Evan's patch to avoid FPreg->intreg copy for cvt; store to mem

llvm-svn: 36693

2 files changed, 38 insertions, 0 deletions

diff --git a/llvm/lib/Target/ARM/ARMInstrVFP.td b/llvm/lib/Target/ARM/ARMInstrVFP.td
index 671e1a48c18..a8a72ed72c8 100644
--- a/llvm/lib/Target/ARM/ARMInstrVFP.td
+++ b/llvm/lib/Target/ARM/ARMInstrVFP.td
@@ -78,6 +78,11 @@ def FSTS  : ASI5<(ops SPR:$src, addrmode5:$addr),
                  [(store SPR:$src, addrmode5:$addr)]>;
 } // isStore
 
+// avoid copying result of FTOUIZ to int reg when we're just storing it
+let AddedComplexity = 1 in
+def : ARMPat<(store (i32 (bitconvert SPR:$src)), addrmode5:$addr),
+             (FSTS SPR:$src, addrmode5:$addr)>;
+
 //===----------------------------------------------------------------------===//
 //  Load / store multiple Instructions.
 //
diff --git a/llvm/test/CodeGen/ARM/fptoint.ll b/llvm/test/CodeGen/ARM/fptoint.ll
new file mode 100644
index 00000000000..45be89c4d87
--- /dev/null
+++ b/llvm/test/CodeGen/ARM/fptoint.ll
@@ -0,0 +1,33 @@
+; RUN: llvm-as < %s | llc -march=arm -mattr=+v6,+vfp2 > %t
+; RUN: not grep fmrs %t
+
+@i = weak global i32 0		; <i32*> [#uses=2]
+@u = weak global i32 0		; <i32*> [#uses=2]
+
+define void @foo5(float %x) {
+entry:
+	%tmp1 = fptosi float %x to i32		; <i32> [#uses=1]
+	store i32 %tmp1, i32* @i
+	ret void
+}
+
+define void @foo6(float %x) {
+entry:
+	%tmp1 = fptoui float %x to i32		; <i32> [#uses=1]
+	store i32 %tmp1, i32* @u
+	ret void
+}
+
+define void @foo7(double %x) {
+entry:
+	%tmp1 = fptosi double %x to i32		; <i32> [#uses=1]
+	store i32 %tmp1, i32* @i
+	ret void
+}
+
+define void @foo8(double %x) {
+entry:
+	%tmp1 = fptoui double %x to i32		; <i32> [#uses=1]
+	store i32 %tmp1, i32* @u
+	ret void
+}