[llvm-branch-commits] [llvm-branch] r204648 - Merging r201097:
Tom Stellard
thomas.stellard at amd.com
Mon Mar 24 11:21:41 PDT 2014
Author: tstellar
Date: Mon Mar 24 13:21:41 2014
New Revision: 204648
URL: http://llvm.org/viewvc/llvm-project?rev=204648&view=rev
Log:
Merging r201097:
------------------------------------------------------------------------
r201097 | thomas.stellard | 2014-02-10 08:58:30 -0800 (Mon, 10 Feb 2014) | 9 lines
R600/SI: Initialize M0 and emit S_WQM_B64 whenever DS instructions are
used
DS instructions that access local memory can only uses addresses that
are less than or equal to the value of M0. When M0 is uninitialized,
then we experience undefined behavior.
This patch also changes the behavior to emit S_WQM_B64 on pixel shaders
no matter what kind of DS instruction is used.
Modified:
llvm/branches/release_34/lib/Target/R600/SILowerControlFlow.cpp
llvm/branches/release_34/test/CodeGen/R600/load.ll
Modified: llvm/branches/release_34/lib/Target/R600/SILowerControlFlow.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/release_34/lib/Target/R600/SILowerControlFlow.cpp?rev=204648&r1=204647&r2=204648&view=diff
==============================================================================
--- llvm/branches/release_34/lib/Target/R600/SILowerControlFlow.cpp (original)
+++ llvm/branches/release_34/lib/Target/R600/SILowerControlFlow.cpp Mon Mar 24 13:21:41 2014
@@ -109,6 +109,23 @@ FunctionPass *llvm::createSILowerControl
return new SILowerControlFlowPass(tm);
}
+static bool isDS(unsigned Opcode) {
+ switch(Opcode) {
+ default: return false;
+ case AMDGPU::DS_ADD_U32_RTN:
+ case AMDGPU::DS_SUB_U32_RTN:
+ case AMDGPU::DS_WRITE_B32:
+ case AMDGPU::DS_WRITE_B8:
+ case AMDGPU::DS_WRITE_B16:
+ case AMDGPU::DS_READ_B32:
+ case AMDGPU::DS_READ_I8:
+ case AMDGPU::DS_READ_U8:
+ case AMDGPU::DS_READ_I16:
+ case AMDGPU::DS_READ_U16:
+ return true;
+ }
+}
+
bool SILowerControlFlowPass::shouldSkip(MachineBasicBlock *From,
MachineBasicBlock *To) {
@@ -435,6 +452,11 @@ bool SILowerControlFlowPass::runOnMachin
Next = llvm::next(I);
MachineInstr &MI = *I;
+ if (isDS(MI.getOpcode())) {
+ NeedM0 = true;
+ NeedWQM = true;
+ }
+
switch (MI.getOpcode()) {
default: break;
case AMDGPU::SI_IF:
@@ -495,14 +517,6 @@ bool SILowerControlFlowPass::runOnMachin
IndirectDst(MI);
break;
- case AMDGPU::DS_READ_B32:
- NeedWQM = true;
- // Fall through
- case AMDGPU::DS_WRITE_B32:
- case AMDGPU::DS_ADD_U32_RTN:
- NeedM0 = true;
- break;
-
case AMDGPU::V_INTERP_P1_F32:
case AMDGPU::V_INTERP_P2_F32:
case AMDGPU::V_INTERP_MOV_F32:
Modified: llvm/branches/release_34/test/CodeGen/R600/load.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/branches/release_34/test/CodeGen/R600/load.ll?rev=204648&r1=204647&r2=204648&view=diff
==============================================================================
--- llvm/branches/release_34/test/CodeGen/R600/load.ll (original)
+++ llvm/branches/release_34/test/CodeGen/R600/load.ll Mon Mar 24 13:21:41 2014
@@ -445,6 +445,7 @@ define void @load_const_addrspace_f32(fl
; R600-CHECK: LDS_UBYTE_READ_RET
; SI-CHECK-LABEL: @load_i8_local
; SI-CHECK-NOT: S_WQM_B64
+; SI-CHECK: S_MOV_B32 m0
; SI-CHECK: DS_READ_U8
define void @load_i8_local(i32 addrspace(1)* %out, i8 addrspace(3)* %in) {
%1 = load i8 addrspace(3)* %in
@@ -458,6 +459,7 @@ define void @load_i8_local(i32 addrspace
; R600-CHECK: ASHR
; SI-CHECK-LABEL: @load_i8_sext_local
; SI-CHECK-NOT: S_WQM_B64
+; SI-CHECK: S_MOV_B32 m0
; SI-CHECK: DS_READ_I8
define void @load_i8_sext_local(i32 addrspace(1)* %out, i8 addrspace(3)* %in) {
entry:
@@ -472,6 +474,7 @@ entry:
; R600-CHECK: LDS_UBYTE_READ_RET
; SI-CHECK-LABEL: @load_v2i8_local
; SI-CHECK-NOT: S_WQM_B64
+; SI-CHECK: S_MOV_B32 m0
; SI-CHECK: DS_READ_U8
; SI-CHECK: DS_READ_U8
define void @load_v2i8_local(<2 x i32> addrspace(1)* %out, <2 x i8> addrspace(3)* %in) {
@@ -489,6 +492,7 @@ entry:
; R600-CHECK-DAG: ASHR
; SI-CHECK-LABEL: @load_v2i8_sext_local
; SI-CHECK-NOT: S_WQM_B64
+; SI-CHECK: S_MOV_B32 m0
; SI-CHECK: DS_READ_I8
; SI-CHECK: DS_READ_I8
define void @load_v2i8_sext_local(<2 x i32> addrspace(1)* %out, <2 x i8> addrspace(3)* %in) {
@@ -506,6 +510,7 @@ entry:
; R600-CHECK: LDS_UBYTE_READ_RET
; SI-CHECK-LABEL: @load_v4i8_local
; SI-CHECK-NOT: S_WQM_B64
+; SI-CHECK: S_MOV_B32 m0
; SI-CHECK: DS_READ_U8
; SI-CHECK: DS_READ_U8
; SI-CHECK: DS_READ_U8
@@ -529,6 +534,7 @@ entry:
; R600-CHECK-DAG: ASHR
; SI-CHECK-LABEL: @load_v4i8_sext_local
; SI-CHECK-NOT: S_WQM_B64
+; SI-CHECK: S_MOV_B32 m0
; SI-CHECK: DS_READ_I8
; SI-CHECK: DS_READ_I8
; SI-CHECK: DS_READ_I8
@@ -546,6 +552,7 @@ entry:
; R600-CHECK: LDS_USHORT_READ_RET
; SI-CHECK-LABEL: @load_i16_local
; SI-CHECK-NOT: S_WQM_B64
+; SI-CHECK: S_MOV_B32 m0
; SI-CHECK: DS_READ_U16
define void @load_i16_local(i32 addrspace(1)* %out, i16 addrspace(3)* %in) {
entry:
@@ -560,6 +567,7 @@ entry:
; R600-CHECK: ASHR
; SI-CHECK-LABEL: @load_i16_sext_local
; SI-CHECK-NOT: S_WQM_B64
+; SI-CHECK: S_MOV_B32 m0
; SI-CHECK: DS_READ_I16
define void @load_i16_sext_local(i32 addrspace(1)* %out, i16 addrspace(3)* %in) {
entry:
@@ -574,6 +582,7 @@ entry:
; R600-CHECK: LDS_USHORT_READ_RET
; SI-CHECK-LABEL: @load_v2i16_local
; SI-CHECK-NOT: S_WQM_B64
+; SI-CHECK: S_MOV_B32 m0
; SI-CHECK: DS_READ_U16
; SI-CHECK: DS_READ_U16
define void @load_v2i16_local(<2 x i32> addrspace(1)* %out, <2 x i16> addrspace(3)* %in) {
@@ -591,6 +600,7 @@ entry:
; R600-CHECK-DAG: ASHR
; SI-CHECK-LABEL: @load_v2i16_sext_local
; SI-CHECK-NOT: S_WQM_B64
+; SI-CHECK: S_MOV_B32 m0
; SI-CHECK: DS_READ_I16
; SI-CHECK: DS_READ_I16
define void @load_v2i16_sext_local(<2 x i32> addrspace(1)* %out, <2 x i16> addrspace(3)* %in) {
@@ -608,6 +618,7 @@ entry:
; R600-CHECK: LDS_USHORT_READ_RET
; SI-CHECK-LABEL: @load_v4i16_local
; SI-CHECK-NOT: S_WQM_B64
+; SI-CHECK: S_MOV_B32 m0
; SI-CHECK: DS_READ_U16
; SI-CHECK: DS_READ_U16
; SI-CHECK: DS_READ_U16
@@ -631,6 +642,7 @@ entry:
; R600-CHECK-DAG: ASHR
; SI-CHECK-LABEL: @load_v4i16_sext_local
; SI-CHECK-NOT: S_WQM_B64
+; SI-CHECK: S_MOV_B32 m0
; SI-CHECK: DS_READ_I16
; SI-CHECK: DS_READ_I16
; SI-CHECK: DS_READ_I16
@@ -643,11 +655,12 @@ entry:
ret void
}
-; load an i32 value from the glocal address space.
+; load an i32 value from the local address space.
; R600-CHECK-LABEL: @load_i32_local
; R600-CHECK: LDS_READ_RET
; SI-CHECK-LABEL: @load_i32_local
; SI-CHECK-NOT: S_WQM_B64
+; SI-CHECK: S_MOV_B32 m0
; SI-CHECK: DS_READ_B32
define void @load_i32_local(i32 addrspace(1)* %out, i32 addrspace(3)* %in) {
entry:
@@ -656,10 +669,11 @@ entry:
ret void
}
-; load a f32 value from the global address space.
+; load a f32 value from the local address space.
; R600-CHECK-LABEL: @load_f32_local
; R600-CHECK: LDS_READ_RET
; SI-CHECK-LABEL: @load_f32_local
+; SI-CHECK: S_MOV_B32 m0
; SI-CHECK: DS_READ_B32
define void @load_f32_local(float addrspace(1)* %out, float addrspace(3)* %in) {
entry:
@@ -673,6 +687,7 @@ entry:
; R600-CHECK: LDS_READ_RET
; R600-CHECK: LDS_READ_RET
; SI-CHECK-LABEL: @load_v2f32_local
+; SI-CHECK: S_MOV_B32 m0
; SI-CHECK: DS_READ_B32
; SI-CHECK: DS_READ_B32
define void @load_v2f32_local(<2 x float> addrspace(1)* %out, <2 x float> addrspace(3)* %in) {
More information about the llvm-branch-commits
mailing list