fix sse1 only codegen in x86-64 mode, which is something we
apparently try to support.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@112168 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index ad43175..c2e37b6 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -1342,12 +1342,18 @@
if (Subtarget->is64Bit()) {
if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
- if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
+ if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
ValToCopy);
+
+ // If we don't have SSE2 available, convert to v4f32 so the generated
+ // register is legal.
+ if (!Subtarget->hasSSE2())
+ ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32,ValToCopy);
+ }
}
}
-
+
Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
Flag = Chain.getValue(1);
}
@@ -3931,10 +3937,9 @@
}
}
- if (NumNonZero == 0) {
- // All undef vector. Return an UNDEF. All zero vectors were handled above.
+ // All undef vector. Return an UNDEF. All zero vectors were handled above.
+ if (NumNonZero == 0)
return DAG.getUNDEF(VT);
- }
// Special case for single non-zero, non-undef, element.
if (NumNonZero == 1) {
@@ -4072,7 +4077,7 @@
if (EVTBits == 16 && NumElems == 8) {
SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
- *this);
+ *this);
if (V.getNode()) return V;
}
diff --git a/test/CodeGen/X86/sse1.ll b/test/CodeGen/X86/sse1.ll
new file mode 100644
index 0000000..2c1acb6
--- /dev/null
+++ b/test/CodeGen/X86/sse1.ll
@@ -0,0 +1,7 @@
+; Tests for SSE1 and below, without SSE2+.
+; RUN: llc < %s -mcpu=pentium3 -O3 | FileCheck %s
+
+define <8 x i16> @test1(<8 x i32> %a) nounwind {
+; CHECK: test1
+ ret <8 x i16> zeroinitializer
+}
