From: Craig Topper <craig.topper@intel.com>
Date: Sun, 8 Sep 2019 19:24:29 +0000 (+0000)
Subject: [X86] Teach materializeVectorConstant to not call getZeroVector/getOnesVector on... 
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=cbda742932a075a5a48ce0c0a6541d8d9d12cfbc;p=llvm

[X86] Teach materializeVectorConstant to not call getZeroVector/getOnesVector on the types we already have isel patterns for.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@371343 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 5e5ec5c6c8b..c7844220dfc 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -9137,7 +9137,8 @@ static SDValue materializeVectorConstant(SDValue Op, SelectionDAG &DAG,
   if (ISD::isBuildVectorAllZeros(Op.getNode())) {
     // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
     // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
-    if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
+    if (VT.isFloatingPoint() ||
+        VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
       return Op;
 
     return getZeroVector(VT, Subtarget, DAG, DL);
@@ -9147,8 +9148,7 @@ static SDValue materializeVectorConstant(SDValue Op, SelectionDAG &DAG,
   // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
   // vpcmpeqd on 256-bit vectors.
   if (Subtarget.hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
-    if (VT == MVT::v4i32 || VT == MVT::v16i32 ||
-        (VT == MVT::v8i32 && Subtarget.hasInt256()))
+    if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
       return Op;
 
     return getOnesVector(VT, DAG, DL);
diff --git a/test/CodeGen/X86/fold-load-vec.ll b/test/CodeGen/X86/fold-load-vec.ll
index 115f2bf7a5b..e8dc8f26ffa 100644
--- a/test/CodeGen/X86/fold-load-vec.ll
+++ b/test/CodeGen/X86/fold-load-vec.ll
@@ -12,7 +12,7 @@ define void @sample_test(<4 x float>* %source, <2 x float>* %dest) nounwind {
 ; CHECK-NEXT:    movq %rsi, {{[0-9]+}}(%rsp)
 ; CHECK-NEXT:    xorps %xmm0, %xmm0
 ; CHECK-NEXT:    movlps %xmm0, (%rsp)
-; CHECK-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[2,3]
+; CHECK-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
 ; CHECK-NEXT:    movlps %xmm0, (%rsp)
 ; CHECK-NEXT:    movlps %xmm0, (%rsi)
 ; CHECK-NEXT:    movq {{[0-9]+}}(%rsp), %rax