[x86] add movddup specialization for build vector lowering (PR37502)

This is admittedly a narrow fix for the problem:
https://bugs.llvm.org/show_bug.cgi?id=37502
...but as the XOP restriction shows, it's a maze to get this right. 
In the motivating example, note that we have movddup before SSE4.1 and 
again with AVX2. That's because insertps isn't available pre-SSE41 and 
vbroadcast is (more generally) available with AVX2 (and the splat is 
reduced to movddup via isel pattern).

Differential Revision: https://reviews.llvm.org/D55898

llvm-svn: 349937

5 files changed, 62 insertions, 62 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 78197191089..43a56f96481 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -6951,6 +6951,26 @@ static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
 /// Custom lower build_vector of v4i32 or v4f32.
 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
                                      const X86Subtarget &Subtarget) {
+  // If this is a splat of a pair of elements, use MOVDDUP (unless the target
+  // has XOP; in that case defer lowering to potentially use VPERMIL2PS).
+  // Because we're creating a less complicated build vector here, we may enable
+  // further folding of the MOVDDUP via shuffle transforms.
+  if (Subtarget.hasSSE3() && !Subtarget.hasXOP() &&
+      Op.getOperand(0) == Op.getOperand(2) &&
+      Op.getOperand(1) == Op.getOperand(3) &&
+      Op.getOperand(0) != Op.getOperand(1)) {
+    SDLoc DL(Op);
+    MVT VT = Op.getSimpleValueType();
+    MVT EltVT = VT.getVectorElementType();
+    // Create a new build vector with the first 2 elements followed by undef
+    // padding, bitcast to v2f64, duplicate, and bitcast back.
+    SDValue Ops[4] = { Op.getOperand(0), Op.getOperand(1),
+                       DAG.getUNDEF(EltVT), DAG.getUNDEF(EltVT) };
+    SDValue NewBV = DAG.getBitcast(MVT::v2f64, DAG.getBuildVector(VT, DL, Ops));
+    SDValue Dup = DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, NewBV);
+    return DAG.getBitcast(VT, Dup);
+  }
+
   // Find all zeroable elements.
   std::bitset<4> Zeroable;
   for (int i=0; i < 4; ++i) {
diff --git a/llvm/test/CodeGen/X86/avx-intrinsics-fast-isel.ll b/llvm/test/CodeGen/X86/avx-intrinsics-fast-isel.ll
index 84b3b007310..1e6d0e8766b 100644
--- a/llvm/test/CodeGen/X86/avx-intrinsics-fast-isel.ll
+++ b/llvm/test/CodeGen/X86/avx-intrinsics-fast-isel.ll
@@ -1956,12 +1956,9 @@ define <4 x i64> @test_mm256_set1_epi32(i32 %a0) nounwind {
 define <4 x i64> @test_mm256_set1_epi64x(i64 %a0) nounwind {
 ; X86-LABEL: test_mm256_set1_epi64x:
 ; X86:       # %bb.0:
-; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
-; X86-NEXT:    vmovd %ecx, %xmm0
-; X86-NEXT:    vpinsrd $1, %eax, %xmm0, %xmm0
-; X86-NEXT:    vpinsrd $2, %ecx, %xmm0, %xmm0
-; X86-NEXT:    vpinsrd $3, %eax, %xmm0, %xmm0
+; X86-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X86-NEXT:    vpinsrd $1, {{[0-9]+}}(%esp), %xmm0, %xmm0
+; X86-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
 ; X86-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; X86-NEXT:    retl
 ;
diff --git a/llvm/test/CodeGen/X86/avx-vbroadcast.ll b/llvm/test/CodeGen/X86/avx-vbroadcast.ll
index b136c72366e..157238622c4 100644
--- a/llvm/test/CodeGen/X86/avx-vbroadcast.ll
+++ b/llvm/test/CodeGen/X86/avx-vbroadcast.ll
@@ -6,12 +6,8 @@ define <4 x i64> @A(i64* %ptr) nounwind uwtable readnone ssp {
 ; X32-LABEL: A:
 ; X32:       ## %bb.0: ## %entry
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-NEXT:    movl (%eax), %ecx
-; X32-NEXT:    movl 4(%eax), %eax
-; X32-NEXT:    vmovd %ecx, %xmm0
-; X32-NEXT:    vpinsrd $1, %eax, %xmm0, %xmm0
-; X32-NEXT:    vpinsrd $2, %ecx, %xmm0, %xmm0
-; X32-NEXT:    vpinsrd $3, %eax, %xmm0, %xmm0
+; X32-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
+; X32-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,1,0,1]
 ; X32-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
 ; X32-NEXT:    retl
 ;
