X86: When expanding PCMPGTQ to PCMPGTD we always want to compare the lower halves...

Take #2 on fixing PR15977.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@182486 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 1f35056f21efb08ab5c61a9e565a18859f429e4a..769f90172ba2685d1ff32b27afa4c5db755299dc 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -9358,12 +9358,19 @@ static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
       Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
 
       // Since SSE has no unsigned integer comparisons, we need to flip the sign
-      // bits of the inputs before performing those operations.
+      // bits of the inputs before performing those operations. The lower
+      // compare is always unsigned.
+      SDValue SB;
       if (FlipSigns) {
-        SDValue SB = DAG.getConstant(0x80000000U, MVT::v4i32);
-        Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
-        Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
+        SB = DAG.getConstant(0x80000000U, MVT::v4i32);
+      } else {
+        SDValue Sign = DAG.getConstant(0x80000000U, MVT::i32);
+        SDValue Zero = DAG.getConstant(0x00000000U, MVT::i32);
+        SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
+                         Sign, Zero, Sign, Zero);
       }
+      Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
+      Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
 
       // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
       SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);

diff --git a/test/CodeGen/X86/vec_compare.ll b/test/CodeGen/X86/vec_compare.ll
index 95324997868066427726a6472663b3645f1942f6..fd5c234bb1609df52095fea0c88aa809349cb01e 100644 (file)
--- a/test/CodeGen/X86/vec_compare.ll
+++ b/test/CodeGen/X86/vec_compare.ll
@@ -67,7 +67,15 @@ define <2 x i64> @test6(<2 x i64> %A, <2 x i64> %B) nounwind {
 }
 
 define <2 x i64> @test7(<2 x i64> %A, <2 x i64> %B) nounwind {
+; CHECK: [[CONSTSEG:[A-Z0-9_]*]]:
+; CHECK:      .long    2147483648
+; CHECK-NEXT: .long    0
+; CHECK-NEXT: .long    2147483648
+; CHECK-NEXT: .long    0
 ; CHECK: test7:
+; CHECK: movdqa [[CONSTSEG]], [[CONSTREG:%xmm[0-9]*]]
+; CHECK: pxor [[CONSTREG]]
+; CHECK: pxor [[CONSTREG]]
 ; CHECK: pcmpgtd %xmm1
 ; CHECK: pshufd $-96
 ; CHECK: pcmpeqd
@@ -83,6 +91,8 @@ define <2 x i64> @test7(<2 x i64> %A, <2 x i64> %B) nounwind {
 
 define <2 x i64> @test8(<2 x i64> %A, <2 x i64> %B) nounwind {
 ; CHECK: test8:
+; CHECK: pxor
+; CHECK: pxor
 ; CHECK: pcmpgtd %xmm0
 ; CHECK: pshufd $-96
 ; CHECK: pcmpeqd
@@ -98,6 +108,8 @@ define <2 x i64> @test8(<2 x i64> %A, <2 x i64> %B) nounwind {
 
 define <2 x i64> @test9(<2 x i64> %A, <2 x i64> %B) nounwind {
 ; CHECK: test9:
+; CHECK: pxor
+; CHECK: pxor
 ; CHECK: pcmpgtd %xmm0
 ; CHECK: pshufd $-96
 ; CHECK: pcmpeqd
@@ -115,6 +127,8 @@ define <2 x i64> @test9(<2 x i64> %A, <2 x i64> %B) nounwind {
 
 define <2 x i64> @test10(<2 x i64> %A, <2 x i64> %B) nounwind {
 ; CHECK: test10:
+; CHECK: pxor
+; CHECK: pxor
 ; CHECK: pcmpgtd %xmm1
 ; CHECK: pshufd $-96
 ; CHECK: pcmpeqd