[X86][AVX] Improved AVX512F/AVX512VL SubVectorBroadcast tests

llvm-svn: 279736

diff --git a/llvm/test/CodeGen/X86/subvector-broadcast.ll b/llvm/test/CodeGen/X86/subvector-broadcast.ll
index 01c63dc..fce8a3e 100644 (file)
--- a/llvm/test/CodeGen/X86/subvector-broadcast.ll
+++ b/llvm/test/CodeGen/X86/subvector-broadcast.ll
@@ -1,12 +1,12 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+avx | FileCheck %s --check-prefix=X32 --check-prefix=X32-AVX --check-prefix=X32-AVX1
 ; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+avx2 | FileCheck %s --check-prefix=X32 --check-prefix=X32-AVX --check-prefix=X32-AVX2
-; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+avx512vl | FileCheck %s --check-prefix=X32 --check-prefix=X32-AVX512
+; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+avx512vl | FileCheck %s --check-prefix=X32 --check-prefix=X32-AVX512 --check-prefix=X32-AVX512F
 ; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+avx512bw,+avx512vl | FileCheck %s --check-prefix=X32 --check-prefix=X32-AVX512 --check-prefix=X32-AVX512BW
 ; RUN: llc < %s -mtriple=i686-apple-darwin -mattr=+avx512dq,+avx512vl | FileCheck %s --check-prefix=X32 --check-prefix=X32-AVX512 --check-prefix=X32-AVX512DQ
 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx | FileCheck %s --check-prefix=X64 --check-prefix=X64-AVX --check-prefix=X64-AVX1
 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx2 | FileCheck %s --check-prefix=X64 --check-prefix=X64-AVX --check-prefix=X64-AVX2
-; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512vl | FileCheck %s --check-prefix=X64 --check-prefix=X64-AVX512
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512vl | FileCheck %s --check-prefix=X64 --check-prefix=X64-AVX512 --check-prefix=X64-AVX512F
 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512bw,+avx512vl | FileCheck %s --check-prefix=X64 --check-prefix=X64-AVX512 --check-prefix=X64-AVX512BW
 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512dq,+avx512vl | FileCheck %s --check-prefix=X64 --check-prefix=X64-AVX512 --check-prefix=X64-AVX512DQ
 
@@ -21,6 +21,12 @@ define <4 x double> @test_broadcast_2f64_4f64(<2 x double> *%p) nounwind {
 ; X32-AVX-NEXT:    vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1]
 ; X32-AVX-NEXT:    retl
 ;
+; X32-AVX512F-LABEL: test_broadcast_2f64_4f64:
+; X32-AVX512F:       ## BB#0:
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX512F-NEXT:    vbroadcastf32x4 (%eax), %ymm0
+; X32-AVX512F-NEXT:    retl
+;
 ; X32-AVX512BW-LABEL: test_broadcast_2f64_4f64:
 ; X32-AVX512BW:       ## BB#0:
 ; X32-AVX512BW-NEXT:    movl {{[0-9]+}}(%esp), %eax
@@ -38,6 +44,11 @@ define <4 x double> @test_broadcast_2f64_4f64(<2 x double> *%p) nounwind {
 ; X64-AVX-NEXT:    vbroadcastf128 {{.*#+}} ymm0 = mem[0,1,0,1]
 ; X64-AVX-NEXT:    retq
 ;
+; X64-AVX512F-LABEL: test_broadcast_2f64_4f64:
