[x86] try harder to use broadcast to load a scalar into vector reg

This is a preliminary step for a preliminary step for D50992. 
I noticed that x86 often misses chances to load a scalar directly 
into a vector register.

So this patch is just allowing more of those cases to match a 
broadcast op in lowerBuildVectorAsBroadcast(). The old code comment 
said it doesn't make sense to use a broadcast when we're loading a 
single element and everything else is undef, but I think that's the 
best case in the improved tests in insert-loaded-scalar.ll. We avoid 
scalar-to-vector-register move and/or less efficient shuffling.

Note that there are some existing types that were already producing 
a broadcast, but that happens semi-accidentally. Ie, it's not 
happening as part of lowerBuildVectorAsBroadcast(). The build vector 
gets expanded into load + shuffle, and then shuffle lowering produces 
the broadcast.

Description of the other test diffs:
1. avx-basic.ll - replacing load+shufle is a win.
2. sse3-avx-addsub-2.ll - vmovddup vs. vbroadcastss is neutral
3. sse41.ll - don't care - we convert that intrinsic to generic IR now, so this test is deprecated
4. vector-shuffle-128-v8.ll / vector-shuffle-256-v16.ll - pshufb alternatives with an extra instruction are not obviously bad

Differential Revision: https://reviews.llvm.org/D51125

llvm-svn: 340685

6 files changed, 101 insertions, 107 deletions

diff --git a/llvm/test/CodeGen/X86/avx-basic.ll b/llvm/test/CodeGen/X86/avx-basic.ll
index d27a641203f..b7c9b692d89 100644
--- a/llvm/test/CodeGen/X86/avx-basic.ll
+++ b/llvm/test/CodeGen/X86/avx-basic.ll
@@ -76,8 +76,7 @@ define <4 x i64> @ISelCrash(<4 x i64> %a) nounwind uwtable readnone ssp {
 define <8 x i32> @VMOVZQI2PQI([0 x float]* nocapture %aFOO) nounwind {
 ; CHECK-LABEL: VMOVZQI2PQI:
 ; CHECK:       ## %bb.0:
-; CHECK-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; CHECK-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,0,1,1]
+; CHECK-NEXT:    vbroadcastss (%rdi), %ymm0
 ; CHECK-NEXT:    retq
   %ptrcast.i33.i = bitcast [0 x float]* %aFOO to i32*
   %val.i34.i = load i32, i32* %ptrcast.i33.i, align 4
diff --git a/llvm/test/CodeGen/X86/insert-loaded-scalar.ll b/llvm/test/CodeGen/X86/insert-loaded-scalar.ll
index ec6b99cf63e..81cb533f442 100644
--- a/llvm/test/CodeGen/X86/insert-loaded-scalar.ll
+++ b/llvm/test/CodeGen/X86/insert-loaded-scalar.ll
@@ -10,11 +10,16 @@ define <16 x i8> @load8_ins_elt0_v16i8(i8* %p) nounwind {
 ; SSE-NEXT:    movd %eax, %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX-LABEL: load8_ins_elt0_v16i8:
-; AVX:       # %bb.0:
-; AVX-NEXT:    movzbl (%rdi), %eax
-; AVX-NEXT:    vmovd %eax, %xmm0
-; AVX-NEXT:    retq
+; AVX1-LABEL: load8_ins_elt0_v16i8:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    movzbl (%rdi), %eax
+; AVX1-NEXT:    vmovd %eax, %xmm0
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: load8_ins_elt0_v16i8:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpbroadcastb (%rdi), %xmm0
+; AVX2-NEXT:    retq
   %x = load i8, i8* %p
   %ins = insertelement <16 x i8> undef, i8 %x, i32 0
   ret <16 x i8> %ins
@@ -27,11 +32,16 @@ define <8 x i16> @load16_ins_elt0_v8i16(i16* %p) nounwind {
