[X86][SSE] Replace insert_vector_elt(vec, -1, idx) with shuffle

Similar to what we already do for zero elt insertion, we can quickly rematerialize 'allbits' vectors so to avoid a unnecessary gpr value and insertion into a vector

llvm-svn: 294162

2 files changed, 24 insertions, 30 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 1aa7c422012..af06cd3a719 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -13844,17 +13844,21 @@ SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
   auto *N2C = cast<ConstantSDNode>(N2);
   unsigned IdxVal = N2C->getZExtValue();
 
-  // If we are clearing out a element, we do this more efficiently with a
-  // blend shuffle than a costly integer insertion.
-  // TODO: would other rematerializable values (e.g. allbits) benefit as well?
+  bool IsZeroElt = X86::isZeroNode(N1);
+  bool IsAllOnesElt = VT.isInteger() && llvm::isAllOnesConstant(N1);
+
+  // If we are inserting a element, see if we can do this more efficiently with
+  // a blend shuffle with a rematerializable vector than a costly integer
+  // insertion.
   // TODO: pre-SSE41 targets will tend to use bit masking - this could still
   // be beneficial if we are inserting several zeros and can combine the masks.
-  if (X86::isZeroNode(N1) && Subtarget.hasSSE41() && NumElts <= 8) {
-    SmallVector<int, 8> ClearMask;
+  if ((IsZeroElt || IsAllOnesElt) && Subtarget.hasSSE41() && NumElts <= 8) {
+    SmallVector<int, 8> BlendMask;
     for (unsigned i = 0; i != NumElts; ++i)
-      ClearMask.push_back(i == IdxVal ? i + NumElts : i);
-    SDValue ZeroVector = getZeroVector(VT, Subtarget, DAG, dl);
-    return DAG.getVectorShuffle(VT, dl, N0, ZeroVector, ClearMask);
+      BlendMask.push_back(i == IdxVal ? i + NumElts : i);
+    SDValue CstVector = IsZeroElt ? getZeroVector(VT, Subtarget, DAG, dl)
+                                  : DAG.getConstant(-1, dl, VT);
+    return DAG.getVectorShuffle(VT, dl, N0, CstVector, BlendMask);
   }
 
   // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
diff --git a/llvm/test/CodeGen/X86/avx-cvt-3.ll b/llvm/test/CodeGen/X86/avx-cvt-3.ll
index 45a6421cf22..066719b3bfe 100644
--- a/llvm/test/CodeGen/X86/avx-cvt-3.ll
+++ b/llvm/test/CodeGen/X86/avx-cvt-3.ll
@@ -48,27 +48,17 @@ define <8 x float> @sitofp_shuffle_zero_v8i32(<8 x i32> %a0) {
 define <8 x float> @sitofp_insert_allbits_v8i32(<8 x i32> %a0) {
 ; X86-LABEL: sitofp_insert_allbits_v8i32:
 ; X86:       # BB#0:
-; X86-NEXT:    movl $-1, %eax
-; X86-NEXT:    vpinsrd $0, %eax, %xmm0, %xmm1
-; X86-NEXT:    vpinsrd $2, %eax, %xmm1, %xmm1
-; X86-NEXT:    vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]
-; X86-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; X86-NEXT:    vpinsrd $0, %eax, %xmm1, %xmm1
-; X86-NEXT:    vpinsrd $1, %eax, %xmm1, %xmm1
-; X86-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; X86-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; X86-NEXT:    vinsertf128 $1, %xmm1, %ymm1, %ymm1
+; X86-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4,5],ymm0[6,7]
 ; X86-NEXT:    vcvtdq2ps %ymm0, %ymm0
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: sitofp_insert_allbits_v8i32:
 ; X64:       # BB#0:
-; X64-NEXT:    movl $-1, %eax
-; X64-NEXT:    vpinsrd $0, %eax, %xmm0, %xmm1
-; X64-NEXT:    vpinsrd $2, %eax, %xmm1, %xmm1
-; X64-NEXT:    vblendpd {{.*#+}} ymm0 = ymm1[0,1],ymm0[2,3]
-; X64-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; X64-NEXT:    vpinsrd $0, %eax, %xmm1, %xmm1
-; X64-NEXT:    vpinsrd $1, %eax, %xmm1, %xmm1
-; X64-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
+; X64-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; X64-NEXT:    vinsertf128 $1, %xmm1, %ymm1, %ymm1
+; X64-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4,5],ymm0[6,7]
 ; X64-NEXT:    vcvtdq2ps %ymm0, %ymm0
 ; X64-NEXT:    retq
   %1 = insertelement <8 x i32> %a0, i32 -1, i32 0
@@ -105,9 +95,9 @@ define <8 x float> @sitofp_insert_constants_v8i32(<8 x i32> %a0) {
 ; X86:       # BB#0:
 ; X86-NEXT:    vxorps %ymm1, %ymm1, %ymm1
 ; X86-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6,7]
-; X86-NEXT:    movl $-1, %eax
-; X86-NEXT:    vpinsrd $2, %eax, %xmm0, %xmm1
-; X86-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
+; X86-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; X86-NEXT:    vinsertf128 $1, %xmm1, %ymm1, %ymm1
+; X86-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6,7]
 ; X86-NEXT:    vextractf128 $1, %ymm0, %xmm1
 ; X86-NEXT:    movl $2, %eax
 ; X86-NEXT:    vpinsrd $0, %eax, %xmm1, %xmm1
@@ -121,9 +111,9 @@ define <8 x float> @sitofp_insert_constants_v8i32(<8 x i32> %a0) {
 ; X64:       # BB#0:
 ; X64-NEXT:    vxorps %ymm1, %ymm1, %ymm1
 ; X64-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1,2,3,4,5,6,7]
-; X64-NEXT:    movl $-1, %eax
-; X64-NEXT:    vpinsrd $2, %eax, %xmm0, %xmm1
-; X64-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
+; X64-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
+; X64-NEXT:    vinsertf128 $1, %xmm1, %ymm1, %ymm1
+; X64-NEXT:    vblendps {{.*#+}} ymm0 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6,7]
 ; X64-NEXT:    vextractf128 $1, %ymm0, %xmm1
 ; X64-NEXT:    movl $2, %eax
 ; X64-NEXT:    vpinsrd $0, %eax, %xmm1, %xmm1