+; X64-AVX512F:       ## BB#0:
+; X64-AVX512F-NEXT:    vbroadcastf32x4 (%rdi), %ymm0
+; X64-AVX512F-NEXT:    retq
+;
 ; X64-AVX512BW-LABEL: test_broadcast_2f64_4f64:
 ; X64-AVX512BW:       ## BB#0:
 ; X64-AVX512BW-NEXT:    vbroadcastf32x4 (%rdi), %ymm0
@@ -60,6 +71,13 @@ define <8 x double> @test_broadcast_2f64_8f64(<2 x double> *%p) nounwind {
 ; X32-AVX-NEXT:    vmovdqa %ymm0, %ymm1
 ; X32-AVX-NEXT:    retl
 ;
+; X32-AVX512F-LABEL: test_broadcast_2f64_8f64:
+; X32-AVX512F:       ## BB#0:
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX512F-NEXT:    vbroadcastf32x4 (%eax), %ymm0
+; X32-AVX512F-NEXT:    vinsertf64x4 $1, %ymm0, %zmm0, %zmm0
+; X32-AVX512F-NEXT:    retl
+;
 ; X32-AVX512BW-LABEL: test_broadcast_2f64_8f64:
 ; X32-AVX512BW:       ## BB#0:
 ; X32-AVX512BW-NEXT:    movl {{[0-9]+}}(%esp), %eax
@@ -80,6 +98,12 @@ define <8 x double> @test_broadcast_2f64_8f64(<2 x double> *%p) nounwind {
 ; X64-AVX-NEXT:    vmovdqa %ymm0, %ymm1
 ; X64-AVX-NEXT:    retq
 ;
+; X64-AVX512F-LABEL: test_broadcast_2f64_8f64:
+; X64-AVX512F:       ## BB#0:
+; X64-AVX512F-NEXT:    vbroadcastf32x4 (%rdi), %ymm0
+; X64-AVX512F-NEXT:    vinsertf64x4 $1, %ymm0, %zmm0, %zmm0
+; X64-AVX512F-NEXT:    retq
+;
 ; X64-AVX512BW-LABEL: test_broadcast_2f64_8f64:
 ; X64-AVX512BW:       ## BB#0:
 ; X64-AVX512BW-NEXT:    vbroadcastf32x4 (%rdi), %ymm0
@@ -140,6 +164,12 @@ define <4 x i64> @test_broadcast_2i64_4i64(<2 x i64> *%p) nounwind {
 ; X32-AVX2-NEXT:    vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1]
 ; X32-AVX2-NEXT:    retl
 ;
+; X32-AVX512F-LABEL: test_broadcast_2i64_4i64:
+; X32-AVX512F:       ## BB#0:
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX512F-NEXT:    vbroadcasti32x4 (%eax), %ymm0
+; X32-AVX512F-NEXT:    retl
+;
 ; X32-AVX512BW-LABEL: test_broadcast_2i64_4i64:
 ; X32-AVX512BW:       ## BB#0:
 ; X32-AVX512BW-NEXT:    movl {{[0-9]+}}(%esp), %eax
@@ -162,6 +192,11 @@ define <4 x i64> @test_broadcast_2i64_4i64(<2 x i64> *%p) nounwind {
 ; X64-AVX2-NEXT:    vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1]
 ; X64-AVX2-NEXT:    retq
 ;
+; X64-AVX512F-LABEL: test_broadcast_2i64_4i64:
+; X64-AVX512F:       ## BB#0:
+; X64-AVX512F-NEXT:    vbroadcasti32x4 (%rdi), %ymm0
+; X64-AVX512F-NEXT:    retq
+;
 ; X64-AVX512BW-LABEL: test_broadcast_2i64_4i64:
 ; X64-AVX512BW:       ## BB#0:
 ; X64-AVX512BW-NEXT:    vbroadcasti32x4 (%rdi), %ymm0
@@ -191,6 +226,13 @@ define <8 x i64> @test_broadcast_2i64_8i64(<2 x i64> *%p) nounwind {
 ; X32-AVX2-NEXT:    vmovdqa %ymm0, %ymm1
 ; X32-AVX2-NEXT:    retl
 ;
+; X32-AVX512F-LABEL: test_broadcast_2i64_8i64:
+; X32-AVX512F:       ## BB#0:
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX512F-NEXT:    vbroadcasti32x4 (%eax), %ymm0
+; X32-AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; X32-AVX512F-NEXT:    retl
+;
