, %ymm1
-; AVX2-NEXT: vpaddq %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovq %xmm0, %rax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+define i64 @test_v16i64_v16i8(<16 x i64> %a0) nounwind {
+; X86-SSE2-LABEL: test_v16i64_v16i8:
+; 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 {{.*#+}} xmm3 = [1,0,1,0]
+; X86-SSE2-NEXT: pand %xmm3, %xmm1
+; X86-SSE2-NEXT: pand %xmm3, %xmm0
+; X86-SSE2-NEXT: pand %xmm3, %xmm2
+; X86-SSE2-NEXT: movdqa 40(%ebp), %xmm4
+; X86-SSE2-NEXT: pand %xmm3, %xmm4
+; X86-SSE2-NEXT: paddq %xmm1, %xmm4
+; X86-SSE2-NEXT: movdqa 72(%ebp), %xmm5
+; X86-SSE2-NEXT: pand %xmm3, %xmm5
+; X86-SSE2-NEXT: movdqa 8(%ebp), %xmm1
+; X86-SSE2-NEXT: pand %xmm3, %xmm1
+; X86-SSE2-NEXT: paddq %xmm5, %xmm1
+; X86-SSE2-NEXT: paddq %xmm4, %xmm1
+; X86-SSE2-NEXT: movdqa 24(%ebp), %xmm4
+; X86-SSE2-NEXT: pand %xmm3, %xmm4
+; X86-SSE2-NEXT: paddq %xmm0, %xmm4
+; X86-SSE2-NEXT: pand 56(%ebp), %xmm3
+; X86-SSE2-NEXT: paddq %xmm2, %xmm3
+; X86-SSE2-NEXT: paddq %xmm4, %xmm3
+; X86-SSE2-NEXT: paddq %xmm1, %xmm3
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
+; 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
+;
+; X64-SSE2-LABEL: test_v16i64_v16i8:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: movdqa {{.*#+}} xmm8 = [1,1]
+; X64-SSE2-NEXT: pand %xmm8, %xmm5
+; X64-SSE2-NEXT: pand %xmm8, %xmm1
+; X64-SSE2-NEXT: paddq %xmm5, %xmm1
+; X64-SSE2-NEXT: pand %xmm8, %xmm7
+; X64-SSE2-NEXT: pand %xmm8, %xmm3
+; X64-SSE2-NEXT: paddq %xmm7, %xmm3
+; X64-SSE2-NEXT: paddq %xmm1, %xmm3
+; X64-SSE2-NEXT: pand %xmm8, %xmm4
+; X64-SSE2-NEXT: pand %xmm8, %xmm0
+; X64-SSE2-NEXT: paddq %xmm4, %xmm0
+; X64-SSE2-NEXT: pand %xmm8, %xmm6
+; X64-SSE2-NEXT: pand %xmm8, %xmm2
+; X64-SSE2-NEXT: paddq %xmm6, %xmm2
+; X64-SSE2-NEXT: paddq %xmm0, %xmm2
+; X64-SSE2-NEXT: paddq %xmm3, %xmm2
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
+; X64-SSE2-NEXT: paddq %xmm2, %xmm0
+; X64-SSE2-NEXT: movq %xmm0, %rax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v16i64_v16i8:
+; 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: pmovsxbq {{.*#+}} xmm3 = [1,1]
+; X86-SSE4-NEXT: pand %xmm3, %xmm1
+; X86-SSE4-NEXT: pand %xmm3, %xmm0
+; X86-SSE4-NEXT: pand %xmm3, %xmm2
+; X86-SSE4-NEXT: movdqa 40(%ebp), %xmm4
+; X86-SSE4-NEXT: pand %xmm3, %xmm4
+; X86-SSE4-NEXT: paddq %xmm1, %xmm4
+; X86-SSE4-NEXT: movdqa 72(%ebp), %xmm5
+; X86-SSE4-NEXT: pand %xmm3, %xmm5
+; X86-SSE4-NEXT: movdqa 8(%ebp), %xmm1
+; X86-SSE4-NEXT: pand %xmm3, %xmm1
+; X86-SSE4-NEXT: paddq %xmm5, %xmm1
+; X86-SSE4-NEXT: paddq %xmm4, %xmm1
+; X86-SSE4-NEXT: movdqa 24(%ebp), %xmm4
+; X86-SSE4-NEXT: pand %xmm3, %xmm4
+; X86-SSE4-NEXT: paddq %xmm0, %xmm4
+; X86-SSE4-NEXT: pand 56(%ebp), %xmm3
+; X86-SSE4-NEXT: paddq %xmm2, %xmm3
+; X86-SSE4-NEXT: paddq %xmm4, %xmm3
+; X86-SSE4-NEXT: paddq %xmm1, %xmm3
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
+; 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
+;
+; X64-SSE4-LABEL: test_v16i64_v16i8:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: pmovsxbq {{.*#+}} xmm8 = [1,1]
+; X64-SSE4-NEXT: pand %xmm8, %xmm5
+; X64-SSE4-NEXT: pand %xmm8, %xmm1
+; X64-SSE4-NEXT: paddq %xmm5, %xmm1
+; X64-SSE4-NEXT: pand %xmm8, %xmm7
+; X64-SSE4-NEXT: pand %xmm8, %xmm3
+; X64-SSE4-NEXT: paddq %xmm7, %xmm3
+; X64-SSE4-NEXT: paddq %xmm1, %xmm3
+; X64-SSE4-NEXT: pand %xmm8, %xmm4
+; X64-SSE4-NEXT: pand %xmm8, %xmm0
+; X64-SSE4-NEXT: paddq %xmm4, %xmm0
+; X64-SSE4-NEXT: pand %xmm8, %xmm6
+; X64-SSE4-NEXT: pand %xmm8, %xmm2
+; X64-SSE4-NEXT: paddq %xmm6, %xmm2
+; X64-SSE4-NEXT: paddq %xmm0, %xmm2
+; X64-SSE4-NEXT: paddq %xmm3, %xmm2
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
+; X64-SSE4-NEXT: paddq %xmm2, %xmm0
+; X64-SSE4-NEXT: movq %xmm0, %rax
+; X64-SSE4-NEXT: retq
+;
+; X86-AVX1-LABEL: test_v16i64_v16i8:
+; 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: vbroadcastsd {{.*#+}} ymm3 = [1,0,1,0,1,0,1,0]
+; X86-AVX1-NEXT: vandps %ymm3, %ymm1, %ymm1
+; X86-AVX1-NEXT: vandps %ymm3, %ymm0, %ymm0
+; X86-AVX1-NEXT: vandps %ymm3, %ymm2, %ymm2
+; X86-AVX1-NEXT: vandps 8(%ebp), %ymm3, %ymm3
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm4
+; X86-AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm3, %xmm1, %xmm2
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm4, %xmm2
+; X86-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X86-AVX1-NEXT: vextractf128 $1, %ymm3, %xmm3
+; X86-AVX1-NEXT: vpaddq %xmm3, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm0, %xmm2, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddq %xmm1, %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_v16i8:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vbroadcastsd {{.*#+}} ymm4 = [1,1,1,1]
+; X64-AVX1-NEXT: vandps %ymm4, %ymm0, %ymm0
+; X64-AVX1-NEXT: vandps %ymm4, %ymm2, %ymm2
+; X64-AVX1-NEXT: vandps %ymm4, %ymm1, %ymm1
+; X64-AVX1-NEXT: vandps %ymm4, %ymm3, %ymm3
+; X64-AVX1-NEXT: vpaddq %xmm3, %xmm1, %xmm4
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm5
+; X64-AVX1-NEXT: vpaddq %xmm4, %xmm5, %xmm4
+; X64-AVX1-NEXT: vextractf128 $1, %ymm3, %xmm3
+; X64-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X64-AVX1-NEXT: vpaddq %xmm3, %xmm1, %xmm1
+; X64-AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm0, %xmm4, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovq %xmm0, %rax
+; X64-AVX1-NEXT: vzeroupper
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v16i64_v16i8:
+; 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: vpbroadcastq {{.*#+}} ymm3 = [1,0,1,0,1,0,1,0]
+; X86-AVX2-NEXT: vpand %ymm3, %ymm1, %ymm1
+; X86-AVX2-NEXT: vpand %ymm3, %ymm2, %ymm2
+; X86-AVX2-NEXT: vpand %ymm3, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpaddq %ymm2, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpand 8(%ebp), %ymm3, %ymm2
+; X86-AVX2-NEXT: vpaddq %ymm2, %ymm1, %ymm1
+; X86-AVX2-NEXT: vpaddq %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddq %xmm1, %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_v16i8:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpbroadcastq {{.*#+}} ymm4 = [1,1,1,1]
+; X64-AVX2-NEXT: vpand %ymm4, %ymm2, %ymm2
+; X64-AVX2-NEXT: vpand %ymm4, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpaddq %ymm2, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpand %ymm4, %ymm3, %ymm2
+; X64-AVX2-NEXT: vpand %ymm4, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpaddq %ymm2, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpaddq %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
;
; AVX512BW-LABEL: test_v16i64_v16i8:
; AVX512BW: # %bb.0:
@@ -331,45 +592,68 @@ define i64 @test_v16i64_v16i8(<16 x i64> %a0) {
; vXi32
;
-define i32 @test_v2i32_v2i16(<2 x i32> %a0) {
-; SSE2-LABEL: test_v2i32_v2i16:
-; SSE2: # %bb.0:
-; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE2-NEXT: paddd %xmm0, %xmm1
-; SSE2-NEXT: movd %xmm1, %eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v2i32_v2i16:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE41-NEXT: paddd %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: retq
-;
-; AVX1-SLOW-LABEL: test_v2i32_v2i16:
-; AVX1-SLOW: # %bb.0:
-; AVX1-SLOW-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
-; AVX1-SLOW-NEXT: retq
-;
-; AVX1-FAST-LABEL: test_v2i32_v2i16:
-; AVX1-FAST: # %bb.0:
-; AVX1-FAST-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
-; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vmovd %xmm0, %eax
-; AVX1-FAST-NEXT: retq
-;
-; AVX2-LABEL: test_v2i32_v2i16:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: retq
+define i32 @test_v2i32_v2i16(<2 x i32> %a0) nounwind {
+; X86-SSE-LABEL: test_v2i32_v2i16:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE-NEXT: paddd %xmm0, %xmm1
+; X86-SSE-NEXT: movd %xmm1, %eax
+; X86-SSE-NEXT: retl
+;
+; X64-SSE-LABEL: test_v2i32_v2i16:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-SSE-NEXT: paddd %xmm0, %xmm1
+; X64-SSE-NEXT: movd %xmm1, %eax
+; X64-SSE-NEXT: retq
+;
+; X86-AVX1-SLOW-LABEL: test_v2i32_v2i16:
+; X86-AVX1-SLOW: # %bb.0:
+; X86-AVX1-SLOW-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-SLOW-NEXT: retl
+;
+; X64-AVX1-SLOW-LABEL: test_v2i32_v2i16:
+; X64-AVX1-SLOW: # %bb.0:
+; X64-AVX1-SLOW-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-SLOW-NEXT: retq
+;
+; X86-AVX1-FAST-LABEL: test_v2i32_v2i16:
+; X86-AVX1-FAST: # %bb.0:
+; X86-AVX1-FAST-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-FAST-NEXT: retl
+;
+; X64-AVX1-FAST-LABEL: test_v2i32_v2i16:
+; X64-AVX1-FAST: # %bb.0:
+; X64-AVX1-FAST-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-FAST-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v2i32_v2i16:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v2i32_v2i16:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovd %xmm0, %eax
+; X64-AVX2-NEXT: retq
;
; AVX512-LABEL: test_v2i32_v2i16:
; AVX512: # %bb.0:
@@ -383,26 +667,16 @@ define i32 @test_v2i32_v2i16(<2 x i32> %a0) {
ret i32 %2
}
-define i32 @test_v4i32(<4 x i32> %a0) {
-; SSE2-LABEL: test_v4i32:
-; SSE2: # %bb.0:
-; SSE2-NEXT: psrld $31, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT: paddd %xmm0, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE2-NEXT: paddd %xmm1, %xmm0
-; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v4i32:
-; SSE41: # %bb.0:
-; SSE41-NEXT: psrld $31, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: paddd %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: paddd %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: retq
+define i32 @test_v4i32(<4 x i32> %a0) nounwind {
+; SSE-LABEL: test_v4i32:
+; SSE: # %bb.0:
+; SSE-NEXT: psrld $31, %xmm0
+; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; SSE-NEXT: paddd %xmm0, %xmm1
+; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; SSE-NEXT: paddd %xmm1, %xmm0
+; SSE-NEXT: movd %xmm0, %eax
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v4i32:
; AVX1-SLOW: # %bb.0:
@@ -412,7 +686,7 @@ define i32 @test_v4i32(<4 x i32> %a0) {
; AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v4i32:
; AVX1-FAST: # %bb.0:
@@ -420,7 +694,7 @@ define i32 @test_v4i32(<4 x i32> %a0) {
; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v4i32:
; AVX2: # %bb.0:
@@ -430,7 +704,7 @@ define i32 @test_v4i32(<4 x i32> %a0) {
; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512BW-LABEL: test_v4i32:
; AVX512BW: # %bb.0:
@@ -454,67 +728,104 @@ define i32 @test_v4i32(<4 x i32> %a0) {
ret i32 %2
}
-define i32 @test_v8i32_v8i8(<8 x i32> %a0) {
-; SSE2-LABEL: test_v8i32_v8i8:
-; SSE2: # %bb.0:
-; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE2-NEXT: por %xmm1, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT: paddd %xmm0, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE2-NEXT: paddd %xmm1, %xmm0
-; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v8i32_v8i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE41-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE41-NEXT: por %xmm1, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: paddd %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: paddd %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: retq
-;
-; AVX1-SLOW-LABEL: test_v8i32_v8i8:
-; AVX1-SLOW: # %bb.0:
-; AVX1-SLOW-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
-; AVX1-SLOW-NEXT: vextractf128 $1, %ymm0, %xmm1
-; AVX1-SLOW-NEXT: vorps %xmm1, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
-; AVX1-SLOW-NEXT: vzeroupper
-; AVX1-SLOW-NEXT: retq
-;
-; AVX1-FAST-LABEL: test_v8i32_v8i8:
-; AVX1-FAST: # %bb.0:
-; AVX1-FAST-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
-; AVX1-FAST-NEXT: vextractf128 $1, %ymm0, %xmm1
-; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm1, %xmm0
-; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vmovd %xmm0, %eax
-; AVX1-FAST-NEXT: vzeroupper
-; AVX1-FAST-NEXT: retq
-;
-; AVX2-LABEL: test_v8i32_v8i8:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpor %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+define i32 @test_v8i32_v8i8(<8 x i32> %a0) nounwind {
+; X86-SSE-LABEL: test_v8i32_v8i8:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm1
+; X86-SSE-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
+; X86-SSE-NEXT: por %xmm1, %xmm0
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE-NEXT: paddd %xmm0, %xmm1
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE-NEXT: paddd %xmm1, %xmm0
+; X86-SSE-NEXT: movd %xmm0, %eax
+; X86-SSE-NEXT: retl
+;
+; X64-SSE-LABEL: test_v8i32_v8i8:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; X64-SSE-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; X64-SSE-NEXT: por %xmm1, %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE-NEXT: paddd %xmm0, %xmm1
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE-NEXT: paddd %xmm1, %xmm0
+; X64-SSE-NEXT: movd %xmm0, %eax
+; X64-SSE-NEXT: retq
+;
+; X86-AVX1-SLOW-LABEL: test_v8i32_v8i8:
+; X86-AVX1-SLOW: # %bb.0:
+; X86-AVX1-SLOW-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}, %ymm0, %ymm0
+; X86-AVX1-SLOW-NEXT: vextractf128 $1, %ymm0, %xmm1
+; X86-AVX1-SLOW-NEXT: vorps %xmm1, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-SLOW-NEXT: vzeroupper
+; X86-AVX1-SLOW-NEXT: retl
+;
+; X64-AVX1-SLOW-LABEL: test_v8i32_v8i8:
+; X64-AVX1-SLOW: # %bb.0:
+; X64-AVX1-SLOW-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
+; X64-AVX1-SLOW-NEXT: vextractf128 $1, %ymm0, %xmm1
+; X64-AVX1-SLOW-NEXT: vorps %xmm1, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-SLOW-NEXT: vzeroupper
+; X64-AVX1-SLOW-NEXT: retq
+;
+; X86-AVX1-FAST-LABEL: test_v8i32_v8i8:
+; X86-AVX1-FAST: # %bb.0:
+; X86-AVX1-FAST-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}, %ymm0, %ymm0
+; X86-AVX1-FAST-NEXT: vextractf128 $1, %ymm0, %xmm1
+; X86-AVX1-FAST-NEXT: vphaddd %xmm0, %xmm1, %xmm0
+; X86-AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-FAST-NEXT: vzeroupper
+; X86-AVX1-FAST-NEXT: retl
+;
+; X64-AVX1-FAST-LABEL: test_v8i32_v8i8:
+; X64-AVX1-FAST: # %bb.0:
+; X64-AVX1-FAST-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
+; X64-AVX1-FAST-NEXT: vextractf128 $1, %ymm0, %xmm1
+; X64-AVX1-FAST-NEXT: vphaddd %xmm0, %xmm1, %xmm0
+; X64-AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-FAST-NEXT: vzeroupper
+; X64-AVX1-FAST-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v8i32_v8i8:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpor %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: vzeroupper
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v8i32_v8i8:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpor %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovd %xmm0, %eax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
;
; AVX512BW-LABEL: test_v8i32_v8i8:
; AVX512BW: # %bb.0:
@@ -540,40 +851,86 @@ define i32 @test_v8i32_v8i8(<8 x i32> %a0) {
ret i32 %2
}
-define i32 @test_v16i32_v16i8(<16 x i32> %a0) {
-; SSE2-LABEL: test_v16i32_v16i8:
-; SSE2: # %bb.0:
-; SSE2-NEXT: movdqa {{.*#+}} xmm4 = [255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0]
-; SSE2-NEXT: pand %xmm4, %xmm2
-; SSE2-NEXT: pand %xmm4, %xmm0
-; SSE2-NEXT: paddd %xmm2, %xmm0
-; SSE2-NEXT: pand %xmm4, %xmm3
-; SSE2-NEXT: pand %xmm4, %xmm1
-; SSE2-NEXT: paddd %xmm3, %xmm1
-; SSE2-NEXT: paddd %xmm0, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE2-NEXT: paddd %xmm1, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE2-NEXT: paddd %xmm0, %xmm1
-; SSE2-NEXT: movd %xmm1, %eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v16i32_v16i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovzxbd {{.*#+}} xmm4 = [255,255,255,255]
-; SSE41-NEXT: pand %xmm4, %xmm2
-; SSE41-NEXT: pand %xmm4, %xmm0
-; SSE41-NEXT: paddd %xmm2, %xmm0
-; SSE41-NEXT: pand %xmm4, %xmm3
-; SSE41-NEXT: pand %xmm4, %xmm1
-; SSE41-NEXT: paddd %xmm3, %xmm1
-; SSE41-NEXT: paddd %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE41-NEXT: paddd %xmm1, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE41-NEXT: paddd %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: retq
+define i32 @test_v16i32_v16i8(<16 x i32> %a0) nounwind {
+; X86-SSE2-LABEL: test_v16i32_v16i8:
+; 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 {{.*#+}} xmm3 = [255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0]
+; X86-SSE2-NEXT: pand %xmm3, %xmm1
+; X86-SSE2-NEXT: pand %xmm3, %xmm2
+; X86-SSE2-NEXT: pand %xmm3, %xmm0
+; X86-SSE2-NEXT: paddd %xmm2, %xmm0
+; X86-SSE2-NEXT: pand 8(%ebp), %xmm3
+; X86-SSE2-NEXT: paddd %xmm1, %xmm3
+; X86-SSE2-NEXT: paddd %xmm0, %xmm3
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
+; X86-SSE2-NEXT: paddd %xmm3, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE2-NEXT: paddd %xmm0, %xmm1
+; X86-SSE2-NEXT: movd %xmm1, %eax
+; X86-SSE2-NEXT: movl %ebp, %esp
+; X86-SSE2-NEXT: popl %ebp
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE2-LABEL: test_v16i32_v16i8:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: movdqa {{.*#+}} xmm4 = [255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0]
+; X64-SSE2-NEXT: pand %xmm4, %xmm2
+; X64-SSE2-NEXT: pand %xmm4, %xmm0
+; X64-SSE2-NEXT: paddd %xmm2, %xmm0
+; X64-SSE2-NEXT: pand %xmm4, %xmm3
+; X64-SSE2-NEXT: pand %xmm4, %xmm1
+; X64-SSE2-NEXT: paddd %xmm3, %xmm1
+; X64-SSE2-NEXT: paddd %xmm0, %xmm1
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X64-SSE2-NEXT: paddd %xmm1, %xmm0
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-SSE2-NEXT: paddd %xmm0, %xmm1
+; X64-SSE2-NEXT: movd %xmm1, %eax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v16i32_v16i8:
+; 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: pmovzxbd {{.*#+}} xmm3 = [255,255,255,255]
+; X86-SSE4-NEXT: pand %xmm3, %xmm1
+; X86-SSE4-NEXT: pand %xmm3, %xmm2
+; X86-SSE4-NEXT: pand %xmm3, %xmm0
+; X86-SSE4-NEXT: paddd %xmm2, %xmm0
+; X86-SSE4-NEXT: pand 8(%ebp), %xmm3
+; X86-SSE4-NEXT: paddd %xmm1, %xmm3
+; X86-SSE4-NEXT: paddd %xmm0, %xmm3
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
+; X86-SSE4-NEXT: paddd %xmm3, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE4-NEXT: paddd %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_v16i8:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: pmovzxbd {{.*#+}} xmm4 = [255,255,255,255]
+; X64-SSE4-NEXT: pand %xmm4, %xmm2
+; X64-SSE4-NEXT: pand %xmm4, %xmm0
+; X64-SSE4-NEXT: paddd %xmm2, %xmm0
+; X64-SSE4-NEXT: pand %xmm4, %xmm3
+; X64-SSE4-NEXT: pand %xmm4, %xmm1
+; X64-SSE4-NEXT: paddd %xmm3, %xmm1
+; X64-SSE4-NEXT: paddd %xmm0, %xmm1
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X64-SSE4-NEXT: paddd %xmm1, %xmm0
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-SSE4-NEXT: paddd %xmm0, %xmm1
+; X64-SSE4-NEXT: movd %xmm1, %eax
+; X64-SSE4-NEXT: retq
;
; AVX1-SLOW-LABEL: test_v16i32_v16i8:
; AVX1-SLOW: # %bb.0:
@@ -591,7 +948,7 @@ define i32 @test_v16i32_v16i8(<16 x i32> %a0) {
; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
; AVX1-SLOW-NEXT: vzeroupper
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v16i32_v16i8:
; AVX1-FAST: # %bb.0:
@@ -608,7 +965,7 @@ define i32 @test_v16i32_v16i8(<16 x i32> %a0) {
; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
; AVX1-FAST-NEXT: vzeroupper
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v16i32_v16i8:
; AVX2: # %bb.0:
@@ -624,7 +981,7 @@ define i32 @test_v16i32_v16i8(<16 x i32> %a0) {
; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v16i32_v16i8:
; AVX512: # %bb.0:
@@ -641,127 +998,286 @@ define i32 @test_v16i32_v16i8(<16 x i32> %a0) {
ret i32 %2
}
-define i32 @test_v32i32_v32i8(<32 x i32> %a0) {
-; SSE2-LABEL: test_v32i32_v32i8:
-; SSE2: # %bb.0:
-; SSE2-NEXT: movdqa {{.*#+}} xmm8 = [255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0]
-; SSE2-NEXT: pand %xmm8, %xmm5
-; SSE2-NEXT: pand %xmm8, %xmm1
-; SSE2-NEXT: paddd %xmm5, %xmm1
-; SSE2-NEXT: pand %xmm8, %xmm7
-; SSE2-NEXT: pand %xmm8, %xmm3
-; SSE2-NEXT: paddd %xmm7, %xmm3
-; SSE2-NEXT: paddd %xmm1, %xmm3
-; SSE2-NEXT: pand %xmm8, %xmm4
-; SSE2-NEXT: pand %xmm8, %xmm0
-; SSE2-NEXT: paddd %xmm4, %xmm0
-; SSE2-NEXT: pand %xmm8, %xmm6
-; SSE2-NEXT: pand %xmm8, %xmm2
-; SSE2-NEXT: paddd %xmm6, %xmm2
-; SSE2-NEXT: paddd %xmm0, %xmm2
-; SSE2-NEXT: paddd %xmm3, %xmm2
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
-; SSE2-NEXT: paddd %xmm2, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE2-NEXT: paddd %xmm0, %xmm1
-; SSE2-NEXT: movd %xmm1, %eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v32i32_v32i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovzxbd {{.*#+}} xmm8 = [255,255,255,255]
-; SSE41-NEXT: pand %xmm8, %xmm5
-; SSE41-NEXT: pand %xmm8, %xmm1
-; SSE41-NEXT: paddd %xmm5, %xmm1
-; SSE41-NEXT: pand %xmm8, %xmm7
-; SSE41-NEXT: pand %xmm8, %xmm3
-; SSE41-NEXT: paddd %xmm7, %xmm3
-; SSE41-NEXT: paddd %xmm1, %xmm3
-; SSE41-NEXT: pand %xmm8, %xmm4
-; SSE41-NEXT: pand %xmm8, %xmm0
-; SSE41-NEXT: paddd %xmm4, %xmm0
-; SSE41-NEXT: pand %xmm8, %xmm6
-; SSE41-NEXT: pand %xmm8, %xmm2
-; SSE41-NEXT: paddd %xmm6, %xmm2
-; SSE41-NEXT: paddd %xmm0, %xmm2
-; SSE41-NEXT: paddd %xmm3, %xmm2
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
-; SSE41-NEXT: paddd %xmm2, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE41-NEXT: paddd %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: retq
-;
-; AVX1-SLOW-LABEL: test_v32i32_v32i8:
-; AVX1-SLOW: # %bb.0:
-; AVX1-SLOW-NEXT: vbroadcastss {{.*#+}} ymm4 = [255,255,255,255,255,255,255,255]
-; AVX1-SLOW-NEXT: vandps %ymm4, %ymm0, %ymm0
-; AVX1-SLOW-NEXT: vandps %ymm4, %ymm2, %ymm2
-; AVX1-SLOW-NEXT: vandps %ymm4, %ymm1, %ymm1
-; AVX1-SLOW-NEXT: vandps %ymm4, %ymm3, %ymm3
-; AVX1-SLOW-NEXT: vpaddd %xmm3, %xmm1, %xmm4
-; AVX1-SLOW-NEXT: vpaddd %xmm2, %xmm0, %xmm5
-; AVX1-SLOW-NEXT: vpaddd %xmm4, %xmm5, %xmm4
-; AVX1-SLOW-NEXT: vextractf128 $1, %ymm3, %xmm3
-; AVX1-SLOW-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-SLOW-NEXT: vpaddd %xmm3, %xmm1, %xmm1
-; AVX1-SLOW-NEXT: vextractf128 $1, %ymm2, %xmm2
-; AVX1-SLOW-NEXT: vextractf128 $1, %ymm0, %xmm0
-; AVX1-SLOW-NEXT: vpaddd %xmm2, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vpaddd %xmm0, %xmm4, %xmm0
-; AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
-; AVX1-SLOW-NEXT: vzeroupper
-; AVX1-SLOW-NEXT: retq
-;
-; AVX1-FAST-LABEL: test_v32i32_v32i8:
-; AVX1-FAST: # %bb.0:
-; AVX1-FAST-NEXT: vbroadcastss {{.*#+}} ymm4 = [255,255,255,255,255,255,255,255]
-; AVX1-FAST-NEXT: vandps %ymm4, %ymm0, %ymm0
-; AVX1-FAST-NEXT: vandps %ymm4, %ymm2, %ymm2
-; AVX1-FAST-NEXT: vandps %ymm4, %ymm1, %ymm1
-; AVX1-FAST-NEXT: vandps %ymm4, %ymm3, %ymm3
-; AVX1-FAST-NEXT: vpaddd %xmm3, %xmm1, %xmm4
-; AVX1-FAST-NEXT: vpaddd %xmm2, %xmm0, %xmm5
-; AVX1-FAST-NEXT: vpaddd %xmm4, %xmm5, %xmm4
-; AVX1-FAST-NEXT: vextractf128 $1, %ymm3, %xmm3
-; AVX1-FAST-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-FAST-NEXT: vpaddd %xmm3, %xmm1, %xmm1
-; AVX1-FAST-NEXT: vextractf128 $1, %ymm2, %xmm2
-; AVX1-FAST-NEXT: vextractf128 $1, %ymm0, %xmm0
-; AVX1-FAST-NEXT: vpaddd %xmm2, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vpaddd %xmm0, %xmm4, %xmm0
-; AVX1-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-FAST-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vmovd %xmm0, %eax
-; AVX1-FAST-NEXT: vzeroupper
-; AVX1-FAST-NEXT: retq
-;
-; AVX2-LABEL: test_v32i32_v32i8:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpbroadcastd {{.*#+}} ymm4 = [255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0]
-; AVX2-NEXT: vpand %ymm4, %ymm2, %ymm2
-; AVX2-NEXT: vpand %ymm4, %ymm0, %ymm0
-; AVX2-NEXT: vpaddd %ymm2, %ymm0, %ymm0
-; AVX2-NEXT: vpand %ymm4, %ymm3, %ymm2
-; AVX2-NEXT: vpand %ymm4, %ymm1, %ymm1
-; AVX2-NEXT: vpaddd %ymm2, %ymm1, %ymm1
-; AVX2-NEXT: vpaddd %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+define i32 @test_v32i32_v32i8(<32 x i32> %a0) nounwind {
+; X86-SSE2-LABEL: test_v32i32_v32i8:
+; 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 {{.*#+}} xmm3 = [255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0]
+; X86-SSE2-NEXT: pand %xmm3, %xmm1
+; X86-SSE2-NEXT: pand %xmm3, %xmm0
+; X86-SSE2-NEXT: pand %xmm3, %xmm2
+; X86-SSE2-NEXT: movdqa 40(%ebp), %xmm4
+; X86-SSE2-NEXT: pand %xmm3, %xmm4
+; X86-SSE2-NEXT: paddd %xmm1, %xmm4
+; X86-SSE2-NEXT: movdqa 72(%ebp), %xmm5
+; X86-SSE2-NEXT: pand %xmm3, %xmm5
+; X86-SSE2-NEXT: movdqa 8(%ebp), %xmm1
+; X86-SSE2-NEXT: pand %xmm3, %xmm1
+; X86-SSE2-NEXT: paddd %xmm5, %xmm1
+; X86-SSE2-NEXT: paddd %xmm4, %xmm1
+; X86-SSE2-NEXT: movdqa 24(%ebp), %xmm4
+; X86-SSE2-NEXT: pand %xmm3, %xmm4
+; X86-SSE2-NEXT: paddd %xmm0, %xmm4
+; X86-SSE2-NEXT: pand 56(%ebp), %xmm3
+; X86-SSE2-NEXT: paddd %xmm2, %xmm3
+; X86-SSE2-NEXT: paddd %xmm4, %xmm3
+; X86-SSE2-NEXT: paddd %xmm1, %xmm3
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
+; X86-SSE2-NEXT: paddd %xmm3, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE2-NEXT: paddd %xmm0, %xmm1
+; X86-SSE2-NEXT: movd %xmm1, %eax
+; X86-SSE2-NEXT: movl %ebp, %esp
+; X86-SSE2-NEXT: popl %ebp
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE2-LABEL: test_v32i32_v32i8:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: movdqa {{.*#+}} xmm8 = [255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0]
+; X64-SSE2-NEXT: pand %xmm8, %xmm5
+; X64-SSE2-NEXT: pand %xmm8, %xmm1
+; X64-SSE2-NEXT: paddd %xmm5, %xmm1
+; X64-SSE2-NEXT: pand %xmm8, %xmm7
+; X64-SSE2-NEXT: pand %xmm8, %xmm3
+; X64-SSE2-NEXT: paddd %xmm7, %xmm3
+; X64-SSE2-NEXT: paddd %xmm1, %xmm3
+; X64-SSE2-NEXT: pand %xmm8, %xmm4
+; X64-SSE2-NEXT: pand %xmm8, %xmm0
+; X64-SSE2-NEXT: paddd %xmm4, %xmm0
+; X64-SSE2-NEXT: pand %xmm8, %xmm6
+; X64-SSE2-NEXT: pand %xmm8, %xmm2
+; X64-SSE2-NEXT: paddd %xmm6, %xmm2
+; X64-SSE2-NEXT: paddd %xmm0, %xmm2
+; X64-SSE2-NEXT: paddd %xmm3, %xmm2
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
+; X64-SSE2-NEXT: paddd %xmm2, %xmm0
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-SSE2-NEXT: paddd %xmm0, %xmm1
+; X64-SSE2-NEXT: movd %xmm1, %eax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v32i32_v32i8:
+; 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: pmovzxbd {{.*#+}} xmm3 = [255,255,255,255]
+; X86-SSE4-NEXT: pand %xmm3, %xmm1
+; X86-SSE4-NEXT: pand %xmm3, %xmm0
+; X86-SSE4-NEXT: pand %xmm3, %xmm2
+; X86-SSE4-NEXT: movdqa 40(%ebp), %xmm4
+; X86-SSE4-NEXT: pand %xmm3, %xmm4
+; X86-SSE4-NEXT: paddd %xmm1, %xmm4
+; X86-SSE4-NEXT: movdqa 72(%ebp), %xmm5
+; X86-SSE4-NEXT: pand %xmm3, %xmm5
+; X86-SSE4-NEXT: movdqa 8(%ebp), %xmm1
+; X86-SSE4-NEXT: pand %xmm3, %xmm1
+; X86-SSE4-NEXT: paddd %xmm5, %xmm1
+; X86-SSE4-NEXT: paddd %xmm4, %xmm1
+; X86-SSE4-NEXT: movdqa 24(%ebp), %xmm4
+; X86-SSE4-NEXT: pand %xmm3, %xmm4
+; X86-SSE4-NEXT: paddd %xmm0, %xmm4
+; X86-SSE4-NEXT: pand 56(%ebp), %xmm3
+; X86-SSE4-NEXT: paddd %xmm2, %xmm3
+; X86-SSE4-NEXT: paddd %xmm4, %xmm3
+; X86-SSE4-NEXT: paddd %xmm1, %xmm3
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
+; X86-SSE4-NEXT: paddd %xmm3, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE4-NEXT: paddd %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_v32i8:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: pmovzxbd {{.*#+}} xmm8 = [255,255,255,255]
+; X64-SSE4-NEXT: pand %xmm8, %xmm5
+; X64-SSE4-NEXT: pand %xmm8, %xmm1
+; X64-SSE4-NEXT: paddd %xmm5, %xmm1
+; X64-SSE4-NEXT: pand %xmm8, %xmm7
+; X64-SSE4-NEXT: pand %xmm8, %xmm3
+; X64-SSE4-NEXT: paddd %xmm7, %xmm3
+; X64-SSE4-NEXT: paddd %xmm1, %xmm3
+; X64-SSE4-NEXT: pand %xmm8, %xmm4
+; X64-SSE4-NEXT: pand %xmm8, %xmm0
+; X64-SSE4-NEXT: paddd %xmm4, %xmm0
+; X64-SSE4-NEXT: pand %xmm8, %xmm6
+; X64-SSE4-NEXT: pand %xmm8, %xmm2
+; X64-SSE4-NEXT: paddd %xmm6, %xmm2
+; X64-SSE4-NEXT: paddd %xmm0, %xmm2
+; X64-SSE4-NEXT: paddd %xmm3, %xmm2
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
+; X64-SSE4-NEXT: paddd %xmm2, %xmm0
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-SSE4-NEXT: paddd %xmm0, %xmm1
+; X64-SSE4-NEXT: movd %xmm1, %eax
+; X64-SSE4-NEXT: retq
+;
+; X86-AVX1-SLOW-LABEL: test_v32i32_v32i8:
+; X86-AVX1-SLOW: # %bb.0:
+; X86-AVX1-SLOW-NEXT: pushl %ebp
+; X86-AVX1-SLOW-NEXT: movl %esp, %ebp
+; X86-AVX1-SLOW-NEXT: andl $-32, %esp
+; X86-AVX1-SLOW-NEXT: subl $32, %esp
+; X86-AVX1-SLOW-NEXT: vbroadcastss {{.*#+}} ymm3 = [255,255,255,255,255,255,255,255]
+; X86-AVX1-SLOW-NEXT: vandps %ymm3, %ymm1, %ymm1
+; X86-AVX1-SLOW-NEXT: vandps %ymm3, %ymm0, %ymm0
+; X86-AVX1-SLOW-NEXT: vandps %ymm3, %ymm2, %ymm2
+; X86-AVX1-SLOW-NEXT: vandps 8(%ebp), %ymm3, %ymm3
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm2, %xmm0, %xmm4
+; X86-AVX1-SLOW-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X86-AVX1-SLOW-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm2, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm3, %xmm1, %xmm2
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm2, %xmm4, %xmm2
+; X86-AVX1-SLOW-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X86-AVX1-SLOW-NEXT: vextractf128 $1, %ymm3, %xmm3
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm3, %xmm1, %xmm1
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm0, %xmm2, %xmm0
+; X86-AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-SLOW-NEXT: movl %ebp, %esp
+; X86-AVX1-SLOW-NEXT: popl %ebp
+; X86-AVX1-SLOW-NEXT: vzeroupper
+; X86-AVX1-SLOW-NEXT: retl
+;
+; X64-AVX1-SLOW-LABEL: test_v32i32_v32i8:
+; X64-AVX1-SLOW: # %bb.0:
+; X64-AVX1-SLOW-NEXT: vbroadcastss {{.*#+}} ymm4 = [255,255,255,255,255,255,255,255]
+; X64-AVX1-SLOW-NEXT: vandps %ymm4, %ymm0, %ymm0
+; X64-AVX1-SLOW-NEXT: vandps %ymm4, %ymm2, %ymm2
+; X64-AVX1-SLOW-NEXT: vandps %ymm4, %ymm1, %ymm1
+; X64-AVX1-SLOW-NEXT: vandps %ymm4, %ymm3, %ymm3
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm3, %xmm1, %xmm4
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm2, %xmm0, %xmm5
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm4, %xmm5, %xmm4
+; X64-AVX1-SLOW-NEXT: vextractf128 $1, %ymm3, %xmm3
+; X64-AVX1-SLOW-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm3, %xmm1, %xmm1
+; X64-AVX1-SLOW-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X64-AVX1-SLOW-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm2, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm0, %xmm4, %xmm0
+; X64-AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-SLOW-NEXT: vzeroupper
+; X64-AVX1-SLOW-NEXT: retq
+;
+; X86-AVX1-FAST-LABEL: test_v32i32_v32i8:
+; X86-AVX1-FAST: # %bb.0:
+; X86-AVX1-FAST-NEXT: pushl %ebp
+; X86-AVX1-FAST-NEXT: movl %esp, %ebp
+; X86-AVX1-FAST-NEXT: andl $-32, %esp
+; X86-AVX1-FAST-NEXT: subl $32, %esp
+; X86-AVX1-FAST-NEXT: vbroadcastss {{.*#+}} ymm3 = [255,255,255,255,255,255,255,255]
+; X86-AVX1-FAST-NEXT: vandps %ymm3, %ymm1, %ymm1
+; X86-AVX1-FAST-NEXT: vandps %ymm3, %ymm0, %ymm0
+; X86-AVX1-FAST-NEXT: vandps %ymm3, %ymm2, %ymm2
+; X86-AVX1-FAST-NEXT: vandps 8(%ebp), %ymm3, %ymm3
+; X86-AVX1-FAST-NEXT: vpaddd %xmm2, %xmm0, %xmm4
+; X86-AVX1-FAST-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X86-AVX1-FAST-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X86-AVX1-FAST-NEXT: vpaddd %xmm2, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vpaddd %xmm3, %xmm1, %xmm2
+; X86-AVX1-FAST-NEXT: vpaddd %xmm2, %xmm4, %xmm2
+; X86-AVX1-FAST-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X86-AVX1-FAST-NEXT: vextractf128 $1, %ymm3, %xmm3