 ; SSE-NEXT:    movd %eax, %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX-LABEL: load16_ins_elt0_v8i16:
-; AVX:       # %bb.0:
-; AVX-NEXT:    movzwl (%rdi), %eax
-; AVX-NEXT:    vmovd %eax, %xmm0
-; AVX-NEXT:    retq
+; AVX1-LABEL: load16_ins_elt0_v8i16:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    movzwl (%rdi), %eax
+; AVX1-NEXT:    vmovd %eax, %xmm0
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: load16_ins_elt0_v8i16:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpbroadcastw (%rdi), %xmm0
+; AVX2-NEXT:    retq
   %x = load i16, i16* %p
   %ins = insertelement <8 x i16> undef, i16 %x, i32 0
   ret <8 x i16> %ins
@@ -105,12 +115,17 @@ define <16 x i8> @load8_ins_eltc_v16i8(i8* %p) nounwind {
 ; SSE-NEXT:    pslld $24, %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX-LABEL: load8_ins_eltc_v16i8:
-; AVX:       # %bb.0:
-; AVX-NEXT:    movzbl (%rdi), %eax
-; AVX-NEXT:    vmovd %eax, %xmm0
-; AVX-NEXT:    vpslld $24, %xmm0, %xmm0
-; AVX-NEXT:    retq
+; AVX1-LABEL: load8_ins_eltc_v16i8:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    movzbl (%rdi), %eax
+; AVX1-NEXT:    vmovd %eax, %xmm0
+; AVX1-NEXT:    vpslld $24, %xmm0, %xmm0
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: load8_ins_eltc_v16i8:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpbroadcastb (%rdi), %xmm0
+; AVX2-NEXT:    retq
   %x = load i8, i8* %p
   %ins = insertelement <16 x i8> undef, i8 %x, i32 3
   ret <16 x i8> %ins
@@ -147,17 +162,10 @@ define <4 x i32> @load32_ins_eltc_v4i32(i32* %p) nounwind {
 ; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
 ; SSE-NEXT:    retq
 ;
-; AVX1-LABEL: load32_ins_eltc_v4i32:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX1-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,1,0,1]
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: load32_ins_eltc_v4i32:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX2-NEXT:    vpbroadcastq %xmm0, %xmm0
-; AVX2-NEXT:    retq
+; AVX-LABEL: load32_ins_eltc_v4i32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vbroadcastss (%rdi), %xmm0
+; AVX-NEXT:    retq
   %x = load i32, i32* %p
   %ins = insertelement <4 x i32> undef, i32 %x, i32 2
   ret <4 x i32> %ins
@@ -223,11 +231,16 @@ define <32 x i8> @load8_ins_elt0_v32i8(i8* %p) nounwind {
 ; SSE-NEXT:    movd %eax, %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX-LABEL: load8_ins_elt0_v32i8:
-; AVX:       # %bb.0:
-; AVX-NEXT:    movzbl (%rdi), %eax
-; AVX-NEXT:    vmovd %eax, %xmm0
-; AVX-NEXT:    retq
+; AVX1-LABEL: load8_ins_elt0_v32i8:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    movzbl (%rdi), %eax
+; AVX1-NEXT:    vmovd %eax, %xmm0
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: load8_ins_elt0_v32i8:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpbroadcastb (%rdi), %ymm0
+; AVX2-NEXT:    retq
   %x = load i8, i8* %p
   %ins = insertelement <32 x i8> undef, i8 %x, i32 0
   ret <32 x i8> %ins
@@ -240,11 +253,16 @@ define <16 x i16> @load16_ins_elt0_v16i16(i16* %p) nounwind {
 ; SSE-NEXT:    movd %eax, %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX-LABEL: load16_ins_elt0_v16i16:
-; AVX:       # %bb.0:
-; AVX-NEXT:    movzwl (%rdi), %eax