 ; X32-AVX512BW-LABEL: test_broadcast_2i64_8i64:
 ; X32-AVX512BW:       ## BB#0:
 ; X32-AVX512BW-NEXT:    movl {{[0-9]+}}(%esp), %eax
@@ -217,6 +259,12 @@ define <8 x i64> @test_broadcast_2i64_8i64(<2 x i64> *%p) nounwind {
 ; X64-AVX2-NEXT:    vmovdqa %ymm0, %ymm1
 ; X64-AVX2-NEXT:    retq
 ;
+; X64-AVX512F-LABEL: test_broadcast_2i64_8i64:
+; X64-AVX512F:       ## BB#0:
+; X64-AVX512F-NEXT:    vbroadcasti32x4 (%rdi), %ymm0
+; X64-AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; X64-AVX512F-NEXT:    retq
+;
 ; X64-AVX512BW-LABEL: test_broadcast_2i64_8i64:
 ; X64-AVX512BW:       ## BB#0:
 ; X64-AVX512BW-NEXT:    vbroadcasti32x4 (%rdi), %ymm0
@@ -299,6 +347,13 @@ define <16 x float> @test_broadcast_4f32_16f32(<4 x float> *%p) nounwind {
 ; X32-AVX-NEXT:    vmovdqa %ymm0, %ymm1
 ; X32-AVX-NEXT:    retl
 ;
+; X32-AVX512F-LABEL: test_broadcast_4f32_16f32:
+; X32-AVX512F:       ## BB#0:
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX512F-NEXT:    vbroadcastf32x4 (%eax), %ymm0
+; X32-AVX512F-NEXT:    vinsertf64x4 $1, %ymm0, %zmm0, %zmm0
+; X32-AVX512F-NEXT:    retl
+;
 ; X32-AVX512BW-LABEL: test_broadcast_4f32_16f32:
 ; X32-AVX512BW:       ## BB#0:
 ; X32-AVX512BW-NEXT:    movl {{[0-9]+}}(%esp), %eax
@@ -319,6 +374,12 @@ define <16 x float> @test_broadcast_4f32_16f32(<4 x float> *%p) nounwind {
 ; X64-AVX-NEXT:    vmovdqa %ymm0, %ymm1
 ; X64-AVX-NEXT:    retq
 ;
+; X64-AVX512F-LABEL: test_broadcast_4f32_16f32:
+; X64-AVX512F:       ## BB#0:
+; X64-AVX512F-NEXT:    vbroadcastf32x4 (%rdi), %ymm0
+; X64-AVX512F-NEXT:    vinsertf64x4 $1, %ymm0, %zmm0, %zmm0
+; X64-AVX512F-NEXT:    retq
+;
 ; X64-AVX512BW-LABEL: test_broadcast_4f32_16f32:
 ; X64-AVX512BW:       ## BB#0:
 ; X64-AVX512BW-NEXT:    vbroadcastf32x4 (%rdi), %ymm0
@@ -343,6 +404,13 @@ define <16 x float> @test_broadcast_8f32_16f32(<8 x float> *%p) nounwind {
 ; X32-AVX-NEXT:    vmovaps %ymm0, %ymm1
 ; X32-AVX-NEXT:    retl
 ;
+; X32-AVX512F-LABEL: test_broadcast_8f32_16f32:
+; X32-AVX512F:       ## BB#0:
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX512F-NEXT:    vmovapd (%eax), %ymm0
+; X32-AVX512F-NEXT:    vinsertf64x4 $1, %ymm0, %zmm0, %zmm0
+; X32-AVX512F-NEXT:    retl
+;
 ; X32-AVX512BW-LABEL: test_broadcast_8f32_16f32:
 ; X32-AVX512BW:       ## BB#0:
 ; X32-AVX512BW-NEXT:    movl {{[0-9]+}}(%esp), %eax
@@ -363,6 +431,12 @@ define <16 x float> @test_broadcast_8f32_16f32(<8 x float> *%p) nounwind {
 ; X64-AVX-NEXT:    vmovaps %ymm0, %ymm1
 ; X64-AVX-NEXT:    retq
 ;
+; X64-AVX512F-LABEL: test_broadcast_8f32_16f32:
+; X64-AVX512F:       ## BB#0:
+; X64-AVX512F-NEXT:    vmovapd (%rdi), %ymm0
+; X64-AVX512F-NEXT:    vinsertf64x4 $1, %ymm0, %zmm0, %zmm0
+; X64-AVX512F-NEXT:    retq
+;
 ; X64-AVX512BW-LABEL: test_broadcast_8f32_16f32:
 ; X64-AVX512BW:       ## BB#0:
 ; X64-AVX512BW-NEXT:    vmovapd (%rdi), %ymm0