+; X86-AVX1-FAST-NEXT: vpaddd %xmm3, %xmm1, %xmm1
+; X86-AVX1-FAST-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vpaddd %xmm0, %xmm2, %xmm0
+; X86-AVX1-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-FAST-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-FAST-NEXT: movl %ebp, %esp
+; X86-AVX1-FAST-NEXT: popl %ebp
+; X86-AVX1-FAST-NEXT: vzeroupper
+; X86-AVX1-FAST-NEXT: retl
+;
+; X64-AVX1-FAST-LABEL: test_v32i32_v32i8:
+; X64-AVX1-FAST: # %bb.0:
+; X64-AVX1-FAST-NEXT: vbroadcastss {{.*#+}} ymm4 = [255,255,255,255,255,255,255,255]
+; X64-AVX1-FAST-NEXT: vandps %ymm4, %ymm0, %ymm0
+; X64-AVX1-FAST-NEXT: vandps %ymm4, %ymm2, %ymm2
+; X64-AVX1-FAST-NEXT: vandps %ymm4, %ymm1, %ymm1
+; X64-AVX1-FAST-NEXT: vandps %ymm4, %ymm3, %ymm3
+; X64-AVX1-FAST-NEXT: vpaddd %xmm3, %xmm1, %xmm4
+; X64-AVX1-FAST-NEXT: vpaddd %xmm2, %xmm0, %xmm5
+; X64-AVX1-FAST-NEXT: vpaddd %xmm4, %xmm5, %xmm4
+; X64-AVX1-FAST-NEXT: vextractf128 $1, %ymm3, %xmm3
+; X64-AVX1-FAST-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X64-AVX1-FAST-NEXT: vpaddd %xmm3, %xmm1, %xmm1
+; X64-AVX1-FAST-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X64-AVX1-FAST-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X64-AVX1-FAST-NEXT: vpaddd %xmm2, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vpaddd %xmm0, %xmm4, %xmm0
+; X64-AVX1-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-FAST-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-FAST-NEXT: vzeroupper
+; X64-AVX1-FAST-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v32i32_v32i8:
+; 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: vpbroadcastd {{.*#+}} ymm3 = [255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0]
+; X86-AVX2-NEXT: vpand %ymm3, %ymm1, %ymm1
+; X86-AVX2-NEXT: vpand %ymm3, %ymm2, %ymm2
+; X86-AVX2-NEXT: vpand %ymm3, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpaddd %ymm2, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpand 8(%ebp), %ymm3, %ymm2
+; X86-AVX2-NEXT: vpaddd %ymm2, %ymm1, %ymm1
+; X86-AVX2-NEXT: vpaddd %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX2-NEXT: vpaddd %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_v32i8:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpbroadcastd {{.*#+}} ymm4 = [255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0]
+; X64-AVX2-NEXT: vpand %ymm4, %ymm2, %ymm2
+; X64-AVX2-NEXT: vpand %ymm4, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpaddd %ymm2, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpand %ymm4, %ymm3, %ymm2
+; X64-AVX2-NEXT: vpand %ymm4, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpaddd %ymm2, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpaddd %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovd %xmm0, %eax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
;
; AVX512-LABEL: test_v32i32_v32i8:
; AVX512: # %bb.0:
@@ -786,52 +1302,78 @@ define i32 @test_v32i32_v32i8(<32 x i32> %a0) {
; vXi16
;
-define i16 @test_v2i16_v2i8(<2 x i16> %a0) {
-; SSE2-LABEL: test_v2i16_v2i8:
-; SSE2: # %bb.0:
-; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE2-NEXT: movdqa %xmm0, %xmm1
-; SSE2-NEXT: psrld $16, %xmm1
-; SSE2-NEXT: paddw %xmm0, %xmm1
-; SSE2-NEXT: movd %xmm1, %eax
-; SSE2-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v2i16_v2i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE41-NEXT: movdqa %xmm0, %xmm1
-; SSE41-NEXT: psrld $16, %xmm1
-; SSE41-NEXT: paddw %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE41-NEXT: retq
-;
-; AVX1-SLOW-LABEL: test_v2i16_v2i8:
-; AVX1-SLOW: # %bb.0:
-; AVX1-SLOW-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vpsrld $16, %xmm0, %xmm1
-; AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
-; AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-SLOW-NEXT: retq
-;
-; AVX1-FAST-LABEL: test_v2i16_v2i8:
-; AVX1-FAST: # %bb.0:
-; AVX1-FAST-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
-; AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vmovd %xmm0, %eax
-; AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-FAST-NEXT: retq
-;
-; AVX2-LABEL: test_v2i16_v2i8:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
-; AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
-; AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX2-NEXT: retq
+define i16 @test_v2i16_v2i8(<2 x i16> %a0) nounwind {
+; X86-SSE-LABEL: test_v2i16_v2i8:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
+; X86-SSE-NEXT: movdqa %xmm0, %xmm1
+; X86-SSE-NEXT: psrld $16, %xmm1
+; X86-SSE-NEXT: paddw %xmm0, %xmm1
+; X86-SSE-NEXT: movd %xmm1, %eax
+; X86-SSE-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-SSE-NEXT: retl
+;
+; X64-SSE-LABEL: test_v2i16_v2i8:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; X64-SSE-NEXT: movdqa %xmm0, %xmm1
+; X64-SSE-NEXT: psrld $16, %xmm1
+; X64-SSE-NEXT: paddw %xmm0, %xmm1
+; X64-SSE-NEXT: movd %xmm1, %eax
+; X64-SSE-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-SSE-NEXT: retq
+;
+; X86-AVX1-SLOW-LABEL: test_v2i16_v2i8:
+; X86-AVX1-SLOW: # %bb.0:
+; X86-AVX1-SLOW-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vpsrld $16, %xmm0, %xmm1
+; X86-AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-AVX1-SLOW-NEXT: retl
+;
+; X64-AVX1-SLOW-LABEL: test_v2i16_v2i8:
+; X64-AVX1-SLOW: # %bb.0:
+; X64-AVX1-SLOW-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vpsrld $16, %xmm0, %xmm1
+; X64-AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-AVX1-SLOW-NEXT: retq
+;
+; X86-AVX1-FAST-LABEL: test_v2i16_v2i8:
+; X86-AVX1-FAST: # %bb.0:
+; X86-AVX1-FAST-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-AVX1-FAST-NEXT: retl
+;
+; X64-AVX1-FAST-LABEL: test_v2i16_v2i8:
+; X64-AVX1-FAST: # %bb.0:
+; X64-AVX1-FAST-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-AVX1-FAST-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v2i16_v2i8:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
+; X86-AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v2i16_v2i8:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; X64-AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
+; X64-AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovd %xmm0, %eax
+; X64-AVX2-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-AVX2-NEXT: retq
;
; AVX512-LABEL: test_v2i16_v2i8:
; AVX512: # %bb.0:
@@ -846,70 +1388,120 @@ define i16 @test_v2i16_v2i8(<2 x i16> %a0) {
ret i16 %2
}
-define i16 @test_v4i16_v4i8(<4 x i16> %a0) {
-; SSE2-LABEL: test_v4i16_v4i8:
-; SSE2: # %bb.0:
-; SSE2-NEXT: movdqa {{.*#+}} xmm1 = [0,65535,65535,65535,65535,65535,65535,65535]
-; SSE2-NEXT: pandn %xmm0, %xmm1
-; SSE2-NEXT: pmulhuw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 # [0,0,0,128,0,64,0,32,u,u,u,u,u,u,u,u]
-; SSE2-NEXT: por %xmm1, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE2-NEXT: paddw %xmm0, %xmm1
-; SSE2-NEXT: movdqa %xmm1, %xmm0
-; SSE2-NEXT: psrld $16, %xmm0
-; SSE2-NEXT: paddw %xmm1, %xmm0
-; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v4i16_v4i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: movq {{.*#+}} xmm1 = [0,32768,16384,8192,0,0,0,0]
-; SSE41-NEXT: pmulhuw %xmm0, %xmm1
-; SSE41-NEXT: pblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3,4,5,6,7]
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: paddw %xmm1, %xmm0
-; SSE41-NEXT: movdqa %xmm0, %xmm1
-; SSE41-NEXT: psrld $16, %xmm1
-; SSE41-NEXT: paddw %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE41-NEXT: retq
-;
-; AVX1-SLOW-LABEL: test_v4i16_v4i8:
-; AVX1-SLOW: # %bb.0:
-; AVX1-SLOW-NEXT: vpmulhuw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1 # [u,32768,16384,8192,u,u,u,u]
-; AVX1-SLOW-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3,4,5,6,7]
-; AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vpsrld $16, %xmm0, %xmm1
-; AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
-; AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-SLOW-NEXT: retq
-;
-; AVX1-FAST-LABEL: test_v4i16_v4i8:
-; AVX1-FAST: # %bb.0:
-; AVX1-FAST-NEXT: vpmulhuw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1 # [u,32768,16384,8192,u,u,u,u]
-; AVX1-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3,4,5,6,7]
-; AVX1-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX1-FAST-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vmovd %xmm0, %eax
-; AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-FAST-NEXT: retq
-;
-; AVX2-LABEL: test_v4i16_v4i8:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpmulhuw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1 # [u,32768,16384,8192,u,u,u,u]
-; AVX2-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3,4,5,6,7]
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
-; AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX2-NEXT: retq
+define i16 @test_v4i16_v4i8(<4 x i16> %a0) nounwind {
+; X86-SSE2-LABEL: test_v4i16_v4i8:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: movdqa {{.*#+}} xmm1 = [0,65535,65535,65535,65535,65535,65535,65535]
+; X86-SSE2-NEXT: pandn %xmm0, %xmm1
+; X86-SSE2-NEXT: pmulhuw {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0 # [0,0,0,128,0,64,0,32,u,u,u,u,u,u,u,u]
+; X86-SSE2-NEXT: por %xmm1, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE2-NEXT: paddw %xmm0, %xmm1
+; X86-SSE2-NEXT: movdqa %xmm1, %xmm0
+; X86-SSE2-NEXT: psrld $16, %xmm0
+; X86-SSE2-NEXT: paddw %xmm1, %xmm0
+; X86-SSE2-NEXT: movd %xmm0, %eax
+; X86-SSE2-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE2-LABEL: test_v4i16_v4i8:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: movdqa {{.*#+}} xmm1 = [0,65535,65535,65535,65535,65535,65535,65535]
+; X64-SSE2-NEXT: pandn %xmm0, %xmm1
+; X64-SSE2-NEXT: pmulhuw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 # [0,0,0,128,0,64,0,32,u,u,u,u,u,u,u,u]
+; X64-SSE2-NEXT: por %xmm1, %xmm0
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-SSE2-NEXT: paddw %xmm0, %xmm1
+; X64-SSE2-NEXT: movdqa %xmm1, %xmm0
+; X64-SSE2-NEXT: psrld $16, %xmm0
+; X64-SSE2-NEXT: paddw %xmm1, %xmm0
+; X64-SSE2-NEXT: movd %xmm0, %eax
+; X64-SSE2-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-SSE2-NEXT: retq
+;
+; SSE4-LABEL: test_v4i16_v4i8:
+; SSE4: # %bb.0:
+; SSE4-NEXT: movq {{.*#+}} xmm1 = [0,32768,16384,8192,0,0,0,0]
+; SSE4-NEXT: pmulhuw %xmm0, %xmm1
+; SSE4-NEXT: pblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3,4,5,6,7]
+; SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; SSE4-NEXT: paddw %xmm1, %xmm0
+; SSE4-NEXT: movdqa %xmm0, %xmm1
+; SSE4-NEXT: psrld $16, %xmm1
+; SSE4-NEXT: paddw %xmm0, %xmm1
+; SSE4-NEXT: movd %xmm1, %eax
+; SSE4-NEXT: # kill: def $ax killed $ax killed $eax
+; SSE4-NEXT: ret{{[l|q]}}
+;
+; X86-AVX1-SLOW-LABEL: test_v4i16_v4i8:
+; X86-AVX1-SLOW: # %bb.0:
+; X86-AVX1-SLOW-NEXT: vpmulhuw {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm1 # [u,32768,16384,8192,u,u,u,u]
+; X86-AVX1-SLOW-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3,4,5,6,7]
+; X86-AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vpsrld $16, %xmm0, %xmm1
+; X86-AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-AVX1-SLOW-NEXT: retl
+;
+; X64-AVX1-SLOW-LABEL: test_v4i16_v4i8:
+; X64-AVX1-SLOW: # %bb.0:
+; X64-AVX1-SLOW-NEXT: vpmulhuw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1 # [u,32768,16384,8192,u,u,u,u]
+; X64-AVX1-SLOW-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3,4,5,6,7]
+; X64-AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vpsrld $16, %xmm0, %xmm1
+; X64-AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-AVX1-SLOW-NEXT: retq
+;
+; X86-AVX1-FAST-LABEL: test_v4i16_v4i8:
+; X86-AVX1-FAST: # %bb.0:
+; X86-AVX1-FAST-NEXT: vpmulhuw {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm1 # [u,32768,16384,8192,u,u,u,u]
+; X86-AVX1-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3,4,5,6,7]
+; X86-AVX1-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX1-FAST-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-AVX1-FAST-NEXT: retl
+;
+; X64-AVX1-FAST-LABEL: test_v4i16_v4i8:
+; X64-AVX1-FAST: # %bb.0:
+; X64-AVX1-FAST-NEXT: vpmulhuw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1 # [u,32768,16384,8192,u,u,u,u]
+; X64-AVX1-FAST-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3,4,5,6,7]
+; X64-AVX1-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX1-FAST-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-AVX1-FAST-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v4i16_v4i8:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpmulhuw {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm1 # [u,32768,16384,8192,u,u,u,u]
+; X86-AVX2-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3,4,5,6,7]
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
+; X86-AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v4i16_v4i8:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpmulhuw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1 # [u,32768,16384,8192,u,u,u,u]
+; X64-AVX2-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3,4,5,6,7]
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
+; X64-AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovd %xmm0, %eax
+; X64-AVX2-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-AVX2-NEXT: retq
;
; AVX512BW-LABEL: test_v4i16_v4i8:
; AVX512BW: # %bb.0:
@@ -940,101 +1532,201 @@ define i16 @test_v4i16_v4i8(<4 x i16> %a0) {
ret i16 %2
}
-define i16 @test_v8i16_v8i8(<8 x i16> %a0) {
-; SSE2-LABEL: test_v8i16_v8i8:
-; SSE2: # %bb.0:
-; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE2-NEXT: packuswb %xmm0, %xmm0
-; SSE2-NEXT: pxor %xmm1, %xmm1
-; SSE2-NEXT: psadbw %xmm0, %xmm1
-; SSE2-NEXT: movd %xmm1, %eax
-; SSE2-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v8i16_v8i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE41-NEXT: packuswb %xmm0, %xmm0
-; SSE41-NEXT: pxor %xmm1, %xmm1
-; SSE41-NEXT: psadbw %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE41-NEXT: retq
-;
-; AVX-LABEL: test_v8i16_v8i8:
-; AVX: # %bb.0:
-; AVX-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
-; AVX-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
-; AVX-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vmovd %xmm0, %eax
-; AVX-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX-NEXT: retq
+define i16 @test_v8i16_v8i8(<8 x i16> %a0) nounwind {
+; X86-SSE-LABEL: test_v8i16_v8i8:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
+; X86-SSE-NEXT: packuswb %xmm0, %xmm0
+; X86-SSE-NEXT: pxor %xmm1, %xmm1
+; X86-SSE-NEXT: psadbw %xmm0, %xmm1
+; X86-SSE-NEXT: movd %xmm1, %eax
+; X86-SSE-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-SSE-NEXT: retl
+;
+; X64-SSE-LABEL: test_v8i16_v8i8:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; X64-SSE-NEXT: packuswb %xmm0, %xmm0
+; X64-SSE-NEXT: pxor %xmm1, %xmm1
+; X64-SSE-NEXT: psadbw %xmm0, %xmm1
+; X64-SSE-NEXT: movd %xmm1, %eax
+; X64-SSE-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-SSE-NEXT: retq
+;
+; X86-AVX1-LABEL: test_v8i16_v8i8:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v8i16_v8i8:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; X64-AVX1-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v8i16_v8i8:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X86-AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v8i16_v8i8:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; X64-AVX2-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X64-AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovd %xmm0, %eax
+; X64-AVX2-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-AVX2-NEXT: retq
+;
+; AVX512-LABEL: test_v8i16_v8i8:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; AVX512-NEXT: vpackuswb %xmm0, %xmm0, %xmm0
+; AVX512-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; AVX512-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: vmovd %xmm0, %eax
+; AVX512-NEXT: # kill: def $ax killed $ax killed $eax
+; AVX512-NEXT: retq
%1 = and <8 x i16> %a0, <i16 0, i16 1, i16 2, i16 4, i16 8, i16 16, i16 32, i16 64>
%2 = call i16 @llvm.vector.reduce.add.v8i16(<8 x i16> %1)
ret i16 %2
}
-define i16 @test_v16i16_v16i8(<16 x i16> %a0) {
-; SSE2-LABEL: test_v16i16_v16i8:
-; SSE2: # %bb.0:
-; SSE2-NEXT: movdqa {{.*#+}} xmm2 = [255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0]
-; SSE2-NEXT: pand %xmm2, %xmm1
-; SSE2-NEXT: pand %xmm2, %xmm0
-; SSE2-NEXT: packuswb %xmm1, %xmm0
-; SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE2-NEXT: pxor %xmm1, %xmm1
-; SSE2-NEXT: psadbw %xmm0, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE2-NEXT: paddq %xmm1, %xmm0
-; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v16i16_v16i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovzxbw {{.*#+}} xmm2 = [255,255,255,255,255,255,255,255]
-; SSE41-NEXT: pand %xmm2, %xmm1
-; SSE41-NEXT: pand %xmm2, %xmm0
-; SSE41-NEXT: packuswb %xmm1, %xmm0
-; SSE41-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE41-NEXT: pxor %xmm1, %xmm1
-; SSE41-NEXT: psadbw %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE41-NEXT: paddq %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE41-NEXT: retq
-;
-; AVX1-LABEL: test_v16i16_v16i8:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
-; AVX1-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
-; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vmovd %xmm0, %eax
-; AVX1-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: test_v16i16_v16i8:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
-; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+define i16 @test_v16i16_v16i8(<16 x i16> %a0) nounwind {
+; X86-SSE2-LABEL: test_v16i16_v16i8:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: movdqa {{.*#+}} xmm2 = [255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0]
+; X86-SSE2-NEXT: pand %xmm2, %xmm1
+; X86-SSE2-NEXT: pand %xmm2, %xmm0
+; X86-SSE2-NEXT: packuswb %xmm1, %xmm0
+; X86-SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
+; X86-SSE2-NEXT: pxor %xmm1, %xmm1
+; X86-SSE2-NEXT: psadbw %xmm0, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE2-NEXT: paddq %xmm1, %xmm0
+; X86-SSE2-NEXT: movd %xmm0, %eax
+; X86-SSE2-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE2-LABEL: test_v16i16_v16i8:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: movdqa {{.*#+}} xmm2 = [255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0]
+; X64-SSE2-NEXT: pand %xmm2, %xmm1
+; X64-SSE2-NEXT: pand %xmm2, %xmm0
+; X64-SSE2-NEXT: packuswb %xmm1, %xmm0
+; X64-SSE2-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; X64-SSE2-NEXT: pxor %xmm1, %xmm1
+; X64-SSE2-NEXT: psadbw %xmm0, %xmm1
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X64-SSE2-NEXT: paddq %xmm1, %xmm0
+; X64-SSE2-NEXT: movd %xmm0, %eax
+; X64-SSE2-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v16i16_v16i8:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pmovzxbw {{.*#+}} xmm2 = [255,255,255,255,255,255,255,255]
+; X86-SSE4-NEXT: pand %xmm2, %xmm1
+; X86-SSE4-NEXT: pand %xmm2, %xmm0
+; X86-SSE4-NEXT: packuswb %xmm1, %xmm0
+; X86-SSE4-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
+; X86-SSE4-NEXT: pxor %xmm1, %xmm1
+; X86-SSE4-NEXT: psadbw %xmm0, %xmm1
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE4-NEXT: paddq %xmm1, %xmm0
+; X86-SSE4-NEXT: movd %xmm0, %eax
+; X86-SSE4-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-SSE4-NEXT: retl
+;
+; X64-SSE4-LABEL: test_v16i16_v16i8:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: pmovzxbw {{.*#+}} xmm2 = [255,255,255,255,255,255,255,255]
+; X64-SSE4-NEXT: pand %xmm2, %xmm1
+; X64-SSE4-NEXT: pand %xmm2, %xmm0
+; X64-SSE4-NEXT: packuswb %xmm1, %xmm0
+; X64-SSE4-NEXT: pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; X64-SSE4-NEXT: pxor %xmm1, %xmm1
+; X64-SSE4-NEXT: psadbw %xmm0, %xmm1
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X64-SSE4-NEXT: paddq %xmm1, %xmm0
+; X64-SSE4-NEXT: movd %xmm0, %eax
+; X64-SSE4-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-SSE4-NEXT: retq
+;
+; X86-AVX1-LABEL: test_v16i16_v16i8:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}, %ymm0, %ymm0
+; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
+; X86-AVX1-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-AVX1-NEXT: vzeroupper
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v16i16_v16i8:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
+; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
+; X64-AVX1-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; X64-AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddq %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_v16i16_v16i8:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X86-AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-AVX2-NEXT: vzeroupper
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v16i16_v16i8:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
+; X64-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X64-AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddq %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
;
; AVX512BW-LABEL: test_v16i16_v16i8:
; AVX512BW: # %bb.0:
@@ -1067,42 +1759,49 @@ define i16 @test_v16i16_v16i8(<16 x i16> %a0) {
ret i16 %2
}
-define i16 @test_v32i16_v32i8(<32 x i16> %a0) {
-; SSE2-LABEL: test_v32i16_v32i8:
-; SSE2: # %bb.0:
-; SSE2-NEXT: psrlw $8, %xmm1
-; SSE2-NEXT: psrlw $8, %xmm0
-; SSE2-NEXT: packuswb %xmm1, %xmm0
-; SSE2-NEXT: psrlw $8, %xmm3
-; SSE2-NEXT: psrlw $8, %xmm2
-; SSE2-NEXT: packuswb %xmm3, %xmm2
-; SSE2-NEXT: pxor %xmm1, %xmm1
-; SSE2-NEXT: psadbw %xmm1, %xmm2
-; SSE2-NEXT: psadbw %xmm1, %xmm0
-; SSE2-NEXT: paddq %xmm2, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT: paddq %xmm0, %xmm1
-; SSE2-NEXT: movd %xmm1, %eax
-; SSE2-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v32i16_v32i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: psrlw $8, %xmm1
-; SSE41-NEXT: psrlw $8, %xmm0
-; SSE41-NEXT: packuswb %xmm1, %xmm0
-; SSE41-NEXT: psrlw $8, %xmm3
-; SSE41-NEXT: psrlw $8, %xmm2
-; SSE41-NEXT: packuswb %xmm3, %xmm2
-; SSE41-NEXT: pxor %xmm1, %xmm1
-; SSE41-NEXT: psadbw %xmm1, %xmm2
-; SSE41-NEXT: psadbw %xmm1, %xmm0
-; SSE41-NEXT: paddq %xmm2, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: paddq %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE41-NEXT: retq
+define i16 @test_v32i16_v32i8(<32 x i16> %a0) nounwind {
+; X86-SSE-LABEL: test_v32i16_v32i8:
+; 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: psrlw $8, %xmm1
+; X86-SSE-NEXT: psrlw $8, %xmm0
+; X86-SSE-NEXT: packuswb %xmm1, %xmm0
+; X86-SSE-NEXT: psrlw $8, %xmm3
+; X86-SSE-NEXT: psrlw $8, %xmm2
+; X86-SSE-NEXT: packuswb %xmm3, %xmm2
+; X86-SSE-NEXT: pxor %xmm1, %xmm1
+; X86-SSE-NEXT: psadbw %xmm1, %xmm2
+; X86-SSE-NEXT: psadbw %xmm1, %xmm0
+; X86-SSE-NEXT: paddq %xmm2, %xmm0
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE-NEXT: paddq %xmm0, %xmm1
+; X86-SSE-NEXT: movd %xmm1, %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_v32i8:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: psrlw $8, %xmm1
+; X64-SSE-NEXT: psrlw $8, %xmm0
+; X64-SSE-NEXT: packuswb %xmm1, %xmm0
+; X64-SSE-NEXT: psrlw $8, %xmm3
+; X64-SSE-NEXT: psrlw $8, %xmm2
+; X64-SSE-NEXT: packuswb %xmm3, %xmm2
+; X64-SSE-NEXT: pxor %xmm1, %xmm1
+; X64-SSE-NEXT: psadbw %xmm1, %xmm2
+; X64-SSE-NEXT: psadbw %xmm1, %xmm0
+; X64-SSE-NEXT: paddq %xmm2, %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE-NEXT: paddq %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_v32i8:
; AVX1: # %bb.0:
@@ -1123,7 +1822,7 @@ define i16 @test_v32i16_v32i8(<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_v32i8:
; AVX2: # %bb.0:
@@ -1140,7 +1839,7 @@ define i16 @test_v32i16_v32i8(<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]}}
;
; AVX512-LABEL: test_v32i16_v32i8:
; AVX512: # %bb.0:
@@ -1161,110 +1860,246 @@ define i16 @test_v32i16_v32i8(<32 x i16> %a0) {
ret i16 %2
}
-define i16 @test_v64i16_v64i8(<64 x i16> %a0) {
-; SSE2-LABEL: test_v64i16_v64i8:
-; SSE2: # %bb.0:
-; SSE2-NEXT: movdqa {{.*#+}} xmm8 = [127,127,127,127,127,127,127,127]
-; SSE2-NEXT: pand %xmm8, %xmm1
-; SSE2-NEXT: pand %xmm8, %xmm0
-; SSE2-NEXT: packuswb %xmm1, %xmm0
-; SSE2-NEXT: pand %xmm8, %xmm5
-; SSE2-NEXT: pand %xmm8, %xmm4
-; SSE2-NEXT: packuswb %xmm5, %xmm4
-; SSE2-NEXT: paddb %xmm0, %xmm4
-; SSE2-NEXT: pand %xmm8, %xmm3
-; SSE2-NEXT: pand %xmm8, %xmm2
-; SSE2-NEXT: packuswb %xmm3, %xmm2
-; SSE2-NEXT: pand %xmm8, %xmm7
-; SSE2-NEXT: pand %xmm8, %xmm6
-; SSE2-NEXT: packuswb %xmm7, %xmm6
-; SSE2-NEXT: paddb %xmm2, %xmm6
-; SSE2-NEXT: pxor %xmm0, %xmm0
-; SSE2-NEXT: psadbw %xmm0, %xmm6
-; SSE2-NEXT: psadbw %xmm0, %xmm4
-; SSE2-NEXT: paddq %xmm6, %xmm4
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]
-; SSE2-NEXT: paddq %xmm4, %xmm0
-; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v64i16_v64i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovsxbw {{.*#+}} xmm8 = [127,127,127,127,127,127,127,127]
-; SSE41-NEXT: pand %xmm8, %xmm1
-; SSE41-NEXT: pand %xmm8, %xmm0
-; SSE41-NEXT: packuswb %xmm1, %xmm0
-; SSE41-NEXT: pand %xmm8, %xmm5
-; SSE41-NEXT: pand %xmm8, %xmm4
-; SSE41-NEXT: packuswb %xmm5, %xmm4
-; SSE41-NEXT: paddb %xmm0, %xmm4
-; SSE41-NEXT: pand %xmm8, %xmm3
-; SSE41-NEXT: pand %xmm8, %xmm2
-; SSE41-NEXT: packuswb %xmm3, %xmm2
-; SSE41-NEXT: pand %xmm8, %xmm7
-; SSE41-NEXT: pand %xmm8, %xmm6
-; SSE41-NEXT: packuswb %xmm7, %xmm6
-; SSE41-NEXT: paddb %xmm2, %xmm6
-; SSE41-NEXT: pxor %xmm0, %xmm0
-; SSE41-NEXT: psadbw %xmm0, %xmm6
-; SSE41-NEXT: psadbw %xmm0, %xmm4
-; SSE41-NEXT: paddq %xmm6, %xmm4
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]
-; SSE41-NEXT: paddq %xmm4, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE41-NEXT: retq
-;
-; AVX1-LABEL: test_v64i16_v64i8:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vbroadcastss {{.*#+}} ymm4 = [127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127]
-; AVX1-NEXT: vandps %ymm4, %ymm0, %ymm0
-; AVX1-NEXT: vandps %ymm4, %ymm2, %ymm2
-; AVX1-NEXT: vandps %ymm4, %ymm1, %ymm1
-; AVX1-NEXT: vandps %ymm4, %ymm3, %ymm3
-; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm4
-; AVX1-NEXT: vpackuswb %xmm4, %xmm3, %xmm3
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm4
-; AVX1-NEXT: vpackuswb %xmm4, %xmm1, %xmm1
-; AVX1-NEXT: vpaddb %xmm3, %xmm1, %xmm1
-; AVX1-NEXT: vpxor %xmm3, %xmm3, %xmm3
-; AVX1-NEXT: vpsadbw %xmm3, %xmm1, %xmm1
-; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm4
-; AVX1-NEXT: vpackuswb %xmm4, %xmm2, %xmm2
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm4
-; AVX1-NEXT: vpackuswb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT: vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vpsadbw %xmm3, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vmovd %xmm0, %eax
-; AVX1-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: test_v64i16_v64i8:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpbroadcastw {{.*#+}} ymm4 = [127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127]
-; AVX2-NEXT: vpand %ymm4, %ymm1, %ymm1
-; AVX2-NEXT: vpand %ymm4, %ymm0, %ymm0
-; AVX2-NEXT: vpackuswb %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vpand %ymm4, %ymm3, %ymm1
-; AVX2-NEXT: vpand %ymm4, %ymm2, %ymm2
-; AVX2-NEXT: vpackuswb %ymm1, %ymm2, %ymm1
-; AVX2-NEXT: vpaddb %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
-; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT: vpsadbw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+define i16 @test_v64i16_v64i8(<64 x i16> %a0) nounwind {
+; X86-SSE2-LABEL: test_v64i16_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 {{.*#+}} xmm3 = [127,127,127,127,127,127,127,127]
+; X86-SSE2-NEXT: pand %xmm3, %xmm1
+; X86-SSE2-NEXT: pand %xmm3, %xmm0
+; X86-SSE2-NEXT: packuswb %xmm1, %xmm0
+; X86-SSE2-NEXT: pand %xmm3, %xmm2
+; X86-SSE2-NEXT: movdqa 40(%ebp), %xmm4
+; X86-SSE2-NEXT: pand %xmm3, %xmm4
+; X86-SSE2-NEXT: movdqa 24(%ebp), %xmm1
+; X86-SSE2-NEXT: pand %xmm3, %xmm1
+; X86-SSE2-NEXT: packuswb %xmm4, %xmm1
+; X86-SSE2-NEXT: paddb %xmm0, %xmm1
+; X86-SSE2-NEXT: movdqa 8(%ebp), %xmm0
+; X86-SSE2-NEXT: pand %xmm3, %xmm0
+; X86-SSE2-NEXT: packuswb %xmm0, %xmm2
+; X86-SSE2-NEXT: movdqa 72(%ebp), %xmm0
+; X86-SSE2-NEXT: pand %xmm3, %xmm0
+; X86-SSE2-NEXT: pand 56(%ebp), %xmm3
+; X86-SSE2-NEXT: packuswb %xmm0, %xmm3
+; X86-SSE2-NEXT: paddb %xmm2, %xmm3
+; X86-SSE2-NEXT: pxor %xmm0, %xmm0
+; X86-SSE2-NEXT: psadbw %xmm0, %xmm3
+; X86-SSE2-NEXT: psadbw %xmm0, %xmm1
+; X86-SSE2-NEXT: paddq %xmm3, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE2-NEXT: paddq %xmm1, %xmm0
+; X86-SSE2-NEXT: movd %xmm0, %eax
+; X86-SSE2-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-SSE2-NEXT: movl %ebp, %esp
+; X86-SSE2-NEXT: popl %ebp
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE2-LABEL: test_v64i16_v64i8:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: movdqa {{.*#+}} xmm8 = [127,127,127,127,127,127,127,127]
+; X64-SSE2-NEXT: pand %xmm8, %xmm1
+; X64-SSE2-NEXT: pand %xmm8, %xmm0
+; X64-SSE2-NEXT: packuswb %xmm1, %xmm0
+; X64-SSE2-NEXT: pand %xmm8, %xmm5
+; X64-SSE2-NEXT: pand %xmm8, %xmm4
+; X64-SSE2-NEXT: packuswb %xmm5, %xmm4
+; X64-SSE2-NEXT: paddb %xmm0, %xmm4
+; X64-SSE2-NEXT: pand %xmm8, %xmm3
+; X64-SSE2-NEXT: pand %xmm8, %xmm2
+; X64-SSE2-NEXT: packuswb %xmm3, %xmm2
+; X64-SSE2-NEXT: pand %xmm8, %xmm7
+; X64-SSE2-NEXT: pand %xmm8, %xmm6
+; X64-SSE2-NEXT: packuswb %xmm7, %xmm6
+; X64-SSE2-NEXT: paddb %xmm2, %xmm6
+; X64-SSE2-NEXT: pxor %xmm0, %xmm0
+; X64-SSE2-NEXT: psadbw %xmm0, %xmm6
+; X64-SSE2-NEXT: psadbw %xmm0, %xmm4
+; X64-SSE2-NEXT: paddq %xmm6, %xmm4
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]
+; X64-SSE2-NEXT: paddq %xmm4, %xmm0
+; X64-SSE2-NEXT: movd %xmm0, %eax
+; X64-SSE2-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v64i16_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: pmovsxbw {{.*#+}} xmm3 = [127,127,127,127,127,127,127,127]
+; X86-SSE4-NEXT: pand %xmm3, %xmm1
+; X86-SSE4-NEXT: pand %xmm3, %xmm0
+; X86-SSE4-NEXT: packuswb %xmm1, %xmm0
+; X86-SSE4-NEXT: pand %xmm3, %xmm2
+; X86-SSE4-NEXT: movdqa 40(%ebp), %xmm4
+; X86-SSE4-NEXT: pand %xmm3, %xmm4
+; X86-SSE4-NEXT: movdqa 24(%ebp), %xmm1
+; X86-SSE4-NEXT: pand %xmm3, %xmm1
+; X86-SSE4-NEXT: packuswb %xmm4, %xmm1
+; X86-SSE4-NEXT: paddb %xmm0, %xmm1
+; X86-SSE4-NEXT: movdqa 8(%ebp), %xmm0
+; X86-SSE4-NEXT: pand %xmm3, %xmm0
+; X86-SSE4-NEXT: packuswb %xmm0, %xmm2
+; X86-SSE4-NEXT: movdqa 72(%ebp), %xmm0
+; X86-SSE4-NEXT: pand %xmm3, %xmm0
+; X86-SSE4-NEXT: pand 56(%ebp), %xmm3
+; X86-SSE4-NEXT: packuswb %xmm0, %xmm3
+; X86-SSE4-NEXT: paddb %xmm2, %xmm3
+; X86-SSE4-NEXT: pxor %xmm0, %xmm0
+; X86-SSE4-NEXT: psadbw %xmm0, %xmm3
+; X86-SSE4-NEXT: psadbw %xmm0, %xmm1
+; X86-SSE4-NEXT: paddq %xmm3, %xmm1
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE4-NEXT: paddq %xmm1, %xmm0
+; X86-SSE4-NEXT: movd %xmm0, %eax
+; X86-SSE4-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-SSE4-NEXT: movl %ebp, %esp
+; X86-SSE4-NEXT: popl %ebp
+; X86-SSE4-NEXT: retl
+;
+; X64-SSE4-LABEL: test_v64i16_v64i8:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: pmovsxbw {{.*#+}} xmm8 = [127,127,127,127,127,127,127,127]
+; X64-SSE4-NEXT: pand %xmm8, %xmm1
+; X64-SSE4-NEXT: pand %xmm8, %xmm0
