R600: Take alignment into account when calculating the stack offset

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@199826 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/R600/AMDGPUFrameLowering.cpp b/lib/Target/R600/AMDGPUFrameLowering.cpp
index 40cc9083c0056dfce74cee4fbf4e8902c0ac8763..0325a00c17728b3fbe972a674663f438c94c1067 100644 (file)
--- a/lib/Target/R600/AMDGPUFrameLowering.cpp
+++ b/lib/Target/R600/AMDGPUFrameLowering.cpp
@@ -77,14 +77,21 @@ int AMDGPUFrameLowering::getFrameIndexOffset(const MachineFunction &MF,
   // Start the offset at 2 so we don't overwrite work group information.
   // XXX: We should only do this when the shader actually uses this
   // information.
-  unsigned Offset = 2;
+  unsigned OffsetBytes = 2 * (getStackWidth(MF) * 4);
   int UpperBound = FI == -1 ? MFI->getNumObjects() : FI;
 
   for (int i = MFI->getObjectIndexBegin(); i < UpperBound; ++i) {
-    unsigned Size = MFI->getObjectSize(i);
-    Offset += (Size / (getStackWidth(MF) * 4));
+    OffsetBytes = RoundUpToAlignment(OffsetBytes, MFI->getObjectAlignment(i));
+    OffsetBytes += MFI->getObjectSize(i);
+    // Each regiter holds 4 bytes, so we must always align the offset to at
+    // least 4 bytes, so that 2 frame objects won't share the same register.
+    OffsetBytes = RoundUpToAlignment(OffsetBytes, 4);
   }
-  return Offset;
+
+  if (FI != -1)
+    OffsetBytes = RoundUpToAlignment(OffsetBytes, MFI->getObjectAlignment(FI));
+
+  return OffsetBytes / (getStackWidth(MF) * 4);
 }
 
 const TargetFrameLowering::SpillSlot *

diff --git a/test/CodeGen/R600/private-memory.ll b/test/CodeGen/R600/private-memory.ll
index b25fc7ba572c812b25b207110212bcda689c46f0..1bd17bf6048280499170039c97445be7404b5e9d 100644 (file)
--- a/test/CodeGen/R600/private-memory.ll
+++ b/test/CodeGen/R600/private-memory.ll
@@ -180,4 +180,35 @@ entry:
   ret void
 }
 
+; Test that two stack objects are not stored in the same register
+; The second stack object should be in T3.X
+; FUNC-LABEL: @no_overlap
+; R600-CHECK: MOV {{\** *}}T3.X
+; SI-CHECK: V_MOV_B32_e32 v3
+define void @no_overlap(i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %0 = alloca [3 x i8], align 1
+  %1 = alloca [2 x i8], align 1
+  %2 = getelementptr [3 x i8]* %0, i32 0, i32 0
+  %3 = getelementptr [3 x i8]* %0, i32 0, i32 1
+  %4 = getelementptr [3 x i8]* %0, i32 0, i32 2
+  %5 = getelementptr [2 x i8]* %1, i32 0, i32 0
+  %6 = getelementptr [2 x i8]* %1, i32 0, i32 1
+  store i8 0, i8* %2
+  store i8 1, i8* %3
+  store i8 2, i8* %4
+  store i8 1, i8* %5
+  store i8 0, i8* %6
+  %7 = getelementptr [3 x i8]* %0, i32 0, i32 %in
+  %8 = getelementptr [2 x i8]* %1, i32 0, i32 %in
+  %9 = load i8* %7
+  %10 = load i8* %8
+  %11 = add i8 %9, %10
+  %12 = sext i8 %11 to i32
+  store i32 %12, i32 addrspace(1)* %out
+  ret void
+}
+
+
+
 declare i32 @llvm.r600.read.tidig.x() nounwind readnone