@@ -432,6 +506,13 @@ define <16 x i32> @test_broadcast_4i32_16i32(<4 x i32> *%p) nounwind {
 ; X32-AVX2-NEXT:    vmovdqa %ymm0, %ymm1
 ; X32-AVX2-NEXT:    retl
 ;
+; X32-AVX512F-LABEL: test_broadcast_4i32_16i32:
+; X32-AVX512F:       ## BB#0:
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX512F-NEXT:    vbroadcasti32x4 (%eax), %ymm0
+; X32-AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; X32-AVX512F-NEXT:    retl
+;
 ; X32-AVX512BW-LABEL: test_broadcast_4i32_16i32:
 ; X32-AVX512BW:       ## BB#0:
 ; X32-AVX512BW-NEXT:    movl {{[0-9]+}}(%esp), %eax
@@ -458,6 +539,12 @@ define <16 x i32> @test_broadcast_4i32_16i32(<4 x i32> *%p) nounwind {
 ; X64-AVX2-NEXT:    vmovdqa %ymm0, %ymm1
 ; X64-AVX2-NEXT:    retq
 ;
+; X64-AVX512F-LABEL: test_broadcast_4i32_16i32:
+; X64-AVX512F:       ## BB#0:
+; X64-AVX512F-NEXT:    vbroadcasti32x4 (%rdi), %ymm0
+; X64-AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; X64-AVX512F-NEXT:    retq
+;
 ; X64-AVX512BW-LABEL: test_broadcast_4i32_16i32:
 ; X64-AVX512BW:       ## BB#0:
 ; X64-AVX512BW-NEXT:    vbroadcasti32x4 (%rdi), %ymm0
@@ -482,6 +569,13 @@ define <16 x i32> @test_broadcast_8i32_16i32(<8 x i32> *%p) nounwind {
 ; X32-AVX-NEXT:    vmovaps %ymm0, %ymm1
 ; X32-AVX-NEXT:    retl
 ;
+; X32-AVX512F-LABEL: test_broadcast_8i32_16i32:
+; X32-AVX512F:       ## BB#0:
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX512F-NEXT:    vmovdqa32 (%eax), %ymm0
+; X32-AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; X32-AVX512F-NEXT:    retl
+;
 ; X32-AVX512BW-LABEL: test_broadcast_8i32_16i32:
 ; X32-AVX512BW:       ## BB#0:
 ; X32-AVX512BW-NEXT:    movl {{[0-9]+}}(%esp), %eax
@@ -502,6 +596,12 @@ define <16 x i32> @test_broadcast_8i32_16i32(<8 x i32> *%p) nounwind {
 ; X64-AVX-NEXT:    vmovaps %ymm0, %ymm1
 ; X64-AVX-NEXT:    retq
 ;
+; X64-AVX512F-LABEL: test_broadcast_8i32_16i32:
+; X64-AVX512F:       ## BB#0:
+; X64-AVX512F-NEXT:    vmovdqa32 (%rdi), %ymm0
+; X64-AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; X64-AVX512F-NEXT:    retq
+;
 ; X64-AVX512BW-LABEL: test_broadcast_8i32_16i32:
 ; X64-AVX512BW:       ## BB#0:
 ; X64-AVX512BW-NEXT:    vmovdqa32 (%rdi), %ymm0
@@ -571,6 +671,13 @@ define <32 x i16> @test_broadcast_8i16_32i16(<8 x i16> *%p) nounwind {
 ; X32-AVX2-NEXT:    vmovdqa %ymm0, %ymm1
 ; X32-AVX2-NEXT:    retl
 ;
+; X32-AVX512F-LABEL: test_broadcast_8i16_32i16:
+; X32-AVX512F:       ## BB#0:
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX512F-NEXT:    vbroadcasti32x4 (%eax), %ymm0
+; X32-AVX512F-NEXT:    vmovdqa64 %ymm0, %ymm1
+; X32-AVX512F-NEXT:    retl
+;
 ; X32-AVX512BW-LABEL: test_broadcast_8i16_32i16:
 ; X32-AVX512BW:       ## BB#0:
 ; X32-AVX512BW-NEXT:    movl {{[0-9]+}}(%esp), %eax
@@ -597,6 +704,12 @@ define <32 x i16> @test_broadcast_8i16_32i16(<8 x i16> *%p) nounwind {
 ; X64-AVX2-NEXT:    vmovdqa %ymm0, %ymm1
 ; X64-AVX2-NEXT:    retq