+; X64-SSE4-NEXT: packuswb %xmm1, %xmm0
+; X64-SSE4-NEXT: pand %xmm8, %xmm5
+; X64-SSE4-NEXT: pand %xmm8, %xmm4
+; X64-SSE4-NEXT: packuswb %xmm5, %xmm4
+; X64-SSE4-NEXT: paddb %xmm0, %xmm4
+; X64-SSE4-NEXT: pand %xmm8, %xmm3
+; X64-SSE4-NEXT: pand %xmm8, %xmm2
+; X64-SSE4-NEXT: packuswb %xmm3, %xmm2
+; X64-SSE4-NEXT: pand %xmm8, %xmm7
+; X64-SSE4-NEXT: pand %xmm8, %xmm6
+; X64-SSE4-NEXT: packuswb %xmm7, %xmm6
+; X64-SSE4-NEXT: paddb %xmm2, %xmm6
+; X64-SSE4-NEXT: pxor %xmm0, %xmm0
+; X64-SSE4-NEXT: psadbw %xmm0, %xmm6
+; X64-SSE4-NEXT: psadbw %xmm0, %xmm4
+; X64-SSE4-NEXT: paddq %xmm6, %xmm4
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]
+; X64-SSE4-NEXT: paddq %xmm4, %xmm0
+; X64-SSE4-NEXT: movd %xmm0, %eax
+; X64-SSE4-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-SSE4-NEXT: retq
+;
+; X86-AVX1-LABEL: test_v64i16_v64i8:
+; 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: vbroadcastss {{.*#+}} ymm3 = [127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127]
+; X86-AVX1-NEXT: vandps %ymm3, %ymm1, %ymm1
+; X86-AVX1-NEXT: vandps %ymm3, %ymm0, %ymm0
+; X86-AVX1-NEXT: vandps %ymm3, %ymm2, %ymm2
+; X86-AVX1-NEXT: vandps 8(%ebp), %ymm3, %ymm3
+; X86-AVX1-NEXT: vextractf128 $1, %ymm2, %xmm4
+; X86-AVX1-NEXT: vpackuswb %xmm4, %xmm2, %xmm2
+; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm4
+; X86-AVX1-NEXT: vpackuswb %xmm4, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddb %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpsadbw %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm4
+; X86-AVX1-NEXT: vpackuswb %xmm4, %xmm1, %xmm1
+; X86-AVX1-NEXT: vextractf128 $1, %ymm3, %xmm4
+; X86-AVX1-NEXT: vpackuswb %xmm4, %xmm3, %xmm3
+; X86-AVX1-NEXT: vpaddb %xmm3, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpsadbw %xmm2, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddq %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_v64i8:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vbroadcastss {{.*#+}} ymm4 = [127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127]
+; X64-AVX1-NEXT: vandps %ymm4, %ymm0, %ymm0
+; X64-AVX1-NEXT: vandps %ymm4, %ymm2, %ymm2
+; X64-AVX1-NEXT: vandps %ymm4, %ymm1, %ymm1
+; X64-AVX1-NEXT: vandps %ymm4, %ymm3, %ymm3
+; X64-AVX1-NEXT: vextractf128 $1, %ymm3, %xmm4
+; X64-AVX1-NEXT: vpackuswb %xmm4, %xmm3, %xmm3
+; X64-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm4
+; X64-AVX1-NEXT: vpackuswb %xmm4, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpaddb %xmm3, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpxor %xmm3, %xmm3, %xmm3
+; X64-AVX1-NEXT: vpsadbw %xmm3, %xmm1, %xmm1
+; X64-AVX1-NEXT: vextractf128 $1, %ymm2, %xmm4
+; X64-AVX1-NEXT: vpackuswb %xmm4, %xmm2, %xmm2
+; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm4
+; X64-AVX1-NEXT: vpackuswb %xmm4, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddb %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpsadbw %xmm3, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddq %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_v64i8:
+; 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: vpbroadcastw {{.*#+}} ymm3 = [127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127]
+; X86-AVX2-NEXT: vpand %ymm3, %ymm2, %ymm2
+; X86-AVX2-NEXT: vpand %ymm3, %ymm1, %ymm1
+; X86-AVX2-NEXT: vpand %ymm3, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpackuswb %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpand 8(%ebp), %ymm3, %ymm1
+; X86-AVX2-NEXT: vpackuswb %ymm1, %ymm2, %ymm1
+; X86-AVX2-NEXT: vpaddb %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
+; X86-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X86-AVX2-NEXT: vpsadbw %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddq %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_v64i8:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpbroadcastw {{.*#+}} ymm4 = [127,127,127,127,127,127,127,127,127,127,127,127,127,127,127,127]
+; X64-AVX2-NEXT: vpand %ymm4, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpand %ymm4, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpackuswb %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpand %ymm4, %ymm3, %ymm1
+; X64-AVX2-NEXT: vpand %ymm4, %ymm2, %ymm2
+; X64-AVX2-NEXT: vpackuswb %ymm1, %ymm2, %ymm1
+; X64-AVX2-NEXT: vpaddb %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
+; X64-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X64-AVX2-NEXT: vpsadbw %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddq %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
;
; AVX512-LABEL: test_v64i16_v64i8:
; AVX512: # %bb.0:
@@ -1314,3 +2149,6 @@ declare i8 @llvm.vector.reduce.add.v16i8(<16 x i8>)
declare i8 @llvm.vector.reduce.add.v32i8(<32 x i8>)
declare i8 @llvm.vector.reduce.add.v64i8(<64 x i8>)
declare i8 @llvm.vector.reduce.add.v128i8(<128 x i8>)
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; AVX: {{.*}}
+; SSE2: {{.*}}
diff --git a/llvm/test/CodeGen/X86/vector-reduce-add-sext.ll b/llvm/test/CodeGen/X86/vector-reduce-add-sext.ll
index dc0ebe6b2e2ef..3b041304e252d 100644
--- a/llvm/test/CodeGen/X86/vector-reduce-add-sext.ll
+++ b/llvm/test/CodeGen/X86/vector-reduce-add-sext.ll
@@ -1,95 +1,210 @@
; 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=SSE2
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefixes=SSE41
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-SLOW
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx,+fast-hops | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-FAST
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefixes=AVX,AVX512
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw,+avx512vl | FileCheck %s --check-prefixes=AVX,AVX512
+; 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,SSE41,X86-SSE4,X86-SSE41
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse4.1 | FileCheck %s --check-prefixes=SSE,X64-SSE,SSE4,SSE41,X64-SSE4,X64-SSE41
+; RUN: llc < %s -mtriple=i686-- -mattr=+sse4.2 | FileCheck %s --check-prefixes=SSE,X86-SSE,SSE4,SSE42,X86-SSE4,X86-SSE42
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse4.2 | FileCheck %s --check-prefixes=SSE,X64-SSE,SSE4,SSE42,X64-SSE4,X64-SSE42
+; RUN: llc < %s -mtriple=i686-- -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-SLOW,X86-AVX1,X86-AVX1-SLOW
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-SLOW,X64-AVX1,X64-AVX1-SLOW
+; RUN: llc < %s -mtriple=i686-- -mattr=+avx,+fast-hops | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-FAST,X86-AVX1,X86-AVX1-FAST
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx,+fast-hops | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-FAST,X64-AVX1,X64-AVX1-FAST
+; RUN: llc < %s -mtriple=i686-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2,X86-AVX2
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2,X64-AVX2
+; 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
;
; vXi64
;
-define i64 @test_v2i64_v2i32(<2 x i32> %a0) {
-; SSE2-LABEL: test_v2i64_v2i32:
-; SSE2: # %bb.0:
-; SSE2-NEXT: pxor %xmm1, %xmm1
-; SSE2-NEXT: pcmpgtd %xmm0, %xmm1
-; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT: paddq %xmm0, %xmm1
-; SSE2-NEXT: movq %xmm1, %rax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v2i64_v2i32:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovsxdq %xmm0, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: paddq %xmm0, %xmm1
-; SSE41-NEXT: movq %xmm1, %rax
-; SSE41-NEXT: retq
-;
-; AVX-LABEL: test_v2i64_v2i32:
-; AVX: # %bb.0:
-; AVX-NEXT: vpmovsxdq %xmm0, %xmm0
-; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vmovq %xmm0, %rax
-; AVX-NEXT: retq
+define i64 @test_v2i64_v2i32(<2 x i32> %a0) nounwind {
+; X86-SSE2-LABEL: test_v2i64_v2i32:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: pxor %xmm1, %xmm1
+; X86-SSE2-NEXT: pcmpgtd %xmm0, %xmm1
+; 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: paddq %xmm0, %xmm1
+; X86-SSE2-NEXT: movd %xmm1, %eax
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE2-NEXT: movd %xmm0, %edx
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE2-LABEL: test_v2i64_v2i32:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: pxor %xmm1, %xmm1
+; X64-SSE2-NEXT: pcmpgtd %xmm0, %xmm1
+; 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: paddq %xmm0, %xmm1
+; X64-SSE2-NEXT: movq %xmm1, %rax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v2i64_v2i32:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pmovsxdq %xmm0, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: paddq %xmm0, %xmm1
+; X86-SSE4-NEXT: movd %xmm1, %eax
+; X86-SSE4-NEXT: pextrd $1, %xmm1, %edx
+; X86-SSE4-NEXT: retl
+;
+; X64-SSE4-LABEL: test_v2i64_v2i32:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: pmovsxdq %xmm0, %xmm0
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE4-NEXT: paddq %xmm0, %xmm1
+; X64-SSE4-NEXT: movq %xmm1, %rax
+; X64-SSE4-NEXT: retq
+;
+; X86-AVX1-LABEL: test_v2i64_v2i32:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vpmovsxdq %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v2i64_v2i32:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vpmovsxdq %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovq %xmm0, %rax
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v2i64_v2i32:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpmovsxdq %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v2i64_v2i32:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpmovsxdq %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: retq
+;
+; AVX512-LABEL: test_v2i64_v2i32:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vpmovsxdq %xmm0, %xmm0
+; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: vmovq %xmm0, %rax
+; AVX512-NEXT: retq
%1 = sext <2 x i32> %a0 to <2 x i64>
%2 = call i64 @llvm.vector.reduce.add.v2i64(<2 x i64> %1)
ret i64 %2
}
-define i64 @test_v4i64_v4i16(<4 x i16> %a0) {
-; SSE2-LABEL: test_v4i64_v4i16:
-; SSE2: # %bb.0:
-; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
-; SSE2-NEXT: psrad $16, %xmm0
-; SSE2-NEXT: pxor %xmm1, %xmm1
-; SSE2-NEXT: pcmpgtd %xmm0, %xmm1
-; SSE2-NEXT: movdqa %xmm0, %xmm2
-; SSE2-NEXT: punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
-; SSE2-NEXT: paddq %xmm2, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT: paddq %xmm0, %xmm1
-; SSE2-NEXT: movq %xmm1, %rax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v4i64_v4i16:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovsxwq %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
-; SSE41-NEXT: pmovsxwq %xmm0, %xmm0
-; SSE41-NEXT: paddq %xmm1, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: paddq %xmm0, %xmm1
-; SSE41-NEXT: movq %xmm1, %rax
-; SSE41-NEXT: retq
-;
-; AVX1-LABEL: test_v4i64_v4i16:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vpmovsxwq %xmm0, %xmm1
-; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
-; AVX1-NEXT: vpmovsxwq %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm0, %xmm1, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vmovq %xmm0, %rax
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: test_v4i64_v4i16:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpmovsxwq %xmm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovq %xmm0, %rax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+define i64 @test_v4i64_v4i16(<4 x i16> %a0) nounwind {
+; X86-SSE2-LABEL: test_v4i64_v4i16:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
+; X86-SSE2-NEXT: psrad $16, %xmm0
+; X86-SSE2-NEXT: pxor %xmm1, %xmm1
+; X86-SSE2-NEXT: pcmpgtd %xmm0, %xmm1
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm2
+; X86-SSE2-NEXT: punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
+; X86-SSE2-NEXT: paddq %xmm2, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE2-NEXT: paddq %xmm0, %xmm1
+; X86-SSE2-NEXT: movd %xmm1, %eax
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE2-NEXT: movd %xmm0, %edx
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE2-LABEL: test_v4i64_v4i16:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
+; X64-SSE2-NEXT: psrad $16, %xmm0
+; X64-SSE2-NEXT: pxor %xmm1, %xmm1
+; X64-SSE2-NEXT: pcmpgtd %xmm0, %xmm1
+; X64-SSE2-NEXT: movdqa %xmm0, %xmm2
+; X64-SSE2-NEXT: punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
+; X64-SSE2-NEXT: paddq %xmm2, %xmm0
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE2-NEXT: paddq %xmm0, %xmm1
+; X64-SSE2-NEXT: movq %xmm1, %rax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v4i64_v4i16:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pmovsxwq %xmm0, %xmm1
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; X86-SSE4-NEXT: pmovsxwq %xmm0, %xmm0
+; X86-SSE4-NEXT: paddq %xmm1, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: paddq %xmm0, %xmm1
+; X86-SSE4-NEXT: movd %xmm1, %eax
+; X86-SSE4-NEXT: pextrd $1, %xmm1, %edx
+; X86-SSE4-NEXT: retl
+;
+; X64-SSE4-LABEL: test_v4i64_v4i16:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: pmovsxwq %xmm0, %xmm1
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; X64-SSE4-NEXT: pmovsxwq %xmm0, %xmm0
+; X64-SSE4-NEXT: paddq %xmm1, %xmm0
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE4-NEXT: paddq %xmm0, %xmm1
+; X64-SSE4-NEXT: movq %xmm1, %rax
+; X64-SSE4-NEXT: retq
+;
+; X86-AVX1-LABEL: test_v4i64_v4i16:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vpmovsxwq %xmm0, %xmm1
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; X86-AVX1-NEXT: vpmovsxwq %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm0, %xmm1, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v4i64_v4i16:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vpmovsxwq %xmm0, %xmm1
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; X64-AVX1-NEXT: vpmovsxwq %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm0, %xmm1, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovq %xmm0, %rax
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v4i64_v4i16:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpmovsxwq %xmm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddq %xmm1, %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_v4i16:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpmovsxwq %xmm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
;
; AVX512-LABEL: test_v4i64_v4i16:
; AVX512: # %bb.0:
@@ -106,80 +221,159 @@ define i64 @test_v4i64_v4i16(<4 x i16> %a0) {
ret i64 %2
}
-define i64 @test_v8i64_v8i8(<8 x i8> %a0) {
-; SSE2-LABEL: test_v8i64_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]
-; SSE2-NEXT: punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; SSE2-NEXT: psrad $24, %xmm1
-; SSE2-NEXT: pxor %xmm2, %xmm2
-; SSE2-NEXT: pxor %xmm3, %xmm3
-; SSE2-NEXT: pcmpgtd %xmm1, %xmm3
-; SSE2-NEXT: movdqa %xmm1, %xmm4
-; SSE2-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]
-; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
-; SSE2-NEXT: psrad $24, %xmm0
-; SSE2-NEXT: pcmpgtd %xmm0, %xmm2
-; SSE2-NEXT: movdqa %xmm0, %xmm5
-; SSE2-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm2[0],xmm5[1],xmm2[1]
-; SSE2-NEXT: paddq %xmm4, %xmm5
-; SSE2-NEXT: punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm3[2],xmm1[3],xmm3[3]
-; SSE2-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3]
-; SSE2-NEXT: paddq %xmm1, %xmm0
-; SSE2-NEXT: paddq %xmm5, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT: paddq %xmm0, %xmm1
-; SSE2-NEXT: movq %xmm1, %rax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v8i64_v8i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovsxbq %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,1,1]
-; SSE41-NEXT: pmovsxbq %xmm2, %xmm2
-; SSE41-NEXT: paddq %xmm1, %xmm2
-; SSE41-NEXT: movdqa %xmm0, %xmm1
-; SSE41-NEXT: psrlq $48, %xmm1
-; SSE41-NEXT: pmovsxbq %xmm1, %xmm1
-; SSE41-NEXT: psrld $16, %xmm0
-; SSE41-NEXT: pmovsxbq %xmm0, %xmm0
-; SSE41-NEXT: paddq %xmm1, %xmm0
-; SSE41-NEXT: paddq %xmm2, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: paddq %xmm0, %xmm1
-; SSE41-NEXT: movq %xmm1, %rax
-; SSE41-NEXT: retq
-;
-; AVX1-LABEL: test_v8i64_v8i8:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vpsrld $16, %xmm0, %xmm1
-; AVX1-NEXT: vpmovsxbq %xmm1, %xmm1
-; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,1,1,1]
-; AVX1-NEXT: vpsrld $16, %xmm2, %xmm3
-; AVX1-NEXT: vpmovsxbq %xmm3, %xmm3
-; AVX1-NEXT: vpaddq %xmm3, %xmm1, %xmm1
-; AVX1-NEXT: vpmovsxbq %xmm0, %xmm0
-; AVX1-NEXT: vpmovsxbq %xmm2, %xmm2
-; AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vmovq %xmm0, %rax
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: test_v8i64_v8i8:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpmovsxbq %xmm0, %ymm1
-; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
-; AVX2-NEXT: vpmovsxbq %xmm0, %ymm0
-; AVX2-NEXT: vpaddq %ymm0, %ymm1, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovq %xmm0, %rax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+define i64 @test_v8i64_v8i8(<8 x i8> %a0) nounwind {
+; X86-SSE2-LABEL: test_v8i64_v8i8:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; X86-SSE2-NEXT: punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
+; X86-SSE2-NEXT: psrad $24, %xmm0
+; X86-SSE2-NEXT: pxor %xmm2, %xmm2
+; X86-SSE2-NEXT: pxor %xmm3, %xmm3
+; X86-SSE2-NEXT: pcmpgtd %xmm0, %xmm3
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm4
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]
+; X86-SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3]
+; X86-SSE2-NEXT: psrad $24, %xmm1
+; X86-SSE2-NEXT: pcmpgtd %xmm1, %xmm2
+; X86-SSE2-NEXT: movdqa %xmm1, %xmm5
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm2[0],xmm5[1],xmm2[1]
+; X86-SSE2-NEXT: paddq %xmm4, %xmm5
+; X86-SSE2-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm3[2],xmm0[3],xmm3[3]
+; X86-SSE2-NEXT: punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm2[2],xmm1[3],xmm2[3]
+; X86-SSE2-NEXT: paddq %xmm0, %xmm1
+; X86-SSE2-NEXT: paddq %xmm5, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE2-NEXT: paddq %xmm1, %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-SSE2-LABEL: test_v8i64_v8i8:
+; X64-SSE2: # %bb.0:
+; 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: punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
+; X64-SSE2-NEXT: psrad $24, %xmm1
+; X64-SSE2-NEXT: pxor %xmm2, %xmm2
+; X64-SSE2-NEXT: pxor %xmm3, %xmm3
+; X64-SSE2-NEXT: pcmpgtd %xmm1, %xmm3
+; X64-SSE2-NEXT: movdqa %xmm1, %xmm4
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1]
+; X64-SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
+; X64-SSE2-NEXT: psrad $24, %xmm0
+; X64-SSE2-NEXT: pcmpgtd %xmm0, %xmm2
+; X64-SSE2-NEXT: movdqa %xmm0, %xmm5
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm2[0],xmm5[1],xmm2[1]
+; X64-SSE2-NEXT: paddq %xmm4, %xmm5
+; X64-SSE2-NEXT: punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm3[2],xmm1[3],xmm3[3]
+; X64-SSE2-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm2[2],xmm0[3],xmm2[3]
+; X64-SSE2-NEXT: paddq %xmm1, %xmm0
+; X64-SSE2-NEXT: paddq %xmm5, %xmm0
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE2-NEXT: paddq %xmm0, %xmm1
+; X64-SSE2-NEXT: movq %xmm1, %rax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v8i64_v8i8:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pmovsxbq %xmm0, %xmm1
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,1,1]
+; X86-SSE4-NEXT: pmovsxbq %xmm2, %xmm2
+; X86-SSE4-NEXT: paddq %xmm1, %xmm2
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm1
+; X86-SSE4-NEXT: psrlq $48, %xmm1
+; X86-SSE4-NEXT: pmovsxbq %xmm1, %xmm1
+; X86-SSE4-NEXT: psrld $16, %xmm0
+; X86-SSE4-NEXT: pmovsxbq %xmm0, %xmm0
+; X86-SSE4-NEXT: paddq %xmm1, %xmm0
+; X86-SSE4-NEXT: paddq %xmm2, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: paddq %xmm0, %xmm1
+; X86-SSE4-NEXT: movd %xmm1, %eax
+; X86-SSE4-NEXT: pextrd $1, %xmm1, %edx
+; X86-SSE4-NEXT: retl
+;
+; X64-SSE4-LABEL: test_v8i64_v8i8:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: pmovsxbq %xmm0, %xmm1
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm2 = xmm0[1,1,1,1]
+; X64-SSE4-NEXT: pmovsxbq %xmm2, %xmm2
+; X64-SSE4-NEXT: paddq %xmm1, %xmm2
+; X64-SSE4-NEXT: movdqa %xmm0, %xmm1
+; X64-SSE4-NEXT: psrlq $48, %xmm1
+; X64-SSE4-NEXT: pmovsxbq %xmm1, %xmm1
+; X64-SSE4-NEXT: psrld $16, %xmm0
+; X64-SSE4-NEXT: pmovsxbq %xmm0, %xmm0
+; X64-SSE4-NEXT: paddq %xmm1, %xmm0
+; X64-SSE4-NEXT: paddq %xmm2, %xmm0
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE4-NEXT: paddq %xmm0, %xmm1
+; X64-SSE4-NEXT: movq %xmm1, %rax
+; X64-SSE4-NEXT: retq
+;
+; X86-AVX1-LABEL: test_v8i64_v8i8:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vpsrld $16, %xmm0, %xmm1
+; X86-AVX1-NEXT: vpmovsxbq %xmm1, %xmm1
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,1,1,1]
+; X86-AVX1-NEXT: vpsrld $16, %xmm2, %xmm3
+; X86-AVX1-NEXT: vpmovsxbq %xmm3, %xmm3
+; X86-AVX1-NEXT: vpaddq %xmm3, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpmovsxbq %xmm0, %xmm0
+; X86-AVX1-NEXT: vpmovsxbq %xmm2, %xmm2
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v8i64_v8i8:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vpsrld $16, %xmm0, %xmm1
+; X64-AVX1-NEXT: vpmovsxbq %xmm1, %xmm1
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[1,1,1,1]
+; X64-AVX1-NEXT: vpsrld $16, %xmm2, %xmm3
+; X64-AVX1-NEXT: vpmovsxbq %xmm3, %xmm3
+; X64-AVX1-NEXT: vpaddq %xmm3, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpmovsxbq %xmm0, %xmm0
+; X64-AVX1-NEXT: vpmovsxbq %xmm2, %xmm2
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovq %xmm0, %rax
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v8i64_v8i8:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpmovsxbq %xmm0, %ymm1
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; X86-AVX2-NEXT: vpmovsxbq %xmm0, %ymm0
+; X86-AVX2-NEXT: vpaddq %ymm0, %ymm1, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddq %xmm1, %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_v8i8:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpmovsxbq %xmm0, %ymm1
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; X64-AVX2-NEXT: vpmovsxbq %xmm0, %ymm0
+; X64-AVX2-NEXT: vpaddq %ymm0, %ymm1, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
;
; AVX512-LABEL: test_v8i64_v8i8:
; AVX512: # %bb.0:
@@ -198,134 +392,275 @@ define i64 @test_v8i64_v8i8(<8 x i8> %a0) {
ret i64 %2
}
-define i64 @test_v16i64_v16i8(<16 x i8> %a0) {
-; SSE2-LABEL: test_v16i64_v16i8:
-; SSE2: # %bb.0:
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm0[8],xmm4[9],xmm0[9],xmm4[10],xmm0[10],xmm4[11],xmm0[11],xmm4[12],xmm0[12],xmm4[13],xmm0[13],xmm4[14],xmm0[14],xmm4[15],xmm0[15]
-; SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
-; SSE2-NEXT: psrad $24, %xmm2
-; SSE2-NEXT: pxor %xmm1, %xmm1
-; SSE2-NEXT: pxor %xmm3, %xmm3
-; SSE2-NEXT: pcmpgtd %xmm2, %xmm3
-; SSE2-NEXT: movdqa %xmm2, %xmm5
-; SSE2-NEXT: punpckhdq {{.*#+}} xmm5 = xmm5[2],xmm3[2],xmm5[3],xmm3[3]
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3],xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7]
-; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3]
-; SSE2-NEXT: psrad $24, %xmm0
-; SSE2-NEXT: pxor %xmm7, %xmm7
-; SSE2-NEXT: pcmpgtd %xmm0, %xmm7
-; SSE2-NEXT: movdqa %xmm0, %xmm8
-; SSE2-NEXT: punpckhdq {{.*#+}} xmm8 = xmm8[2],xmm7[2],xmm8[3],xmm7[3]
-; SSE2-NEXT: paddq %xmm5, %xmm8
-; SSE2-NEXT: punpckhwd {{.*#+}} xmm4 = xmm4[4,4,5,5,6,6,7,7]
-; SSE2-NEXT: psrad $24, %xmm4
-; SSE2-NEXT: pxor %xmm5, %xmm5
-; SSE2-NEXT: pcmpgtd %xmm4, %xmm5
-; SSE2-NEXT: movdqa %xmm4, %xmm9
-; SSE2-NEXT: punpckhdq {{.*#+}} xmm9 = xmm9[2],xmm5[2],xmm9[3],xmm5[3]
-; SSE2-NEXT: punpckhwd {{.*#+}} xmm6 = xmm6[4,4,5,5,6,6,7,7]
-; SSE2-NEXT: psrad $24, %xmm6
-; SSE2-NEXT: pcmpgtd %xmm6, %xmm1
-; SSE2-NEXT: movdqa %xmm6, %xmm10
-; SSE2-NEXT: punpckhdq {{.*#+}} xmm10 = xmm10[2],xmm1[2],xmm10[3],xmm1[3]
-; SSE2-NEXT: paddq %xmm9, %xmm10
-; SSE2-NEXT: paddq %xmm8, %xmm10
-; SSE2-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]
-; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1]
-; SSE2-NEXT: paddq %xmm2, %xmm0
-; SSE2-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1]
-; SSE2-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm1[0],xmm6[1],xmm1[1]
-; SSE2-NEXT: paddq %xmm4, %xmm6
-; SSE2-NEXT: paddq %xmm0, %xmm6
-; SSE2-NEXT: paddq %xmm10, %xmm6
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm6[2,3,2,3]
-; SSE2-NEXT: paddq %xmm6, %xmm0
-; SSE2-NEXT: movq %xmm0, %rax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v16i64_v16i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: movdqa %xmm0, %xmm1
-; SSE41-NEXT: movdqa %xmm0, %xmm2
-; SSE41-NEXT: movdqa %xmm0, %xmm3
-; SSE41-NEXT: pmovsxbq %xmm0, %xmm4
-; SSE41-NEXT: pshufd {{.*#+}} xmm5 = xmm0[2,3,2,3]
-; SSE41-NEXT: pshufd {{.*#+}} xmm6 = xmm0[3,3,3,3]
-; SSE41-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,1,1]
-; SSE41-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; SSE41-NEXT: pmovsxbq %xmm0, %xmm0
-; SSE41-NEXT: psrld $16, %xmm1
-; SSE41-NEXT: pmovsxbq %xmm1, %xmm1
-; SSE41-NEXT: paddq %xmm0, %xmm1
-; SSE41-NEXT: psrldq {{.*#+}} xmm2 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
-; SSE41-NEXT: pmovsxbq %xmm2, %xmm0
-; SSE41-NEXT: psrlq $48, %xmm3
-; SSE41-NEXT: pmovsxbq %xmm3, %xmm2
-; SSE41-NEXT: paddq %xmm0, %xmm2
-; SSE41-NEXT: paddq %xmm1, %xmm2
-; SSE41-NEXT: pmovsxbq %xmm5, %xmm0
-; SSE41-NEXT: paddq %xmm4, %xmm0
-; SSE41-NEXT: pmovsxbq %xmm6, %xmm1
-; SSE41-NEXT: pmovsxbq %xmm7, %xmm3
-; SSE41-NEXT: paddq %xmm1, %xmm3
-; SSE41-NEXT: paddq %xmm0, %xmm3
-; SSE41-NEXT: paddq %xmm2, %xmm3
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
-; SSE41-NEXT: paddq %xmm3, %xmm0
-; SSE41-NEXT: movq %xmm0, %rax
-; SSE41-NEXT: retq
-;
-; AVX1-LABEL: test_v16i64_v16i8:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vpsrld $16, %xmm0, %xmm1
-; AVX1-NEXT: vpmovsxbq %xmm1, %xmm1
-; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpmovsxbw %xmm2, %xmm3
-; AVX1-NEXT: vpshufd {{.*#+}} xmm4 = xmm3[1,1,1,1]
-; AVX1-NEXT: vpmovsxwq %xmm4, %xmm4
-; AVX1-NEXT: vpaddq %xmm4, %xmm1, %xmm1
-; AVX1-NEXT: vpmovsxbw %xmm0, %xmm4
-; AVX1-NEXT: vpshufd {{.*#+}} xmm5 = xmm4[3,3,3,3]
-; AVX1-NEXT: vpmovsxwq %xmm5, %xmm5
-; AVX1-NEXT: vpshufd {{.*#+}} xmm6 = xmm3[3,3,3,3]
-; AVX1-NEXT: vpmovsxwq %xmm6, %xmm6
-; AVX1-NEXT: vpaddq %xmm6, %xmm5, %xmm5
-; AVX1-NEXT: vpaddq %xmm5, %xmm1, %xmm1
-; AVX1-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,3,2,3]
-; AVX1-NEXT: vpmovsxwq %xmm4, %xmm4
-; AVX1-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,3,2,3]
-; AVX1-NEXT: vpmovsxwq %xmm3, %xmm3
-; AVX1-NEXT: vpaddq %xmm3, %xmm4, %xmm3
-; AVX1-NEXT: vpmovsxbq %xmm0, %xmm0
-; AVX1-NEXT: vpmovsxbq %xmm2, %xmm2
-; AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm3, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vmovq %xmm0, %rax
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: test_v16i64_v16i8:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpmovsxbw %xmm0, %ymm1
-; AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
-; AVX2-NEXT: vpshufd {{.*#+}} xmm3 = xmm2[2,3,2,3]
-; AVX2-NEXT: vpmovsxwq %xmm3, %ymm3
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
-; AVX2-NEXT: vpmovsxwq %xmm1, %ymm1
-; AVX2-NEXT: vpaddq %ymm3, %ymm1, %ymm1
-; AVX2-NEXT: vpmovsxbq %xmm0, %ymm0
-; AVX2-NEXT: vpmovsxwq %xmm2, %ymm2
-; AVX2-NEXT: vpaddq %ymm2, %ymm0, %ymm0
-; AVX2-NEXT: vpaddq %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovq %xmm0, %rax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+define i64 @test_v16i64_v16i8(<16 x i8> %a0) nounwind {
+; X86-SSE2-LABEL: test_v16i64_v16i8:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: subl $48, %esp
+; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm0[8],xmm3[9],xmm0[9],xmm3[10],xmm0[10],xmm3[11],xmm0[11],xmm3[12],xmm0[12],xmm3[13],xmm0[13],xmm3[14],xmm0[14],xmm3[15],xmm0[15]
+; X86-SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
+; X86-SSE2-NEXT: psrad $24, %xmm2
+; X86-SSE2-NEXT: pxor %xmm1, %xmm1
+; X86-SSE2-NEXT: pcmpgtd %xmm2, %xmm1
+; X86-SSE2-NEXT: movdqu %xmm1, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
+; X86-SSE2-NEXT: movdqa %xmm2, %xmm4
+; X86-SSE2-NEXT: punpckhdq {{.*#+}} xmm4 = xmm4[2],xmm1[2],xmm4[3],xmm1[3]
+; 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: punpcklwd {{.*#+}} xmm7 = xmm7[0],xmm0[0],xmm7[1],xmm0[1],xmm7[2],xmm0[2],xmm7[3],xmm0[3]
+; X86-SSE2-NEXT: psrad $24, %xmm7
+; X86-SSE2-NEXT: pxor %xmm5, %xmm5
+; X86-SSE2-NEXT: pcmpgtd %xmm7, %xmm5
+; X86-SSE2-NEXT: movdqu %xmm5, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
+; X86-SSE2-NEXT: movdqa %xmm7, %xmm1
+; X86-SSE2-NEXT: punpckhdq {{.*#+}} xmm1 = xmm1[2],xmm5[2],xmm1[3],xmm5[3]
+; X86-SSE2-NEXT: paddq %xmm4, %xmm1
+; X86-SSE2-NEXT: punpckhwd {{.*#+}} xmm3 = xmm3[4,4,5,5,6,6,7,7]
+; X86-SSE2-NEXT: psrad $24, %xmm3
+; X86-SSE2-NEXT: pxor %xmm4, %xmm4
+; X86-SSE2-NEXT: pcmpgtd %xmm3, %xmm4
+; X86-SSE2-NEXT: movdqu %xmm4, (%esp) # 16-byte Spill
+; X86-SSE2-NEXT: movdqa %xmm3, %xmm6
+; X86-SSE2-NEXT: punpckhdq {{.*#+}} xmm6 = xmm6[2],xmm4[2],xmm6[3],xmm4[3]
+; X86-SSE2-NEXT: punpckhwd {{.*#+}} xmm4 = xmm4[4],xmm0[4],xmm4[5],xmm0[5],xmm4[6],xmm0[6],xmm4[7],xmm0[7]
+; X86-SSE2-NEXT: psrad $24, %xmm4
+; X86-SSE2-NEXT: pxor %xmm5, %xmm5
+; X86-SSE2-NEXT: pcmpgtd %xmm4, %xmm5
+; X86-SSE2-NEXT: movdqa %xmm4, %xmm0
+; X86-SSE2-NEXT: punpckhdq {{.*#+}} xmm0 = xmm0[2],xmm5[2],xmm0[3],xmm5[3]
+; X86-SSE2-NEXT: paddq %xmm6, %xmm0
+; X86-SSE2-NEXT: paddq %xmm1, %xmm0
+; X86-SSE2-NEXT: movdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 # 16-byte Reload
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
+; X86-SSE2-NEXT: movdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 # 16-byte Reload
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm7 = xmm7[0],xmm1[0],xmm7[1],xmm1[1]
+; X86-SSE2-NEXT: paddq %xmm2, %xmm7
+; X86-SSE2-NEXT: movdqu (%esp), %xmm1 # 16-byte Reload
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1]
+; X86-SSE2-NEXT: paddq %xmm3, %xmm4
+; X86-SSE2-NEXT: paddq %xmm7, %xmm4
+; X86-SSE2-NEXT: paddq %xmm0, %xmm4
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]
+; X86-SSE2-NEXT: paddq %xmm4, %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: addl $48, %esp
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE2-LABEL: test_v16i64_v16i8:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm0[8],xmm4[9],xmm0[9],xmm4[10],xmm0[10],xmm4[11],xmm0[11],xmm4[12],xmm0[12],xmm4[13],xmm0[13],xmm4[14],xmm0[14],xmm4[15],xmm0[15]
+; X64-SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
+; X64-SSE2-NEXT: psrad $24, %xmm2
+; X64-SSE2-NEXT: pxor %xmm1, %xmm1
+; X64-SSE2-NEXT: pxor %xmm3, %xmm3
+; X64-SSE2-NEXT: pcmpgtd %xmm2, %xmm3
+; X64-SSE2-NEXT: movdqa %xmm2, %xmm5
+; X64-SSE2-NEXT: punpckhdq {{.*#+}} xmm5 = xmm5[2],xmm3[2],xmm5[3],xmm3[3]
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm6 = xmm6[0],xmm0[0],xmm6[1],xmm0[1],xmm6[2],xmm0[2],xmm6[3],xmm0[3],xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7]
+; X64-SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm6[0],xmm0[1],xmm6[1],xmm0[2],xmm6[2],xmm0[3],xmm6[3]
+; X64-SSE2-NEXT: psrad $24, %xmm0
+; X64-SSE2-NEXT: pxor %xmm7, %xmm7
+; X64-SSE2-NEXT: pcmpgtd %xmm0, %xmm7
+; X64-SSE2-NEXT: movdqa %xmm0, %xmm8
+; X64-SSE2-NEXT: punpckhdq {{.*#+}} xmm8 = xmm8[2],xmm7[2],xmm8[3],xmm7[3]
+; X64-SSE2-NEXT: paddq %xmm5, %xmm8
+; X64-SSE2-NEXT: punpckhwd {{.*#+}} xmm4 = xmm4[4,4,5,5,6,6,7,7]
+; X64-SSE2-NEXT: psrad $24, %xmm4
+; X64-SSE2-NEXT: pxor %xmm5, %xmm5
+; X64-SSE2-NEXT: pcmpgtd %xmm4, %xmm5
+; X64-SSE2-NEXT: movdqa %xmm4, %xmm9
+; X64-SSE2-NEXT: punpckhdq {{.*#+}} xmm9 = xmm9[2],xmm5[2],xmm9[3],xmm5[3]
+; X64-SSE2-NEXT: punpckhwd {{.*#+}} xmm6 = xmm6[4,4,5,5,6,6,7,7]
+; X64-SSE2-NEXT: psrad $24, %xmm6
+; X64-SSE2-NEXT: pcmpgtd %xmm6, %xmm1
+; X64-SSE2-NEXT: movdqa %xmm6, %xmm10
+; X64-SSE2-NEXT: punpckhdq {{.*#+}} xmm10 = xmm10[2],xmm1[2],xmm10[3],xmm1[3]
+; X64-SSE2-NEXT: paddq %xmm9, %xmm10
+; X64-SSE2-NEXT: paddq %xmm8, %xmm10
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm7[0],xmm0[1],xmm7[1]
+; X64-SSE2-NEXT: paddq %xmm2, %xmm0
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm5[0],xmm4[1],xmm5[1]
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm6 = xmm6[0],xmm1[0],xmm6[1],xmm1[1]
+; X64-SSE2-NEXT: paddq %xmm4, %xmm6
+; X64-SSE2-NEXT: paddq %xmm0, %xmm6
+; X64-SSE2-NEXT: paddq %xmm10, %xmm6
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm6[2,3,2,3]
+; X64-SSE2-NEXT: paddq %xmm6, %xmm0
+; X64-SSE2-NEXT: movq %xmm0, %rax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v16i64_v16i8:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm1
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm2
+; X86-SSE4-NEXT: movdqa %xmm0, %xmm3
+; X86-SSE4-NEXT: pmovsxbq %xmm0, %xmm4
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm5 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm6 = xmm0[3,3,3,3]
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,1,1]
+; X86-SSE4-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE4-NEXT: pmovsxbq %xmm0, %xmm0
+; X86-SSE4-NEXT: psrld $16, %xmm1
+; X86-SSE4-NEXT: pmovsxbq %xmm1, %xmm1
+; X86-SSE4-NEXT: paddq %xmm0, %xmm1
+; X86-SSE4-NEXT: psrldq {{.*#+}} xmm2 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X86-SSE4-NEXT: pmovsxbq %xmm2, %xmm0
+; X86-SSE4-NEXT: psrlq $48, %xmm3
+; X86-SSE4-NEXT: pmovsxbq %xmm3, %xmm2
+; X86-SSE4-NEXT: paddq %xmm0, %xmm2
+; X86-SSE4-NEXT: paddq %xmm1, %xmm2
+; X86-SSE4-NEXT: pmovsxbq %xmm5, %xmm0
+; X86-SSE4-NEXT: paddq %xmm4, %xmm0
+; X86-SSE4-NEXT: pmovsxbq %xmm6, %xmm1
+; X86-SSE4-NEXT: pmovsxbq %xmm7, %xmm3
+; X86-SSE4-NEXT: paddq %xmm1, %xmm3
+; X86-SSE4-NEXT: paddq %xmm0, %xmm3
+; X86-SSE4-NEXT: paddq %xmm2, %xmm3
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
+; X86-SSE4-NEXT: paddq %xmm3, %xmm0
+; X86-SSE4-NEXT: movd %xmm0, %eax
+; X86-SSE4-NEXT: pextrd $1, %xmm0, %edx
+; X86-SSE4-NEXT: retl
+;
+; X64-SSE4-LABEL: test_v16i64_v16i8:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: movdqa %xmm0, %xmm1
+; X64-SSE4-NEXT: movdqa %xmm0, %xmm2
+; X64-SSE4-NEXT: movdqa %xmm0, %xmm3
+; X64-SSE4-NEXT: pmovsxbq %xmm0, %xmm4
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm5 = xmm0[2,3,2,3]
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm6 = xmm0[3,3,3,3]
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm7 = xmm0[1,1,1,1]
+; X64-SSE4-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE4-NEXT: pmovsxbq %xmm0, %xmm0
+; X64-SSE4-NEXT: psrld $16, %xmm1
+; X64-SSE4-NEXT: pmovsxbq %xmm1, %xmm1
+; X64-SSE4-NEXT: paddq %xmm0, %xmm1
+; X64-SSE4-NEXT: psrldq {{.*#+}} xmm2 = xmm2[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
+; X64-SSE4-NEXT: pmovsxbq %xmm2, %xmm0
+; X64-SSE4-NEXT: psrlq $48, %xmm3
+; X64-SSE4-NEXT: pmovsxbq %xmm3, %xmm2
+; X64-SSE4-NEXT: paddq %xmm0, %xmm2
+; X64-SSE4-NEXT: paddq %xmm1, %xmm2
+; X64-SSE4-NEXT: pmovsxbq %xmm5, %xmm0
+; X64-SSE4-NEXT: paddq %xmm4, %xmm0
+; X64-SSE4-NEXT: pmovsxbq %xmm6, %xmm1
+; X64-SSE4-NEXT: pmovsxbq %xmm7, %xmm3
+; X64-SSE4-NEXT: paddq %xmm1, %xmm3
+; X64-SSE4-NEXT: paddq %xmm0, %xmm3
+; X64-SSE4-NEXT: paddq %xmm2, %xmm3
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
+; X64-SSE4-NEXT: paddq %xmm3, %xmm0
+; X64-SSE4-NEXT: movq %xmm0, %rax
+; X64-SSE4-NEXT: retq
+;
+; X86-AVX1-LABEL: test_v16i64_v16i8:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vpsrld $16, %xmm0, %xmm1
+; X86-AVX1-NEXT: vpmovsxbq %xmm1, %xmm2
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpmovsxbw %xmm1, %xmm3
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm4 = xmm3[1,1,1,1]
+; X86-AVX1-NEXT: vpmovsxwq %xmm4, %xmm4
+; X86-AVX1-NEXT: vpaddq %xmm4, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpmovsxbw %xmm0, %xmm4
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm5 = xmm4[3,3,3,3]
+; X86-AVX1-NEXT: vpmovsxwq %xmm5, %xmm5
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm6 = xmm3[3,3,3,3]
+; X86-AVX1-NEXT: vpmovsxwq %xmm6, %xmm6
+; X86-AVX1-NEXT: vpaddq %xmm6, %xmm5, %xmm5
+; X86-AVX1-NEXT: vpaddq %xmm5, %xmm2, %xmm2
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,3,2,3]
+; X86-AVX1-NEXT: vpmovsxwq %xmm4, %xmm4
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,3,2,3]
+; X86-AVX1-NEXT: vpmovsxwq %xmm3, %xmm3
+; X86-AVX1-NEXT: vpaddq %xmm3, %xmm4, %xmm3
+; X86-AVX1-NEXT: vpmovsxbq %xmm0, %xmm0
+; X86-AVX1-NEXT: vpmovsxbq %xmm1, %xmm1
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm3, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v16i64_v16i8:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vpsrld $16, %xmm0, %xmm1
+; X64-AVX1-NEXT: vpmovsxbq %xmm1, %xmm1
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpmovsxbw %xmm2, %xmm3
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm4 = xmm3[1,1,1,1]
+; X64-AVX1-NEXT: vpmovsxwq %xmm4, %xmm4
+; X64-AVX1-NEXT: vpaddq %xmm4, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpmovsxbw %xmm0, %xmm4
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm5 = xmm4[3,3,3,3]
+; X64-AVX1-NEXT: vpmovsxwq %xmm5, %xmm5
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm6 = xmm3[3,3,3,3]
+; X64-AVX1-NEXT: vpmovsxwq %xmm6, %xmm6
+; X64-AVX1-NEXT: vpaddq %xmm6, %xmm5, %xmm5
+; X64-AVX1-NEXT: vpaddq %xmm5, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[2,3,2,3]
+; X64-AVX1-NEXT: vpmovsxwq %xmm4, %xmm4
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm3 = xmm3[2,3,2,3]
+; X64-AVX1-NEXT: vpmovsxwq %xmm3, %xmm3
+; X64-AVX1-NEXT: vpaddq %xmm3, %xmm4, %xmm3
+; X64-AVX1-NEXT: vpmovsxbq %xmm0, %xmm0
+; X64-AVX1-NEXT: vpmovsxbq %xmm2, %xmm2
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm3, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovq %xmm0, %rax
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v16i64_v16i8:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpmovsxbw %xmm0, %ymm1
+; X86-AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm3 = xmm2[2,3,2,3]
+; X86-AVX2-NEXT: vpmovsxwq %xmm3, %ymm3
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
+; X86-AVX2-NEXT: vpmovsxwq %xmm1, %ymm1
+; X86-AVX2-NEXT: vpaddq %ymm3, %ymm1, %ymm1
+; X86-AVX2-NEXT: vpmovsxbq %xmm0, %ymm0
+; X86-AVX2-NEXT: vpmovsxwq %xmm2, %ymm2
+; X86-AVX2-NEXT: vpaddq %ymm2, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpaddq %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddq %xmm1, %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_v16i64_v16i8:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpmovsxbw %xmm0, %ymm1
+; X64-AVX2-NEXT: vextracti128 $1, %ymm1, %xmm2
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm3 = xmm2[2,3,2,3]
+; X64-AVX2-NEXT: vpmovsxwq %xmm3, %ymm3
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
+; X64-AVX2-NEXT: vpmovsxwq %xmm1, %ymm1
+; X64-AVX2-NEXT: vpaddq %ymm3, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpmovsxbq %xmm0, %ymm0
+; X64-AVX2-NEXT: vpmovsxwq %xmm2, %ymm2
+; X64-AVX2-NEXT: vpaddq %ymm2, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpaddq %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
;
; AVX512-LABEL: test_v16i64_v16i8:
; AVX512: # %bb.0:
@@ -352,7 +687,7 @@ define i64 @test_v16i64_v16i8(<16 x i8> %a0) {
; vXi32
;
-define i32 @test_v2i32_v2i16(<2 x i16> %a0) {
+define i32 @test_v2i32_v2i16(<2 x i16> %a0) nounwind {
; SSE2-LABEL: test_v2i32_v2i16:
; SSE2: # %bb.0:
; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
@@ -360,15 +695,15 @@ define i32 @test_v2i32_v2i16(<2 x i16> %a0) {
; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; SSE2-NEXT: paddd %xmm0, %xmm1
; SSE2-NEXT: movd %xmm1, %eax
-; SSE2-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
;
-; SSE41-LABEL: test_v2i32_v2i16:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovsxwd %xmm0, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE41-NEXT: paddd %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: retq
+; SSE4-LABEL: test_v2i32_v2i16:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pmovsxwd %xmm0, %xmm0
+; SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; SSE4-NEXT: paddd %xmm0, %xmm1
+; SSE4-NEXT: movd %xmm1, %eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v2i32_v2i16:
; AVX1-SLOW: # %bb.0:
@@ -376,14 +711,14 @@ define i32 @test_v2i32_v2i16(<2 x i16> %a0) {
; AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v2i32_v2i16:
; AVX1-FAST: # %bb.0:
; AVX1-FAST-NEXT: vpmovsxwd %xmm0, %xmm0
; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v2i32_v2i16:
; AVX2: # %bb.0:
@@ -391,7 +726,7 @@ define i32 @test_v2i32_v2i16(<2 x i16> %a0) {
; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v2i32_v2i16:
; AVX512: # %bb.0:
@@ -405,7 +740,7 @@ define i32 @test_v2i32_v2i16(<2 x i16> %a0) {
ret i32 %2
}
-define i32 @test_v4i32(<4 x i8> %a0) {
+define i32 @test_v4i32(<4 x i8> %a0) nounwind {
; SSE2-LABEL: test_v4i32:
; SSE2: # %bb.0:
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
@@ -416,17 +751,17 @@ define i32 @test_v4i32(<4 x i8> %a0) {
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
; SSE2-NEXT: paddd %xmm1, %xmm0
; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v4i32:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovsxbd %xmm0, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: paddd %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: paddd %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
+;
+; SSE4-LABEL: test_v4i32:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pmovsxbd %xmm0, %xmm0
+; SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; SSE4-NEXT: paddd %xmm0, %xmm1
+; SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; SSE4-NEXT: paddd %xmm1, %xmm0
+; SSE4-NEXT: movd %xmm0, %eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v4i32:
; AVX1-SLOW: # %bb.0:
@@ -436,7 +771,7 @@ define i32 @test_v4i32(<4 x i8> %a0) {
; AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v4i32:
; AVX1-FAST: # %bb.0:
@@ -444,7 +779,7 @@ define i32 @test_v4i32(<4 x i8> %a0) {
; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v4i32:
; AVX2: # %bb.0:
@@ -454,7 +789,7 @@ define i32 @test_v4i32(<4 x i8> %a0) {
; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v4i32:
; AVX512: # %bb.0:
@@ -470,7 +805,7 @@ define i32 @test_v4i32(<4 x i8> %a0) {
ret i32 %2
}
-define i32 @test_v8i32_v8i8(<8 x i8> %a0) {
+define i32 @test_v8i32_v8i8(<8 x i8> %a0) nounwind {
; SSE2-LABEL: test_v8i32_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]
@@ -484,20 +819,20 @@ define i32 @test_v8i32_v8i8(<8 x i8> %a0) {
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
; SSE2-NEXT: paddd %xmm1, %xmm0
; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v8i32_v8i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovsxbd %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
-; SSE41-NEXT: pmovsxbd %xmm0, %xmm0
-; SSE41-NEXT: paddd %xmm1, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: paddd %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: paddd %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
+;
+; SSE4-LABEL: test_v8i32_v8i8:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pmovsxbd %xmm0, %xmm1
+; SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; SSE4-NEXT: pmovsxbd %xmm0, %xmm0
+; SSE4-NEXT: paddd %xmm1, %xmm0
+; SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; SSE4-NEXT: paddd %xmm0, %xmm1
+; SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; SSE4-NEXT: paddd %xmm1, %xmm0
+; SSE4-NEXT: movd %xmm0, %eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v8i32_v8i8:
; AVX1-SLOW: # %bb.0:
@@ -510,7 +845,7 @@ define i32 @test_v8i32_v8i8(<8 x i8> %a0) {
; AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v8i32_v8i8:
; AVX1-FAST: # %bb.0:
@@ -521,7 +856,7 @@ define i32 @test_v8i32_v8i8(<8 x i8> %a0) {
; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v8i32_v8i8:
; AVX2: # %bb.0:
@@ -534,7 +869,7 @@ define i32 @test_v8i32_v8i8(<8 x i8> %a0) {
; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v8i32_v8i8:
; AVX512: # %bb.0:
@@ -553,7 +888,7 @@ define i32 @test_v8i32_v8i8(<8 x i8> %a0) {
ret i32 %2
}
-define i32 @test_v16i32_v16i8(<16 x i8> %a0) {
+define i32 @test_v16i32_v16i8(<16 x i8> %a0) nounwind {
; SSE2-LABEL: test_v16i32_v16i8:
; SSE2: # %bb.0:
; SSE2-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
@@ -574,26 +909,26 @@ define i32 @test_v16i32_v16i8(<16 x i8> %a0) {
; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
; SSE2-NEXT: paddd %xmm1, %xmm0
; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v16i32_v16i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovsxbd %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]
-; SSE41-NEXT: pmovsxbd %xmm2, %xmm2
-; SSE41-NEXT: paddd %xmm1, %xmm2
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[3,3,3,3]
-; SSE41-NEXT: pmovsxbd %xmm1, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
-; SSE41-NEXT: pmovsxbd %xmm0, %xmm0
-; SSE41-NEXT: paddd %xmm1, %xmm0
-; SSE41-NEXT: paddd %xmm2, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: paddd %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: paddd %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
+;
+; SSE4-LABEL: test_v16i32_v16i8:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pmovsxbd %xmm0, %xmm1
+; SSE4-NEXT: pshufd {{.*#+}} xmm2 = xmm0[2,3,2,3]
+; SSE4-NEXT: pmovsxbd %xmm2, %xmm2
+; SSE4-NEXT: paddd %xmm1, %xmm2
+; SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[3,3,3,3]
+; SSE4-NEXT: pmovsxbd %xmm1, %xmm1
+; SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; SSE4-NEXT: pmovsxbd %xmm0, %xmm0
+; SSE4-NEXT: paddd %xmm1, %xmm0
+; SSE4-NEXT: paddd %xmm2, %xmm0
+; SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; SSE4-NEXT: paddd %xmm0, %xmm1
+; SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; SSE4-NEXT: paddd %xmm1, %xmm0
+; SSE4-NEXT: movd %xmm0, %eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v16i32_v16i8:
; AVX1-SLOW: # %bb.0:
@@ -612,7 +947,7 @@ define i32 @test_v16i32_v16i8(<16 x i8> %a0) {
; AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v16i32_v16i8:
; AVX1-FAST: # %bb.0:
@@ -630,7 +965,7 @@ define i32 @test_v16i32_v16i8(<16 x i8> %a0) {
; AVX1-FAST-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v16i32_v16i8:
; AVX2: # %bb.0:
@@ -646,7 +981,7 @@ define i32 @test_v16i32_v16i8(<16 x i8> %a0) {
; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v16i32_v16i8:
; AVX512: # %bb.0:
@@ -667,72 +1002,138 @@ define i32 @test_v16i32_v16i8(<16 x i8> %a0) {
ret i32 %2
}
-define i32 @test_v32i32_v32i8(<32 x i8> %a0) {
-; SSE2-LABEL: test_v32i32_v32i8:
-; SSE2: # %bb.0:
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
-; SSE2-NEXT: punpckhwd {{.*#+}} xmm3 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
-; SSE2-NEXT: psrad $24, %xmm3
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3],xmm4[4],xmm0[4],xmm4[5],xmm0[5],xmm4[6],xmm0[6],xmm4[7],xmm0[7]
-; SSE2-NEXT: punpckhwd {{.*#+}} xmm5 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
-; SSE2-NEXT: psrad $24, %xmm5
-; SSE2-NEXT: paddd %xmm3, %xmm5
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE2-NEXT: punpckhwd {{.*#+}} xmm3 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
-; SSE2-NEXT: psrad $24, %xmm3
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm0 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
-; SSE2-NEXT: punpckhwd {{.*#+}} xmm6 = xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7]
-; SSE2-NEXT: psrad $24, %xmm6
-; SSE2-NEXT: paddd %xmm3, %xmm6
-; SSE2-NEXT: paddd %xmm5, %xmm6
-; SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3]
-; SSE2-NEXT: psrad $24, %xmm2
-; SSE2-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
-; SSE2-NEXT: psrad $24, %xmm3
-; SSE2-NEXT: paddd %xmm2, %xmm3
-; SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3]
-; SSE2-NEXT: psrad $24, %xmm1
-; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
-; SSE2-NEXT: psrad $24, %xmm0
-; SSE2-NEXT: paddd %xmm1, %xmm0
-; SSE2-NEXT: paddd %xmm3, %xmm0
-; SSE2-NEXT: paddd %xmm6, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT: paddd %xmm0, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE2-NEXT: paddd %xmm1, %xmm0
-; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v32i32_v32i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm1[1,1,1,1]
-; SSE41-NEXT: pmovsxbd %xmm2, %xmm2
-; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,1,1]
-; SSE41-NEXT: pmovsxbd %xmm3, %xmm3
-; SSE41-NEXT: paddd %xmm2, %xmm3
-; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm1[3,3,3,3]
-; SSE41-NEXT: pmovsxbd %xmm2, %xmm2
-; SSE41-NEXT: pshufd {{.*#+}} xmm4 = xmm0[3,3,3,3]
-; SSE41-NEXT: pmovsxbd %xmm4, %xmm4
-; SSE41-NEXT: paddd %xmm2, %xmm4
-; SSE41-NEXT: paddd %xmm3, %xmm4
-; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm1[2,3,2,3]
-; SSE41-NEXT: pmovsxbd %xmm2, %xmm2
-; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm0[2,3,2,3]
-; SSE41-NEXT: pmovsxbd %xmm3, %xmm3
-; SSE41-NEXT: paddd %xmm2, %xmm3
-; SSE41-NEXT: pmovsxbd %xmm1, %xmm1
-; SSE41-NEXT: pmovsxbd %xmm0, %xmm0
-; SSE41-NEXT: paddd %xmm1, %xmm0
-; SSE41-NEXT: paddd %xmm3, %xmm0
-; SSE41-NEXT: paddd %xmm4, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: paddd %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: paddd %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: retq
+define i32 @test_v32i32_v32i8(<32 x i8> %a0) nounwind {
+; X86-SSE2-LABEL: test_v32i32_v32i8:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; X86-SSE2-NEXT: punpckhwd {{.*#+}} xmm4 = xmm4[4],xmm2[4],xmm4[5],xmm2[5],xmm4[6],xmm2[6],xmm4[7],xmm2[7]
+; X86-SSE2-NEXT: psrad $24, %xmm4
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1],xmm3[2],xmm0[2],xmm3[3],xmm0[3],xmm3[4],xmm0[4],xmm3[5],xmm0[5],xmm3[6],xmm0[6],xmm3[7],xmm0[7]
+; X86-SSE2-NEXT: punpckhwd {{.*#+}} xmm5 = xmm5[4],xmm3[4],xmm5[5],xmm3[5],xmm5[6],xmm3[6],xmm5[7],xmm3[7]
+; X86-SSE2-NEXT: psrad $24, %xmm5
+; X86-SSE2-NEXT: paddd %xmm4, %xmm5
+; 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: punpckhwd {{.*#+}} xmm6 = xmm6[4],xmm1[4],xmm6[5],xmm1[5],xmm6[6],xmm1[6],xmm6[7],xmm1[7]
+; X86-SSE2-NEXT: psrad $24, %xmm6
+; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm0[8],xmm4[9],xmm0[9],xmm4[10],xmm0[10],xmm4[11],xmm0[11],xmm4[12],xmm0[12],xmm4[13],xmm0[13],xmm4[14],xmm0[14],xmm4[15],xmm0[15]
+; X86-SSE2-NEXT: punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm4[4],xmm0[5],xmm4[5],xmm0[6],xmm4[6],xmm0[7],xmm4[7]
+; X86-SSE2-NEXT: psrad $24, %xmm0
+; X86-SSE2-NEXT: paddd %xmm6, %xmm0
+; X86-SSE2-NEXT: paddd %xmm5, %xmm0
+; X86-SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3]
+; X86-SSE2-NEXT: psrad $24, %xmm2
+; X86-SSE2-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0,0,1,1,2,2,3,3]
+; X86-SSE2-NEXT: psrad $24, %xmm3
+; X86-SSE2-NEXT: paddd %xmm2, %xmm3
+; X86-SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3]
+; X86-SSE2-NEXT: psrad $24, %xmm1
+; X86-SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
+; X86-SSE2-NEXT: psrad $24, %xmm2
+; X86-SSE2-NEXT: paddd %xmm1, %xmm2
+; X86-SSE2-NEXT: paddd %xmm3, %xmm2
+; X86-SSE2-NEXT: paddd %xmm0, %xmm2
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
+; X86-SSE2-NEXT: paddd %xmm2, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE2-NEXT: paddd %xmm0, %xmm1
+; X86-SSE2-NEXT: movd %xmm1, %eax
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE2-LABEL: test_v32i32_v32i8:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
+; X64-SSE2-NEXT: punpckhwd {{.*#+}} xmm3 = xmm3[4],xmm2[4],xmm3[5],xmm2[5],xmm3[6],xmm2[6],xmm3[7],xmm2[7]
+; X64-SSE2-NEXT: psrad $24, %xmm3
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm4 = xmm4[0],xmm0[0],xmm4[1],xmm0[1],xmm4[2],xmm0[2],xmm4[3],xmm0[3],xmm4[4],xmm0[4],xmm4[5],xmm0[5],xmm4[6],xmm0[6],xmm4[7],xmm0[7]
+; X64-SSE2-NEXT: punpckhwd {{.*#+}} xmm5 = xmm5[4],xmm4[4],xmm5[5],xmm4[5],xmm5[6],xmm4[6],xmm5[7],xmm4[7]
+; X64-SSE2-NEXT: psrad $24, %xmm5
+; X64-SSE2-NEXT: paddd %xmm3, %xmm5
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE2-NEXT: punpckhwd {{.*#+}} xmm3 = xmm3[4],xmm1[4],xmm3[5],xmm1[5],xmm3[6],xmm1[6],xmm3[7],xmm1[7]
+; X64-SSE2-NEXT: psrad $24, %xmm3
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm0 = xmm0[8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15]
+; X64-SSE2-NEXT: punpckhwd {{.*#+}} xmm6 = xmm6[4],xmm0[4],xmm6[5],xmm0[5],xmm6[6],xmm0[6],xmm6[7],xmm0[7]
+; X64-SSE2-NEXT: psrad $24, %xmm6
+; X64-SSE2-NEXT: paddd %xmm3, %xmm6
+; X64-SSE2-NEXT: paddd %xmm5, %xmm6
+; X64-SSE2-NEXT: punpcklwd {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3]
+; X64-SSE2-NEXT: psrad $24, %xmm2
+; X64-SSE2-NEXT: punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm4[0],xmm3[1],xmm4[1],xmm3[2],xmm4[2],xmm3[3],xmm4[3]
+; X64-SSE2-NEXT: psrad $24, %xmm3
+; X64-SSE2-NEXT: paddd %xmm2, %xmm3
+; X64-SSE2-NEXT: punpcklwd {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3]
+; X64-SSE2-NEXT: psrad $24, %xmm1
+; X64-SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3]
+; X64-SSE2-NEXT: psrad $24, %xmm0
+; X64-SSE2-NEXT: paddd %xmm1, %xmm0
+; X64-SSE2-NEXT: paddd %xmm3, %xmm0
+; X64-SSE2-NEXT: paddd %xmm6, %xmm0
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE2-NEXT: paddd %xmm0, %xmm1
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE2-NEXT: paddd %xmm1, %xmm0
+; X64-SSE2-NEXT: movd %xmm0, %eax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v32i32_v32i8:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm2 = xmm1[1,1,1,1]
+; X86-SSE4-NEXT: pmovsxbd %xmm2, %xmm2
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,1,1]
+; X86-SSE4-NEXT: pmovsxbd %xmm3, %xmm3
+; X86-SSE4-NEXT: paddd %xmm2, %xmm3
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm2 = xmm1[3,3,3,3]
+; X86-SSE4-NEXT: pmovsxbd %xmm2, %xmm4
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm2 = xmm0[3,3,3,3]
+; X86-SSE4-NEXT: pmovsxbd %xmm2, %xmm2
+; X86-SSE4-NEXT: paddd %xmm4, %xmm2
+; X86-SSE4-NEXT: paddd %xmm3, %xmm2
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm3 = xmm1[2,3,2,3]
+; X86-SSE4-NEXT: pmovsxbd %xmm3, %xmm3
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm4 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: pmovsxbd %xmm4, %xmm4
+; X86-SSE4-NEXT: paddd %xmm3, %xmm4
+; X86-SSE4-NEXT: pmovsxbd %xmm1, %xmm1
+; X86-SSE4-NEXT: pmovsxbd %xmm0, %xmm0
+; X86-SSE4-NEXT: paddd %xmm1, %xmm0
+; X86-SSE4-NEXT: paddd %xmm4, %xmm0
+; X86-SSE4-NEXT: paddd %xmm2, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: paddd %xmm0, %xmm1
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE4-NEXT: paddd %xmm1, %xmm0
+; X86-SSE4-NEXT: movd %xmm0, %eax
+; X86-SSE4-NEXT: retl
+;
+; X64-SSE4-LABEL: test_v32i32_v32i8:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm2 = xmm1[1,1,1,1]
+; X64-SSE4-NEXT: pmovsxbd %xmm2, %xmm2
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm3 = xmm0[1,1,1,1]
+; X64-SSE4-NEXT: pmovsxbd %xmm3, %xmm3
+; X64-SSE4-NEXT: paddd %xmm2, %xmm3
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm2 = xmm1[3,3,3,3]
+; X64-SSE4-NEXT: pmovsxbd %xmm2, %xmm2
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm4 = xmm0[3,3,3,3]
+; X64-SSE4-NEXT: pmovsxbd %xmm4, %xmm4
+; X64-SSE4-NEXT: paddd %xmm2, %xmm4
+; X64-SSE4-NEXT: paddd %xmm3, %xmm4
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm2 = xmm1[2,3,2,3]
+; X64-SSE4-NEXT: pmovsxbd %xmm2, %xmm2
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm3 = xmm0[2,3,2,3]
+; X64-SSE4-NEXT: pmovsxbd %xmm3, %xmm3
+; X64-SSE4-NEXT: paddd %xmm2, %xmm3
+; X64-SSE4-NEXT: pmovsxbd %xmm1, %xmm1
+; X64-SSE4-NEXT: pmovsxbd %xmm0, %xmm0
+; X64-SSE4-NEXT: paddd %xmm1, %xmm0
+; X64-SSE4-NEXT: paddd %xmm3, %xmm0
+; X64-SSE4-NEXT: paddd %xmm4, %xmm0
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE4-NEXT: paddd %xmm0, %xmm1
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE4-NEXT: paddd %xmm1, %xmm0
+; X64-SSE4-NEXT: movd %xmm0, %eax
+; X64-SSE4-NEXT: retq
;
; AVX1-SLOW-LABEL: test_v32i32_v32i8:
; AVX1-SLOW: # %bb.0:
@@ -764,7 +1165,7 @@ define i32 @test_v32i32_v32i8(<32 x i8> %a0) {
; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
; AVX1-SLOW-NEXT: vzeroupper
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v32i32_v32i8:
; AVX1-FAST: # %bb.0:
@@ -795,7 +1196,7 @@ define i32 @test_v32i32_v32i8(<32 x i8> %a0) {
; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
; AVX1-FAST-NEXT: vzeroupper
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v32i32_v32i8:
; AVX2: # %bb.0:
@@ -817,7 +1218,7 @@ define i32 @test_v32i32_v32i8(<32 x i8> %a0) {
; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v32i32_v32i8:
; AVX512: # %bb.0:
@@ -845,7 +1246,7 @@ define i32 @test_v32i32_v32i8(<32 x i8> %a0) {
; vXi16
;
-define i16 @test_v2i16_v2i8(<2 x i8> %a0) {
+define i16 @test_v2i16_v2i8(<2 x i8> %a0) nounwind {
; SSE2-LABEL: test_v2i16_v2i8:
; SSE2: # %bb.0:
; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
@@ -855,17 +1256,17 @@ define i16 @test_v2i16_v2i8(<2 x i8> %a0) {
; SSE2-NEXT: paddw %xmm0, %xmm1
; SSE2-NEXT: movd %xmm1, %eax
; SSE2-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v2i16_v2i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovsxbw %xmm0, %xmm0
-; SSE41-NEXT: movdqa %xmm0, %xmm1
-; SSE41-NEXT: psrld $16, %xmm1
-; SSE41-NEXT: paddw %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE41-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
+;
+; SSE4-LABEL: test_v2i16_v2i8:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pmovsxbw %xmm0, %xmm0
+; SSE4-NEXT: movdqa %xmm0, %xmm1
+; SSE4-NEXT: psrld $16, %xmm1
+; SSE4-NEXT: paddw %xmm0, %xmm1
+; SSE4-NEXT: movd %xmm1, %eax
+; SSE4-NEXT: # kill: def $ax killed $ax killed $eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v2i16_v2i8:
; AVX1-SLOW: # %bb.0:
@@ -874,7 +1275,7 @@ define i16 @test_v2i16_v2i8(<2 x i8> %a0) {
; AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
; AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v2i16_v2i8:
; AVX1-FAST: # %bb.0:
@@ -882,7 +1283,7 @@ define i16 @test_v2i16_v2i8(<2 x i8> %a0) {
; AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
; AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v2i16_v2i8:
; AVX2: # %bb.0:
@@ -891,7 +1292,7 @@ define i16 @test_v2i16_v2i8(<2 x i8> %a0) {
; AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v2i16_v2i8:
; AVX512: # %bb.0:
@@ -906,7 +1307,7 @@ define i16 @test_v2i16_v2i8(<2 x i8> %a0) {
ret i16 %2
}
-define i16 @test_v4i16_v4i8(<4 x i8> %a0) {
+define i16 @test_v4i16_v4i8(<4 x i8> %a0) nounwind {
; SSE2-LABEL: test_v4i16_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]
@@ -918,19 +1319,19 @@ define i16 @test_v4i16_v4i8(<4 x i8> %a0) {
; SSE2-NEXT: paddw %xmm1, %xmm0
; SSE2-NEXT: movd %xmm0, %eax
; SSE2-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v4i16_v4i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovsxbw %xmm0, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE41-NEXT: paddw %xmm0, %xmm1
-; SSE41-NEXT: movdqa %xmm1, %xmm0
-; SSE41-NEXT: psrld $16, %xmm0
-; SSE41-NEXT: paddw %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE41-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
+;
+; SSE4-LABEL: test_v4i16_v4i8:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pmovsxbw %xmm0, %xmm0
+; SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; SSE4-NEXT: paddw %xmm0, %xmm1
+; SSE4-NEXT: movdqa %xmm1, %xmm0
+; SSE4-NEXT: psrld $16, %xmm0
+; SSE4-NEXT: paddw %xmm1, %xmm0
+; SSE4-NEXT: movd %xmm0, %eax
+; SSE4-NEXT: # kill: def $ax killed $ax killed $eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v4i16_v4i8:
; AVX1-SLOW: # %bb.0:
@@ -941,7 +1342,7 @@ define i16 @test_v4i16_v4i8(<4 x i8> %a0) {
; AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
; AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v4i16_v4i8:
; AVX1-FAST: # %bb.0:
@@ -951,7 +1352,7 @@ define i16 @test_v4i16_v4i8(<4 x i8> %a0) {
; AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
; AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v4i16_v4i8:
; AVX2: # %bb.0:
@@ -962,7 +1363,7 @@ define i16 @test_v4i16_v4i8(<4 x i8> %a0) {
; AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v4i16_v4i8:
; AVX512: # %bb.0:
@@ -980,7 +1381,7 @@ define i16 @test_v4i16_v4i8(<4 x i8> %a0) {
}
-define i16 @test_v8i16_v8i8(<8 x i8> %a0) {
+define i16 @test_v8i16_v8i8(<8 x i8> %a0) nounwind {
; SSE2-LABEL: test_v8i16_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]
@@ -994,21 +1395,21 @@ define i16 @test_v8i16_v8i8(<8 x i8> %a0) {
; SSE2-NEXT: paddw %xmm0, %xmm1
; SSE2-NEXT: movd %xmm1, %eax
; SSE2-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v8i16_v8i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovsxbw %xmm0, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: paddw %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: paddw %xmm1, %xmm0
-; SSE41-NEXT: movdqa %xmm0, %xmm1
-; SSE41-NEXT: psrld $16, %xmm1
-; SSE41-NEXT: paddw %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE41-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
+;
+; SSE4-LABEL: test_v8i16_v8i8:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pmovsxbw %xmm0, %xmm0
+; SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; SSE4-NEXT: paddw %xmm0, %xmm1
+; SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; SSE4-NEXT: paddw %xmm1, %xmm0
+; SSE4-NEXT: movdqa %xmm0, %xmm1
+; SSE4-NEXT: psrld $16, %xmm1
+; SSE4-NEXT: paddw %xmm0, %xmm1
+; SSE4-NEXT: movd %xmm1, %eax
+; SSE4-NEXT: # kill: def $ax killed $ax killed $eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v8i16_v8i8:
; AVX1-SLOW: # %bb.0:
@@ -1021,7 +1422,7 @@ define i16 @test_v8i16_v8i8(<8 x i8> %a0) {
; AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
; AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v8i16_v8i8:
; AVX1-FAST: # %bb.0:
@@ -1031,7 +1432,7 @@ define i16 @test_v8i16_v8i8(<8 x i8> %a0) {
; AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
; AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v8i16_v8i8:
; AVX2: # %bb.0:
@@ -1044,7 +1445,7 @@ define i16 @test_v8i16_v8i8(<8 x i8> %a0) {
; AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v8i16_v8i8:
; AVX512: # %bb.0:
@@ -1063,7 +1464,7 @@ define i16 @test_v8i16_v8i8(<8 x i8> %a0) {
ret i16 %2
}
-define i16 @test_v16i16_v16i8(<16 x i8> %a0) {
+define i16 @test_v16i16_v16i8(<16 x i8> %a0) nounwind {
; SSE2-LABEL: test_v16i16_v16i8:
; SSE2: # %bb.0:
; SSE2-NEXT: punpckhbw {{.*#+}} xmm1 = xmm1[8],xmm0[8],xmm1[9],xmm0[9],xmm1[10],xmm0[10],xmm1[11],xmm0[11],xmm1[12],xmm0[12],xmm1[13],xmm0[13],xmm1[14],xmm0[14],xmm1[15],xmm0[15]
@@ -1080,24 +1481,24 @@ define i16 @test_v16i16_v16i8(<16 x i8> %a0) {
; SSE2-NEXT: paddw %xmm0, %xmm1
; SSE2-NEXT: movd %xmm1, %eax
; SSE2-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v16i16_v16i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovsxbw %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
-; SSE41-NEXT: pmovsxbw %xmm0, %xmm0
-; SSE41-NEXT: paddw %xmm1, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: paddw %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: paddw %xmm1, %xmm0
-; SSE41-NEXT: movdqa %xmm0, %xmm1
-; SSE41-NEXT: psrld $16, %xmm1
-; SSE41-NEXT: paddw %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE41-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
+;
+; SSE4-LABEL: test_v16i16_v16i8:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pmovsxbw %xmm0, %xmm1
+; SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
+; SSE4-NEXT: pmovsxbw %xmm0, %xmm0
+; SSE4-NEXT: paddw %xmm1, %xmm0
+; SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; SSE4-NEXT: paddw %xmm0, %xmm1
+; SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; SSE4-NEXT: paddw %xmm1, %xmm0
+; SSE4-NEXT: movdqa %xmm0, %xmm1
+; SSE4-NEXT: psrld $16, %xmm1
+; SSE4-NEXT: paddw %xmm0, %xmm1
+; SSE4-NEXT: movd %xmm1, %eax
+; SSE4-NEXT: # kill: def $ax killed $ax killed $eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v16i16_v16i8:
; AVX1-SLOW: # %bb.0:
@@ -1113,7 +1514,7 @@ define i16 @test_v16i16_v16i8(<16 x i8> %a0) {
; AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
; AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v16i16_v16i8:
; AVX1-FAST: # %bb.0:
@@ -1126,7 +1527,7 @@ define i16 @test_v16i16_v16i8(<16 x i8> %a0) {
; AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
; AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v16i16_v16i8:
; AVX2: # %bb.0:
@@ -1142,7 +1543,7 @@ define i16 @test_v16i16_v16i8(<16 x i8> %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_v16i8:
; AVX512: # %bb.0:
@@ -1164,7 +1565,7 @@ define i16 @test_v16i16_v16i8(<16 x i8> %a0) {
ret i16 %2
}
-define i16 @test_v32i16_v32i8(<32 x i8> %a0) {
+define i16 @test_v32i16_v32i8(<32 x i8> %a0) nounwind {
; SSE2-LABEL: test_v32i16_v32i8:
; SSE2: # %bb.0:
; SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3],xmm2[4],xmm1[4],xmm2[5],xmm1[5],xmm2[6],xmm1[6],xmm2[7],xmm1[7]
@@ -1187,29 +1588,29 @@ define i16 @test_v32i16_v32i8(<32 x i8> %a0) {
; SSE2-NEXT: paddw %xmm0, %xmm1
; SSE2-NEXT: movd %xmm1, %eax
; SSE2-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v32i16_v32i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm1[2,3,2,3]
-; SSE41-NEXT: pmovsxbw %xmm2, %xmm2
-; SSE41-NEXT: pshufd {{.*#+}} xmm3 = xmm0[2,3,2,3]
-; SSE41-NEXT: pmovsxbw %xmm3, %xmm3
-; SSE41-NEXT: paddw %xmm2, %xmm3
-; SSE41-NEXT: pmovsxbw %xmm1, %xmm1
-; SSE41-NEXT: pmovsxbw %xmm0, %xmm0
-; SSE41-NEXT: paddw %xmm1, %xmm0
-; SSE41-NEXT: paddw %xmm3, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: paddw %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE41-NEXT: paddw %xmm1, %xmm0
-; SSE41-NEXT: movdqa %xmm0, %xmm1
-; SSE41-NEXT: psrld $16, %xmm1
-; SSE41-NEXT: paddw %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE41-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
+;
+; SSE4-LABEL: test_v32i16_v32i8:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pshufd {{.*#+}} xmm2 = xmm1[2,3,2,3]
+; SSE4-NEXT: pmovsxbw %xmm2, %xmm2
+; SSE4-NEXT: pshufd {{.*#+}} xmm3 = xmm0[2,3,2,3]
+; SSE4-NEXT: pmovsxbw %xmm3, %xmm3
+; SSE4-NEXT: paddw %xmm2, %xmm3
+; SSE4-NEXT: pmovsxbw %xmm1, %xmm1
+; SSE4-NEXT: pmovsxbw %xmm0, %xmm0
+; SSE4-NEXT: paddw %xmm1, %xmm0
+; SSE4-NEXT: paddw %xmm3, %xmm0
+; SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; SSE4-NEXT: paddw %xmm0, %xmm1
+; SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; SSE4-NEXT: paddw %xmm1, %xmm0
+; SSE4-NEXT: movdqa %xmm0, %xmm1
+; SSE4-NEXT: psrld $16, %xmm1
+; SSE4-NEXT: paddw %xmm0, %xmm1
+; SSE4-NEXT: movd %xmm1, %eax
+; SSE4-NEXT: # kill: def $ax killed $ax killed $eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v32i16_v32i8:
; AVX1-SLOW: # %bb.0:
@@ -1232,7 +1633,7 @@ define i16 @test_v32i16_v32i8(<32 x i8> %a0) {
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
; AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
; AVX1-SLOW-NEXT: vzeroupper
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v32i16_v32i8:
; AVX1-FAST: # %bb.0:
@@ -1254,7 +1655,7 @@ define i16 @test_v32i16_v32i8(<32 x i8> %a0) {
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
; AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
; AVX1-FAST-NEXT: vzeroupper
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v32i16_v32i8:
; AVX2: # %bb.0:
@@ -1273,7 +1674,7 @@ define i16 @test_v32i16_v32i8(<32 x i8> %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_v32i16_v32i8:
; AVX512: # %bb.0:
@@ -1297,74 +1698,156 @@ define i16 @test_v32i16_v32i8(<32 x i8> %a0) {
ret i16 %2
}
-define i16 @test_v64i16_v64i8(<64 x i8> %a0) {
-; SSE2-LABEL: test_v64i16_v64i8:
-; SSE2: # %bb.0:
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm2[8],xmm4[9],xmm2[9],xmm4[10],xmm2[10],xmm4[11],xmm2[11],xmm4[12],xmm2[12],xmm4[13],xmm2[13],xmm4[14],xmm2[14],xmm4[15],xmm2[15]
-; SSE2-NEXT: psraw $8, %xmm4
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm5 = xmm5[8],xmm0[8],xmm5[9],xmm0[9],xmm5[10],xmm0[10],xmm5[11],xmm0[11],xmm5[12],xmm0[12],xmm5[13],xmm0[13],xmm5[14],xmm0[14],xmm5[15],xmm0[15]
-; SSE2-NEXT: psraw $8, %xmm5
-; SSE2-NEXT: paddw %xmm4, %xmm5
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm3[8],xmm4[9],xmm3[9],xmm4[10],xmm3[10],xmm4[11],xmm3[11],xmm4[12],xmm3[12],xmm4[13],xmm3[13],xmm4[14],xmm3[14],xmm4[15],xmm3[15]
-; SSE2-NEXT: psraw $8, %xmm4
-; SSE2-NEXT: punpckhbw {{.*#+}} xmm6 = xmm6[8],xmm1[8],xmm6[9],xmm1[9],xmm6[10],xmm1[10],xmm6[11],xmm1[11],xmm6[12],xmm1[12],xmm6[13],xmm1[13],xmm6[14],xmm1[14],xmm6[15],xmm1[15]
-; SSE2-NEXT: psraw $8, %xmm6
-; SSE2-NEXT: paddw %xmm4, %xmm6
-; SSE2-NEXT: paddw %xmm5, %xmm6
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: psraw $8, %xmm2
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: psraw $8, %xmm0
-; SSE2-NEXT: paddw %xmm2, %xmm0
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3],xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
-; SSE2-NEXT: psraw $8, %xmm2
-; SSE2-NEXT: punpcklbw {{.*#+}} xmm1 = xmm1[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
-; SSE2-NEXT: psraw $8, %xmm1
-; SSE2-NEXT: paddw %xmm2, %xmm1
-; SSE2-NEXT: paddw %xmm0, %xmm1
-; SSE2-NEXT: paddw %xmm6, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE2-NEXT: paddw %xmm1, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE2-NEXT: paddw %xmm0, %xmm1
-; SSE2-NEXT: movdqa %xmm1, %xmm0
-; SSE2-NEXT: psrld $16, %xmm0
-; SSE2-NEXT: paddw %xmm1, %xmm0
-; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v64i16_v64i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovsxbw %xmm2, %xmm4
-; SSE41-NEXT: pmovsxbw %xmm0, %xmm5
-; SSE41-NEXT: paddw %xmm4, %xmm5
-; SSE41-NEXT: pmovsxbw %xmm3, %xmm4
-; SSE41-NEXT: pmovsxbw %xmm1, %xmm6
-; SSE41-NEXT: paddw %xmm4, %xmm6
-; SSE41-NEXT: paddw %xmm5, %xmm6
-; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
-; SSE41-NEXT: pmovsxbw %xmm2, %xmm2
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
-; SSE41-NEXT: pmovsxbw %xmm0, %xmm0
-; SSE41-NEXT: paddw %xmm2, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm2 = xmm3[2,3,2,3]
-; SSE41-NEXT: pmovsxbw %xmm2, %xmm2
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
-; SSE41-NEXT: pmovsxbw %xmm1, %xmm1
-; SSE41-NEXT: paddw %xmm2, %xmm1
-; SSE41-NEXT: paddw %xmm0, %xmm1
-; SSE41-NEXT: paddw %xmm6, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE41-NEXT: paddw %xmm1, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE41-NEXT: paddw %xmm0, %xmm1
-; SSE41-NEXT: movdqa %xmm1, %xmm0
-; SSE41-NEXT: psrld $16, %xmm0
-; SSE41-NEXT: paddw %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE41-NEXT: retq
+define i16 @test_v64i16_v64i8(<64 x i8> %a0) nounwind {
+; X86-SSE2-LABEL: test_v64i16_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: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm2[8],xmm4[9],xmm2[9],xmm4[10],xmm2[10],xmm4[11],xmm2[11],xmm4[12],xmm2[12],xmm4[13],xmm2[13],xmm4[14],xmm2[14],xmm4[15],xmm2[15]
+; X86-SSE2-NEXT: psraw $8, %xmm4
+; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm5 = xmm5[8],xmm0[8],xmm5[9],xmm0[9],xmm5[10],xmm0[10],xmm5[11],xmm0[11],xmm5[12],xmm0[12],xmm5[13],xmm0[13],xmm5[14],xmm0[14],xmm5[15],xmm0[15]
+; X86-SSE2-NEXT: psraw $8, %xmm5
+; X86-SSE2-NEXT: paddw %xmm4, %xmm5
+; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm6 = xmm6[8],xmm3[8],xmm6[9],xmm3[9],xmm6[10],xmm3[10],xmm6[11],xmm3[11],xmm6[12],xmm3[12],xmm6[13],xmm3[13],xmm6[14],xmm3[14],xmm6[15],xmm3[15]
+; X86-SSE2-NEXT: psraw $8, %xmm6
+; X86-SSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm1[8],xmm4[9],xmm1[9],xmm4[10],xmm1[10],xmm4[11],xmm1[11],xmm4[12],xmm1[12],xmm4[13],xmm1[13],xmm4[14],xmm1[14],xmm4[15],xmm1[15]
+; X86-SSE2-NEXT: psraw $8, %xmm4
+; X86-SSE2-NEXT: paddw %xmm6, %xmm4
+; X86-SSE2-NEXT: paddw %xmm5, %xmm4
+; 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: psraw $8, %xmm2
+; 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: psraw $8, %xmm0
+; X86-SSE2-NEXT: paddw %xmm2, %xmm0
+; X86-SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3],xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
+; X86-SSE2-NEXT: psraw $8, %xmm2
+; 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: psraw $8, %xmm1
+; X86-SSE2-NEXT: paddw %xmm2, %xmm1
+; X86-SSE2-NEXT: paddw %xmm0, %xmm1
+; X86-SSE2-NEXT: paddw %xmm4, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE2-NEXT: paddw %xmm1, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE2-NEXT: paddw %xmm0, %xmm1
+; X86-SSE2-NEXT: movdqa %xmm1, %xmm0
+; X86-SSE2-NEXT: psrld $16, %xmm0
+; X86-SSE2-NEXT: paddw %xmm1, %xmm0
+; X86-SSE2-NEXT: movd %xmm0, %eax
+; X86-SSE2-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-SSE2-NEXT: movl %ebp, %esp
+; X86-SSE2-NEXT: popl %ebp
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE2-LABEL: test_v64i16_v64i8:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm2[8],xmm4[9],xmm2[9],xmm4[10],xmm2[10],xmm4[11],xmm2[11],xmm4[12],xmm2[12],xmm4[13],xmm2[13],xmm4[14],xmm2[14],xmm4[15],xmm2[15]
+; X64-SSE2-NEXT: psraw $8, %xmm4
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm5 = xmm5[8],xmm0[8],xmm5[9],xmm0[9],xmm5[10],xmm0[10],xmm5[11],xmm0[11],xmm5[12],xmm0[12],xmm5[13],xmm0[13],xmm5[14],xmm0[14],xmm5[15],xmm0[15]
+; X64-SSE2-NEXT: psraw $8, %xmm5
+; X64-SSE2-NEXT: paddw %xmm4, %xmm5
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm3[8],xmm4[9],xmm3[9],xmm4[10],xmm3[10],xmm4[11],xmm3[11],xmm4[12],xmm3[12],xmm4[13],xmm3[13],xmm4[14],xmm3[14],xmm4[15],xmm3[15]
+; X64-SSE2-NEXT: psraw $8, %xmm4
+; X64-SSE2-NEXT: punpckhbw {{.*#+}} xmm6 = xmm6[8],xmm1[8],xmm6[9],xmm1[9],xmm6[10],xmm1[10],xmm6[11],xmm1[11],xmm6[12],xmm1[12],xmm6[13],xmm1[13],xmm6[14],xmm1[14],xmm6[15],xmm1[15]
+; X64-SSE2-NEXT: psraw $8, %xmm6
+; X64-SSE2-NEXT: paddw %xmm4, %xmm6
+; X64-SSE2-NEXT: paddw %xmm5, %xmm6
+; 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: psraw $8, %xmm2
+; 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: psraw $8, %xmm0
+; X64-SSE2-NEXT: paddw %xmm2, %xmm0
+; X64-SSE2-NEXT: punpcklbw {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3],xmm2[4],xmm3[4],xmm2[5],xmm3[5],xmm2[6],xmm3[6],xmm2[7],xmm3[7]
+; X64-SSE2-NEXT: psraw $8, %xmm2
+; 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: psraw $8, %xmm1
+; X64-SSE2-NEXT: paddw %xmm2, %xmm1
+; X64-SSE2-NEXT: paddw %xmm0, %xmm1
+; X64-SSE2-NEXT: paddw %xmm6, %xmm1
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X64-SSE2-NEXT: paddw %xmm1, %xmm0
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-SSE2-NEXT: paddw %xmm0, %xmm1
+; X64-SSE2-NEXT: movdqa %xmm1, %xmm0
+; X64-SSE2-NEXT: psrld $16, %xmm0
+; X64-SSE2-NEXT: paddw %xmm1, %xmm0
+; X64-SSE2-NEXT: movd %xmm0, %eax
+; X64-SSE2-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v64i16_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), %xmm4
+; X86-SSE4-NEXT: pmovsxbw %xmm2, %xmm3
+; X86-SSE4-NEXT: pmovsxbw %xmm0, %xmm5
+; X86-SSE4-NEXT: paddw %xmm3, %xmm5
+; X86-SSE4-NEXT: pmovsxbw %xmm4, %xmm6
+; X86-SSE4-NEXT: pmovsxbw %xmm1, %xmm3
+; X86-SSE4-NEXT: paddw %xmm6, %xmm3
+; X86-SSE4-NEXT: paddw %xmm5, %xmm3
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
+; X86-SSE4-NEXT: pmovsxbw %xmm2, %xmm2
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: pmovsxbw %xmm0, %xmm0
+; X86-SSE4-NEXT: paddw %xmm2, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm2 = xmm4[2,3,2,3]
+; X86-SSE4-NEXT: pmovsxbw %xmm2, %xmm2
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
+; X86-SSE4-NEXT: pmovsxbw %xmm1, %xmm1
+; X86-SSE4-NEXT: paddw %xmm2, %xmm1
+; X86-SSE4-NEXT: paddw %xmm0, %xmm1
+; X86-SSE4-NEXT: paddw %xmm3, %xmm1
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE4-NEXT: paddw %xmm1, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE4-NEXT: paddw %xmm0, %xmm1
+; X86-SSE4-NEXT: movdqa %xmm1, %xmm0
+; X86-SSE4-NEXT: psrld $16, %xmm0
+; X86-SSE4-NEXT: paddw %xmm1, %xmm0
+; X86-SSE4-NEXT: movd %xmm0, %eax
+; X86-SSE4-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-SSE4-NEXT: movl %ebp, %esp
+; X86-SSE4-NEXT: popl %ebp
+; X86-SSE4-NEXT: retl
+;
+; X64-SSE4-LABEL: test_v64i16_v64i8:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: pmovsxbw %xmm2, %xmm4
+; X64-SSE4-NEXT: pmovsxbw %xmm0, %xmm5
+; X64-SSE4-NEXT: paddw %xmm4, %xmm5
+; X64-SSE4-NEXT: pmovsxbw %xmm3, %xmm4
+; X64-SSE4-NEXT: pmovsxbw %xmm1, %xmm6
+; X64-SSE4-NEXT: paddw %xmm4, %xmm6
+; X64-SSE4-NEXT: paddw %xmm5, %xmm6
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
+; X64-SSE4-NEXT: pmovsxbw %xmm2, %xmm2
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
+; X64-SSE4-NEXT: pmovsxbw %xmm0, %xmm0
+; X64-SSE4-NEXT: paddw %xmm2, %xmm0
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm2 = xmm3[2,3,2,3]
+; X64-SSE4-NEXT: pmovsxbw %xmm2, %xmm2
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
+; X64-SSE4-NEXT: pmovsxbw %xmm1, %xmm1
+; X64-SSE4-NEXT: paddw %xmm2, %xmm1
+; X64-SSE4-NEXT: paddw %xmm0, %xmm1
+; X64-SSE4-NEXT: paddw %xmm6, %xmm1
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X64-SSE4-NEXT: paddw %xmm1, %xmm0
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-SSE4-NEXT: paddw %xmm0, %xmm1
+; X64-SSE4-NEXT: movdqa %xmm1, %xmm0
+; X64-SSE4-NEXT: psrld $16, %xmm0
+; X64-SSE4-NEXT: paddw %xmm1, %xmm0
+; X64-SSE4-NEXT: movd %xmm0, %eax
+; X64-SSE4-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-SSE4-NEXT: retq
;
; AVX1-SLOW-LABEL: test_v64i16_v64i8:
; AVX1-SLOW: # %bb.0:
@@ -1398,7 +1881,7 @@ define i16 @test_v64i16_v64i8(<64 x i8> %a0) {
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
; AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
; AVX1-SLOW-NEXT: vzeroupper
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v64i16_v64i8:
; AVX1-FAST: # %bb.0:
@@ -1431,7 +1914,7 @@ define i16 @test_v64i16_v64i8(<64 x i8> %a0) {
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
; AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
; AVX1-FAST-NEXT: vzeroupper
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v64i16_v64i8:
; AVX2: # %bb.0:
@@ -1455,7 +1938,7 @@ define i16 @test_v64i16_v64i8(<64 x i8> %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_v64i16_v64i8:
; AVX512: # %bb.0:
@@ -1487,63 +1970,132 @@ define i16 @test_v64i16_v64i8(<64 x i8> %a0) {
; vXi1 - sum of extended bool vectors
;
-define i64 @test_v2i64_v2i1(<2 x i64> %a0) {
-; SSE2-LABEL: test_v2i64_v2i1:
-; SSE2: # %bb.0:
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE2-NEXT: pxor %xmm1, %xmm1
-; SSE2-NEXT: pcmpgtd %xmm0, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE2-NEXT: paddq %xmm1, %xmm0
-; SSE2-NEXT: movq %xmm0, %rax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v2i64_v2i1:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
-; SSE41-NEXT: pxor %xmm1, %xmm1
-; SSE41-NEXT: pcmpgtd %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE41-NEXT: paddq %xmm1, %xmm0
-; SSE41-NEXT: movq %xmm0, %rax
-; SSE41-NEXT: retq
-;
-; AVX-LABEL: test_v2i64_v2i1:
-; AVX: # %bb.0:
-; AVX-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0
-; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vmovq %xmm0, %rax
-; AVX-NEXT: retq
+define i64 @test_v2i64_v2i1(<2 x i64> %a0) nounwind {
+; X86-SSE2-LABEL: test_v2i64_v2i1:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: pxor %xmm1, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; X86-SSE2-NEXT: pcmpgtd %xmm0, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE2-NEXT: paddq %xmm1, %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-SSE2-LABEL: test_v2i64_v2i1:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; X64-SSE2-NEXT: pxor %xmm1, %xmm1
+; X64-SSE2-NEXT: pcmpgtd %xmm0, %xmm1
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X64-SSE2-NEXT: paddq %xmm1, %xmm0
+; X64-SSE2-NEXT: movq %xmm0, %rax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE41-LABEL: test_v2i64_v2i1:
+; X86-SSE41: # %bb.0:
+; X86-SSE41-NEXT: pxor %xmm1, %xmm1
+; X86-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; X86-SSE41-NEXT: pcmpgtd %xmm0, %xmm1
+; X86-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE41-NEXT: paddq %xmm1, %xmm0
+; X86-SSE41-NEXT: movd %xmm0, %eax
+; X86-SSE41-NEXT: pextrd $1, %xmm0, %edx
+; X86-SSE41-NEXT: retl
+;
+; X64-SSE41-LABEL: test_v2i64_v2i1:
+; X64-SSE41: # %bb.0:
+; X64-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]
+; X64-SSE41-NEXT: pxor %xmm1, %xmm1
+; X64-SSE41-NEXT: pcmpgtd %xmm0, %xmm1
+; X64-SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X64-SSE41-NEXT: paddq %xmm1, %xmm0
+; X64-SSE41-NEXT: movq %xmm0, %rax
+; X64-SSE41-NEXT: retq
+;
+; X86-SSE42-LABEL: test_v2i64_v2i1:
+; X86-SSE42: # %bb.0:
+; X86-SSE42-NEXT: pxor %xmm1, %xmm1
+; X86-SSE42-NEXT: pcmpgtq %xmm0, %xmm1
+; X86-SSE42-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE42-NEXT: paddq %xmm1, %xmm0
+; X86-SSE42-NEXT: movd %xmm0, %eax
+; X86-SSE42-NEXT: pextrd $1, %xmm0, %edx
+; X86-SSE42-NEXT: retl
+;
+; X64-SSE42-LABEL: test_v2i64_v2i1:
+; X64-SSE42: # %bb.0:
+; X64-SSE42-NEXT: pxor %xmm1, %xmm1
+; X64-SSE42-NEXT: pcmpgtq %xmm0, %xmm1
+; X64-SSE42-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X64-SSE42-NEXT: paddq %xmm1, %xmm0
+; X64-SSE42-NEXT: movq %xmm0, %rax
+; X64-SSE42-NEXT: retq
+;
+; X86-AVX1-LABEL: test_v2i64_v2i1:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v2i64_v2i1:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovq %xmm0, %rax
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v2i64_v2i1:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X86-AVX2-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v2i64_v2i1:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X64-AVX2-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: retq
+;
+; AVX512-LABEL: test_v2i64_v2i1:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; AVX512-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0
+; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: vmovq %xmm0, %rax
+; AVX512-NEXT: retq
%1 = icmp slt <2 x i64> %a0, zeroinitializer
%2 = sext <2 x i1> %1 to <2 x i64>
%3 = call i64 @llvm.vector.reduce.add.v2i64(<2 x i64> %2)
ret i64 %3
}
-define i32 @test_v4i32_v4i1(<4 x i32> %a0) {
-; SSE2-LABEL: test_v4i32_v4i1:
-; SSE2: # %bb.0:
-; SSE2-NEXT: pxor %xmm1, %xmm1
-; SSE2-NEXT: pcmpgtd %xmm0, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE2-NEXT: paddd %xmm1, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE2-NEXT: paddd %xmm0, %xmm1
-; SSE2-NEXT: movd %xmm1, %eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v4i32_v4i1:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pxor %xmm1, %xmm1
-; SSE41-NEXT: pcmpgtd %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE41-NEXT: paddd %xmm1, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE41-NEXT: paddd %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: retq
+define i32 @test_v4i32_v4i1(<4 x i32> %a0) nounwind {
+; SSE-LABEL: test_v4i32_v4i1:
+; SSE: # %bb.0:
+; SSE-NEXT: pxor %xmm1, %xmm1
+; SSE-NEXT: pcmpgtd %xmm0, %xmm1
+; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; SSE-NEXT: paddd %xmm1, %xmm0
+; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; SSE-NEXT: paddd %xmm0, %xmm1
+; SSE-NEXT: movd %xmm1, %eax
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v4i32_v4i1:
; AVX1-SLOW: # %bb.0:
@@ -1554,7 +2106,7 @@ define i32 @test_v4i32_v4i1(<4 x i32> %a0) {
; AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v4i32_v4i1:
; AVX1-FAST: # %bb.0:
@@ -1563,7 +2115,7 @@ define i32 @test_v4i32_v4i1(<4 x i32> %a0) {
; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v4i32_v4i1:
; AVX2: # %bb.0:
@@ -1574,7 +2126,7 @@ define i32 @test_v4i32_v4i1(<4 x i32> %a0) {
; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v4i32_v4i1:
; AVX512: # %bb.0:
@@ -1592,36 +2144,21 @@ define i32 @test_v4i32_v4i1(<4 x i32> %a0) {
ret i32 %3
}
-define i16 @test_v8i16_v8i1(<8 x i16> %a0) {
-; SSE2-LABEL: test_v8i16_v8i1:
-; SSE2: # %bb.0:
-; SSE2-NEXT: pxor %xmm1, %xmm1
-; SSE2-NEXT: pcmpgtw %xmm0, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE2-NEXT: paddw %xmm1, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE2-NEXT: paddw %xmm0, %xmm1
-; SSE2-NEXT: movdqa %xmm1, %xmm0
-; SSE2-NEXT: psrld $16, %xmm0
-; SSE2-NEXT: paddw %xmm1, %xmm0
-; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v8i16_v8i1:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pxor %xmm1, %xmm1
-; SSE41-NEXT: pcmpgtw %xmm0, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE41-NEXT: paddw %xmm1, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE41-NEXT: paddw %xmm0, %xmm1
-; SSE41-NEXT: movdqa %xmm1, %xmm0
-; SSE41-NEXT: psrld $16, %xmm0
-; SSE41-NEXT: paddw %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE41-NEXT: retq
+define i16 @test_v8i16_v8i1(<8 x i16> %a0) nounwind {
+; SSE-LABEL: test_v8i16_v8i1:
+; SSE: # %bb.0:
+; SSE-NEXT: pxor %xmm1, %xmm1
+; SSE-NEXT: pcmpgtw %xmm0, %xmm1
+; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; SSE-NEXT: paddw %xmm1, %xmm0
+; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; SSE-NEXT: paddw %xmm0, %xmm1
+; SSE-NEXT: movdqa %xmm1, %xmm0
+; SSE-NEXT: psrld $16, %xmm0
+; SSE-NEXT: paddw %xmm1, %xmm0
+; SSE-NEXT: movd %xmm0, %eax
+; SSE-NEXT: # kill: def $ax killed $ax killed $eax
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v8i16_v8i1:
; AVX1-SLOW: # %bb.0:
@@ -1635,7 +2172,7 @@ define i16 @test_v8i16_v8i1(<8 x i16> %a0) {
; AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
; AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v8i16_v8i1:
; AVX1-FAST: # %bb.0:
@@ -1646,7 +2183,7 @@ define i16 @test_v8i16_v8i1(<8 x i16> %a0) {
; AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
; AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v8i16_v8i1:
; AVX2: # %bb.0:
@@ -1660,7 +2197,7 @@ define i16 @test_v8i16_v8i1(<8 x i16> %a0) {
; AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v8i16_v8i1:
; AVX512: # %bb.0:
@@ -1681,30 +2218,18 @@ define i16 @test_v8i16_v8i1(<8 x i16> %a0) {
ret i16 %3
}
-define i8 @test_v16i8_v16i1(<16 x i8> %a0) {
-; SSE2-LABEL: test_v16i8_v16i1:
-; SSE2: # %bb.0:
-; SSE2-NEXT: pxor %xmm1, %xmm1
-; SSE2-NEXT: pxor %xmm2, %xmm2
-; SSE2-NEXT: pcmpgtb %xmm0, %xmm2
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
-; SSE2-NEXT: paddb %xmm2, %xmm0
-; SSE2-NEXT: psadbw %xmm1, %xmm0
-; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: # kill: def $al killed $al killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v16i8_v16i1:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pxor %xmm1, %xmm1
-; SSE41-NEXT: pxor %xmm2, %xmm2
-; SSE41-NEXT: pcmpgtb %xmm0, %xmm2
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
-; SSE41-NEXT: paddb %xmm2, %xmm0
-; SSE41-NEXT: psadbw %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: # kill: def $al killed $al killed $eax
-; SSE41-NEXT: retq
+define i8 @test_v16i8_v16i1(<16 x i8> %a0) nounwind {
+; SSE-LABEL: test_v16i8_v16i1:
+; SSE: # %bb.0:
+; SSE-NEXT: pxor %xmm1, %xmm1
+; SSE-NEXT: pxor %xmm2, %xmm2
+; SSE-NEXT: pcmpgtb %xmm0, %xmm2
+; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,2,3]
+; SSE-NEXT: paddb %xmm2, %xmm0
+; SSE-NEXT: psadbw %xmm1, %xmm0
+; SSE-NEXT: movd %xmm0, %eax
+; SSE-NEXT: # kill: def $al killed $al killed $eax
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX-LABEL: test_v16i8_v16i1:
; AVX: # %bb.0:
@@ -1715,43 +2240,28 @@ define i8 @test_v16i8_v16i1(<16 x i8> %a0) {
; AVX-NEXT: vpsadbw %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 = icmp slt <16 x i8> %a0, zeroinitializer
%2 = sext <16 x i1> %1 to <16 x i8>
%3 = call i8 @llvm.vector.reduce.add.v16i8(<16 x i8> %2)
ret i8 %3
}
-define i8 @test_v32i8_v32i1(<32 x i8> %a0) {
-; SSE2-LABEL: test_v32i8_v32i1:
-; SSE2: # %bb.0:
-; SSE2-NEXT: pxor %xmm2, %xmm2
-; SSE2-NEXT: pxor %xmm3, %xmm3
-; SSE2-NEXT: pcmpgtb %xmm1, %xmm3
-; SSE2-NEXT: pxor %xmm1, %xmm1
-; SSE2-NEXT: pcmpgtb %xmm0, %xmm1
-; SSE2-NEXT: paddb %xmm3, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE2-NEXT: paddb %xmm1, %xmm0
-; SSE2-NEXT: psadbw %xmm2, %xmm0
-; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: # kill: def $al killed $al killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v32i8_v32i1:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pxor %xmm2, %xmm2
-; SSE41-NEXT: pxor %xmm3, %xmm3
-; SSE41-NEXT: pcmpgtb %xmm1, %xmm3
-; SSE41-NEXT: pxor %xmm1, %xmm1
-; SSE41-NEXT: pcmpgtb %xmm0, %xmm1
-; SSE41-NEXT: paddb %xmm3, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE41-NEXT: paddb %xmm1, %xmm0
-; SSE41-NEXT: psadbw %xmm2, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: # kill: def $al killed $al killed $eax
-; SSE41-NEXT: retq
+define i8 @test_v32i8_v32i1(<32 x i8> %a0) nounwind {
+; SSE-LABEL: test_v32i8_v32i1:
+; SSE: # %bb.0:
+; SSE-NEXT: pxor %xmm2, %xmm2
+; SSE-NEXT: pxor %xmm3, %xmm3
+; SSE-NEXT: pcmpgtb %xmm1, %xmm3
+; SSE-NEXT: pxor %xmm1, %xmm1
+; SSE-NEXT: pcmpgtb %xmm0, %xmm1
+; SSE-NEXT: paddb %xmm3, %xmm1
+; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; SSE-NEXT: paddb %xmm1, %xmm0
+; SSE-NEXT: psadbw %xmm2, %xmm0
+; SSE-NEXT: movd %xmm0, %eax
+; SSE-NEXT: # kill: def $al killed $al killed $eax
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX1-LABEL: test_v32i8_v32i1:
; AVX1: # %bb.0:
@@ -1766,7 +2276,7 @@ define i8 @test_v32i8_v32i1(<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_v32i1:
; AVX2: # %bb.0:
@@ -1781,7 +2291,7 @@ define i8 @test_v32i8_v32i1(<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_v32i1:
; AVX512: # %bb.0:
@@ -1803,48 +2313,54 @@ define i8 @test_v32i8_v32i1(<32 x i8> %a0) {
ret i8 %3
}
-define i8 @test_v64i8_v64i1(<64 x i8> %a0) {
-; SSE2-LABEL: test_v64i8_v64i1:
-; SSE2: # %bb.0:
-; SSE2-NEXT: pxor %xmm4, %xmm4
-; SSE2-NEXT: pxor %xmm5, %xmm5
-; SSE2-NEXT: pcmpgtb %xmm2, %xmm5
-; SSE2-NEXT: pxor %xmm2, %xmm2
-; SSE2-NEXT: pcmpgtb %xmm0, %xmm2
-; SSE2-NEXT: paddb %xmm5, %xmm2
-; SSE2-NEXT: pxor %xmm0, %xmm0
-; SSE2-NEXT: pcmpgtb %xmm3, %xmm0
-; SSE2-NEXT: pxor %xmm3, %xmm3
-; SSE2-NEXT: pcmpgtb %xmm1, %xmm3
-; SSE2-NEXT: paddb %xmm0, %xmm3
-; SSE2-NEXT: paddb %xmm2, %xmm3
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
-; SSE2-NEXT: paddb %xmm3, %xmm0
-; SSE2-NEXT: psadbw %xmm4, %xmm0
-; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: # kill: def $al killed $al killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v64i8_v64i1:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pxor %xmm4, %xmm4
-; SSE41-NEXT: pxor %xmm5, %xmm5
-; SSE41-NEXT: pcmpgtb %xmm2, %xmm5
-; SSE41-NEXT: pxor %xmm2, %xmm2
-; SSE41-NEXT: pcmpgtb %xmm0, %xmm2
-; SSE41-NEXT: paddb %xmm5, %xmm2
-; SSE41-NEXT: pxor %xmm0, %xmm0
-; SSE41-NEXT: pcmpgtb %xmm3, %xmm0
-; SSE41-NEXT: pxor %xmm3, %xmm3
-; SSE41-NEXT: pcmpgtb %xmm1, %xmm3
-; SSE41-NEXT: paddb %xmm0, %xmm3
-; SSE41-NEXT: paddb %xmm2, %xmm3
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
-; SSE41-NEXT: paddb %xmm3, %xmm0
-; SSE41-NEXT: psadbw %xmm4, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: # kill: def $al killed $al killed $eax
-; SSE41-NEXT: retq
+define i8 @test_v64i8_v64i1(<64 x i8> %a0) nounwind {
+; X86-SSE-LABEL: test_v64i8_v64i1:
+; 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: pxor %xmm3, %xmm3
+; X86-SSE-NEXT: pxor %xmm4, %xmm4
+; X86-SSE-NEXT: pcmpgtb %xmm1, %xmm4
+; X86-SSE-NEXT: pxor %xmm1, %xmm1
+; X86-SSE-NEXT: pcmpgtb %xmm2, %xmm1
+; X86-SSE-NEXT: pxor %xmm2, %xmm2
+; X86-SSE-NEXT: pcmpgtb %xmm0, %xmm2
+; X86-SSE-NEXT: paddb %xmm1, %xmm2
+; X86-SSE-NEXT: pxor %xmm0, %xmm0
+; X86-SSE-NEXT: pcmpgtb 8(%ebp), %xmm0
+; X86-SSE-NEXT: paddb %xmm4, %xmm0
+; X86-SSE-NEXT: paddb %xmm2, %xmm0
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE-NEXT: paddb %xmm0, %xmm1
+; X86-SSE-NEXT: psadbw %xmm3, %xmm1
+; X86-SSE-NEXT: movd %xmm1, %eax
+; X86-SSE-NEXT: # kill: def $al killed $al killed $eax
+; X86-SSE-NEXT: movl %ebp, %esp
+; X86-SSE-NEXT: popl %ebp
+; X86-SSE-NEXT: retl
+;
+; X64-SSE-LABEL: test_v64i8_v64i1:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: pxor %xmm4, %xmm4
+; X64-SSE-NEXT: pxor %xmm5, %xmm5
+; X64-SSE-NEXT: pcmpgtb %xmm2, %xmm5
+; X64-SSE-NEXT: pxor %xmm2, %xmm2
+; X64-SSE-NEXT: pcmpgtb %xmm0, %xmm2
+; X64-SSE-NEXT: paddb %xmm5, %xmm2
+; X64-SSE-NEXT: pxor %xmm0, %xmm0
+; X64-SSE-NEXT: pcmpgtb %xmm3, %xmm0
+; X64-SSE-NEXT: pxor %xmm3, %xmm3
+; X64-SSE-NEXT: pcmpgtb %xmm1, %xmm3
+; X64-SSE-NEXT: paddb %xmm0, %xmm3
+; X64-SSE-NEXT: paddb %xmm2, %xmm3
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
+; X64-SSE-NEXT: paddb %xmm3, %xmm0
+; X64-SSE-NEXT: psadbw %xmm4, %xmm0
+; X64-SSE-NEXT: movd %xmm0, %eax
+; X64-SSE-NEXT: # kill: def $al killed $al killed $eax
+; X64-SSE-NEXT: retq
;
; AVX1-LABEL: test_v64i8_v64i1:
; AVX1: # %bb.0:
@@ -1864,7 +2380,7 @@ define i8 @test_v64i8_v64i1(<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_v64i1:
; AVX2: # %bb.0:
@@ -1881,7 +2397,7 @@ define i8 @test_v64i8_v64i1(<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]}}
;
; AVX512-LABEL: test_v64i8_v64i1:
; AVX512: # %bb.0:
@@ -1905,123 +2421,191 @@ define i8 @test_v64i8_v64i1(<64 x i8> %a0) {
ret i8 %3
}
-define i8 @test_v128i8_v128i1(<128 x i8> %a0) {
-; SSE2-LABEL: test_v128i8_v128i1:
-; SSE2: # %bb.0:
-; SSE2-NEXT: pxor %xmm8, %xmm8
-; SSE2-NEXT: pxor %xmm9, %xmm9
-; SSE2-NEXT: pcmpgtb %xmm4, %xmm9
-; SSE2-NEXT: pxor %xmm4, %xmm4
-; SSE2-NEXT: pcmpgtb %xmm0, %xmm4
-; SSE2-NEXT: paddb %xmm9, %xmm4
-; SSE2-NEXT: pxor %xmm0, %xmm0
-; SSE2-NEXT: pcmpgtb %xmm6, %xmm0
-; SSE2-NEXT: pxor %xmm6, %xmm6
-; SSE2-NEXT: pcmpgtb %xmm2, %xmm6
-; SSE2-NEXT: paddb %xmm0, %xmm6
-; SSE2-NEXT: paddb %xmm4, %xmm6
-; SSE2-NEXT: pxor %xmm0, %xmm0
-; SSE2-NEXT: pcmpgtb %xmm5, %xmm0
-; SSE2-NEXT: pxor %xmm2, %xmm2
-; SSE2-NEXT: pcmpgtb %xmm1, %xmm2
