aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorTom Stellard <thomas.stellard@amd.com>2013-11-18 18:50:15 +0000
committerTom Stellard <thomas.stellard@amd.com>2013-11-18 18:50:15 +0000
commitbf9ddd5e8ff207e9b9a1cd484242e8d460f3834e (patch)
tree1ddea3680a13a87a160d55ce337c83bb7292f329
parent7c5498ee55e241dea91cca022bd9a5831a1ba4b1 (diff)
downloadexternal_llvm-bf9ddd5e8ff207e9b9a1cd484242e8d460f3834e.zip
external_llvm-bf9ddd5e8ff207e9b9a1cd484242e8d460f3834e.tar.gz
external_llvm-bf9ddd5e8ff207e9b9a1cd484242e8d460f3834e.tar.bz2
R600/SI: Fix another case of illegal VGPR->SGPR copy
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@195025 91177308-0d34-0410-b5e6-96231b3b80d8
-rw-r--r--lib/Target/R600/SIFixSGPRCopies.cpp5
-rw-r--r--test/CodeGen/R600/sgpr-copy.ll26
2 files changed, 28 insertions, 3 deletions
diff --git a/lib/Target/R600/SIFixSGPRCopies.cpp b/lib/Target/R600/SIFixSGPRCopies.cpp
index 6116b4e..b49fda9 100644
--- a/lib/Target/R600/SIFixSGPRCopies.cpp
+++ b/lib/Target/R600/SIFixSGPRCopies.cpp
@@ -181,14 +181,13 @@ bool SIFixSGPRCopies::isVGPRToSGPRCopy(const MachineInstr &Copy,
unsigned SrcReg = Copy.getOperand(1).getReg();
unsigned SrcSubReg = Copy.getOperand(1).getSubReg();
const TargetRegisterClass *DstRC = MRI.getRegClass(DstReg);
+ const TargetRegisterClass *SrcRC;
if (!TargetRegisterInfo::isVirtualRegister(SrcReg) ||
DstRC == &AMDGPU::M0RegRegClass)
return false;
- const TargetRegisterClass *SrcRC = TRI->getSubRegClass(
- MRI.getRegClass(SrcReg), SrcSubReg);
-
+ SrcRC = inferRegClassFromDef(TRI, MRI, SrcReg, SrcSubReg);
return TRI->isSGPRClass(DstRC) &&
!TRI->getCommonSubClass(DstRC, SrcRC);
}
diff --git a/test/CodeGen/R600/sgpr-copy.ll b/test/CodeGen/R600/sgpr-copy.ll
index 67cedd3..5ec1a08 100644
--- a/test/CodeGen/R600/sgpr-copy.ll
+++ b/test/CodeGen/R600/sgpr-copy.ll
@@ -268,3 +268,29 @@ endif:
}
!2 = metadata !{metadata !"const", null, i32 1}
+
+; CHECK-LABEL: @copy1
+; CHECK: BUFFER_LOAD_DWORD
+; CHECK: V_ADD
+; CHECK: S_ENDPGM
+define void @copy1(float addrspace(1)* %out, float addrspace(1)* %in0) {
+entry:
+ %0 = load float addrspace(1)* %in0
+ %1 = fcmp oeq float %0, 0.0
+ br i1 %1, label %if0, label %endif
+
+if0:
+ %2 = bitcast float %0 to i32
+ %3 = fcmp olt float %0, 0.0
+ br i1 %3, label %if1, label %endif
+
+if1:
+ %4 = add i32 %2, 1
+ br label %endif
+
+endif:
+ %5 = phi i32 [ 0, %entry ], [ %2, %if0 ], [ %4, %if1 ]
+ %6 = bitcast i32 %5 to float
+ store float %6, float addrspace(1)* %out
+ ret void
+}