-; AVX-NEXT:    vmovd %eax, %xmm0
-; AVX-NEXT:    retq
+; AVX1-LABEL: load16_ins_elt0_v16i16:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    movzwl (%rdi), %eax
+; AVX1-NEXT:    vmovd %eax, %xmm0
+; AVX1-NEXT:    retq
+;
+; AVX2-LABEL: load16_ins_elt0_v16i16:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    vpbroadcastw (%rdi), %ymm0
+; AVX2-NEXT:    retq
   %x = load i16, i16* %p
   %ins = insertelement <16 x i16> undef, i16 %x, i32 0
   ret <16 x i16> %ins
@@ -328,10 +346,7 @@ define <32 x i8> @load8_ins_eltc_v32i8(i8* %p) nounwind {
 ;
 ; AVX2-LABEL: load8_ins_eltc_v32i8:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    movzbl (%rdi), %eax
-; AVX2-NEXT:    vmovd %eax, %xmm0
-; AVX2-NEXT:    vpsllq $40, %xmm0, %xmm0
-; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
+; AVX2-NEXT:    vpbroadcastb (%rdi), %ymm0
 ; AVX2-NEXT:    retq
   %x = load i8, i8* %p
   %ins = insertelement <32 x i8> undef, i8 %x, i32 21
@@ -356,10 +371,7 @@ define <16 x i16> @load16_ins_eltc_v16i16(i16* %p) nounwind {
 ;
 ; AVX2-LABEL: load16_ins_eltc_v16i16:
 ; AVX2:       # %bb.0:
-; AVX2-NEXT:    movzwl (%rdi), %eax
-; AVX2-NEXT:    vmovd %eax, %xmm0
-; AVX2-NEXT:    vpbroadcastw %xmm0, %xmm0
-; AVX2-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm0
+; AVX2-NEXT:    vpbroadcastw (%rdi), %ymm0
 ; AVX2-NEXT:    retq
   %x = load i16, i16* %p
   %ins = insertelement <16 x i16> undef, i16 %x, i32 11
@@ -373,18 +385,10 @@ define <8 x i32> @load32_ins_eltc_v8i32(i32* %p) nounwind {
 ; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,1,2,0]
 ; SSE-NEXT:    retq
 ;
-; AVX1-LABEL: load32_ins_eltc_v8i32:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX1-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,1,2,0]
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: load32_ins_eltc_v8i32:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vbroadcastss (%rdi), %xmm0
-; AVX2-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
-; AVX2-NEXT:    retq
+; AVX-LABEL: load32_ins_eltc_v8i32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vbroadcastss (%rdi), %ymm0
+; AVX-NEXT:    retq
   %x = load i32, i32* %p
   %ins = insertelement <8 x i32> undef, i32 %x, i32 7
   ret <8 x i32> %ins
@@ -397,17 +401,10 @@ define <4 x i64> @load64_ins_eltc_v4i64(i64* %p) nounwind {
 ; SSE-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[0,1,0,1]
 ; SSE-NEXT:    retq
 ;
-; AVX1-LABEL: load64_ins_eltc_v4i64:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
-; AVX1-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,1,0,1]
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: load64_ins_eltc_v4i64:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vbroadcastsd (%rdi), %ymm0
-; AVX2-NEXT:    retq
+; AVX-LABEL: load64_ins_eltc_v4i64:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vbroadcastsd (%rdi), %ymm0
+; AVX-NEXT:    retq
   %x = load i64, i64* %p
   %ins = insertelement <4 x i64> undef, i64 %x, i32 3
   ret <4 x i64> %ins
@@ -420,18 +417,10 @@ define <8 x float> @load32_ins_eltc_v8f32(float* %p) nounwind {
 ; SSE-NEXT:    movsldup {{.*#+}} xmm1 = xmm0[0,0,2,2]
 ; SSE-NEXT:    retq
 ;
-; AVX1-LABEL: load32_ins_eltc_v8f32:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX1-NEXT:    vmovsldup {{.*#+}} xmm0 = xmm0[0,0,2,2]
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: load32_ins_eltc_v8f32:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vbroadcastss (%rdi), %xmm0
-; AVX2-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
-; AVX2-NEXT:    retq
+; AVX-LABEL: load32_ins_eltc_v8f32:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vbroadcastss (%rdi), %ymm0
+; AVX-NEXT:    retq
   %x = load float, float* %p
   %ins = insertelement <8 x float> undef, float %x, i32 5
   ret <8 x float> %ins
@@ -443,16 +432,10 @@ define <4 x double> @load64_ins_eltc_v4f64(double* %p) nounwind {
 ; SSE-NEXT:    movddup {{.*#+}} xmm1 = mem[0,0]
 ; SSE-NEXT:    retq
 ;
-; AVX1-LABEL: load64_ins_eltc_v4f64:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovddup {{.*#+}} xmm0 = mem[0,0]
-; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
-; AVX1-NEXT:    retq
-;
-; AVX2-LABEL: load64_ins_eltc_v4f64:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vbroadcastsd (%rdi), %ymm0
-; AVX2-NEXT:    retq
+; AVX-LABEL: load64_ins_eltc_v4f64:
+; AVX:       # %bb.0:
+; AVX-NEXT:    vbroadcastsd (%rdi), %ymm0
+; AVX-NEXT:    retq
   %x = load double, double* %p
   %ins = insertelement <4 x double> undef, double %x, i32 3
   ret <4 x double> %ins
diff --git a/llvm/test/CodeGen/X86/sse3-avx-addsub-2.ll b/llvm/test/CodeGen/X86/sse3-avx-addsub-2.ll
index 9f9fe237ad4..d5631801e39 100644
--- a/llvm/test/CodeGen/X86/sse3-avx-addsub-2.ll
+++ b/llvm/test/CodeGen/X86/sse3-avx-addsub-2.ll
@@ -274,13 +274,21 @@ define <4 x float> @test11(<4 x float> %A, <4 x float> %B) {
 ; SSE-NEXT:    movddup {{.*#+}} xmm0 = xmm0[0,0]
 ; SSE-NEXT:    retq
 ;
-; AVX-LABEL: test11:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
-; AVX-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm1[1,0]
-; AVX-NEXT:    vsubss %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
-; AVX-NEXT:    retq
+; AVX1-LABEL: test11:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX1-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm1[1,0]
+; AVX1-NEXT:    vsubss %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
+; AVX1-NEXT:    retq
+;
+; AVX512-LABEL: test11:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    vpermilpd {{.*#+}} xmm0 = xmm0[1,0]
+; AVX512-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm1[1,0]
+; AVX512-NEXT:    vsubss %xmm1, %xmm0, %xmm0
+; AVX512-NEXT:    vbroadcastss %xmm0, %xmm0
+; AVX512-NEXT:    retq
   %1 = extractelement <4 x float> %A, i32 2
   %2 = extractelement <4 x float> %B, i32 2
   %sub = fsub float %1, %2
diff --git a/llvm/test/CodeGen/X86/sse41.ll b/llvm/test/CodeGen/X86/sse41.ll
index c466ee7ed18..b5c37263423 100644
--- a/llvm/test/CodeGen/X86/sse41.ll
+++ b/llvm/test/CodeGen/X86/sse41.ll
@@ -97,8 +97,9 @@ define <2 x i64> @pmovzxbq_1() nounwind {
 ; X86-AVX512:       ## %bb.0: ## %entry
 ; X86-AVX512-NEXT:    movl L_g16$non_lazy_ptr, %eax ## encoding: [0xa1,A,A,A,A]
 ; X86-AVX512-NEXT:    ## fixup A - offset: 1, value: L_g16$non_lazy_ptr, kind: FK_Data_4
-; X86-AVX512-NEXT:    vpmovzxbq (%eax), %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x32,0x00]
