[llvm] [AMDGPU][SIInsertWaitcnts][NFC] Drop MaxCounter member variable (PR #181092)
via llvm-commits
llvm-commits at lists.llvm.org
Mon Feb 16 11:18:51 PST 2026
https://github.com/vporpo updated https://github.com/llvm/llvm-project/pull/181092
>From 3507aa0c95d4b2ae4eab1552a4a9cd6cec03e577 Mon Sep 17 00:00:00 2001
From: Vasileios Porpodas <vasileios.porpodas at amd.com>
Date: Thu, 12 Feb 2026 03:18:41 +0000
Subject: [PATCH 1/4] [AMDGPU][SIInsertWaitcnts][NFC] Drop MaxCounter member
variable
IsExpertMode and MaxCounter can be trivially computed from MF and ST.
So keeping them as member variables is kind of redundant.
Replacing them with functions that compute them from MF and ST help reduce
the amount of state in the class.
---
llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp | 87 ++++++++++++---------
1 file changed, 49 insertions(+), 38 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
index 111867583fde3..4251916f94c45 100644
--- a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
@@ -194,6 +194,22 @@ template <> struct enum_iteration_traits<WaitEventType> {
namespace {
+bool isExpertMode(const MachineFunction &MF, const GCNSubtarget &ST) {
+ return ST.hasExpertSchedulingMode() &&
+ (ExpertSchedulingModeFlag.getNumOccurrences()
+ ? ExpertSchedulingModeFlag
+ : MF.getFunction()
+ .getFnAttribute("amdgpu-expert-scheduling-mode")
+ .getValueAsBool());
+
+}
+
+InstCounterType getMaxCounter(const GCNSubtarget &ST, bool IsExpertMode) {
+ if (ST.hasExtendedWaitCounts())
+ return IsExpertMode ? NUM_EXPERT_INST_CNTS : NUM_EXTENDED_INST_CNTS;
+ return NUM_NORMAL_INST_CNTS;
+}
+
/// Return an iterator over all events between VMEM_ACCESS (the first event)
/// and \c MaxEvent (exclusive, default value yields an enumeration over
/// all counters).
@@ -354,23 +370,23 @@ class WaitcntGenerator {
const GCNSubtarget &ST;
const SIInstrInfo &TII;
AMDGPU::IsaVersion IV;
- InstCounterType MaxCounter;
bool OptNone;
bool ExpandWaitcntProfiling = false;
const AMDGPU::HardwareLimits *Limits = nullptr;
+ bool IsExpertMode = false;
public:
WaitcntGenerator() = delete;
WaitcntGenerator(const WaitcntGenerator &) = delete;
- WaitcntGenerator(const MachineFunction &MF, InstCounterType MaxCounter,
+ WaitcntGenerator(const MachineFunction &MF,
const AMDGPU::HardwareLimits *Limits)
: ST(MF.getSubtarget<GCNSubtarget>()), TII(*ST.getInstrInfo()),
- IV(AMDGPU::getIsaVersion(ST.getCPU())), MaxCounter(MaxCounter),
+ IV(AMDGPU::getIsaVersion(ST.getCPU())),
OptNone(MF.getFunction().hasOptNone() ||
MF.getTarget().getOptLevel() == CodeGenOptLevel::None),
ExpandWaitcntProfiling(
MF.getFunction().hasFnAttribute("amdgpu-expand-waitcnt-profiling")),
- Limits(Limits) {}
+ Limits(Limits), IsExpertMode(isExpertMode(MF, ST)) {}
// Return true if the current function should be compiled with no
// optimization.
@@ -441,7 +457,9 @@ class WaitcntGeneratorPreGFX12 final : public WaitcntGenerator {
WaitEventSet()};
public:
- using WaitcntGenerator::WaitcntGenerator;
+ WaitcntGeneratorPreGFX12(MachineFunction *MF,
+ const AMDGPU::HardwareLimits *Limits)
+ : WaitcntGenerator(*MF, Limits) {}
bool
applyPreexistingWaitcnt(WaitcntBrackets &ScoreBrackets,
MachineInstr &OldWaitcntInstr, AMDGPU::Waitcnt &Wait,
@@ -461,7 +479,6 @@ class WaitcntGeneratorPreGFX12 final : public WaitcntGenerator {
class WaitcntGeneratorGFX12Plus final : public WaitcntGenerator {
protected:
- bool IsExpertMode;
static constexpr const WaitEventSet
WaitEventMaskForInstGFX12Plus[NUM_INST_CNTS] = {
WaitEventSet({VMEM_ACCESS, GLOBAL_INV_ACCESS}),
@@ -480,10 +497,8 @@ class WaitcntGeneratorGFX12Plus final : public WaitcntGenerator {
public:
WaitcntGeneratorGFX12Plus() = delete;
WaitcntGeneratorGFX12Plus(const MachineFunction &MF,
- InstCounterType MaxCounter,
- const AMDGPU::HardwareLimits *Limits,
- bool IsExpertMode)
- : WaitcntGenerator(MF, MaxCounter, Limits), IsExpertMode(IsExpertMode) {}
+ const AMDGPU::HardwareLimits *Limits)
+ : WaitcntGenerator(MF, Limits) {}
bool
applyPreexistingWaitcnt(WaitcntBrackets &ScoreBrackets,
@@ -515,8 +530,7 @@ class SIInsertWaitcnts {
const SIRegisterInfo *TRI = nullptr;
const MachineRegisterInfo *MRI = nullptr;
InstCounterType SmemAccessCounter;
- InstCounterType MaxCounter;
- bool IsExpertMode = false;
+ const MachineFunction *MF = nullptr;
private:
DenseMap<const Value *, MachineBasicBlock *> SLoadAddresses;
@@ -1067,7 +1081,9 @@ bool WaitcntBrackets::hasPointSamplePendingVmemTypes(const MachineInstr &MI,
void WaitcntBrackets::updateByEvent(WaitEventType E, MachineInstr &Inst) {
InstCounterType T = Context->getCounterFromEvent(E);
- assert(T < Context->MaxCounter);
+ assert(T < getMaxCounter(*Context->ST,
+ isExpertMode(*Context->MF, *Context->ST)) &&
+ "T not supported in current mode!");
unsigned UB = getScoreUB(T);
unsigned CurrScore = UB + 1;
@@ -1287,7 +1303,8 @@ void WaitcntBrackets::recordAsyncMark(MachineInstr &Inst) {
void WaitcntBrackets::print(raw_ostream &OS) const {
const GCNSubtarget *ST = Context->ST;
- for (auto T : inst_counter_types(Context->MaxCounter)) {
+ for (auto T : inst_counter_types(
+ getMaxCounter(*ST, isExpertMode(*Context->MF, *Context->ST)))) {
unsigned SR = getScoreRange(T);
switch (T) {
case LOAD_CNT:
@@ -1745,7 +1762,7 @@ bool WaitcntGenerator::promoteSoftWaitCnt(MachineInstr *Waitcnt) const {
bool WaitcntGeneratorPreGFX12::applyPreexistingWaitcnt(
WaitcntBrackets &ScoreBrackets, MachineInstr &OldWaitcntInstr,
AMDGPU::Waitcnt &Wait, MachineBasicBlock::instr_iterator It) const {
- assert(isNormalMode(MaxCounter));
+ assert(isNormalMode(getMaxCounter(ST, IsExpertMode)));
bool Modified = false;
MachineInstr *WaitcntInstr = nullptr;
@@ -1867,7 +1884,7 @@ bool WaitcntGeneratorPreGFX12::applyPreexistingWaitcnt(
bool WaitcntGeneratorPreGFX12::createNewWaitcnt(
MachineBasicBlock &Block, MachineBasicBlock::instr_iterator It,
AMDGPU::Waitcnt Wait, const WaitcntBrackets &ScoreBrackets) {
- assert(isNormalMode(MaxCounter));
+ assert(isNormalMode(getMaxCounter(ST, IsExpertMode)));
bool Modified = false;
const DebugLoc &DL = Block.findDebugLoc(It);
@@ -1984,7 +2001,7 @@ WaitcntGeneratorGFX12Plus::getAllZeroWaitcnt(bool IncludeVSCnt) const {
bool WaitcntGeneratorGFX12Plus::applyPreexistingWaitcnt(
WaitcntBrackets &ScoreBrackets, MachineInstr &OldWaitcntInstr,