-; SSE2-NEXT: paddb %xmm0, %xmm2
-; SSE2-NEXT: pxor %xmm0, %xmm0
-; SSE2-NEXT: pcmpgtb %xmm7, %xmm0
-; SSE2-NEXT: pxor %xmm1, %xmm1
-; SSE2-NEXT: pcmpgtb %xmm3, %xmm1
-; SSE2-NEXT: paddb %xmm0, %xmm1
-; SSE2-NEXT: paddb %xmm2, %xmm1
-; SSE2-NEXT: paddb %xmm6, %xmm1
-; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE2-NEXT: paddb %xmm1, %xmm0
-; SSE2-NEXT: psadbw %xmm8, %xmm0
-; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: # kill: def $al killed $al killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v128i8_v128i1:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pxor %xmm8, %xmm8
-; SSE41-NEXT: pxor %xmm9, %xmm9
-; SSE41-NEXT: pcmpgtb %xmm4, %xmm9
-; SSE41-NEXT: pxor %xmm4, %xmm4
-; SSE41-NEXT: pcmpgtb %xmm0, %xmm4
-; SSE41-NEXT: paddb %xmm9, %xmm4
-; SSE41-NEXT: pxor %xmm0, %xmm0
-; SSE41-NEXT: pcmpgtb %xmm6, %xmm0
-; SSE41-NEXT: pxor %xmm6, %xmm6
-; SSE41-NEXT: pcmpgtb %xmm2, %xmm6
-; SSE41-NEXT: paddb %xmm0, %xmm6
-; SSE41-NEXT: paddb %xmm4, %xmm6
-; SSE41-NEXT: pxor %xmm0, %xmm0
-; SSE41-NEXT: pcmpgtb %xmm5, %xmm0
-; SSE41-NEXT: pxor %xmm2, %xmm2
-; SSE41-NEXT: pcmpgtb %xmm1, %xmm2
-; SSE41-NEXT: paddb %xmm0, %xmm2
-; SSE41-NEXT: pxor %xmm0, %xmm0
-; SSE41-NEXT: pcmpgtb %xmm7, %xmm0
-; SSE41-NEXT: pxor %xmm1, %xmm1
-; SSE41-NEXT: pcmpgtb %xmm3, %xmm1
-; SSE41-NEXT: paddb %xmm0, %xmm1
-; SSE41-NEXT: paddb %xmm2, %xmm1
-; SSE41-NEXT: paddb %xmm6, %xmm1
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE41-NEXT: paddb %xmm1, %xmm0
-; SSE41-NEXT: psadbw %xmm8, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: # kill: def $al killed $al killed $eax
-; SSE41-NEXT: retq
-;
-; AVX1-LABEL: test_v128i8_v128i1:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vpxor %xmm4, %xmm4, %xmm4
-; AVX1-NEXT: vpcmpgtb %xmm2, %xmm4, %xmm5
-; AVX1-NEXT: vpcmpgtb %xmm0, %xmm4, %xmm6
-; AVX1-NEXT: vpaddb %xmm5, %xmm6, %xmm5
-; AVX1-NEXT: vpcmpgtb %xmm3, %xmm4, %xmm6
-; AVX1-NEXT: vpcmpgtb %xmm1, %xmm4, %xmm7
-; AVX1-NEXT: vpaddb %xmm6, %xmm7, %xmm6
-; AVX1-NEXT: vpaddb %xmm6, %xmm5, %xmm5
-; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2
-; AVX1-NEXT: vpcmpgtb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT: vpcmpgtb %xmm0, %xmm4, %xmm0
-; AVX1-NEXT: vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm2
-; AVX1-NEXT: vpcmpgtb %xmm2, %xmm4, %xmm2
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT: vpcmpgtb %xmm1, %xmm4, %xmm1
-; AVX1-NEXT: vpaddb %xmm2, %xmm1, %xmm1
-; AVX1-NEXT: vpaddb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpaddb %xmm0, %xmm5, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpaddb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpsadbw %xmm4, %xmm0, %xmm0
-; AVX1-NEXT: vmovd %xmm0, %eax
-; AVX1-NEXT: # kill: def $al killed $al killed $eax
-; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: test_v128i8_v128i1:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpxor %xmm4, %xmm4, %xmm4
-; AVX2-NEXT: vpcmpgtb %ymm2, %ymm4, %ymm2
-; AVX2-NEXT: vpcmpgtb %ymm0, %ymm4, %ymm0
-; AVX2-NEXT: vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT: vpcmpgtb %ymm3, %ymm4, %ymm2
-; AVX2-NEXT: vpcmpgtb %ymm1, %ymm4, %ymm1
-; AVX2-NEXT: vpaddb %ymm2, %ymm1, %ymm1
-; AVX2-NEXT: vpaddb %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpaddb %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddb %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: # kill: def $al killed $al killed $eax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+define i8 @test_v128i8_v128i1(<128 x i8> %a0) nounwind {
+; X86-SSE-LABEL: test_v128i8_v128i1:
+; 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: pxor %xmm3, %xmm3
+; X86-SSE-NEXT: pxor %xmm4, %xmm4
+; X86-SSE-NEXT: pcmpgtb %xmm0, %xmm4
+; X86-SSE-NEXT: pxor %xmm5, %xmm5
+; X86-SSE-NEXT: pcmpgtb %xmm2, %xmm5
+; X86-SSE-NEXT: pxor %xmm2, %xmm2
+; X86-SSE-NEXT: pcmpgtb %xmm1, %xmm2
+; X86-SSE-NEXT: pxor %xmm1, %xmm1
+; X86-SSE-NEXT: pcmpgtb 24(%ebp), %xmm1
+; X86-SSE-NEXT: paddb %xmm4, %xmm1
+; X86-SSE-NEXT: pxor %xmm0, %xmm0
+; X86-SSE-NEXT: pcmpgtb 56(%ebp), %xmm0
+; X86-SSE-NEXT: paddb %xmm5, %xmm0
+; X86-SSE-NEXT: paddb %xmm1, %xmm0
+; X86-SSE-NEXT: pxor %xmm1, %xmm1
+; X86-SSE-NEXT: pcmpgtb 40(%ebp), %xmm1
+; X86-SSE-NEXT: paddb %xmm2, %xmm1
+; X86-SSE-NEXT: pxor %xmm2, %xmm2
+; X86-SSE-NEXT: pcmpgtb 72(%ebp), %xmm2
+; X86-SSE-NEXT: pxor %xmm4, %xmm4
+; X86-SSE-NEXT: pcmpgtb 8(%ebp), %xmm4
+; X86-SSE-NEXT: paddb %xmm2, %xmm4
+; X86-SSE-NEXT: paddb %xmm1, %xmm4
+; X86-SSE-NEXT: paddb %xmm0, %xmm4
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm4[2,3,2,3]
+; X86-SSE-NEXT: paddb %xmm4, %xmm0
+; X86-SSE-NEXT: psadbw %xmm3, %xmm0
+; X86-SSE-NEXT: movd %xmm0, %eax
+; X86-SSE-NEXT: # kill: def $al killed $al killed $eax
+; X86-SSE-NEXT: movl %ebp, %esp
+; X86-SSE-NEXT: popl %ebp
+; X86-SSE-NEXT: retl
+;
+; X64-SSE-LABEL: test_v128i8_v128i1:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: pxor %xmm8, %xmm8
+; X64-SSE-NEXT: pxor %xmm9, %xmm9
+; X64-SSE-NEXT: pcmpgtb %xmm4, %xmm9
+; X64-SSE-NEXT: pxor %xmm4, %xmm4
+; X64-SSE-NEXT: pcmpgtb %xmm0, %xmm4
+; X64-SSE-NEXT: paddb %xmm9, %xmm4
+; X64-SSE-NEXT: pxor %xmm0, %xmm0
+; X64-SSE-NEXT: pcmpgtb %xmm6, %xmm0
+; X64-SSE-NEXT: pxor %xmm6, %xmm6
+; X64-SSE-NEXT: pcmpgtb %xmm2, %xmm6
+; X64-SSE-NEXT: paddb %xmm0, %xmm6
+; X64-SSE-NEXT: paddb %xmm4, %xmm6
+; X64-SSE-NEXT: pxor %xmm0, %xmm0
+; X64-SSE-NEXT: pcmpgtb %xmm5, %xmm0
+; X64-SSE-NEXT: pxor %xmm2, %xmm2
+; X64-SSE-NEXT: pcmpgtb %xmm1, %xmm2
+; X64-SSE-NEXT: paddb %xmm0, %xmm2
+; X64-SSE-NEXT: pxor %xmm0, %xmm0
+; X64-SSE-NEXT: pcmpgtb %xmm7, %xmm0
+; X64-SSE-NEXT: pxor %xmm1, %xmm1
+; X64-SSE-NEXT: pcmpgtb %xmm3, %xmm1
+; X64-SSE-NEXT: paddb %xmm0, %xmm1
+; X64-SSE-NEXT: paddb %xmm2, %xmm1
+; X64-SSE-NEXT: paddb %xmm6, %xmm1
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X64-SSE-NEXT: paddb %xmm1, %xmm0
+; X64-SSE-NEXT: psadbw %xmm8, %xmm0
+; X64-SSE-NEXT: movd %xmm0, %eax
+; X64-SSE-NEXT: # kill: def $al killed $al killed $eax
+; X64-SSE-NEXT: retq
+;
+; X86-AVX1-LABEL: test_v128i8_v128i1:
+; 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: vpxor %xmm3, %xmm3, %xmm3
+; X86-AVX1-NEXT: vpcmpgtb %xmm1, %xmm3, %xmm4
+; X86-AVX1-NEXT: vpcmpgtb %xmm2, %xmm3, %xmm5
+; X86-AVX1-NEXT: vpcmpgtb %xmm0, %xmm3, %xmm6
+; X86-AVX1-NEXT: vpaddb %xmm5, %xmm6, %xmm5
+; X86-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X86-AVX1-NEXT: vpcmpgtb %xmm1, %xmm3, %xmm1
+; X86-AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X86-AVX1-NEXT: vpcmpgtb %xmm2, %xmm3, %xmm2
+; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X86-AVX1-NEXT: vpcmpgtb %xmm0, %xmm3, %xmm0
+; X86-AVX1-NEXT: vpaddb %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpcmpgtb 8(%ebp), %xmm3, %xmm2
+; X86-AVX1-NEXT: vpaddb %xmm2, %xmm4, %xmm2
+; X86-AVX1-NEXT: vpaddb %xmm2, %xmm5, %xmm2
+; X86-AVX1-NEXT: vpcmpgtb 24(%ebp), %xmm3, %xmm4
+; X86-AVX1-NEXT: vpaddb %xmm4, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpaddb %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddb %xmm0, %xmm2, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddb %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpsadbw %xmm3, %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_v128i1:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vpxor %xmm4, %xmm4, %xmm4
+; X64-AVX1-NEXT: vpcmpgtb %xmm2, %xmm4, %xmm5
+; X64-AVX1-NEXT: vpcmpgtb %xmm0, %xmm4, %xmm6
+; X64-AVX1-NEXT: vpaddb %xmm5, %xmm6, %xmm5
+; X64-AVX1-NEXT: vpcmpgtb %xmm3, %xmm4, %xmm6
+; X64-AVX1-NEXT: vpcmpgtb %xmm1, %xmm4, %xmm7
+; X64-AVX1-NEXT: vpaddb %xmm6, %xmm7, %xmm6
+; X64-AVX1-NEXT: vpaddb %xmm6, %xmm5, %xmm5
+; X64-AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X64-AVX1-NEXT: vpcmpgtb %xmm2, %xmm4, %xmm2
+; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X64-AVX1-NEXT: vpcmpgtb %xmm0, %xmm4, %xmm0
+; X64-AVX1-NEXT: vpaddb %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vextractf128 $1, %ymm3, %xmm2
+; X64-AVX1-NEXT: vpcmpgtb %xmm2, %xmm4, %xmm2
+; X64-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X64-AVX1-NEXT: vpcmpgtb %xmm1, %xmm4, %xmm1
+; X64-AVX1-NEXT: vpaddb %xmm2, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpaddb %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddb %xmm0, %xmm5, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddb %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpsadbw %xmm4, %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_v128i1:
+; 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: vpxor %xmm3, %xmm3, %xmm3
+; X86-AVX2-NEXT: vpcmpgtb %ymm1, %ymm3, %ymm1
+; X86-AVX2-NEXT: vpcmpgtb %ymm2, %ymm3, %ymm2
+; X86-AVX2-NEXT: vpcmpgtb %ymm0, %ymm3, %ymm0
+; X86-AVX2-NEXT: vpaddb %ymm2, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpcmpgtb 8(%ebp), %ymm3, %ymm2
+; X86-AVX2-NEXT: vpaddb %ymm2, %ymm1, %ymm1
+; X86-AVX2-NEXT: vpaddb %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpaddb %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddb %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X86-AVX2-NEXT: vpsadbw %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_v128i1:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpxor %xmm4, %xmm4, %xmm4
+; X64-AVX2-NEXT: vpcmpgtb %ymm2, %ymm4, %ymm2
+; X64-AVX2-NEXT: vpcmpgtb %ymm0, %ymm4, %ymm0
+; X64-AVX2-NEXT: vpaddb %ymm2, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpcmpgtb %ymm3, %ymm4, %ymm2
+; X64-AVX2-NEXT: vpcmpgtb %ymm1, %ymm4, %ymm1
+; X64-AVX2-NEXT: vpaddb %ymm2, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpaddb %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpaddb %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddb %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X64-AVX2-NEXT: vpsadbw %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
;
; AVX512-LABEL: test_v128i8_v128i1:
; AVX512: # %bb.0:
@@ -2073,3 +2657,12 @@ declare i8 @llvm.vector.reduce.add.v16i8(<16 x i8>)
declare i8 @llvm.vector.reduce.add.v32i8(<32 x i8>)
declare i8 @llvm.vector.reduce.add.v64i8(<64 x i8>)
declare i8 @llvm.vector.reduce.add.v128i8(<128 x i8>)
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; AVX512BW: {{.*}}
+; AVX512BWVL: {{.*}}
+; SSE41: {{.*}}
+; SSE42: {{.*}}
+; X64-AVX1-FAST: {{.*}}
+; X64-AVX1-SLOW: {{.*}}
+; X86-AVX1-FAST: {{.*}}
+; X86-AVX1-SLOW: {{.*}}
diff --git a/llvm/test/CodeGen/X86/vector-reduce-add-zext.ll b/llvm/test/CodeGen/X86/vector-reduce-add-zext.ll
index d99b200385585..9d13de751812f 100644
--- a/llvm/test/CodeGen/X86/vector-reduce-add-zext.ll
+++ b/llvm/test/CodeGen/X86/vector-reduce-add-zext.ll
@@ -1,92 +1,204 @@
; 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,AVX1,AVX1-SLOW
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx,+fast-hops | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-FAST
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefixes=AVX,AVX512
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw,+avx512vl | FileCheck %s --check-prefixes=AVX,AVX512
+; 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,SSE41
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+sse4.1 | FileCheck %s --check-prefixes=SSE,X64-SSE,SSE4,X64-SSE4,SSE41
+; 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,AVX1,AVX1-SLOW,X86-AVX1,X86-AVX1-SLOW
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-SLOW,X64-AVX1,X64-AVX1-SLOW
+; RUN: llc < %s -mtriple=i686-- -mattr=+avx,+fast-hops | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-FAST,X86-AVX1,X86-AVX1-FAST
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx,+fast-hops | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-FAST,X64-AVX1,X64-AVX1-FAST
+; RUN: llc < %s -mtriple=i686-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2,X86-AVX2
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2,X64-AVX2
+; 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
;
; vXi64
;
-define i64 @test_v2i64_v2i32(<2 x i32> %a0) {
-; SSE2-LABEL: test_v2i64_v2i32:
-; SSE2: # %bb.0:
-; SSE2-NEXT: xorps %xmm1, %xmm1
-; SSE2-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]
-; SSE2-NEXT: psrlq $32, %xmm0
-; SSE2-NEXT: paddq %xmm1, %xmm0
-; SSE2-NEXT: movq %xmm0, %rax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v2i64_v2i32:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: paddq %xmm0, %xmm1
-; SSE41-NEXT: movq %xmm1, %rax
-; SSE41-NEXT: retq
-;
-; AVX-LABEL: test_v2i64_v2i32:
-; AVX: # %bb.0:
-; AVX-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
-; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vmovq %xmm0, %rax
-; AVX-NEXT: retq
+define i64 @test_v2i64_v2i32(<2 x i32> %a0) nounwind {
+; X86-SSE2-LABEL: test_v2i64_v2i32:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: pxor %xmm1, %xmm1
+; 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: paddq %xmm0, %xmm1
+; X86-SSE2-NEXT: movd %xmm1, %eax
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE2-NEXT: movd %xmm0, %edx
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE2-LABEL: test_v2i64_v2i32:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: xorps %xmm1, %xmm1
+; X64-SSE2-NEXT: movss {{.*#+}} xmm1 = xmm0[0],xmm1[1,2,3]
+; X64-SSE2-NEXT: psrlq $32, %xmm0
+; X64-SSE2-NEXT: paddq %xmm1, %xmm0
+; X64-SSE2-NEXT: movq %xmm0, %rax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v2i64_v2i32:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: paddq %xmm0, %xmm1
+; X86-SSE4-NEXT: movd %xmm1, %eax
+; X86-SSE4-NEXT: pextrd $1, %xmm1, %edx
+; X86-SSE4-NEXT: retl
+;
+; X64-SSE4-LABEL: test_v2i64_v2i32:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: pmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE4-NEXT: paddq %xmm0, %xmm1
+; X64-SSE4-NEXT: movq %xmm1, %rax
+; X64-SSE4-NEXT: retq
+;
+; X86-AVX1-LABEL: test_v2i64_v2i32:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v2i64_v2i32:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovq %xmm0, %rax
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v2i64_v2i32:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v2i64_v2i32:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: retq
+;
+; AVX512-LABEL: test_v2i64_v2i32:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero
+; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: vmovq %xmm0, %rax
+; AVX512-NEXT: retq
%1 = zext <2 x i32> %a0 to <2 x i64>
%2 = call i64 @llvm.vector.reduce.add.v2i64(<2 x i64> %1)
ret i64 %2
}
-define i64 @test_v4i64_v4i16(<4 x i16> %a0) {
-; SSE2-LABEL: test_v4i64_v4i16:
-; SSE2: # %bb.0:
-; SSE2-NEXT: pxor %xmm1, %xmm1
-; SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; SSE2-NEXT: movdqa %xmm0, %xmm2
-; SSE2-NEXT: punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]
-; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
-; SSE2-NEXT: paddq %xmm2, %xmm0
-; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE2-NEXT: paddq %xmm0, %xmm1
-; SSE2-NEXT: movq %xmm1, %rax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v4i64_v4i16:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovzxwq {{.*#+}} xmm1 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; SSE41-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
-; SSE41-NEXT: pmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; SSE41-NEXT: paddq %xmm1, %xmm0
-; SSE41-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE41-NEXT: paddq %xmm0, %xmm1
-; SSE41-NEXT: movq %xmm1, %rax
-; SSE41-NEXT: retq
-;
-; AVX1-LABEL: test_v4i64_v4i16:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
-; AVX1-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
-; AVX1-NEXT: vpaddq %xmm0, %xmm1, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vmovq %xmm0, %rax
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: test_v4i64_v4i16:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpmovzxwq {{.*#+}} ymm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovq %xmm0, %rax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+define i64 @test_v4i64_v4i16(<4 x i16> %a0) nounwind {
+; X86-SSE2-LABEL: test_v4i64_v4i16:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: pxor %xmm1, %xmm1
+; X86-SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; X86-SSE2-NEXT: movdqa %xmm0, %xmm2
+; X86-SSE2-NEXT: punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; X86-SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
+; X86-SSE2-NEXT: paddq %xmm2, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE2-NEXT: paddq %xmm0, %xmm1
+; X86-SSE2-NEXT: movd %xmm1, %eax
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE2-NEXT: movd %xmm0, %edx
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE2-LABEL: test_v4i64_v4i16:
+; X64-SSE2: # %bb.0:
+; X64-SSE2-NEXT: pxor %xmm1, %xmm1
+; X64-SSE2-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; X64-SSE2-NEXT: movdqa %xmm0, %xmm2
+; X64-SSE2-NEXT: punpckhdq {{.*#+}} xmm2 = xmm2[2],xmm1[2],xmm2[3],xmm1[3]
+; X64-SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
+; X64-SSE2-NEXT: paddq %xmm2, %xmm0
+; X64-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE2-NEXT: paddq %xmm0, %xmm1
+; X64-SSE2-NEXT: movq %xmm1, %rax
+; X64-SSE2-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v4i64_v4i16:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pmovzxwq {{.*#+}} xmm1 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; X86-SSE4-NEXT: pmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
+; X86-SSE4-NEXT: paddq %xmm1, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: paddq %xmm0, %xmm1
+; X86-SSE4-NEXT: movd %xmm1, %eax
+; X86-SSE4-NEXT: pextrd $1, %xmm1, %edx
+; X86-SSE4-NEXT: retl
+;
+; X64-SSE4-LABEL: test_v4i64_v4i16:
+; X64-SSE4: # %bb.0:
+; X64-SSE4-NEXT: pmovzxwq {{.*#+}} xmm1 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; X64-SSE4-NEXT: pmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
+; X64-SSE4-NEXT: paddq %xmm1, %xmm0
+; X64-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE4-NEXT: paddq %xmm0, %xmm1
+; X64-SSE4-NEXT: movq %xmm1, %rax
+; X64-SSE4-NEXT: retq
+;
+; X86-AVX1-LABEL: test_v4i64_v4i16:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; X86-AVX1-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
+; X86-AVX1-NEXT: vpaddq %xmm0, %xmm1, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v4i64_v4i16:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vpmovzxwq {{.*#+}} xmm1 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; X64-AVX1-NEXT: vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
+; X64-AVX1-NEXT: vpaddq %xmm0, %xmm1, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovq %xmm0, %rax
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v4i64_v4i16:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpmovzxwq {{.*#+}} ymm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddq %xmm1, %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_v4i16:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpmovzxwq {{.*#+}} ymm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
;
; AVX512-LABEL: test_v4i64_v4i16:
; AVX512: # %bb.0:
@@ -103,43 +215,148 @@ define i64 @test_v4i64_v4i16(<4 x i16> %a0) {
ret i64 %2
}
-define i64 @test_v8i64_v8i8(<8 x i8> %a0) {
-; SSE-LABEL: test_v8i64_v8i8:
-; SSE: # %bb.0:
-; SSE-NEXT: pxor %xmm1, %xmm1
-; SSE-NEXT: psadbw %xmm0, %xmm1
-; SSE-NEXT: movq %xmm1, %rax
-; SSE-NEXT: retq
-;
-; AVX-LABEL: test_v8i64_v8i8:
-; AVX: # %bb.0:
-; AVX-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vmovq %xmm0, %rax
-; AVX-NEXT: retq
+define i64 @test_v8i64_v8i8(<8 x i8> %a0) nounwind {
+; X86-SSE2-LABEL: test_v8i64_v8i8:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: pxor %xmm1, %xmm1
+; X86-SSE2-NEXT: psadbw %xmm0, %xmm1
+; X86-SSE2-NEXT: movd %xmm1, %eax
+; X86-SSE2-NEXT: xorl %edx, %edx
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE-LABEL: test_v8i64_v8i8:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: pxor %xmm1, %xmm1
+; X64-SSE-NEXT: psadbw %xmm0, %xmm1
+; X64-SSE-NEXT: movq %xmm1, %rax
+; X64-SSE-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v8i64_v8i8:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pxor %xmm1, %xmm1
+; X86-SSE4-NEXT: psadbw %xmm0, %xmm1
+; X86-SSE4-NEXT: movd %xmm1, %eax
+; X86-SSE4-NEXT: pextrd $1, %xmm1, %edx
+; X86-SSE4-NEXT: retl
+;
+; X86-AVX1-LABEL: test_v8i64_v8i8:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v8i64_v8i8:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovq %xmm0, %rax
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v8i64_v8i8:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X86-AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v8i64_v8i8:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X64-AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: retq
+;
+; AVX512-LABEL: test_v8i64_v8i8:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; AVX512-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: vmovq %xmm0, %rax
+; AVX512-NEXT: retq
%1 = zext <8 x i8> %a0 to <8 x i64>
%2 = call i64 @llvm.vector.reduce.add.v8i64(<8 x i64> %1)
ret i64 %2
}
-define i64 @test_v16i64_v16i8(<16 x i8> %a0) {
-; SSE-LABEL: test_v16i64_v16i8:
-; SSE: # %bb.0:
-; SSE-NEXT: pxor %xmm1, %xmm1
-; SSE-NEXT: psadbw %xmm0, %xmm1
-; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE-NEXT: paddq %xmm1, %xmm0
-; SSE-NEXT: movq %xmm0, %rax
-; SSE-NEXT: retq
-;
-; AVX-LABEL: test_v16i64_v16i8:
-; AVX: # %bb.0:
-; AVX-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vmovq %xmm0, %rax
-; AVX-NEXT: retq
+define i64 @test_v16i64_v16i8(<16 x i8> %a0) nounwind {
+; X86-SSE2-LABEL: test_v16i64_v16i8:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: pxor %xmm1, %xmm1
+; X86-SSE2-NEXT: psadbw %xmm0, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE2-NEXT: paddq %xmm1, %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_v16i64_v16i8:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: pxor %xmm1, %xmm1
+; X64-SSE-NEXT: psadbw %xmm0, %xmm1
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X64-SSE-NEXT: paddq %xmm1, %xmm0
+; X64-SSE-NEXT: movq %xmm0, %rax
+; X64-SSE-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v16i64_v16i8:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: pxor %xmm1, %xmm1
+; X86-SSE4-NEXT: psadbw %xmm0, %xmm1
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE4-NEXT: paddq %xmm1, %xmm0
+; X86-SSE4-NEXT: movd %xmm0, %eax
+; X86-SSE4-NEXT: pextrd $1, %xmm0, %edx
+; X86-SSE4-NEXT: retl
+;
+; X86-AVX1-LABEL: test_v16i64_v16i8:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v16i64_v16i8:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vmovq %xmm0, %rax
+; X64-AVX1-NEXT: retq
+;
+; X86-AVX2-LABEL: test_v16i64_v16i8:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X86-AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: vpextrd $1, %xmm0, %edx
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v16i64_v16i8:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X64-AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: retq
+;
+; AVX512-LABEL: test_v16i64_v16i8:
+; AVX512: # %bb.0:
+; AVX512-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; AVX512-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; AVX512-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; AVX512-NEXT: vmovq %xmm0, %rax
+; AVX512-NEXT: retq
%1 = zext <16 x i8> %a0 to <16 x i64>
%2 = call i64 @llvm.vector.reduce.add.v16i64(<16 x i64> %1)
ret i64 %2
@@ -149,7 +366,7 @@ define i64 @test_v16i64_v16i8(<16 x i8> %a0) {
; vXi32
;
-define i32 @test_v2i32_v2i16(<2 x i16> %a0) {
+define i32 @test_v2i32_v2i16(<2 x i16> %a0) nounwind {
; SSE2-LABEL: test_v2i32_v2i16:
; SSE2: # %bb.0:
; SSE2-NEXT: pxor %xmm1, %xmm1
@@ -158,15 +375,15 @@ define i32 @test_v2i32_v2i16(<2 x i16> %a0) {
; SSE2-NEXT: psrld $16, %xmm0
; SSE2-NEXT: paddd %xmm2, %xmm0
; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
;
-; SSE41-LABEL: test_v2i32_v2i16:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pmovzxwd {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
-; SSE41-NEXT: psrld $16, %xmm0
-; SSE41-NEXT: paddd %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: retq
+; SSE4-LABEL: test_v2i32_v2i16:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pmovzxwd {{.*#+}} xmm1 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
+; SSE4-NEXT: psrld $16, %xmm0
+; SSE4-NEXT: paddd %xmm1, %xmm0
+; SSE4-NEXT: movd %xmm0, %eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v2i32_v2i16:
; AVX1-SLOW: # %bb.0:
@@ -174,14 +391,14 @@ define i32 @test_v2i32_v2i16(<2 x i16> %a0) {
; AVX1-SLOW-NEXT: vpsrld $16, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vpaddd %xmm0, %xmm1, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v2i32_v2i16:
; AVX1-FAST: # %bb.0:
; AVX1-FAST-NEXT: vpmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v2i32_v2i16:
; AVX2: # %bb.0:
@@ -189,7 +406,7 @@ define i32 @test_v2i32_v2i16(<2 x i16> %a0) {
; AVX2-NEXT: vpsrld $16, %xmm0, %xmm0
; AVX2-NEXT: vpaddd %xmm0, %xmm1, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v2i32_v2i16:
; AVX512: # %bb.0:
@@ -203,22 +420,22 @@ define i32 @test_v2i32_v2i16(<2 x i16> %a0) {
ret i32 %2
}
-define i32 @test_v4i32(<4 x i8> %a0) {
+define i32 @test_v4i32(<4 x i8> %a0) nounwind {
; SSE2-LABEL: test_v4i32:
; SSE2: # %bb.0:
; SSE2-NEXT: pxor %xmm1, %xmm1
; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
; SSE2-NEXT: psadbw %xmm1, %xmm0
; SSE2-NEXT: movd %xmm0, %eax
-; SSE2-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
;
-; SSE41-LABEL: test_v4i32:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pxor %xmm1, %xmm1
-; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3,4,5,6,7]
-; SSE41-NEXT: psadbw %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: retq
+; SSE4-LABEL: test_v4i32:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pxor %xmm1, %xmm1
+; SSE4-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3,4,5,6,7]
+; SSE4-NEXT: psadbw %xmm1, %xmm0
+; SSE4-NEXT: movd %xmm0, %eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX1-LABEL: test_v4i32:
; AVX1: # %bb.0:
@@ -226,7 +443,7 @@ define i32 @test_v4i32(<4 x i8> %a0) {
; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3,4,5,6,7]
; AVX1-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vmovd %xmm0, %eax
-; AVX1-NEXT: retq
+; AVX1-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v4i32:
; AVX2: # %bb.0:
@@ -234,7 +451,7 @@ define i32 @test_v4i32(<4 x i8> %a0) {
; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
; AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v4i32:
; AVX512: # %bb.0:
@@ -249,26 +466,26 @@ define i32 @test_v4i32(<4 x i8> %a0) {
ret i32 %2
}
-define i32 @test_v8i32_v8i8(<8 x i8> %a0) {
+define i32 @test_v8i32_v8i8(<8 x i8> %a0) nounwind {
; SSE-LABEL: test_v8i32_v8i8:
; SSE: # %bb.0:
; SSE-NEXT: pxor %xmm1, %xmm1
; SSE-NEXT: psadbw %xmm0, %xmm1
; SSE-NEXT: movd %xmm1, %eax
-; SSE-NEXT: retq
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX-LABEL: test_v8i32_v8i8:
; AVX: # %bb.0:
; AVX-NEXT: vpxor %xmm1, %xmm1, %xmm1
; AVX-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
; AVX-NEXT: vmovd %xmm0, %eax
-; AVX-NEXT: retq
+; AVX-NEXT: ret{{[l|q]}}
%1 = zext <8 x i8> %a0 to <8 x i32>
%2 = call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> %1)
ret i32 %2
}
-define i32 @test_v16i32_v16i8(<16 x i8> %a0) {
+define i32 @test_v16i32_v16i8(<16 x i8> %a0) nounwind {
; SSE-LABEL: test_v16i32_v16i8:
; SSE: # %bb.0:
; SSE-NEXT: pxor %xmm1, %xmm1
@@ -276,7 +493,7 @@ define i32 @test_v16i32_v16i8(<16 x i8> %a0) {
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
; SSE-NEXT: paddq %xmm1, %xmm0
; SSE-NEXT: movd %xmm0, %eax
-; SSE-NEXT: retq
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX-LABEL: test_v16i32_v16i8:
; AVX: # %bb.0:
@@ -285,13 +502,13 @@ define i32 @test_v16i32_v16i8(<16 x i8> %a0) {
; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
; AVX-NEXT: vpaddq %xmm1, %xmm0, %xmm0
; AVX-NEXT: vmovd %xmm0, %eax
-; AVX-NEXT: retq
+; AVX-NEXT: ret{{[l|q]}}
%1 = zext <16 x i8> %a0 to <16 x i32>
%2 = call i32 @llvm.vector.reduce.add.v16i32(<16 x i32> %1)
ret i32 %2
}
-define i32 @test_v32i32_v32i8(<32 x i8> %a0) {
+define i32 @test_v32i32_v32i8(<32 x i8> %a0) nounwind {
; SSE-LABEL: test_v32i32_v32i8:
; SSE: # %bb.0:
; SSE-NEXT: pxor %xmm2, %xmm2
@@ -301,7 +518,7 @@ define i32 @test_v32i32_v32i8(<32 x i8> %a0) {
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
; SSE-NEXT: paddq %xmm0, %xmm1
; SSE-NEXT: movd %xmm1, %eax
-; SSE-NEXT: retq
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX1-LABEL: test_v32i32_v32i8:
; AVX1: # %bb.0:
@@ -314,7 +531,7 @@ define i32 @test_v32i32_v32i8(<32 x i8> %a0) {
; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vmovd %xmm0, %eax
; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
+; AVX1-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v32i32_v32i8:
; AVX2: # %bb.0:
@@ -326,7 +543,7 @@ define i32 @test_v32i32_v32i8(<32 x i8> %a0) {
; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v32i32_v32i8:
; AVX512: # %bb.0:
@@ -348,7 +565,7 @@ define i32 @test_v32i32_v32i8(<32 x i8> %a0) {
; vXi16
;
-define i16 @test_v2i16_v2i8(<2 x i8> %a0) {
+define i16 @test_v2i16_v2i8(<2 x i8> %a0) nounwind {
; SSE2-LABEL: test_v2i16_v2i8:
; SSE2: # %bb.0:
; SSE2-NEXT: pxor %xmm1, %xmm1
@@ -358,17 +575,17 @@ define i16 @test_v2i16_v2i8(<2 x i8> %a0) {
; SSE2-NEXT: paddw %xmm0, %xmm1
; SSE2-NEXT: movd %xmm1, %eax
; SSE2-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v2i16_v2i8:
-; 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: movdqa %xmm0, %xmm1
-; SSE41-NEXT: psrld $16, %xmm1
-; SSE41-NEXT: paddw %xmm0, %xmm1
-; SSE41-NEXT: movd %xmm1, %eax
-; SSE41-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE41-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
+;
+; SSE4-LABEL: test_v2i16_v2i8:
+; 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: movdqa %xmm0, %xmm1
+; SSE4-NEXT: psrld $16, %xmm1
+; SSE4-NEXT: paddw %xmm0, %xmm1
+; SSE4-NEXT: movd %xmm1, %eax
+; SSE4-NEXT: # kill: def $ax killed $ax killed $eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v2i16_v2i8:
; AVX1-SLOW: # %bb.0:
@@ -377,7 +594,7 @@ define i16 @test_v2i16_v2i8(<2 x i8> %a0) {
; AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
; AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v2i16_v2i8:
; AVX1-FAST: # %bb.0:
@@ -385,7 +602,7 @@ define i16 @test_v2i16_v2i8(<2 x i8> %a0) {
; AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
; AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v2i16_v2i8:
; AVX2: # %bb.0:
@@ -394,7 +611,7 @@ define i16 @test_v2i16_v2i8(<2 x i8> %a0) {
; AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v2i16_v2i8:
; AVX512: # %bb.0:
@@ -409,7 +626,7 @@ define i16 @test_v2i16_v2i8(<2 x i8> %a0) {
ret i16 %2
}
-define i16 @test_v4i16_v4i8(<4 x i8> %a0) {
+define i16 @test_v4i16_v4i8(<4 x i8> %a0) nounwind {
; SSE2-LABEL: test_v4i16_v4i8:
; SSE2: # %bb.0:
; SSE2-NEXT: pxor %xmm1, %xmm1
@@ -421,19 +638,19 @@ define i16 @test_v4i16_v4i8(<4 x i8> %a0) {
; SSE2-NEXT: paddw %xmm1, %xmm0
; SSE2-NEXT: movd %xmm0, %eax
; SSE2-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE2-NEXT: retq
-;
-; SSE41-LABEL: test_v4i16_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: paddw %xmm0, %xmm1