 ;
+; X64-AVX512F-LABEL: test_broadcast_8i16_32i16:
+; X64-AVX512F:       ## BB#0:
+; X64-AVX512F-NEXT:    vbroadcasti32x4 (%rdi), %ymm0
+; X64-AVX512F-NEXT:    vmovdqa64 %ymm0, %ymm1
+; X64-AVX512F-NEXT:    retq
+;
 ; X64-AVX512BW-LABEL: test_broadcast_8i16_32i16:
 ; X64-AVX512BW:       ## BB#0:
 ; X64-AVX512BW-NEXT:    vbroadcasti32x4 (%rdi), %ymm0
@@ -621,6 +734,13 @@ define <32 x i16> @test_broadcast_16i16_32i16(<16 x i16> *%p) nounwind {
 ; X32-AVX-NEXT:    vmovaps %ymm0, %ymm1
 ; X32-AVX-NEXT:    retl
 ;
+; X32-AVX512F-LABEL: test_broadcast_16i16_32i16:
+; X32-AVX512F:       ## BB#0:
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX512F-NEXT:    vmovaps (%eax), %ymm0
+; X32-AVX512F-NEXT:    vmovaps %ymm0, %ymm1
+; X32-AVX512F-NEXT:    retl
+;
 ; X32-AVX512BW-LABEL: test_broadcast_16i16_32i16:
 ; X32-AVX512BW:       ## BB#0:
 ; X32-AVX512BW-NEXT:    movl {{[0-9]+}}(%esp), %eax
@@ -641,6 +761,12 @@ define <32 x i16> @test_broadcast_16i16_32i16(<16 x i16> *%p) nounwind {
 ; X64-AVX-NEXT:    vmovaps %ymm0, %ymm1
 ; X64-AVX-NEXT:    retq
 ;
+; X64-AVX512F-LABEL: test_broadcast_16i16_32i16:
+; X64-AVX512F:       ## BB#0:
+; X64-AVX512F-NEXT:    vmovaps (%rdi), %ymm0
+; X64-AVX512F-NEXT:    vmovaps %ymm0, %ymm1
+; X64-AVX512F-NEXT:    retq
+;
 ; X64-AVX512BW-LABEL: test_broadcast_16i16_32i16:
 ; X64-AVX512BW:       ## BB#0:
 ; X64-AVX512BW-NEXT:    vmovdqu16 (%rdi), %ymm0
@@ -710,6 +836,13 @@ define <64 x i8> @test_broadcast_16i8_64i8(<16 x i8> *%p) nounwind {
 ; X32-AVX2-NEXT:    vmovdqa %ymm0, %ymm1
 ; X32-AVX2-NEXT:    retl
 ;
+; X32-AVX512F-LABEL: test_broadcast_16i8_64i8:
+; X32-AVX512F:       ## BB#0:
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX512F-NEXT:    vbroadcasti32x4 (%eax), %ymm0
+; X32-AVX512F-NEXT:    vmovdqa64 %ymm0, %ymm1
+; X32-AVX512F-NEXT:    retl
+;
 ; X32-AVX512BW-LABEL: test_broadcast_16i8_64i8:
 ; X32-AVX512BW:       ## BB#0:
 ; X32-AVX512BW-NEXT:    movl {{[0-9]+}}(%esp), %eax
@@ -736,6 +869,12 @@ define <64 x i8> @test_broadcast_16i8_64i8(<16 x i8> *%p) nounwind {
 ; X64-AVX2-NEXT:    vmovdqa %ymm0, %ymm1
 ; X64-AVX2-NEXT:    retq
 ;
+; X64-AVX512F-LABEL: test_broadcast_16i8_64i8:
+; X64-AVX512F:       ## BB#0:
+; X64-AVX512F-NEXT:    vbroadcasti32x4 (%rdi), %ymm0
+; X64-AVX512F-NEXT:    vmovdqa64 %ymm0, %ymm1
+; X64-AVX512F-NEXT:    retq
+;
 ; X64-AVX512BW-LABEL: test_broadcast_16i8_64i8:
 ; X64-AVX512BW:       ## BB#0:
 ; X64-AVX512BW-NEXT:    vbroadcasti32x4 (%rdi), %ymm0
@@ -760,6 +899,13 @@ define <64 x i8> @test_broadcast_32i8_64i8(<32 x i8> *%p) nounwind {
 ; X32-AVX-NEXT:    vmovaps %ymm0, %ymm1
 ; X32-AVX-NEXT:    retl
 ;
+; X32-AVX512F-LABEL: test_broadcast_32i8_64i8:
+; X32-AVX512F:       ## BB#0:
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX512F-NEXT:    vmovaps (%eax), %ymm0
+; X32-AVX512F-NEXT:    vmovaps %ymm0, %ymm1