@@ -31,17 +27,19 @@ entry:
 define <4 x i64> @A2(i64* %ptr, i64* %ptr2) nounwind uwtable readnone ssp {
 ; X32-LABEL: A2:
 ; X32:       ## %bb.0: ## %entry
+; X32-NEXT:    pushl %esi
+; X32-NEXT:    .cfi_def_cfa_offset 8
+; X32-NEXT:    .cfi_offset %esi, -8
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %ecx
 ; X32-NEXT:    movl (%ecx), %edx
-; X32-NEXT:    movl 4(%ecx), %ecx
-; X32-NEXT:    movl %ecx, 4(%eax)
+; X32-NEXT:    movl 4(%ecx), %esi
+; X32-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
 ; X32-NEXT:    movl %edx, (%eax)
-; X32-NEXT:    vmovd %edx, %xmm0
-; X32-NEXT:    vpinsrd $1, %ecx, %xmm0, %xmm0
-; X32-NEXT:    vpinsrd $2, %edx, %xmm0, %xmm0
-; X32-NEXT:    vpinsrd $3, %ecx, %xmm0, %xmm0
+; X32-NEXT:    movl %esi, 4(%eax)
+; X32-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,1,0,1]
 ; X32-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
+; X32-NEXT:    popl %esi
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: A2:
@@ -592,12 +590,8 @@ define <2 x i64> @G(i64* %ptr) nounwind uwtable readnone ssp {
 ; X32-LABEL: G:
 ; X32:       ## %bb.0: ## %entry
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-NEXT:    movl (%eax), %ecx
-; X32-NEXT:    movl 4(%eax), %eax
-; X32-NEXT:    vmovd %ecx, %xmm0
-; X32-NEXT:    vpinsrd $1, %eax, %xmm0, %xmm0
-; X32-NEXT:    vpinsrd $2, %ecx, %xmm0, %xmm0
-; X32-NEXT:    vpinsrd $3, %eax, %xmm0, %xmm0
+; X32-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
+; X32-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,1,0,1]
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: G:
@@ -615,16 +609,18 @@ entry:
 define <2 x i64> @G2(i64* %ptr, i64* %ptr2) nounwind uwtable readnone ssp {
 ; X32-LABEL: G2:
 ; X32:       ## %bb.0: ## %entry
+; X32-NEXT:    pushl %esi
+; X32-NEXT:    .cfi_def_cfa_offset 8
+; X32-NEXT:    .cfi_offset %esi, -8
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X32-NEXT:    movl {{[0-9]+}}(%esp), %ecx
 ; X32-NEXT:    movl (%ecx), %edx
-; X32-NEXT:    movl 4(%ecx), %ecx
-; X32-NEXT:    movl %ecx, 4(%eax)
+; X32-NEXT:    movl 4(%ecx), %esi
+; X32-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
 ; X32-NEXT:    movl %edx, (%eax)
-; X32-NEXT:    vmovd %edx, %xmm0
-; X32-NEXT:    vpinsrd $1, %ecx, %xmm0, %xmm0
-; X32-NEXT:    vpinsrd $2, %edx, %xmm0, %xmm0
-; X32-NEXT:    vpinsrd $3, %ecx, %xmm0, %xmm0
+; X32-NEXT:    movl %esi, 4(%eax)
+; X32-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,1,0,1]
+; X32-NEXT:    popl %esi
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: G2:
@@ -879,11 +875,11 @@ define float @broadcast_lifetime() nounwind {
 ; X32-NEXT:    movl %esi, (%esp)
 ; X32-NEXT:    calll _gfunc
 ; X32-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; X32-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp) ## 4-byte Spill
+; X32-NEXT:    vmovss %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) ## 4-byte Spill
 ; X32-NEXT:    movl %esi, (%esp)
 ; X32-NEXT:    calll _gfunc
 ; X32-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; X32-NEXT:    vsubss {{[0-9]+}}(%esp), %xmm0, %xmm0 ## 4-byte Folded Reload
+; X32-NEXT:    vsubss {{[-0-9]+}}(%e{{[sb]}}p), %xmm0, %xmm0 ## 4-byte Folded Reload
 ; X32-NEXT:    vmovss %xmm0, {{[0-9]+}}(%esp)
 ; X32-NEXT:    flds {{[0-9]+}}(%esp)
 ; X32-NEXT:    addl $40, %esp
@@ -896,11 +892,11 @@ define float @broadcast_lifetime() nounwind {
 ; X64-NEXT:    leaq {{[0-9]+}}(%rsp), %rdi
 ; X64-NEXT:    callq _gfunc
 ; X64-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; X64-NEXT:    vmovss %xmm0, {{[0-9]+}}(%rsp) ## 4-byte Spill
+; X64-NEXT:    vmovss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 4-byte Spill
 ; X64-NEXT:    leaq {{[0-9]+}}(%rsp), %rdi
 ; X64-NEXT:    callq _gfunc
 ; X64-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; X64-NEXT:    vsubss {{[0-9]+}}(%rsp), %xmm0, %xmm0 ## 4-byte Folded Reload
+; X64-NEXT:    vsubss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0, %xmm0 ## 4-byte Folded Reload
