Use movaps to load a v4f32 build_vector of all-constant values into a
register instead of loading each element individually.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40478 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index a25710b..23f9e95 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -2495,6 +2495,7 @@
unsigned NumZero = 0;
unsigned NumNonZero = 0;
unsigned NonZeros = 0;
+ unsigned NumNonZeroImms = 0;
std::set<SDOperand> Values;
for (unsigned i = 0; i < NumElems; ++i) {
SDOperand Elt = Op.getOperand(i);
@@ -2505,6 +2506,9 @@
else {
NonZeros |= (1 << i);
NumNonZero++;
+ if (Elt.getOpcode() == ISD::Constant ||
+ Elt.getOpcode() == ISD::ConstantFP)
+ NumNonZeroImms++;
}
}
}
@@ -2548,6 +2552,11 @@
}
}
+ // A vector full of immediates; various special cases are already
+ // handled, so this is best done with a single constant-pool load.
+ if (NumNonZero == NumNonZeroImms)
+ return SDOperand();
+
// Let legalizer expand 2-wide build_vectors.
if (EVTBits == 64)
return SDOperand();
diff --git a/test/CodeGen/X86/v4f32-immediate.ll b/test/CodeGen/X86/v4f32-immediate.ll
new file mode 100644
index 0000000..67b5e79
--- /dev/null
+++ b/test/CodeGen/X86/v4f32-immediate.ll
@@ -0,0 +1,5 @@
+; RUN: llvm-as < %s | llc -march=x86 -mattr=+sse | grep movaps
+
+define <4 x float> @foo() {
+ ret <4 x float> <float 3.223542354, float 2.3, float 1.2, float 0.1>
+}