-; SSE41-NEXT: movdqa %xmm1, %xmm0
-; SSE41-NEXT: psrld $16, %xmm0
-; SSE41-NEXT: paddw %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE41-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
+;
+; SSE4-LABEL: test_v4i16_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: paddw %xmm0, %xmm1
+; SSE4-NEXT: movdqa %xmm1, %xmm0
+; SSE4-NEXT: psrld $16, %xmm0
+; SSE4-NEXT: paddw %xmm1, %xmm0
+; SSE4-NEXT: movd %xmm0, %eax
+; SSE4-NEXT: # kill: def $ax killed $ax killed $eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v4i16_v4i8:
; AVX1-SLOW: # %bb.0:
@@ -444,7 +661,7 @@ define i16 @test_v4i16_v4i8(<4 x i8> %a0) {
; AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
; AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v4i16_v4i8:
; AVX1-FAST: # %bb.0:
@@ -454,7 +671,7 @@ define i16 @test_v4i16_v4i8(<4 x i8> %a0) {
; AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
; AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v4i16_v4i8:
; AVX2: # %bb.0:
@@ -465,7 +682,7 @@ define i16 @test_v4i16_v4i8(<4 x i8> %a0) {
; AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v4i16_v4i8:
; AVX512: # %bb.0:
@@ -483,14 +700,14 @@ define i16 @test_v4i16_v4i8(<4 x i8> %a0) {
}
-define i16 @test_v8i16_v8i8(<8 x i8> %a0) {
+define i16 @test_v8i16_v8i8(<8 x i8> %a0) nounwind {
; SSE-LABEL: test_v8i16_v8i8:
; SSE: # %bb.0:
; SSE-NEXT: pxor %xmm1, %xmm1
; SSE-NEXT: psadbw %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_v8i8:
; AVX: # %bb.0:
@@ -498,13 +715,13 @@ define i16 @test_v8i16_v8i8(<8 x i8> %a0) {
; AVX-NEXT: vpsadbw %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 = zext <8 x i8> %a0 to <8 x i16>
%2 = call i16 @llvm.vector.reduce.add.v8i16(<8 x i16> %1)
ret i16 %2
}
-define i16 @test_v16i16_v16i8(<16 x i8> %a0) {
+define i16 @test_v16i16_v16i8(<16 x i8> %a0) nounwind {
; SSE-LABEL: test_v16i16_v16i8:
; SSE: # %bb.0:
; SSE-NEXT: pxor %xmm1, %xmm1
@@ -513,7 +730,7 @@ define i16 @test_v16i16_v16i8(<16 x i8> %a0) {
; SSE-NEXT: paddq %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_v16i16_v16i8:
; AVX: # %bb.0:
@@ -523,13 +740,13 @@ define i16 @test_v16i16_v16i8(<16 x i8> %a0) {
; AVX-NEXT: vpaddq %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 = zext <16 x i8> %a0 to <16 x i16>
%2 = call i16 @llvm.vector.reduce.add.v16i16(<16 x i16> %1)
ret i16 %2
}
-define i16 @test_v32i16_v32i8(<32 x i8> %a0) {
+define i16 @test_v32i16_v32i8(<32 x i8> %a0) nounwind {
; SSE-LABEL: test_v32i16_v32i8:
; SSE: # %bb.0:
; SSE-NEXT: pxor %xmm2, %xmm2
@@ -540,7 +757,7 @@ define i16 @test_v32i16_v32i8(<32 x i8> %a0) {
; SSE-NEXT: paddq %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_v32i16_v32i8:
; AVX1: # %bb.0:
@@ -554,7 +771,7 @@ define i16 @test_v32i16_v32i8(<32 x i8> %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_v32i8:
; AVX2: # %bb.0:
@@ -567,7 +784,7 @@ define i16 @test_v32i16_v32i8(<32 x i8> %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_v32i16_v32i8:
; AVX512: # %bb.0:
@@ -586,22 +803,44 @@ define i16 @test_v32i16_v32i8(<32 x i8> %a0) {
ret i16 %2
}
-define i16 @test_v64i16_v64i8(<64 x i8> %a0) {
-; SSE-LABEL: test_v64i16_v64i8:
-; SSE: # %bb.0:
-; SSE-NEXT: pxor %xmm4, %xmm4
-; SSE-NEXT: psadbw %xmm4, %xmm3
-; SSE-NEXT: psadbw %xmm4, %xmm1
-; SSE-NEXT: paddq %xmm3, %xmm1
-; SSE-NEXT: psadbw %xmm4, %xmm2
-; SSE-NEXT: psadbw %xmm4, %xmm0
-; SSE-NEXT: paddq %xmm2, %xmm0
-; SSE-NEXT: paddq %xmm1, %xmm0
-; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE-NEXT: paddq %xmm0, %xmm1
-; SSE-NEXT: movd %xmm1, %eax
-; SSE-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE-NEXT: retq
+define i16 @test_v64i16_v64i8(<64 x i8> %a0) nounwind {
+; X86-SSE-LABEL: test_v64i16_v64i8:
+; 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: pxor %xmm3, %xmm3
+; X86-SSE-NEXT: psadbw %xmm3, %xmm2
+; X86-SSE-NEXT: psadbw %xmm3, %xmm0
+; X86-SSE-NEXT: paddq %xmm2, %xmm0
+; X86-SSE-NEXT: psadbw %xmm3, %xmm1
+; X86-SSE-NEXT: psadbw 8(%ebp), %xmm3
+; X86-SSE-NEXT: paddq %xmm1, %xmm3
+; X86-SSE-NEXT: paddq %xmm0, %xmm3
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,2,3]
+; X86-SSE-NEXT: paddq %xmm3, %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_v64i16_v64i8:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: pxor %xmm4, %xmm4
+; X64-SSE-NEXT: psadbw %xmm4, %xmm3
+; X64-SSE-NEXT: psadbw %xmm4, %xmm1
+; X64-SSE-NEXT: paddq %xmm3, %xmm1
+; X64-SSE-NEXT: psadbw %xmm4, %xmm2
+; X64-SSE-NEXT: psadbw %xmm4, %xmm0
+; X64-SSE-NEXT: paddq %xmm2, %xmm0
+; X64-SSE-NEXT: paddq %xmm1, %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE-NEXT: paddq %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_v64i16_v64i8:
; AVX1: # %bb.0:
@@ -620,7 +859,7 @@ define i16 @test_v64i16_v64i8(<64 x i8> %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_v64i16_v64i8:
; AVX2: # %bb.0:
@@ -635,7 +874,7 @@ define i16 @test_v64i16_v64i8(<64 x i8> %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_v64i16_v64i8:
; AVX512: # %bb.0:
@@ -682,3 +921,11 @@ declare i8 @llvm.vector.reduce.add.v16i8(<16 x i8>)
declare i8 @llvm.vector.reduce.add.v32i8(<32 x i8>)
declare i8 @llvm.vector.reduce.add.v64i8(<64 x i8>)
declare i8 @llvm.vector.reduce.add.v128i8(<128 x i8>)
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; AVX512BW: {{.*}}
+; AVX512BWVL: {{.*}}
+; SSE41: {{.*}}
+; X64-AVX1-FAST: {{.*}}
+; X64-AVX1-SLOW: {{.*}}
+; X86-AVX1-FAST: {{.*}}
+; X86-AVX1-SLOW: {{.*}}
diff --git a/llvm/test/CodeGen/X86/vector-reduce-add.ll b/llvm/test/CodeGen/X86/vector-reduce-add.ll
index aed4e023e340c..ff5de0e587505 100644
--- a/llvm/test/CodeGen/X86/vector-reduce-add.ll
+++ b/llvm/test/CodeGen/X86/vector-reduce-add.ll
@@ -1,30 +1,77 @@
; 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,AVX1,AVX1-SLOW
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx,+fast-hops | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-FAST
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefix=AVX512
-; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw,+avx512vl | FileCheck %s --check-prefix=AVX512
+; 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,AVX1,AVX1-SLOW,X86-AVX1,X86-AVX1-SLOW
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-SLOW,X64-AVX1,X64-AVX1-SLOW
+; RUN: llc < %s -mtriple=i686-- -mattr=+avx,+fast-hops | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-FAST,X86-AVX1,X86-AVX1-FAST
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx,+fast-hops | FileCheck %s --check-prefixes=AVX,AVX1,AVX1-FAST,X64-AVX1,X64-AVX1-FAST
+; RUN: llc < %s -mtriple=i686-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2,X86-AVX2
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2,X64-AVX2
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefix=AVX512
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f,+avx512bw,+avx512vl | FileCheck %s --check-prefix=AVX512
;
; 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: paddq %xmm0, %xmm1
-; SSE-NEXT: movq %xmm1, %rax
-; SSE-NEXT: retq
-;
-; AVX-LABEL: test_v2i64:
-; AVX: # %bb.0:
-; AVX-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vmovq %xmm0, %rax
-; AVX-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: paddq %xmm0, %xmm1
+; X86-SSE2-NEXT: movd %xmm1, %eax
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[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: paddq %xmm0, %xmm1
+; X64-SSE-NEXT: movq %xmm1, %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: paddq %xmm0, %xmm1
+; X86-SSE4-NEXT: movd %xmm1, %eax
+; X86-SSE4-NEXT: pextrd $1, %xmm1, %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: vpaddq %xmm1, %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: vpaddq %xmm1, %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: vpaddq %xmm1, %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: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: retq
;
; AVX512-LABEL: test_v2i64:
; AVX512: # %bb.0:
@@ -36,34 +83,75 @@ 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: paddq %xmm1, %xmm0
-; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE-NEXT: paddq %xmm0, %xmm1
-; SSE-NEXT: movq %xmm1, %rax
-; SSE-NEXT: retq
-;
-; AVX1-LABEL: test_v4i64:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vmovq %xmm0, %rax
-; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: test_v4i64:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovq %xmm0, %rax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+define i64 @test_v4i64(<4 x i64> %a0) nounwind {
+; X86-SSE2-LABEL: test_v4i64:
+; X86-SSE2: # %bb.0:
+; X86-SSE2-NEXT: paddq %xmm1, %xmm0
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE2-NEXT: paddq %xmm0, %xmm1
+; X86-SSE2-NEXT: movd %xmm1, %eax
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X86-SSE2-NEXT: movd %xmm0, %edx
+; X86-SSE2-NEXT: retl
+;
+; X64-SSE-LABEL: test_v4i64:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: paddq %xmm1, %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE-NEXT: paddq %xmm0, %xmm1
+; X64-SSE-NEXT: movq %xmm1, %rax
+; X64-SSE-NEXT: retq
+;
+; X86-SSE4-LABEL: test_v4i64:
+; X86-SSE4: # %bb.0:
+; X86-SSE4-NEXT: paddq %xmm1, %xmm0
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE4-NEXT: paddq %xmm0, %xmm1
+; X86-SSE4-NEXT: movd %xmm1, %eax
+; X86-SSE4-NEXT: pextrd $1, %xmm1, %edx
+; X86-SSE4-NEXT: retl
+;
+; X86-AVX1-LABEL: test_v4i64:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddq %xmm1, %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: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddq %xmm1, %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: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddq %xmm1, %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: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
;
; AVX512-LABEL: test_v4i64:
; AVX512: # %bb.0:
@@ -78,40 +166,101 @@ 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: paddq %xmm3, %xmm1
-; SSE-NEXT: paddq %xmm2, %xmm0
-; SSE-NEXT: paddq %xmm1, %xmm0
-; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE-NEXT: paddq %xmm0, %xmm1
-; SSE-NEXT: movq %xmm1, %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: vpaddq %xmm2, %xmm3, %xmm2
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vmovq %xmm0, %rax
-; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: test_v8i64:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpaddq %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovq %xmm0, %rax
-; AVX2-NEXT: vzeroupper
-; AVX2-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: paddq %xmm2, %xmm0
+; X86-SSE2-NEXT: paddq 8(%ebp), %xmm1
+; X86-SSE2-NEXT: paddq %xmm0, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE2-NEXT: paddq %xmm1, %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
+;
+; X64-SSE-LABEL: test_v8i64:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: paddq %xmm3, %xmm1
+; X64-SSE-NEXT: paddq %xmm2, %xmm0
+; X64-SSE-NEXT: paddq %xmm1, %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE-NEXT: paddq %xmm0, %xmm1
+; X64-SSE-NEXT: movq %xmm1, %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: paddq %xmm2, %xmm0
+; X86-SSE4-NEXT: paddq 8(%ebp), %xmm1
+; X86-SSE4-NEXT: paddq %xmm0, %xmm1
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE4-NEXT: paddq %xmm1, %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: vpaddq %xmm2, %xmm3, %xmm2
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddq %xmm1, %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: vpaddq %xmm2, %xmm3, %xmm2
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddq %xmm1, %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: vpaddq %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddq %xmm1, %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: vpaddq %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
;
; AVX512-LABEL: test_v8i64:
; AVX512: # %bb.0:
@@ -128,52 +277,142 @@ 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: paddq %xmm6, %xmm2
-; SSE-NEXT: paddq %xmm4, %xmm0
-; SSE-NEXT: paddq %xmm2, %xmm0
-; SSE-NEXT: paddq %xmm7, %xmm3
-; SSE-NEXT: paddq %xmm5, %xmm1
-; SSE-NEXT: paddq %xmm3, %xmm1
-; SSE-NEXT: paddq %xmm0, %xmm1
-; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE-NEXT: paddq %xmm1, %xmm0
-; SSE-NEXT: movq %xmm0, %rax
-; SSE-NEXT: retq
-;
-; AVX1-LABEL: test_v16i64:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vpaddq %xmm3, %xmm1, %xmm4
-; AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm5
-; AVX1-NEXT: vpaddq %xmm4, %xmm5, %xmm4
-; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm3
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-NEXT: vpaddq %xmm3, %xmm1, %xmm1
-; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpaddq %xmm0, %xmm4, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vmovq %xmm0, %rax
-; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: test_v16i64:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpaddq %ymm3, %ymm1, %ymm1
-; AVX2-NEXT: vpaddq %ymm2, %ymm0, %ymm0
-; AVX2-NEXT: vpaddq %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovq %xmm0, %rax
-; AVX2-NEXT: vzeroupper
-; AVX2-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 8(%ebp), %xmm3
+; X86-SSE2-NEXT: paddq 56(%ebp), %xmm2
+; X86-SSE2-NEXT: paddq 24(%ebp), %xmm0
+; X86-SSE2-NEXT: paddq %xmm2, %xmm0
+; X86-SSE2-NEXT: paddq 72(%ebp), %xmm3
+; X86-SSE2-NEXT: paddq 40(%ebp), %xmm1
+; X86-SSE2-NEXT: paddq %xmm3, %xmm1
+; X86-SSE2-NEXT: paddq %xmm0, %xmm1
+; X86-SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE2-NEXT: paddq %xmm1, %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
+;
+; X64-SSE-LABEL: test_v16i64:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: paddq %xmm6, %xmm2
+; X64-SSE-NEXT: paddq %xmm4, %xmm0
+; X64-SSE-NEXT: paddq %xmm2, %xmm0
+; X64-SSE-NEXT: paddq %xmm7, %xmm3
+; X64-SSE-NEXT: paddq %xmm5, %xmm1
+; X64-SSE-NEXT: paddq %xmm3, %xmm1
+; X64-SSE-NEXT: paddq %xmm0, %xmm1
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X64-SSE-NEXT: paddq %xmm1, %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 8(%ebp), %xmm3
+; X86-SSE4-NEXT: paddq 56(%ebp), %xmm2
+; X86-SSE4-NEXT: paddq 24(%ebp), %xmm0
+; X86-SSE4-NEXT: paddq %xmm2, %xmm0
+; X86-SSE4-NEXT: paddq 72(%ebp), %xmm3
+; X86-SSE4-NEXT: paddq 40(%ebp), %xmm1
+; X86-SSE4-NEXT: paddq %xmm3, %xmm1
+; X86-SSE4-NEXT: paddq %xmm0, %xmm1
+; X86-SSE4-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE4-NEXT: paddq %xmm1, %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: vpaddq %xmm2, %xmm0, %xmm3
+; X86-AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq 8(%ebp), %xmm1, %xmm2
+; X86-AVX1-NEXT: vpaddq %xmm2, %xmm3, %xmm2
+; X86-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X86-AVX1-NEXT: vpaddq 24(%ebp), %xmm1, %xmm1
+; X86-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddq %xmm0, %xmm2, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddq %xmm1, %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: vpaddq %xmm3, %xmm1, %xmm4
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm5
+; X64-AVX1-NEXT: vpaddq %xmm4, %xmm5, %xmm4
+; X64-AVX1-NEXT: vextractf128 $1, %ymm3, %xmm3
+; X64-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X64-AVX1-NEXT: vpaddq %xmm3, %xmm1, %xmm1
+; X64-AVX1-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddq %xmm0, %xmm4, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddq %xmm1, %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: vpaddq %ymm2, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpaddq 8(%ebp), %ymm1, %ymm1
+; X86-AVX2-NEXT: vpaddq %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddq %xmm1, %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: vpaddq %ymm3, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpaddq %ymm2, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpaddq %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddq %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovq %xmm0, %rax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
;
; AVX512-LABEL: test_v16i64:
; AVX512: # %bb.0:
@@ -195,33 +434,33 @@ 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 {
; SSE-LABEL: test_v2i32:
; SSE: # %bb.0:
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; SSE-NEXT: paddd %xmm0, %xmm1
; SSE-NEXT: movd %xmm1, %eax
-; SSE-NEXT: retq
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v2i32:
; AVX1-SLOW: # %bb.0:
; AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v2i32:
; AVX1-FAST: # %bb.0:
; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v2i32:
; AVX2: # %bb.0:
; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v2i32:
; AVX512: # %bb.0:
@@ -233,7 +472,7 @@ define i32 @test_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 {
; SSE-LABEL: test_v4i32:
; SSE: # %bb.0:
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
@@ -241,7 +480,7 @@ define i32 @test_v4i32(<4 x i32> %a0) {
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
; SSE-NEXT: paddd %xmm1, %xmm0
; SSE-NEXT: movd %xmm0, %eax
-; SSE-NEXT: retq
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v4i32:
; AVX1-SLOW: # %bb.0:
@@ -250,14 +489,14 @@ define i32 @test_v4i32(<4 x i32> %a0) {
; AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v4i32:
; AVX1-FAST: # %bb.0:
; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v4i32:
; AVX2: # %bb.0:
@@ -266,7 +505,7 @@ define i32 @test_v4i32(<4 x i32> %a0) {
; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v4i32:
; AVX512: # %bb.0:
@@ -280,7 +519,7 @@ define i32 @test_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 {
; SSE-LABEL: test_v8i32:
; SSE: # %bb.0:
; SSE-NEXT: paddd %xmm1, %xmm0
@@ -289,7 +528,7 @@ define i32 @test_v8i32(<8 x i32> %a0) {
; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
; SSE-NEXT: paddd %xmm1, %xmm0
; SSE-NEXT: movd %xmm0, %eax
-; SSE-NEXT: retq
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v8i32:
; AVX1-SLOW: # %bb.0:
@@ -301,7 +540,7 @@ define i32 @test_v8i32(<8 x i32> %a0) {
; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
; AVX1-SLOW-NEXT: vzeroupper
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v8i32:
; AVX1-FAST: # %bb.0:
@@ -311,7 +550,7 @@ define i32 @test_v8i32(<8 x i32> %a0) {
; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
; AVX1-FAST-NEXT: vzeroupper
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v8i32:
; AVX2: # %bb.0:
@@ -323,7 +562,7 @@ define i32 @test_v8i32(<8 x i32> %a0) {
; AVX2-NEXT: vpaddd %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:
@@ -340,18 +579,36 @@ define i32 @test_v8i32(<8 x i32> %a0) {
ret i32 %1
}
-define i32 @test_v16i32(<16 x i32> %a0) {
-; SSE-LABEL: test_v16i32:
-; SSE: # %bb.0:
-; SSE-NEXT: paddd %xmm3, %xmm1
-; SSE-NEXT: paddd %xmm2, %xmm0
-; SSE-NEXT: paddd %xmm1, %xmm0
-; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE-NEXT: paddd %xmm0, %xmm1
-; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE-NEXT: paddd %xmm1, %xmm0
-; SSE-NEXT: movd %xmm0, %eax
-; SSE-NEXT: retq
+define i32 @test_v16i32(<16 x i32> %a0) nounwind {
+; X86-SSE-LABEL: test_v16i32:
+; 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: paddd %xmm2, %xmm0
+; X86-SSE-NEXT: paddd 8(%ebp), %xmm1
+; X86-SSE-NEXT: paddd %xmm0, %xmm1
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE-NEXT: paddd %xmm1, %xmm0
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE-NEXT: paddd %xmm0, %xmm1
+; X86-SSE-NEXT: movd %xmm1, %eax
+; X86-SSE-NEXT: movl %ebp, %esp
+; X86-SSE-NEXT: popl %ebp
+; X86-SSE-NEXT: retl
+;
+; X64-SSE-LABEL: test_v16i32:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: paddd %xmm3, %xmm1
+; X64-SSE-NEXT: paddd %xmm2, %xmm0
+; X64-SSE-NEXT: paddd %xmm1, %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE-NEXT: paddd %xmm0, %xmm1
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE-NEXT: paddd %xmm1, %xmm0
+; X64-SSE-NEXT: movd %xmm0, %eax
+; X64-SSE-NEXT: retq
;
; AVX1-SLOW-LABEL: test_v16i32:
; AVX1-SLOW: # %bb.0:
@@ -366,7 +623,7 @@ define i32 @test_v16i32(<16 x i32> %a0) {
; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
; AVX1-SLOW-NEXT: vzeroupper
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v16i32:
; AVX1-FAST: # %bb.0:
@@ -380,7 +637,7 @@ define i32 @test_v16i32(<16 x i32> %a0) {
; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
; AVX1-FAST-NEXT: vzeroupper
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v16i32:
; AVX2: # %bb.0:
@@ -393,7 +650,7 @@ define i32 @test_v16i32(<16 x i32> %a0) {
; AVX2-NEXT: vpaddd %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:
@@ -412,78 +669,173 @@ define i32 @test_v16i32(<16 x i32> %a0) {
ret i32 %1
}
-define i32 @test_v32i32(<32 x i32> %a0) {
-; SSE-LABEL: test_v32i32:
-; SSE: # %bb.0:
-; SSE-NEXT: paddd %xmm6, %xmm2
-; SSE-NEXT: paddd %xmm4, %xmm0
-; SSE-NEXT: paddd %xmm2, %xmm0
-; SSE-NEXT: paddd %xmm7, %xmm3
-; SSE-NEXT: paddd %xmm5, %xmm1
-; SSE-NEXT: paddd %xmm3, %xmm1
-; SSE-NEXT: paddd %xmm0, %xmm1
-; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE-NEXT: paddd %xmm1, %xmm0
-; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE-NEXT: paddd %xmm0, %xmm1
-; SSE-NEXT: movd %xmm1, %eax
-; SSE-NEXT: retq
-;
-; AVX1-SLOW-LABEL: test_v32i32:
-; AVX1-SLOW: # %bb.0:
-; AVX1-SLOW-NEXT: vpaddd %xmm3, %xmm1, %xmm4
-; AVX1-SLOW-NEXT: vpaddd %xmm2, %xmm0, %xmm5
-; AVX1-SLOW-NEXT: vpaddd %xmm4, %xmm5, %xmm4
-; AVX1-SLOW-NEXT: vextractf128 $1, %ymm3, %xmm3
-; AVX1-SLOW-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-SLOW-NEXT: vpaddd %xmm3, %xmm1, %xmm1
-; AVX1-SLOW-NEXT: vextractf128 $1, %ymm2, %xmm2
-; AVX1-SLOW-NEXT: vextractf128 $1, %ymm0, %xmm0
-; AVX1-SLOW-NEXT: vpaddd %xmm2, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vpaddd %xmm0, %xmm4, %xmm0
-; AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
-; AVX1-SLOW-NEXT: vzeroupper
-; AVX1-SLOW-NEXT: retq
-;
-; AVX1-FAST-LABEL: test_v32i32:
-; AVX1-FAST: # %bb.0:
-; AVX1-FAST-NEXT: vpaddd %xmm3, %xmm1, %xmm4
-; AVX1-FAST-NEXT: vpaddd %xmm2, %xmm0, %xmm5
-; AVX1-FAST-NEXT: vpaddd %xmm4, %xmm5, %xmm4
-; AVX1-FAST-NEXT: vextractf128 $1, %ymm3, %xmm3
-; AVX1-FAST-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-FAST-NEXT: vpaddd %xmm3, %xmm1, %xmm1
-; AVX1-FAST-NEXT: vextractf128 $1, %ymm2, %xmm2
-; AVX1-FAST-NEXT: vextractf128 $1, %ymm0, %xmm0
-; AVX1-FAST-NEXT: vpaddd %xmm2, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vpaddd %xmm0, %xmm4, %xmm0
-; AVX1-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-FAST-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vmovd %xmm0, %eax
-; AVX1-FAST-NEXT: vzeroupper
-; AVX1-FAST-NEXT: retq
-;
-; AVX2-LABEL: test_v32i32:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpaddd %ymm3, %ymm1, %ymm1
-; AVX2-NEXT: vpaddd %ymm2, %ymm0, %ymm0
-; AVX2-NEXT: vpaddd %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+define i32 @test_v32i32(<32 x i32> %a0) nounwind {
+; X86-SSE-LABEL: test_v32i32:
+; 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: paddd 56(%ebp), %xmm2
+; X86-SSE-NEXT: paddd 24(%ebp), %xmm0
+; X86-SSE-NEXT: paddd %xmm2, %xmm0
+; X86-SSE-NEXT: paddd 72(%ebp), %xmm3
+; X86-SSE-NEXT: paddd 40(%ebp), %xmm1
+; X86-SSE-NEXT: paddd %xmm3, %xmm1
+; X86-SSE-NEXT: paddd %xmm0, %xmm1
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE-NEXT: paddd %xmm1, %xmm0
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE-NEXT: paddd %xmm0, %xmm1
+; X86-SSE-NEXT: movd %xmm1, %eax
+; X86-SSE-NEXT: movl %ebp, %esp
+; X86-SSE-NEXT: popl %ebp
+; X86-SSE-NEXT: retl
+;
+; X64-SSE-LABEL: test_v32i32:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: paddd %xmm6, %xmm2
+; X64-SSE-NEXT: paddd %xmm4, %xmm0
+; X64-SSE-NEXT: paddd %xmm2, %xmm0
+; X64-SSE-NEXT: paddd %xmm7, %xmm3
+; X64-SSE-NEXT: paddd %xmm5, %xmm1
+; X64-SSE-NEXT: paddd %xmm3, %xmm1
+; X64-SSE-NEXT: paddd %xmm0, %xmm1
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X64-SSE-NEXT: paddd %xmm1, %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-SSE-NEXT: paddd %xmm0, %xmm1
+; X64-SSE-NEXT: movd %xmm1, %eax
+; X64-SSE-NEXT: retq
+;
+; X86-AVX1-SLOW-LABEL: test_v32i32:
+; X86-AVX1-SLOW: # %bb.0:
+; X86-AVX1-SLOW-NEXT: pushl %ebp
+; X86-AVX1-SLOW-NEXT: movl %esp, %ebp
+; X86-AVX1-SLOW-NEXT: andl $-32, %esp
+; X86-AVX1-SLOW-NEXT: subl $32, %esp
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm2, %xmm0, %xmm3
+; X86-AVX1-SLOW-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X86-AVX1-SLOW-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm2, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vpaddd 8(%ebp), %xmm1, %xmm2
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm2, %xmm3, %xmm2
+; X86-AVX1-SLOW-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X86-AVX1-SLOW-NEXT: vpaddd 24(%ebp), %xmm1, %xmm1
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm0, %xmm2, %xmm0
+; X86-AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-SLOW-NEXT: movl %ebp, %esp
+; X86-AVX1-SLOW-NEXT: popl %ebp
+; X86-AVX1-SLOW-NEXT: vzeroupper
+; X86-AVX1-SLOW-NEXT: retl
+;
+; X64-AVX1-SLOW-LABEL: test_v32i32:
+; X64-AVX1-SLOW: # %bb.0:
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm3, %xmm1, %xmm4
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm2, %xmm0, %xmm5
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm4, %xmm5, %xmm4
+; X64-AVX1-SLOW-NEXT: vextractf128 $1, %ymm3, %xmm3
+; X64-AVX1-SLOW-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm3, %xmm1, %xmm1
+; X64-AVX1-SLOW-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X64-AVX1-SLOW-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm2, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm0, %xmm4, %xmm0
+; X64-AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX1-SLOW-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-SLOW-NEXT: vzeroupper
+; X64-AVX1-SLOW-NEXT: retq
+;
+; X86-AVX1-FAST-LABEL: test_v32i32:
+; X86-AVX1-FAST: # %bb.0:
+; X86-AVX1-FAST-NEXT: pushl %ebp
+; X86-AVX1-FAST-NEXT: movl %esp, %ebp
+; X86-AVX1-FAST-NEXT: andl $-32, %esp
+; X86-AVX1-FAST-NEXT: subl $32, %esp
+; X86-AVX1-FAST-NEXT: vpaddd %xmm2, %xmm0, %xmm3
+; X86-AVX1-FAST-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X86-AVX1-FAST-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X86-AVX1-FAST-NEXT: vpaddd %xmm2, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vpaddd 8(%ebp), %xmm1, %xmm2
+; X86-AVX1-FAST-NEXT: vpaddd %xmm2, %xmm3, %xmm2
+; X86-AVX1-FAST-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X86-AVX1-FAST-NEXT: vpaddd 24(%ebp), %xmm1, %xmm1
+; X86-AVX1-FAST-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vpaddd %xmm0, %xmm2, %xmm0
+; X86-AVX1-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-FAST-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-FAST-NEXT: movl %ebp, %esp
+; X86-AVX1-FAST-NEXT: popl %ebp
+; X86-AVX1-FAST-NEXT: vzeroupper
+; X86-AVX1-FAST-NEXT: retl
+;
+; X64-AVX1-FAST-LABEL: test_v32i32:
+; X64-AVX1-FAST: # %bb.0:
+; X64-AVX1-FAST-NEXT: vpaddd %xmm3, %xmm1, %xmm4
+; X64-AVX1-FAST-NEXT: vpaddd %xmm2, %xmm0, %xmm5
+; X64-AVX1-FAST-NEXT: vpaddd %xmm4, %xmm5, %xmm4
+; X64-AVX1-FAST-NEXT: vextractf128 $1, %ymm3, %xmm3
+; X64-AVX1-FAST-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X64-AVX1-FAST-NEXT: vpaddd %xmm3, %xmm1, %xmm1
+; X64-AVX1-FAST-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X64-AVX1-FAST-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X64-AVX1-FAST-NEXT: vpaddd %xmm2, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vpaddd %xmm0, %xmm4, %xmm0
+; X64-AVX1-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-FAST-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vphaddd %xmm0, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-FAST-NEXT: vzeroupper
+; X64-AVX1-FAST-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: vpaddd %ymm2, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpaddd 8(%ebp), %ymm1, %ymm1
+; X86-AVX2-NEXT: vpaddd %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX2-NEXT: vpaddd %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: vpaddd %ymm3, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpaddd %ymm2, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpaddd %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX2-NEXT: vpaddd %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vmovd %xmm0, %eax
+; X64-AVX2-NEXT: vzeroupper
+; X64-AVX2-NEXT: retq
;
; AVX512-LABEL: test_v32i32:
; AVX512: # %bb.0:
@@ -507,7 +859,7 @@ 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
@@ -515,7 +867,7 @@ define i16 @test_v2i16(<2 x i16> %a0) {
; SSE-NEXT: paddw %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-SLOW-LABEL: test_v2i16:
; AVX1-SLOW: # %bb.0:
@@ -523,14 +875,14 @@ define i16 @test_v2i16(<2 x i16> %a0) {
; AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
; AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v2i16:
; AVX1-FAST: # %bb.0:
; AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
; AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v2i16:
; AVX2: # %bb.0:
@@ -538,7 +890,7 @@ define i16 @test_v2i16(<2 x i16> %a0) {
; AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v2i16:
; AVX512: # %bb.0:
@@ -551,7 +903,7 @@ define i16 @test_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]
@@ -561,7 +913,7 @@ define i16 @test_v4i16(<4 x i16> %a0) {
; SSE-NEXT: paddw %xmm1, %xmm0
; SSE-NEXT: movd %xmm0, %eax
; SSE-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE-NEXT: retq
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX1-SLOW-LABEL: test_v4i16:
; AVX1-SLOW: # %bb.0:
@@ -571,7 +923,7 @@ define i16 @test_v4i16(<4 x i16> %a0) {
; AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
; AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v4i16:
; AVX1-FAST: # %bb.0:
@@ -580,7 +932,7 @@ define i16 @test_v4i16(<4 x i16> %a0) {
; AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
; AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v4i16:
; AVX2: # %bb.0:
@@ -590,7 +942,7 @@ define i16 @test_v4i16(<4 x i16> %a0) {
; AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v4i16:
; AVX512: # %bb.0:
@@ -605,7 +957,7 @@ define i16 @test_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]
@@ -617,7 +969,7 @@ define i16 @test_v8i16(<8 x i16> %a0) {
; SSE-NEXT: paddw %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-SLOW-LABEL: test_v8i16:
; AVX1-SLOW: # %bb.0:
@@ -629,7 +981,7 @@ define i16 @test_v8i16(<8 x i16> %a0) {
; AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
; AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v8i16:
; AVX1-FAST: # %bb.0:
@@ -638,7 +990,7 @@ define i16 @test_v8i16(<8 x i16> %a0) {
; AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
; AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v8i16:
; AVX2: # %bb.0:
@@ -650,7 +1002,7 @@ define i16 @test_v8i16(<8 x i16> %a0) {
; AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v8i16:
; AVX512: # %bb.0:
@@ -667,7 +1019,7 @@ define i16 @test_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: paddw %xmm1, %xmm0
@@ -680,7 +1032,7 @@ define i16 @test_v16i16(<16 x i16> %a0) {
; SSE-NEXT: paddw %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-SLOW-LABEL: test_v16i16:
; AVX1-SLOW: # %bb.0:
@@ -695,7 +1047,7 @@ define i16 @test_v16i16(<16 x i16> %a0) {
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
; AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
; AVX1-SLOW-NEXT: vzeroupper
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v16i16:
; AVX1-FAST: # %bb.0:
@@ -707,7 +1059,7 @@ define i16 @test_v16i16(<16 x i16> %a0) {
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
; AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
; AVX1-FAST-NEXT: vzeroupper
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v16i16:
; AVX2: # %bb.0:
@@ -722,7 +1074,7 @@ 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:
@@ -742,22 +1094,44 @@ 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: paddw %xmm3, %xmm1
-; SSE-NEXT: paddw %xmm2, %xmm0
-; SSE-NEXT: paddw %xmm1, %xmm0
-; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE-NEXT: paddw %xmm0, %xmm1
-; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
-; SSE-NEXT: paddw %xmm1, %xmm0
-; SSE-NEXT: movdqa %xmm0, %xmm1
-; SSE-NEXT: psrld $16, %xmm1
-; SSE-NEXT: paddw %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: paddw %xmm2, %xmm0
+; X86-SSE-NEXT: paddw 8(%ebp), %xmm1
+; X86-SSE-NEXT: paddw %xmm0, %xmm1
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE-NEXT: paddw %xmm1, %xmm0
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE-NEXT: paddw %xmm0, %xmm1
+; X86-SSE-NEXT: movdqa %xmm1, %xmm0
+; X86-SSE-NEXT: psrld $16, %xmm0
+; X86-SSE-NEXT: paddw %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: paddw %xmm3, %xmm1
+; X64-SSE-NEXT: paddw %xmm2, %xmm0
+; X64-SSE-NEXT: paddw %xmm1, %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE-NEXT: paddw %xmm0, %xmm1
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,1,1,1]
+; X64-SSE-NEXT: paddw %xmm1, %xmm0
+; X64-SSE-NEXT: movdqa %xmm0, %xmm1
+; X64-SSE-NEXT: psrld $16, %xmm1
+; X64-SSE-NEXT: paddw %xmm0, %xmm1
+; X64-SSE-NEXT: movd %xmm1, %eax
+; X64-SSE-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-SSE-NEXT: retq
;
; AVX1-SLOW-LABEL: test_v32i16:
; AVX1-SLOW: # %bb.0:
@@ -775,7 +1149,7 @@ define i16 @test_v32i16(<32 x i16> %a0) {
; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
; AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
; AVX1-SLOW-NEXT: vzeroupper
-; AVX1-SLOW-NEXT: retq
+; AVX1-SLOW-NEXT: ret{{[l|q]}}
;
; AVX1-FAST-LABEL: test_v32i16:
; AVX1-FAST: # %bb.0:
@@ -792,7 +1166,7 @@ define i16 @test_v32i16(<32 x i16> %a0) {
; AVX1-FAST-NEXT: vmovd %xmm0, %eax
; AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
; AVX1-FAST-NEXT: vzeroupper
-; AVX1-FAST-NEXT: retq
+; AVX1-FAST-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v32i16:
; AVX2: # %bb.0:
@@ -808,7 +1182,7 @@ 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]}}
;
; AVX512-LABEL: test_v32i16:
; AVX512: # %bb.0:
@@ -830,91 +1204,199 @@ 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: paddw %xmm6, %xmm2
-; SSE-NEXT: paddw %xmm4, %xmm0
-; SSE-NEXT: paddw %xmm2, %xmm0
-; SSE-NEXT: paddw %xmm7, %xmm3
-; SSE-NEXT: paddw %xmm5, %xmm1
-; SSE-NEXT: paddw %xmm3, %xmm1
-; SSE-NEXT: paddw %xmm0, %xmm1
-; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
-; SSE-NEXT: paddw %xmm1, %xmm0
-; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; SSE-NEXT: paddw %xmm0, %xmm1
-; SSE-NEXT: movdqa %xmm1, %xmm0
-; SSE-NEXT: psrld $16, %xmm0
-; SSE-NEXT: paddw %xmm1, %xmm0
-; SSE-NEXT: movd %xmm0, %eax
-; SSE-NEXT: # kill: def $ax killed $ax killed $eax
-; SSE-NEXT: retq
-;
-; AVX1-SLOW-LABEL: test_v64i16:
-; AVX1-SLOW: # %bb.0:
-; AVX1-SLOW-NEXT: vpaddw %xmm3, %xmm1, %xmm4
-; AVX1-SLOW-NEXT: vpaddw %xmm2, %xmm0, %xmm5
-; AVX1-SLOW-NEXT: vpaddw %xmm4, %xmm5, %xmm4
-; AVX1-SLOW-NEXT: vextractf128 $1, %ymm3, %xmm3
-; AVX1-SLOW-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-SLOW-NEXT: vpaddw %xmm3, %xmm1, %xmm1
-; AVX1-SLOW-NEXT: vextractf128 $1, %ymm2, %xmm2
-; AVX1-SLOW-NEXT: vextractf128 $1, %ymm0, %xmm0
-; AVX1-SLOW-NEXT: vpaddw %xmm2, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vpaddw %xmm0, %xmm4, %xmm0
-; AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vpsrld $16, %xmm0, %xmm1
-; AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-SLOW-NEXT: vmovd %xmm0, %eax
-; AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-SLOW-NEXT: vzeroupper
-; AVX1-SLOW-NEXT: retq
-;
-; AVX1-FAST-LABEL: test_v64i16:
-; AVX1-FAST: # %bb.0:
-; AVX1-FAST-NEXT: vpaddw %xmm3, %xmm1, %xmm4
-; AVX1-FAST-NEXT: vpaddw %xmm2, %xmm0, %xmm5
-; AVX1-FAST-NEXT: vpaddw %xmm4, %xmm5, %xmm4
-; AVX1-FAST-NEXT: vextractf128 $1, %ymm3, %xmm3
-; AVX1-FAST-NEXT: vextractf128 $1, %ymm1, %xmm1
-; AVX1-FAST-NEXT: vpaddw %xmm3, %xmm1, %xmm1
-; AVX1-FAST-NEXT: vextractf128 $1, %ymm2, %xmm2
-; AVX1-FAST-NEXT: vextractf128 $1, %ymm0, %xmm0
-; AVX1-FAST-NEXT: vpaddw %xmm2, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vpaddw %xmm0, %xmm4, %xmm0
-; AVX1-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-FAST-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX1-FAST-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
-; AVX1-FAST-NEXT: vmovd %xmm0, %eax
-; AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX1-FAST-NEXT: vzeroupper
-; AVX1-FAST-NEXT: retq
-;
-; AVX2-LABEL: test_v64i16:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpaddw %ymm3, %ymm1, %ymm1
-; AVX2-NEXT: vpaddw %ymm2, %ymm0, %ymm0
-; AVX2-NEXT: vpaddw %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
-; AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
-; AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
-; AVX2-NEXT: vzeroupper
-; AVX2-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: paddw 56(%ebp), %xmm2
+; X86-SSE-NEXT: paddw 24(%ebp), %xmm0
+; X86-SSE-NEXT: paddw %xmm2, %xmm0
+; X86-SSE-NEXT: paddw 72(%ebp), %xmm3
+; X86-SSE-NEXT: paddw 40(%ebp), %xmm1
+; X86-SSE-NEXT: paddw %xmm3, %xmm1
+; X86-SSE-NEXT: paddw %xmm0, %xmm1
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE-NEXT: paddw %xmm1, %xmm0
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-SSE-NEXT: paddw %xmm0, %xmm1
+; X86-SSE-NEXT: movdqa %xmm1, %xmm0
+; X86-SSE-NEXT: psrld $16, %xmm0
+; X86-SSE-NEXT: paddw %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_v64i16:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: paddw %xmm6, %xmm2
+; X64-SSE-NEXT: paddw %xmm4, %xmm0
+; X64-SSE-NEXT: paddw %xmm2, %xmm0
+; X64-SSE-NEXT: paddw %xmm7, %xmm3
+; X64-SSE-NEXT: paddw %xmm5, %xmm1
+; X64-SSE-NEXT: paddw %xmm3, %xmm1
+; X64-SSE-NEXT: paddw %xmm0, %xmm1
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X64-SSE-NEXT: paddw %xmm1, %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-SSE-NEXT: paddw %xmm0, %xmm1
+; X64-SSE-NEXT: movdqa %xmm1, %xmm0
+; X64-SSE-NEXT: psrld $16, %xmm0
+; X64-SSE-NEXT: paddw %xmm1, %xmm0
+; X64-SSE-NEXT: movd %xmm0, %eax
+; X64-SSE-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-SSE-NEXT: retq
+;
+; X86-AVX1-SLOW-LABEL: test_v64i16:
+; X86-AVX1-SLOW: # %bb.0:
+; X86-AVX1-SLOW-NEXT: pushl %ebp
+; X86-AVX1-SLOW-NEXT: movl %esp, %ebp
+; X86-AVX1-SLOW-NEXT: andl $-32, %esp
+; X86-AVX1-SLOW-NEXT: subl $32, %esp
+; X86-AVX1-SLOW-NEXT: vpaddw %xmm2, %xmm0, %xmm3
+; X86-AVX1-SLOW-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X86-AVX1-SLOW-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vpaddw %xmm2, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vpaddw 8(%ebp), %xmm1, %xmm2
+; X86-AVX1-SLOW-NEXT: vpaddw %xmm2, %xmm3, %xmm2
+; X86-AVX1-SLOW-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X86-AVX1-SLOW-NEXT: vpaddw 24(%ebp), %xmm1, %xmm1
+; X86-AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vpaddw %xmm0, %xmm2, %xmm0
+; X86-AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vpsrld $16, %xmm0, %xmm1
+; X86-AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X86-AVX1-SLOW-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-AVX1-SLOW-NEXT: movl %ebp, %esp
+; X86-AVX1-SLOW-NEXT: popl %ebp
+; X86-AVX1-SLOW-NEXT: vzeroupper
+; X86-AVX1-SLOW-NEXT: retl
+;
+; X64-AVX1-SLOW-LABEL: test_v64i16:
+; X64-AVX1-SLOW: # %bb.0:
+; X64-AVX1-SLOW-NEXT: vpaddw %xmm3, %xmm1, %xmm4
+; X64-AVX1-SLOW-NEXT: vpaddw %xmm2, %xmm0, %xmm5
+; X64-AVX1-SLOW-NEXT: vpaddw %xmm4, %xmm5, %xmm4
+; X64-AVX1-SLOW-NEXT: vextractf128 $1, %ymm3, %xmm3
+; X64-AVX1-SLOW-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X64-AVX1-SLOW-NEXT: vpaddw %xmm3, %xmm1, %xmm1
+; X64-AVX1-SLOW-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X64-AVX1-SLOW-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vpaddw %xmm2, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vpaddw %xmm0, %xmm4, %xmm0
+; X64-AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vpsrld $16, %xmm0, %xmm1
+; X64-AVX1-SLOW-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X64-AVX1-SLOW-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-SLOW-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-AVX1-SLOW-NEXT: vzeroupper
+; X64-AVX1-SLOW-NEXT: retq
+;
+; X86-AVX1-FAST-LABEL: test_v64i16:
+; X86-AVX1-FAST: # %bb.0:
+; X86-AVX1-FAST-NEXT: pushl %ebp
+; X86-AVX1-FAST-NEXT: movl %esp, %ebp
+; X86-AVX1-FAST-NEXT: andl $-32, %esp
+; X86-AVX1-FAST-NEXT: subl $32, %esp
+; X86-AVX1-FAST-NEXT: vpaddw %xmm2, %xmm0, %xmm3
+; X86-AVX1-FAST-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X86-AVX1-FAST-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X86-AVX1-FAST-NEXT: vpaddw %xmm2, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vpaddw 8(%ebp), %xmm1, %xmm2
+; X86-AVX1-FAST-NEXT: vpaddw %xmm2, %xmm3, %xmm2
+; X86-AVX1-FAST-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X86-AVX1-FAST-NEXT: vpaddw 24(%ebp), %xmm1, %xmm1
+; X86-AVX1-FAST-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vpaddw %xmm0, %xmm2, %xmm0
+; X86-AVX1-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-FAST-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX1-FAST-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
+; X86-AVX1-FAST-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
+; X86-AVX1-FAST-NEXT: movl %ebp, %esp
+; X86-AVX1-FAST-NEXT: popl %ebp
+; X86-AVX1-FAST-NEXT: vzeroupper
+; X86-AVX1-FAST-NEXT: retl
+;
+; X64-AVX1-FAST-LABEL: test_v64i16:
+; X64-AVX1-FAST: # %bb.0:
+; X64-AVX1-FAST-NEXT: vpaddw %xmm3, %xmm1, %xmm4
+; X64-AVX1-FAST-NEXT: vpaddw %xmm2, %xmm0, %xmm5
+; X64-AVX1-FAST-NEXT: vpaddw %xmm4, %xmm5, %xmm4
+; X64-AVX1-FAST-NEXT: vextractf128 $1, %ymm3, %xmm3
+; X64-AVX1-FAST-NEXT: vextractf128 $1, %ymm1, %xmm1
+; X64-AVX1-FAST-NEXT: vpaddw %xmm3, %xmm1, %xmm1
+; X64-AVX1-FAST-NEXT: vextractf128 $1, %ymm2, %xmm2
+; X64-AVX1-FAST-NEXT: vextractf128 $1, %ymm0, %xmm0
+; X64-AVX1-FAST-NEXT: vpaddw %xmm2, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vpaddw %xmm0, %xmm4, %xmm0
+; X64-AVX1-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-FAST-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX1-FAST-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vphaddw %xmm0, %xmm0, %xmm0
+; X64-AVX1-FAST-NEXT: vmovd %xmm0, %eax
+; X64-AVX1-FAST-NEXT: # kill: def $ax killed $ax killed $eax
+; X64-AVX1-FAST-NEXT: vzeroupper
+; X64-AVX1-FAST-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: vpaddw %ymm2, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpaddw 8(%ebp), %ymm1, %ymm1
+; X86-AVX2-NEXT: vpaddw %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X86-AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
+; X86-AVX2-NEXT: vpaddw %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: vpaddw %ymm3, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpaddw %ymm2, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpaddw %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,1,1]
+; X64-AVX2-NEXT: vpaddw %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpsrld $16, %xmm0, %xmm1
+; X64-AVX2-NEXT: vpaddw %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
;
; AVX512-LABEL: test_v64i16:
; AVX512: # %bb.0:
@@ -941,7 +1423,7 @@ 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
@@ -949,7 +1431,7 @@ define i8 @test_v2i8(<2 x i8> %a0) {
; SSE-NEXT: paddb %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:
@@ -957,7 +1439,7 @@ define i8 @test_v2i8(<2 x i8> %a0) {
; AVX-NEXT: vpaddb %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]}}
;
; AVX512-LABEL: test_v2i8:
; AVX512: # %bb.0:
@@ -970,27 +1452,71 @@ define i8 @test_v2i8(<2 x i8> %a0) {
ret i8 %1
}
-define i8 @test_v2i8_load(ptr %p) {
-; SSE-LABEL: test_v2i8_load:
-; SSE: # %bb.0:
-; SSE-NEXT: movzwl (%rdi), %eax
-; SSE-NEXT: movd %eax, %xmm0
-; SSE-NEXT: movdqa %xmm0, %xmm1
-; SSE-NEXT: psrlw $8, %xmm1
-; SSE-NEXT: paddb %xmm0, %xmm1
-; SSE-NEXT: movd %xmm1, %eax
-; SSE-NEXT: # kill: def $al killed $al killed $eax
-; SSE-NEXT: retq
-;
-; AVX-LABEL: test_v2i8_load:
-; AVX: # %bb.0:
-; AVX-NEXT: movzwl (%rdi), %eax
-; AVX-NEXT: vmovd %eax, %xmm0
-; AVX-NEXT: vpsrlw $8, %xmm0, %xmm1
-; AVX-NEXT: vpaddb %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vmovd %xmm0, %eax
-; AVX-NEXT: # kill: def $al killed $al killed $eax
-; AVX-NEXT: retq
+define i8 @test_v2i8_load(ptr %p) nounwind {
+; X86-SSE-LABEL: test_v2i8_load:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-SSE-NEXT: movzwl (%eax), %eax
+; X86-SSE-NEXT: movd %eax, %xmm0
+; X86-SSE-NEXT: movdqa %xmm0, %xmm1
+; X86-SSE-NEXT: psrlw $8, %xmm1
+; X86-SSE-NEXT: paddb %xmm0, %xmm1
+; X86-SSE-NEXT: movd %xmm1, %eax
+; X86-SSE-NEXT: # kill: def $al killed $al killed $eax
+; X86-SSE-NEXT: retl
+;
+; X64-SSE-LABEL: test_v2i8_load:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: movzwl (%rdi), %eax
+; X64-SSE-NEXT: movd %eax, %xmm0
+; X64-SSE-NEXT: movdqa %xmm0, %xmm1
+; X64-SSE-NEXT: psrlw $8, %xmm1
+; X64-SSE-NEXT: paddb %xmm0, %xmm1
+; X64-SSE-NEXT: movd %xmm1, %eax
+; X64-SSE-NEXT: # kill: def $al killed $al killed $eax
+; X64-SSE-NEXT: retq
+;
+; X86-AVX1-LABEL: test_v2i8_load:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX1-NEXT: movzwl (%eax), %eax
+; X86-AVX1-NEXT: vmovd %eax, %xmm0
+; X86-AVX1-NEXT: vpsrlw $8, %xmm0, %xmm1
+; X86-AVX1-NEXT: vpaddb %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: # kill: def $al killed $al killed $eax
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v2i8_load:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: movzwl (%rdi), %eax
+; X64-AVX1-NEXT: vmovd %eax, %xmm0
+; X64-AVX1-NEXT: vpsrlw $8, %xmm0, %xmm1
+; X64-AVX1-NEXT: vpaddb %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: test_v2i8_load:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX2-NEXT: movzwl (%eax), %eax
+; X86-AVX2-NEXT: vmovd %eax, %xmm0
+; X86-AVX2-NEXT: vpsrlw $8, %xmm0, %xmm1
+; X86-AVX2-NEXT: vpaddb %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: # kill: def $al killed $al killed $eax
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v2i8_load:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: movzwl (%rdi), %eax
+; X64-AVX2-NEXT: vmovd %eax, %xmm0
+; X64-AVX2-NEXT: vpsrlw $8, %xmm0, %xmm1
+; X64-AVX2-NEXT: vpaddb %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: test_v2i8_load:
; AVX512: # %bb.0:
@@ -1006,7 +1532,7 @@ define i8 @test_v2i8_load(ptr %p) {
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: pxor %xmm1, %xmm1
@@ -1014,16 +1540,16 @@ define i8 @test_v4i8(<4 x i8> %a0) {
; SSE2-NEXT: psadbw %xmm1, %xmm0
; SSE2-NEXT: movd %xmm0, %eax
; SSE2-NEXT: # kill: def $al killed $al killed $eax
-; SSE2-NEXT: retq
+; SSE2-NEXT: ret{{[l|q]}}
;
-; SSE41-LABEL: test_v4i8:
-; SSE41: # %bb.0:
-; SSE41-NEXT: pxor %xmm1, %xmm1
-; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3,4,5,6,7]
-; SSE41-NEXT: psadbw %xmm1, %xmm0
-; SSE41-NEXT: movd %xmm0, %eax
-; SSE41-NEXT: # kill: def $al killed $al killed $eax
-; SSE41-NEXT: retq
+; SSE4-LABEL: test_v4i8:
+; SSE4: # %bb.0:
+; SSE4-NEXT: pxor %xmm1, %xmm1
+; SSE4-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0,1],xmm1[2,3,4,5,6,7]
+; SSE4-NEXT: psadbw %xmm1, %xmm0
+; SSE4-NEXT: movd %xmm0, %eax
+; SSE4-NEXT: # kill: def $al killed $al killed $eax
+; SSE4-NEXT: ret{{[l|q]}}
;
; AVX1-LABEL: test_v4i8:
; AVX1: # %bb.0:
@@ -1032,7 +1558,7 @@ define i8 @test_v4i8(<4 x i8> %a0) {
; AVX1-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
; AVX1-NEXT: vmovd %xmm0, %eax
; AVX1-NEXT: # kill: def $al killed $al killed $eax
-; AVX1-NEXT: retq
+; AVX1-NEXT: ret{{[l|q]}}
;
; AVX2-LABEL: test_v4i8:
; AVX2: # %bb.0:
@@ -1041,7 +1567,7 @@ define i8 @test_v4i8(<4 x i8> %a0) {
; AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
; AVX2-NEXT: vmovd %xmm0, %eax
; AVX2-NEXT: # kill: def $al killed $al killed $eax
-; AVX2-NEXT: retq
+; AVX2-NEXT: ret{{[l|q]}}
;
; AVX512-LABEL: test_v4i8:
; AVX512: # %bb.0:
@@ -1056,24 +1582,63 @@ define i8 @test_v4i8(<4 x i8> %a0) {
ret i8 %1
}
-define i8 @test_v4i8_load(ptr %p) {
-; SSE-LABEL: test_v4i8_load:
-; SSE: # %bb.0:
-; SSE-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; SSE-NEXT: pxor %xmm1, %xmm1
-; SSE-NEXT: psadbw %xmm0, %xmm1
-; SSE-NEXT: movd %xmm1, %eax
-; SSE-NEXT: # kill: def $al killed $al killed $eax
-; SSE-NEXT: retq
-;
-; AVX-LABEL: test_v4i8_load:
-; AVX: # %bb.0:
-; AVX-NEXT: vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vmovd %xmm0, %eax
-; AVX-NEXT: # kill: def $al killed $al killed $eax
-; AVX-NEXT: retq
+define i8 @test_v4i8_load(ptr %p) nounwind {
+; X86-SSE-LABEL: test_v4i8_load:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-SSE-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X86-SSE-NEXT: pxor %xmm1, %xmm1
+; X86-SSE-NEXT: psadbw %xmm0, %xmm1
+; X86-SSE-NEXT: movd %xmm1, %eax
+; X86-SSE-NEXT: # kill: def $al killed $al killed $eax
+; X86-SSE-NEXT: retl
+;
+; X64-SSE-LABEL: test_v4i8_load:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X64-SSE-NEXT: pxor %xmm1, %xmm1
+; X64-SSE-NEXT: psadbw %xmm0, %xmm1
+; X64-SSE-NEXT: movd %xmm1, %eax
+; X64-SSE-NEXT: # kill: def $al killed $al killed $eax
+; X64-SSE-NEXT: retq
+;
+; X86-AVX1-LABEL: test_v4i8_load:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX1-NEXT: vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X86-AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: # kill: def $al killed $al killed $eax
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v4i8_load:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X64-AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpsadbw %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: test_v4i8_load:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX2-NEXT: vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X86-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X86-AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: # kill: def $al killed $al killed $eax
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v4i8_load:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X64-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X64-AVX2-NEXT: vpsadbw %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: test_v4i8_load:
; AVX512: # %bb.0:
@@ -1088,14 +1653,14 @@ define i8 @test_v4i8_load(ptr %p) {
ret i8 %1
}
-define i8 @test_v8i8(<8 x i8> %a0) {
+define i8 @test_v8i8(<8 x i8> %a0) nounwind {
; SSE-LABEL: test_v8i8:
; SSE: # %bb.0:
; SSE-NEXT: pxor %xmm1, %xmm1
; SSE-NEXT: psadbw %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_v8i8:
; AVX: # %bb.0:
@@ -1103,7 +1668,7 @@ define i8 @test_v8i8(<8 x i8> %a0) {
; AVX-NEXT: vpsadbw %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]}}
;
; AVX512-LABEL: test_v8i8:
; AVX512: # %bb.0:
@@ -1116,24 +1681,63 @@ define i8 @test_v8i8(<8 x i8> %a0) {
ret i8 %1
}
-define i8 @test_v8i8_load(ptr %p) {
-; SSE-LABEL: test_v8i8_load:
-; SSE: # %bb.0:
-; SSE-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
-; SSE-NEXT: pxor %xmm1, %xmm1
-; SSE-NEXT: psadbw %xmm0, %xmm1
-; SSE-NEXT: movd %xmm1, %eax
-; SSE-NEXT: # kill: def $al killed $al killed $eax
-; SSE-NEXT: retq
-;
-; AVX-LABEL: test_v8i8_load:
-; AVX: # %bb.0:
-; AVX-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
-; AVX-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
-; AVX-NEXT: vmovd %xmm0, %eax
-; AVX-NEXT: # kill: def $al killed $al killed $eax
-; AVX-NEXT: retq
+define i8 @test_v8i8_load(ptr %p) nounwind {
+; X86-SSE-LABEL: test_v8i8_load:
+; X86-SSE: # %bb.0:
+; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-SSE-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
+; X86-SSE-NEXT: pxor %xmm1, %xmm1
+; X86-SSE-NEXT: psadbw %xmm0, %xmm1
+; X86-SSE-NEXT: movd %xmm1, %eax
+; X86-SSE-NEXT: # kill: def $al killed $al killed $eax
+; X86-SSE-NEXT: retl
+;
+; X64-SSE-LABEL: test_v8i8_load:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
+; X64-SSE-NEXT: pxor %xmm1, %xmm1
+; X64-SSE-NEXT: psadbw %xmm0, %xmm1
+; X64-SSE-NEXT: movd %xmm1, %eax
+; X64-SSE-NEXT: # kill: def $al killed $al killed $eax
+; X64-SSE-NEXT: retq
+;
+; X86-AVX1-LABEL: test_v8i8_load:
+; X86-AVX1: # %bb.0:
+; X86-AVX1-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX1-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
+; X86-AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vmovd %xmm0, %eax
+; X86-AVX1-NEXT: # kill: def $al killed $al killed $eax
+; X86-AVX1-NEXT: retl
+;
+; X64-AVX1-LABEL: test_v8i8_load:
+; X64-AVX1: # %bb.0:
+; X64-AVX1-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
+; X64-AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpsadbw %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: test_v8i8_load:
+; X86-AVX2: # %bb.0:
+; X86-AVX2-NEXT: movl {{[0-9]+}}(%esp), %eax
+; X86-AVX2-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
+; X86-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X86-AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vmovd %xmm0, %eax
+; X86-AVX2-NEXT: # kill: def $al killed $al killed $eax
+; X86-AVX2-NEXT: retl
+;
+; X64-AVX2-LABEL: test_v8i8_load:
+; X64-AVX2: # %bb.0:
+; X64-AVX2-NEXT: vmovq {{.*#+}} xmm0 = mem[0],zero
+; X64-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X64-AVX2-NEXT: vpsadbw %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: test_v8i8_load:
; AVX512: # %bb.0:
@@ -1148,7 +1752,7 @@ define i8 @test_v8i8_load(ptr %p) {
ret i8 %1
}
-define i8 @test_v16i8(<16 x i8> %a0) {
+define i8 @test_v16i8(<16 x i8> %a0) nounwind {
; SSE-LABEL: test_v16i8:
; SSE: # %bb.0:
; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
@@ -1157,7 +1761,7 @@ define i8 @test_v16i8(<16 x i8> %a0) {
; SSE-NEXT: psadbw %xmm1, %xmm0
; SSE-NEXT: movd %xmm0, %eax
; SSE-NEXT: # kill: def $al killed $al killed $eax
-; SSE-NEXT: retq
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX-LABEL: test_v16i8:
; AVX: # %bb.0:
@@ -1167,7 +1771,7 @@ define i8 @test_v16i8(<16 x i8> %a0) {
; AVX-NEXT: vpsadbw %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]}}
;
; AVX512-LABEL: test_v16i8:
; AVX512: # %bb.0:
@@ -1182,7 +1786,7 @@ define i8 @test_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 {
; SSE-LABEL: test_v32i8:
; SSE: # %bb.0:
; SSE-NEXT: paddb %xmm1, %xmm0
@@ -1192,7 +1796,7 @@ define i8 @test_v32i8(<32 x i8> %a0) {
; SSE-NEXT: psadbw %xmm1, %xmm0
; SSE-NEXT: movd %xmm0, %eax
; SSE-NEXT: # kill: def $al killed $al killed $eax
-; SSE-NEXT: retq
+; SSE-NEXT: ret{{[l|q]}}
;
; AVX1-LABEL: test_v32i8:
; AVX1: # %bb.0:
@@ -1205,7 +1809,7 @@ 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:
@@ -1218,7 +1822,7 @@ 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:
@@ -1236,19 +1840,38 @@ define i8 @test_v32i8(<32 x i8> %a0) {
ret i8 %1
}
-define i8 @test_v64i8(<64 x i8> %a0) {
-; SSE-LABEL: test_v64i8:
-; SSE: # %bb.0:
-; SSE-NEXT: paddb %xmm3, %xmm1
-; SSE-NEXT: paddb %xmm2, %xmm0
-; SSE-NEXT: paddb %xmm1, %xmm0
-; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE-NEXT: paddb %xmm0, %xmm1
-; SSE-NEXT: pxor %xmm0, %xmm0
-; SSE-NEXT: psadbw %xmm1, %xmm0
-; SSE-NEXT: movd %xmm0, %eax
-; SSE-NEXT: # kill: def $al killed $al killed $eax
-; SSE-NEXT: retq
+define i8 @test_v64i8(<64 x i8> %a0) nounwind {
+; X86-SSE-LABEL: test_v64i8:
+; 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: paddb %xmm2, %xmm0
+; X86-SSE-NEXT: paddb 8(%ebp), %xmm1
+; X86-SSE-NEXT: paddb %xmm0, %xmm1
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; X86-SSE-NEXT: paddb %xmm1, %xmm0
+; X86-SSE-NEXT: pxor %xmm1, %xmm1
+; X86-SSE-NEXT: psadbw %xmm0, %xmm1
+; X86-SSE-NEXT: movd %xmm1, %eax
+; X86-SSE-NEXT: # kill: def $al killed $al killed $eax
+; X86-SSE-NEXT: movl %ebp, %esp
+; X86-SSE-NEXT: popl %ebp
+; X86-SSE-NEXT: retl
+;
+; X64-SSE-LABEL: test_v64i8:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: paddb %xmm3, %xmm1
+; X64-SSE-NEXT: paddb %xmm2, %xmm0
+; X64-SSE-NEXT: paddb %xmm1, %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE-NEXT: paddb %xmm0, %xmm1
+; X64-SSE-NEXT: pxor %xmm0, %xmm0
+; X64-SSE-NEXT: psadbw %xmm1, %xmm0
+; X64-SSE-NEXT: movd %xmm0, %eax
+; X64-SSE-NEXT: # kill: def $al killed $al killed $eax
+; X64-SSE-NEXT: retq
;
; AVX1-LABEL: test_v64i8:
; AVX1: # %bb.0:
@@ -1264,7 +1887,7 @@ 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:
@@ -1278,7 +1901,7 @@ 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]}}
;
; AVX512-LABEL: test_v64i8:
; AVX512: # %bb.0:
@@ -1298,61 +1921,134 @@ define i8 @test_v64i8(<64 x i8> %a0) {
ret i8 %1
}
-define i8 @test_v128i8(<128 x i8> %a0) {
-; SSE-LABEL: test_v128i8:
-; SSE: # %bb.0:
-; SSE-NEXT: paddb %xmm7, %xmm3
-; SSE-NEXT: paddb %xmm5, %xmm1
-; SSE-NEXT: paddb %xmm3, %xmm1
-; SSE-NEXT: paddb %xmm6, %xmm2
-; SSE-NEXT: paddb %xmm4, %xmm0
-; SSE-NEXT: paddb %xmm2, %xmm0
-; SSE-NEXT: paddb %xmm1, %xmm0
-; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; SSE-NEXT: paddb %xmm0, %xmm1
-; SSE-NEXT: pxor %xmm0, %xmm0
-; SSE-NEXT: psadbw %xmm1, %xmm0
-; SSE-NEXT: movd %xmm0, %eax
-; SSE-NEXT: # kill: def $al killed $al killed $eax
-; SSE-NEXT: retq
-;
-; AVX1-LABEL: test_v128i8:
-; AVX1: # %bb.0:
-; AVX1-NEXT: vextractf128 $1, %ymm3, %xmm4
-; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm5
-; AVX1-NEXT: vpaddb %xmm4, %xmm5, %xmm4
-; AVX1-NEXT: vextractf128 $1, %ymm2, %xmm5
-; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm6
-; AVX1-NEXT: vpaddb %xmm5, %xmm6, %xmm5
-; AVX1-NEXT: vpaddb %xmm4, %xmm5, %xmm4
-; AVX1-NEXT: vpaddb %xmm3, %xmm1, %xmm1
-; AVX1-NEXT: vpaddb %xmm2, %xmm0, %xmm0
-; AVX1-NEXT: vpaddb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpaddb %xmm4, %xmm0, %xmm0
-; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX1-NEXT: vpaddb %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX1-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
-; AVX1-NEXT: vmovd %xmm0, %eax
-; AVX1-NEXT: # kill: def $al killed $al killed $eax
-; AVX1-NEXT: vzeroupper
-; AVX1-NEXT: retq
-;
-; AVX2-LABEL: test_v128i8:
-; AVX2: # %bb.0:
-; AVX2-NEXT: vpaddb %ymm3, %ymm1, %ymm1
-; AVX2-NEXT: vpaddb %ymm2, %ymm0, %ymm0
-; AVX2-NEXT: vpaddb %ymm1, %ymm0, %ymm0
-; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
-; AVX2-NEXT: vpaddb %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
-; AVX2-NEXT: vpaddb %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
-; AVX2-NEXT: vpsadbw %xmm1, %xmm0, %xmm0
-; AVX2-NEXT: vmovd %xmm0, %eax
-; AVX2-NEXT: # kill: def $al killed $al killed $eax
-; AVX2-NEXT: vzeroupper
-; AVX2-NEXT: retq
+define i8 @test_v128i8(<128 x i8> %a0) nounwind {
+; X86-SSE-LABEL: test_v128i8:
+; 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: paddb 72(%ebp), %xmm3
+; X86-SSE-NEXT: paddb 40(%ebp), %xmm1
+; X86-SSE-NEXT: paddb %xmm3, %xmm1
+; X86-SSE-NEXT: paddb 56(%ebp), %xmm2
+; X86-SSE-NEXT: paddb 24(%ebp), %xmm0
+; X86-SSE-NEXT: paddb %xmm2, %xmm0
+; X86-SSE-NEXT: paddb %xmm1, %xmm0
+; X86-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-SSE-NEXT: paddb %xmm0, %xmm1
+; X86-SSE-NEXT: pxor %xmm0, %xmm0
+; X86-SSE-NEXT: psadbw %xmm1, %xmm0
+; X86-SSE-NEXT: movd %xmm0, %eax
+; X86-SSE-NEXT: # kill: def $al killed $al killed $eax
+; X86-SSE-NEXT: movl %ebp, %esp
+; X86-SSE-NEXT: popl %ebp
+; X86-SSE-NEXT: retl
+;
+; X64-SSE-LABEL: test_v128i8:
+; X64-SSE: # %bb.0:
+; X64-SSE-NEXT: paddb %xmm7, %xmm3
+; X64-SSE-NEXT: paddb %xmm5, %xmm1
+; X64-SSE-NEXT: paddb %xmm3, %xmm1
+; X64-SSE-NEXT: paddb %xmm6, %xmm2
+; X64-SSE-NEXT: paddb %xmm4, %xmm0
+; X64-SSE-NEXT: paddb %xmm2, %xmm0
+; X64-SSE-NEXT: paddb %xmm1, %xmm0
+; X64-SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-SSE-NEXT: paddb %xmm0, %xmm1
+; X64-SSE-NEXT: pxor %xmm0, %xmm0
+; X64-SSE-NEXT: psadbw %xmm1, %xmm0
+; X64-SSE-NEXT: movd %xmm0, %eax
+; X64-SSE-NEXT: # kill: def $al killed $al killed $eax
+; X64-SSE-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: vextractf128 $1, %ymm2, %xmm3
+; X86-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm4
+; X86-AVX1-NEXT: vpaddb %xmm3, %xmm4, %xmm3
+; X86-AVX1-NEXT: vextractf128 $1, %ymm1, %xmm4
+; X86-AVX1-NEXT: vpaddb 24(%ebp), %xmm4, %xmm4
+; X86-AVX1-NEXT: vpaddb %xmm4, %xmm3, %xmm3
+; X86-AVX1-NEXT: vpaddb %xmm2, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddb 8(%ebp), %xmm1, %xmm1
+; X86-AVX1-NEXT: vpaddb %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpaddb %xmm3, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX1-NEXT: vpaddb %xmm1, %xmm0, %xmm0
+; X86-AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X86-AVX1-NEXT: vpsadbw %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: vextractf128 $1, %ymm1, %xmm5
+; X64-AVX1-NEXT: vpaddb %xmm4, %xmm5, %xmm4
+; X64-AVX1-NEXT: vextractf128 $1, %ymm2, %xmm5
+; X64-AVX1-NEXT: vextractf128 $1, %ymm0, %xmm6
+; X64-AVX1-NEXT: vpaddb %xmm5, %xmm6, %xmm5
+; X64-AVX1-NEXT: vpaddb %xmm4, %xmm5, %xmm4
+; X64-AVX1-NEXT: vpaddb %xmm3, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpaddb %xmm2, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddb %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpaddb %xmm4, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX1-NEXT: vpaddb %xmm1, %xmm0, %xmm0
+; X64-AVX1-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X64-AVX1-NEXT: vpsadbw %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: vpaddb %ymm2, %ymm0, %ymm0
+; X86-AVX2-NEXT: vpaddb 8(%ebp), %ymm1, %ymm1
+; X86-AVX2-NEXT: vpaddb %ymm1, %ymm0, %ymm0
+; X86-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X86-AVX2-NEXT: vpaddb %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X86-AVX2-NEXT: vpaddb %xmm1, %xmm0, %xmm0
+; X86-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X86-AVX2-NEXT: vpsadbw %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: vpaddb %ymm3, %ymm1, %ymm1
+; X64-AVX2-NEXT: vpaddb %ymm2, %ymm0, %ymm0
+; X64-AVX2-NEXT: vpaddb %ymm1, %ymm0, %ymm0
+; X64-AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
+; X64-AVX2-NEXT: vpaddb %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[2,3,2,3]
+; X64-AVX2-NEXT: vpaddb %xmm1, %xmm0, %xmm0
+; X64-AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
+; X64-AVX2-NEXT: vpsadbw %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
;
; AVX512-LABEL: test_v128i8:
; AVX512: # %bb.0:
@@ -1398,3 +2094,6 @@ declare i8 @llvm.vector.reduce.add.v16i8(<16 x i8>)
declare i8 @llvm.vector.reduce.add.v32i8(<32 x i8>)
declare i8 @llvm.vector.reduce.add.v64i8(<64 x i8>)
declare i8 @llvm.vector.reduce.add.v128i8(<128 x i8>)
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; X64-SSE2: {{.*}}
+; X64-SSE4: {{.*}}
More information about the llvm-commits
mailing list