[X86][AVX] Match broadcast loads through a bitcast

[oota-llvm.git] / test / CodeGen / X86 / fp-stack-ret.ll

diff --git a/test/CodeGen/X86/fp-stack-ret.ll b/test/CodeGen/X86/fp-stack-ret.ll
index 6ed8115945ca8099355b9f24e0622d0cb3583cf0..9635e2d2511a815ce7256c2c9b2d9535e098c704 100644 (file)
--- a/test/CodeGen/X86/fp-stack-ret.ll
+++ b/test/CodeGen/X86/fp-stack-ret.ll
@@ -1,25 +1,40 @@
-; RUN: llvm-as < %s | llc -mtriple=i686-apple-darwin8 -mcpu=yonah -march=x86 > %t &&
-; RUN: grep fldl %t | wc -l | grep 1 &&
-; RUN: not grep xmm %t &&
-; RUN: grep 'sub.*esp' %t | wc -l | grep 1
+; RUN: llc < %s -mtriple=i686-apple-darwin8 -mcpu=yonah -march=x86 | FileCheck %s
 
 ; These testcases shouldn't require loading into an XMM register then storing 
 ; to memory, then reloading into an FPStack reg.
 
+; CHECK: test1
+; CHECK: fldl
+; CHECK-NEXT: ret
 define double @test1(double *%P) {
-        %A = load double* %P
+        %A = load double, double* %P
         ret double %A
 }
 
-; fastcc should return a value 
+; fastcc should return a value
+; CHECK: test2
+; CHECK-NOT: xmm
+; CHECK: ret
 define fastcc double @test2(<2 x double> %A) {
        %B = extractelement <2 x double> %A, i32 0
        ret double %B
 }
 
+; CHECK: test3
+; CHECK: sub{{.*}}%esp
+; CHECK-NOT: xmm
 define fastcc double @test3(<4 x float> %A) {
        %B = bitcast <4 x float> %A to <2 x double>
        %C = call fastcc double @test2(<2 x double> %B)
        ret double %C
 }
-       
+
+; Clear the stack when not using a return value.
+; CHECK: test4
+; CHECK: call
+; CHECK: fstp
+; CHECK: ret
+define void @test4(double *%P) {
+  %A = call double @test1(double *%P)
+  ret void
+}