+; X32-AVX512F-NEXT:    retl
+;
 ; X32-AVX512BW-LABEL: test_broadcast_32i8_64i8:
 ; X32-AVX512BW:       ## BB#0:
 ; X32-AVX512BW-NEXT:    movl {{[0-9]+}}(%esp), %eax
@@ -780,6 +926,12 @@ define <64 x i8> @test_broadcast_32i8_64i8(<32 x i8> *%p) nounwind {
 ; X64-AVX-NEXT:    vmovaps %ymm0, %ymm1
 ; X64-AVX-NEXT:    retq
 ;
+; X64-AVX512F-LABEL: test_broadcast_32i8_64i8:
+; X64-AVX512F:       ## BB#0:
+; X64-AVX512F-NEXT:    vmovaps (%rdi), %ymm0
+; X64-AVX512F-NEXT:    vmovaps %ymm0, %ymm1
+; X64-AVX512F-NEXT:    retq
+;
 ; X64-AVX512BW-LABEL: test_broadcast_32i8_64i8:
 ; X64-AVX512BW:       ## BB#0:
 ; X64-AVX512BW-NEXT:    vmovdqu8 (%rdi), %ymm0
@@ -810,6 +962,15 @@ define <4 x double> @test_broadcast_2f64_4f64_reuse(<2 x double>* %p0, <2 x doub
 ; X32-AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; X32-AVX-NEXT:    retl
 ;
+; X32-AVX512F-LABEL: test_broadcast_2f64_4f64_reuse:
+; X32-AVX512F:       ## BB#0:
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-AVX512F-NEXT:    vmovaps (%ecx), %xmm0
+; X32-AVX512F-NEXT:    vmovaps %xmm0, (%eax)
+; X32-AVX512F-NEXT:    vinsertf32x4 $1, %xmm0, %ymm0, %ymm0
+; X32-AVX512F-NEXT:    retl
+;
 ; X32-AVX512BW-LABEL: test_broadcast_2f64_4f64_reuse:
 ; X32-AVX512BW:       ## BB#0:
 ; X32-AVX512BW-NEXT:    movl {{[0-9]+}}(%esp), %eax
@@ -835,6 +996,13 @@ define <4 x double> @test_broadcast_2f64_4f64_reuse(<2 x double>* %p0, <2 x doub
 ; X64-AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; X64-AVX-NEXT:    retq
 ;
+; X64-AVX512F-LABEL: test_broadcast_2f64_4f64_reuse:
+; X64-AVX512F:       ## BB#0:
+; X64-AVX512F-NEXT:    vmovaps (%rdi), %xmm0
+; X64-AVX512F-NEXT:    vmovaps %xmm0, (%rsi)
+; X64-AVX512F-NEXT:    vinsertf32x4 $1, %xmm0, %ymm0, %ymm0
+; X64-AVX512F-NEXT:    retq
+;
 ; X64-AVX512BW-LABEL: test_broadcast_2f64_4f64_reuse:
 ; X64-AVX512BW:       ## BB#0:
 ; X64-AVX512BW-NEXT:    vmovaps (%rdi), %xmm0
@@ -864,6 +1032,15 @@ define <4 x i64> @test_broadcast_2i64_4i64_reuse(<2 x i64>* %p0, <2 x i64>* %p1)
 ; X32-AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; X32-AVX-NEXT:    retl
 ;
+; X32-AVX512F-LABEL: test_broadcast_2i64_4i64_reuse:
+; X32-AVX512F:       ## BB#0:
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-AVX512F-NEXT:    vmovdqa64 (%ecx), %xmm0
+; X32-AVX512F-NEXT:    vmovdqa64 %xmm0, (%eax)
+; X32-AVX512F-NEXT:    vinserti32x4 $1, %xmm0, %ymm0, %ymm0
+; X32-AVX512F-NEXT:    retl
+;
 ; X32-AVX512BW-LABEL: test_broadcast_2i64_4i64_reuse:
 ; X32-AVX512BW:       ## BB#0:
 ; X32-AVX512BW-NEXT:    movl {{[0-9]+}}(%esp), %eax
@@ -889,6 +1066,13 @@ define <4 x i64> @test_broadcast_2i64_4i64_reuse(<2 x i64>* %p0, <2 x i64>* %p1)
 ; X64-AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; X64-AVX-NEXT:    retq
 ;
+; X64-AVX512F-LABEL: test_broadcast_2i64_4i64_reuse:
+; X64-AVX512F:       ## BB#0:
+; X64-AVX512F-NEXT:    vmovdqa64 (%rdi), %xmm0