 ; X64-NEXT:    addq $40, %rsp
 ; X64-NEXT:    retq
   %1 = alloca <4 x float>, align 16
diff --git a/llvm/test/CodeGen/X86/build-vector-128.ll b/llvm/test/CodeGen/X86/build-vector-128.ll
index 9702e4be22a..1c1dab64047 100644
--- a/llvm/test/CodeGen/X86/build-vector-128.ll
+++ b/llvm/test/CodeGen/X86/build-vector-128.ll
@@ -527,39 +527,27 @@ define <4 x float> @PR37502(float %x, float %y) {
 ; SSE41-32-LABEL: PR37502:
 ; SSE41-32:       # %bb.0:
 ; SSE41-32-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
-; SSE41-32-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
-; SSE41-32-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0,1,2],mem[0]
+; SSE41-32-NEXT:    movddup {{.*#+}} xmm0 = xmm0[0,0]
 ; SSE41-32-NEXT:    retl
 ;
 ; SSE41-64-LABEL: PR37502:
 ; SSE41-64:       # %bb.0:
-; SSE41-64-NEXT:    movaps %xmm0, %xmm2
-; SSE41-64-NEXT:    insertps {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[2,3]
-; SSE41-64-NEXT:    insertps {{.*#+}} xmm2 = xmm2[0,1],xmm0[0],xmm2[3]
-; SSE41-64-NEXT:    insertps {{.*#+}} xmm2 = xmm2[0,1,2],xmm1[0]
-; SSE41-64-NEXT:    movaps %xmm2, %xmm0
+; SSE41-64-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[2,3]
+; SSE41-64-NEXT:    movddup {{.*#+}} xmm0 = xmm0[0,0]
 ; SSE41-64-NEXT:    retq
 ;
-; AVX1-32-LABEL: PR37502:
-; AVX1-32:       # %bb.0:
-; AVX1-32-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
-; AVX1-32-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
-; AVX1-32-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],mem[0]
-; AVX1-32-NEXT:    retl
+; AVX-32-LABEL: PR37502:
+; AVX-32:       # %bb.0:
+; AVX-32-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
+; AVX-32-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
+; AVX-32-NEXT:    retl
 ;
 ; AVX1-64-LABEL: PR37502:
 ; AVX1-64:       # %bb.0:
-; AVX1-64-NEXT:    vinsertps {{.*#+}} xmm2 = xmm0[0],xmm1[0],xmm0[2,3]
-; AVX1-64-NEXT:    vinsertps {{.*#+}} xmm0 = xmm2[0,1],xmm0[0],xmm2[3]
-; AVX1-64-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[0]
+; AVX1-64-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[2,3]
+; AVX1-64-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
 ; AVX1-64-NEXT:    retq
 ;
-; AVX2-32-LABEL: PR37502:
-; AVX2-32:       # %bb.0:
-; AVX2-32-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
-; AVX2-32-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
-; AVX2-32-NEXT:    retl
-;
 ; AVX2-64-LABEL: PR37502:
 ; AVX2-64:       # %bb.0:
 ; AVX2-64-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
diff --git a/llvm/test/CodeGen/X86/sse2-intrinsics-fast-isel.ll b/llvm/test/CodeGen/X86/sse2-intrinsics-fast-isel.ll
index ba98b97a5e0..72a4bf34676 100644
--- a/llvm/test/CodeGen/X86/sse2-intrinsics-fast-isel.ll
+++ b/llvm/test/CodeGen/X86/sse2-intrinsics-fast-isel.ll
@@ -3979,12 +3979,11 @@ define <2 x i64> @test_mm_set1_epi64x(i64 %a0) nounwind {
 ;
 ; X86-AVX1-LABEL: test_mm_set1_epi64x:
 ; X86-AVX1:       # %bb.0:
-; X86-AVX1-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x08]
-; X86-AVX1-NEXT:    movl {{[0-9]+}}(%esp), %ecx # encoding: [0x8b,0x4c,0x24,0x04]
-; X86-AVX1-NEXT:    vmovd %ecx, %xmm0 # encoding: [0xc5,0xf9,0x6e,0xc1]
-; X86-AVX1-NEXT:    vpinsrd $1, %eax, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0x79,0x22,0xc0,0x01]
-; X86-AVX1-NEXT:    vpinsrd $2, %ecx, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0x79,0x22,0xc1,0x02]
-; X86-AVX1-NEXT:    vpinsrd $3, %eax, %xmm0, %xmm0 # encoding: [0xc4,0xe3,0x79,0x22,0xc0,0x03]
+; X86-AVX1-NEXT:    vmovd {{[0-9]+}}(%esp), %xmm0 # encoding: [0xc5,0xf9,0x6e,0x44,0x24,0x04]
+; X86-AVX1-NEXT:    # xmm0 = mem[0],zero,zero,zero
+; X86-AVX1-NEXT:    vpinsrd $1, {{[0-9]+}}(%esp), %xmm0, %xmm0 # encoding: [0xc4,0xe3,0x79,0x22,0x44,0x24,0x08,0x01]
+; X86-AVX1-NEXT:    vpshufd $68, %xmm0, %xmm0 # encoding: [0xc5,0xf9,0x70,0xc0,0x44]
+; X86-AVX1-NEXT:    # xmm0 = xmm0[0,1,0,1]
 ; X86-AVX1-NEXT:    retl # encoding: [0xc3]
 ;
 ; X86-AVX512-LABEL: test_mm_set1_epi64x: