maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

R600: Begin private memory at the second GPR. 

This way private memory does not over-write work group information
stored in GPRs 0 and 1.

llvm-svn: 199824
This commit is contained in:
Tom Stellard 2014-01-22 19:24:19 +00:00
parent e93736057f
commit 27982b1d4a
2 changed files with 29 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
llvm/lib/Target/R600/AMDGPUFrameLowering.cpp
View File

@ -74,7 +74,10 @@ unsigned AMDGPUFrameLowering::getStackWidth(const MachineFunction &MF) const {
int AMDGPUFrameLowering::getFrameIndexOffset(const MachineFunction &MF, int AMDGPUFrameLowering::getFrameIndexOffset(const MachineFunction &MF,
int FI) const { int FI) const {
const MachineFrameInfo *MFI = MF.getFrameInfo(); const MachineFrameInfo *MFI = MF.getFrameInfo();
unsigned Offset = 0; // 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;
int UpperBound = FI == -1 ? MFI->getNumObjects() : FI; int UpperBound = FI == -1 ? MFI->getNumObjects() : FI;
for (int i = MFI->getObjectIndexBegin(); i < UpperBound; ++i) { for (int i = MFI->getObjectIndexBegin(); i < UpperBound; ++i) {

25
llvm/test/CodeGen/R600/private-memory.ll
View File

@ -156,3 +156,28 @@ entry:
ret void ret void
} }
; Make sure we don't overwrite workitem information with private memory
; FUNC-LABEL: @work_item_info
; R600-CHECK-NOT: MOV T0.X
; Additional check in case the move ends up in the last slot
; R600-CHECK-NOT: MOV * TO.X
; SI-CHECK-NOT: V_MOV_B32_e{{(32|64)}} v0
define void @work_item_info(i32 addrspace(1)* %out, i32 %in) {
entry:
%0 = alloca [2 x i32]
%1 = getelementptr [2 x i32]* %0, i32 0, i32 0
%2 = getelementptr [2 x i32]* %0, i32 0, i32 1
store i32 0, i32* %1
store i32 1, i32* %2
%3 = getelementptr [2 x i32]* %0, i32 0, i32 %in
%4 = load i32* %3
%5 = call i32 @llvm.r600.read.tidig.x()
%6 = add i32 %4, %5
store i32 %6, i32 addrspace(1)* %out
ret void
}
declare i32 @llvm.r600.read.tidig.x() nounwind readnone