+; X64-AVX512F-NEXT:    vmovdqa64 %xmm0, (%rsi)
+; X64-AVX512F-NEXT:    vinserti32x4 $1, %xmm0, %ymm0, %ymm0
+; X64-AVX512F-NEXT:    retq
+;
 ; X64-AVX512BW-LABEL: test_broadcast_2i64_4i64_reuse:
 ; X64-AVX512BW:       ## BB#0:
 ; X64-AVX512BW-NEXT:    vmovdqa64 (%rdi), %xmm0
@@ -994,6 +1178,15 @@ define <16 x i16> @test_broadcast_8i16_16i16_reuse(<8 x i16> *%p0, <8 x i16> *%p
 ; X32-AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; X32-AVX-NEXT:    retl
 ;
+; X32-AVX512F-LABEL: test_broadcast_8i16_16i16_reuse:
+; X32-AVX512F:       ## BB#0:
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-AVX512F-NEXT:    vmovdqa64 (%ecx), %xmm0
+; X32-AVX512F-NEXT:    vmovdqa32 %xmm0, (%eax)
+; X32-AVX512F-NEXT:    vinserti32x4 $1, %xmm0, %ymm0, %ymm0
+; X32-AVX512F-NEXT:    retl
+;
 ; X32-AVX512BW-LABEL: test_broadcast_8i16_16i16_reuse:
 ; X32-AVX512BW:       ## BB#0:
 ; X32-AVX512BW-NEXT:    movl {{[0-9]+}}(%esp), %eax
@@ -1019,6 +1212,13 @@ define <16 x i16> @test_broadcast_8i16_16i16_reuse(<8 x i16> *%p0, <8 x i16> *%p
 ; X64-AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; X64-AVX-NEXT:    retq
 ;
+; X64-AVX512F-LABEL: test_broadcast_8i16_16i16_reuse:
+; X64-AVX512F:       ## BB#0:
+; X64-AVX512F-NEXT:    vmovdqa64 (%rdi), %xmm0
+; X64-AVX512F-NEXT:    vmovdqa32 %xmm0, (%rsi)
+; X64-AVX512F-NEXT:    vinserti32x4 $1, %xmm0, %ymm0, %ymm0
+; X64-AVX512F-NEXT:    retq
+;
 ; X64-AVX512BW-LABEL: test_broadcast_8i16_16i16_reuse:
 ; X64-AVX512BW:       ## BB#0:
 ; X64-AVX512BW-NEXT:    vmovdqu16 (%rdi), %xmm0
@@ -1048,6 +1248,15 @@ define <32 x i8> @test_broadcast_16i8_32i8_reuse(<16 x i8> *%p0, <16 x i8> *%p1)
 ; X32-AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; X32-AVX-NEXT:    retl
 ;
+; X32-AVX512F-LABEL: test_broadcast_16i8_32i8_reuse:
+; X32-AVX512F:       ## BB#0:
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-AVX512F-NEXT:    vmovdqa64 (%ecx), %xmm0
+; X32-AVX512F-NEXT:    vmovdqa32 %xmm0, (%eax)
+; X32-AVX512F-NEXT:    vinserti32x4 $1, %xmm0, %ymm0, %ymm0
+; X32-AVX512F-NEXT:    retl
+;
 ; X32-AVX512BW-LABEL: test_broadcast_16i8_32i8_reuse:
 ; X32-AVX512BW:       ## BB#0:
 ; X32-AVX512BW-NEXT:    movl {{[0-9]+}}(%esp), %eax
@@ -1073,6 +1282,13 @@ define <32 x i8> @test_broadcast_16i8_32i8_reuse(<16 x i8> *%p0, <16 x i8> *%p1)
 ; X64-AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; X64-AVX-NEXT:    retq
 ;
+; X64-AVX512F-LABEL: test_broadcast_16i8_32i8_reuse:
+; X64-AVX512F:       ## BB#0:
+; X64-AVX512F-NEXT:    vmovdqa64 (%rdi), %xmm0
+; X64-AVX512F-NEXT:    vmovdqa32 %xmm0, (%rsi)
+; X64-AVX512F-NEXT:    vinserti32x4 $1, %xmm0, %ymm0, %ymm0
+; X64-AVX512F-NEXT:    retq
+;
 ; X64-AVX512BW-LABEL: test_broadcast_16i8_32i8_reuse:
 ; X64-AVX512BW:       ## BB#0:
 ; X64-AVX512BW-NEXT:    vmovdqu8 (%rdi), %xmm0