-; X86-AVX512-NEXT:    ## xmm0 = mem[0],zero,zero,zero,zero,zero,zero,zero,mem[1],zero,zero,zero,zero,zero,zero,zero
+; X86-AVX512-NEXT:    vpbroadcastw (%eax), %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x79,0x00]
+; X86-AVX512-NEXT:    vpmovzxbq %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x32,0xc0]
+; X86-AVX512-NEXT:    ## xmm0 = xmm0[0],zero,zero,zero,zero,zero,zero,zero,xmm0[1],zero,zero,zero,zero,zero,zero,zero
 ; X86-AVX512-NEXT:    retl ## encoding: [0xc3]
 ;
 ; X64-SSE-LABEL: pmovzxbq_1:
@@ -121,8 +122,9 @@ define <2 x i64> @pmovzxbq_1() nounwind {
 ; X64-AVX512:       ## %bb.0: ## %entry
 ; X64-AVX512-NEXT:    movq _g16@{{.*}}(%rip), %rax ## encoding: [0x48,0x8b,0x05,A,A,A,A]
 ; X64-AVX512-NEXT:    ## fixup A - offset: 3, value: _g16@GOTPCREL-4, kind: reloc_riprel_4byte_movq_load
-; X64-AVX512-NEXT:    vpmovzxbq (%rax), %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x32,0x00]
-; X64-AVX512-NEXT:    ## xmm0 = mem[0],zero,zero,zero,zero,zero,zero,zero,mem[1],zero,zero,zero,zero,zero,zero,zero
+; X64-AVX512-NEXT:    vpbroadcastw (%rax), %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x79,0x00]
+; X64-AVX512-NEXT:    vpmovzxbq %xmm0, %xmm0 ## EVEX TO VEX Compression encoding: [0xc4,0xe2,0x79,0x32,0xc0]
+; X64-AVX512-NEXT:    ## xmm0 = xmm0[0],zero,zero,zero,zero,zero,zero,zero,xmm0[1],zero,zero,zero,zero,zero,zero,zero
 ; X64-AVX512-NEXT:    retq ## encoding: [0xc3]
 entry:
 	%0 = load i16, i16* @g16, align 2		; <i16> [#uses=1]
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-128-v8.ll b/llvm/test/CodeGen/X86/vector-shuffle-128-v8.ll
index 367a72f56c4..c59db9b2f0a 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-128-v8.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-128-v8.ll
@@ -2619,8 +2619,9 @@ define <8 x i16> @insert_dup_elt3_mem_v8i16_i32(i32* %ptr) {
 ;
 ; AVX1-LABEL: insert_dup_elt3_mem_v8i16_i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX1-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[2,3,2,3,2,3,2,3,2,3,2,3,2,3,2,3]
+; AVX1-NEXT:    vbroadcastss (%rdi), %xmm0
+; AVX1-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[3,3,2,3,4,5,6,7]
+; AVX1-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
 ; AVX1-NEXT:    retq
 ;
 ; AVX2OR512VL-LABEL: insert_dup_elt3_mem_v8i16_i32:
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-256-v16.ll b/llvm/test/CodeGen/X86/vector-shuffle-256-v16.ll
index 01998a5ac01..3566ebde720 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-256-v16.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-256-v16.ll
@@ -4597,8 +4597,9 @@ define <16 x i16> @insert_dup_elt1_mem_v16i16_i32(i32* %ptr) #0 {
 define <16 x i16> @insert_dup_elt3_mem_v16i16_i32(i32* %ptr) #0 {
 ; AVX1-LABEL: insert_dup_elt3_mem_v16i16_i32:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX1-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[2,3,2,3,2,3,2,3,2,3,2,3,2,3,2,3]
+; AVX1-NEXT:    vbroadcastss (%rdi), %xmm0
+; AVX1-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[3,3,2,3,4,5,6,7]
+; AVX1-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,0,0,0]
 ; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; AVX1-NEXT:    retq
 ;