RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-23 11:13:28 +01:00
Code Issues Projects Releases Wiki Activity

AMDGPU: Don't track lgkmcnt for global_/scratch_ instructions

Browse Source 
llvm-svn: 308766
This commit is contained in:
Matt Arsenault 2017-07-21 18:34:51 +00:00
parent 77d8531464
commit 9193402adb
3 changed files with 17 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
lib/Target/AMDGPU/FLATInstructions.td
View File

@ -25,11 +25,6 @@ class FLAT_Pseudo<string opName, dag outs, dag ins,
let SubtargetPredicate = isCIVI;
let FLAT = 1;
// Internally, FLAT instruction are executed as both an LDS and a
// Buffer instruction; so, they increment both VM_CNT and LGKM_CNT
// and are not considered done until both have been decremented.
let VM_CNT = 1;
let LGKM_CNT = 1;
let UseNamedOperandTable = 1;
let hasSideEffects = 0;
@ -59,6 +54,12 @@ class FLAT_Pseudo<string opName, dag outs, dag ins,
// TODO: M0 if it could possibly access LDS (before gfx9? only)?
let Uses = !if(is_flat_global, [EXEC], [EXEC, FLAT_SCR]);
// Internally, FLAT instruction are executed as both an LDS and a
// Buffer instruction; so, they increment both VM_CNT and LGKM_CNT
// and are not considered done until both have been decremented.
let VM_CNT = 1;
let LGKM_CNT = !if(!or(is_flat_global, is_flat_scratch), 0, 1);
}
class FLAT_Real <bits<7> op, FLAT_Pseudo ps> :

11
lib/Target/AMDGPU/SIInsertWaitcnts.cpp
View File

@ -1151,8 +1151,7 @@ void SIInsertWaitcnts::updateEventWaitCntAfter(
// instruction, update the upper-bound of the appropriate counter's
// bracket and the destination operand scores.
// TODO: Use the (TSFlags & SIInstrFlags::LGKM_CNT) property everywhere.
uint64_t TSFlags = Inst.getDesc().TSFlags;
if (TII->isDS(Inst) && (TSFlags & SIInstrFlags::LGKM_CNT)) {
if (TII->isDS(Inst) && TII->usesLGKM_CNT(Inst)) {
if (TII->getNamedOperand(Inst, AMDGPU::OpName::gds) &&
TII->getNamedOperand(Inst, AMDGPU::OpName::gds)->getImm() != 0) {
ScoreBrackets->updateByEvent(TII, TRI, MRI, GDS_ACCESS, Inst);
@ -1162,8 +1161,12 @@ void SIInsertWaitcnts::updateEventWaitCntAfter(
}
} else if (TII->isFLAT(Inst)) {
assert(Inst.mayLoad() || Inst.mayStore());
ScoreBrackets->updateByEvent(TII, TRI, MRI, VMEM_ACCESS, Inst);
ScoreBrackets->updateByEvent(TII, TRI, MRI, LDS_ACCESS, Inst);
if (TII->usesVM_CNT(Inst))
ScoreBrackets->updateByEvent(TII, TRI, MRI, VMEM_ACCESS, Inst);
if (TII->usesLGKM_CNT(Inst))
ScoreBrackets->updateByEvent(TII, TRI, MRI, LDS_ACCESS, Inst);
// This is a flat memory operation. Check to see if it has memory
// tokens for both LDS and Memory, and if so mark it as a flat.

4
lib/Target/AMDGPU/SIInstrInfo.h
View File

@ -496,6 +496,10 @@ public:
return MI.getDesc().TSFlags & SIInstrFlags::VM_CNT;
}
static bool usesLGKM_CNT(const MachineInstr &MI) {
return MI.getDesc().TSFlags & SIInstrFlags::LGKM_CNT;
}
static bool sopkIsZext(const MachineInstr &MI) {
return MI.getDesc().TSFlags & SIInstrFlags::SOPK_ZEXT;
}