@@ -1107,6 +1323,16 @@ define <8 x i32> @test_broadcast_4i32_8i32_chain(<4 x i32>* %p0, <4 x float>* %p
 ; X32-AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; X32-AVX-NEXT:    retl
 ;
+; X32-AVX512F-LABEL: test_broadcast_4i32_8i32_chain:
+; X32-AVX512F:       ## BB#0:
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-AVX512F-NEXT:    vmovdqa64 (%ecx), %xmm0
+; X32-AVX512F-NEXT:    vpxord %xmm1, %xmm1, %xmm1
+; X32-AVX512F-NEXT:    vmovdqa32 %xmm1, (%eax)
+; X32-AVX512F-NEXT:    vinserti32x4 $1, %xmm0, %ymm0, %ymm0
+; X32-AVX512F-NEXT:    retl
+;
 ; X32-AVX512BW-LABEL: test_broadcast_4i32_8i32_chain:
 ; X32-AVX512BW:       ## BB#0:
 ; X32-AVX512BW-NEXT:    movl {{[0-9]+}}(%esp), %eax
@@ -1135,6 +1361,14 @@ define <8 x i32> @test_broadcast_4i32_8i32_chain(<4 x i32>* %p0, <4 x float>* %p
 ; X64-AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; X64-AVX-NEXT:    retq
 ;
+; X64-AVX512F-LABEL: test_broadcast_4i32_8i32_chain:
+; X64-AVX512F:       ## BB#0:
+; X64-AVX512F-NEXT:    vmovdqa64 (%rdi), %xmm0
+; X64-AVX512F-NEXT:    vpxord %xmm1, %xmm1, %xmm1
+; X64-AVX512F-NEXT:    vmovdqa32 %xmm1, (%rsi)
+; X64-AVX512F-NEXT:    vinserti32x4 $1, %xmm0, %ymm0, %ymm0
+; X64-AVX512F-NEXT:    retq
+;
 ; X64-AVX512BW-LABEL: test_broadcast_4i32_8i32_chain:
 ; X64-AVX512BW:       ## BB#0:
 ; X64-AVX512BW-NEXT:    vmovdqa64 (%rdi), %xmm0
@@ -1168,6 +1402,17 @@ define <16 x i32> @test_broadcast_4i32_16i32_chain(<4 x i32>* %p0, <4 x float>*
 ; X32-AVX-NEXT:    vmovaps %ymm0, %ymm1
 ; X32-AVX-NEXT:    retl
 ;
+; X32-AVX512F-LABEL: test_broadcast_4i32_16i32_chain:
+; X32-AVX512F:       ## BB#0:
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-AVX512F-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X32-AVX512F-NEXT:    vmovdqa64 (%ecx), %xmm0
+; X32-AVX512F-NEXT:    vpxord %xmm1, %xmm1, %xmm1
+; X32-AVX512F-NEXT:    vmovdqa32 %xmm1, (%eax)
+; X32-AVX512F-NEXT:    vinserti32x4 $1, %xmm0, %ymm0, %ymm0
+; X32-AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; X32-AVX512F-NEXT:    retl
+;
 ; X32-AVX512BW-LABEL: test_broadcast_4i32_16i32_chain:
 ; X32-AVX512BW:       ## BB#0:
 ; X32-AVX512BW-NEXT:    movl {{[0-9]+}}(%esp), %eax
@@ -1199,6 +1444,15 @@ define <16 x i32> @test_broadcast_4i32_16i32_chain(<4 x i32>* %p0, <4 x float>*
 ; X64-AVX-NEXT:    vmovaps %ymm0, %ymm1
 ; X64-AVX-NEXT:    retq
 ;
+; X64-AVX512F-LABEL: test_broadcast_4i32_16i32_chain:
+; X64-AVX512F:       ## BB#0:
+; X64-AVX512F-NEXT:    vmovdqa64 (%rdi), %xmm0
+; X64-AVX512F-NEXT:    vpxord %xmm1, %xmm1, %xmm1
+; X64-AVX512F-NEXT:    vmovdqa32 %xmm1, (%rsi)
+; X64-AVX512F-NEXT:    vinserti32x4 $1, %xmm0, %ymm0, %ymm0
+; X64-AVX512F-NEXT:    vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; X64-AVX512F-NEXT:    retq
+;
 ; X64-AVX512BW-LABEL: test_broadcast_4i32_16i32_chain:
 ; X64-AVX512BW:       ## BB#0:
 ; X64-AVX512BW-NEXT:    vmovdqa64 (%rdi), %xmm0