AMDGPU::Waitcnt &Wait, MachineBasicBlock::instr_iterator It) const {
- assert(!isNormalMode(MaxCounter));
+ assert(!isNormalMode(getMaxCounter(ST, IsExpertMode)));
bool Modified = false;
MachineInstr *CombinedLoadDsCntInstr = nullptr;
@@ -2258,7 +2275,7 @@ bool WaitcntGeneratorGFX12Plus::applyPreexistingWaitcnt(
bool WaitcntGeneratorGFX12Plus::createNewWaitcnt(
MachineBasicBlock &Block, MachineBasicBlock::instr_iterator It,
AMDGPU::Waitcnt Wait, const WaitcntBrackets &ScoreBrackets) {
- assert(!isNormalMode(MaxCounter));
+ assert(!isNormalMode(getMaxCounter(ST, IsExpertMode)));
bool Modified = false;
const DebugLoc &DL = Block.findDebugLoc(It);
@@ -2825,7 +2842,7 @@ void SIInsertWaitcnts::updateEventWaitcntAfter(MachineInstr &Inst,
bool IsVMEMAccess = false;
bool IsSMEMAccess = false;
- if (IsExpertMode) {
+ if (isExpertMode(*MF, *ST)) {
if (const auto ET = getExpertSchedulingEventType(Inst))
ScoreBrackets->updateByEvent(*ET, Inst);
}
@@ -2999,7 +3016,8 @@ bool WaitcntBrackets::mergeAsyncMarks(ArrayRef<MergeInfo> MergeInfos,
unsigned MergeCount = std::min(OtherSize, OurSize);
assert(OurSize == MaxSize);
for (unsigned Idx = 1; Idx <= MergeCount; ++Idx) {
- for (auto T : inst_counter_types(Context->MaxCounter)) {
+ for (auto T : inst_counter_types(getMaxCounter(
+ *Context->ST, isExpertMode(*Context->MF, *Context->ST)))) {
StrictDom |= mergeScore(MergeInfos[T], AsyncMarks[OurSize - Idx][T],
OtherMarks[OtherSize - Idx][T]);
}
@@ -3034,7 +3052,8 @@ bool WaitcntBrackets::merge(const WaitcntBrackets &Other) {
// Array to store MergeInfo for each counter type
MergeInfo MergeInfos[NUM_INST_CNTS];
- for (auto T : inst_counter_types(Context->MaxCounter)) {
+ for (auto T : inst_counter_types(getMaxCounter(
+ *Context->ST, isExpertMode(*Context->MF, *Context->ST)))) {
// Merge event flags for this counter
const WaitEventSet &EventsForT = Context->getWaitEvents(T);
const WaitEventSet OldEvents = PendingEvents & EventsForT;
@@ -3097,7 +3116,8 @@ bool WaitcntBrackets::merge(const WaitcntBrackets &Other) {
}
StrictDom |= mergeAsyncMarks(MergeInfos, Other.AsyncMarks);
- for (auto T : inst_counter_types(Context->MaxCounter))
+ for (auto T : inst_counter_types(getMaxCounter(
+ *Context->ST, isExpertMode(*Context->MF, *Context->ST))))
StrictDom |= mergeScore(MergeInfos[T], AsyncScore[T], Other.AsyncScore[T]);
purgeEmptyTrackingData();
@@ -3202,8 +3222,8 @@ bool SIInsertWaitcnts::insertWaitcntInBlock(MachineFunction &MF,
// Track pre-existing waitcnts that were added in earlier iterations or by
// the memory legalizer.
- if (isWaitInstr(Inst) ||
- (IsExpertMode && Inst.getOpcode() == AMDGPU::S_WAITCNT_DEPCTR)) {
+ if (isWaitInstr(Inst) || (isExpertMode(MF, *ST) &&
+ Inst.getOpcode() == AMDGPU::S_WAITCNT_DEPCTR)) {
++Iter;
bool IsSoftXcnt = isSoftXcnt(Inst);
// The Memory Legalizer conservatively inserts a soft xcnt before each
@@ -3557,6 +3577,7 @@ bool SIInsertWaitcnts::run(MachineFunction &MF) {
TRI = &TII->getRegisterInfo();
MRI = &MF.getRegInfo();
const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
+ this->MF = &MF;
AMDGPU::IsaVersion IV = AMDGPU::getIsaVersion(ST->getCPU());
@@ -3564,21 +3585,11 @@ bool SIInsertWaitcnts::run(MachineFunction &MF) {
Limits = AMDGPU::HardwareLimits(IV);
if (ST->hasExtendedWaitCounts()) {
- IsExpertMode = ST->hasExpertSchedulingMode() &&
- (ExpertSchedulingModeFlag.getNumOccurrences()
- ? ExpertSchedulingModeFlag
- : MF.getFunction()
- .getFnAttribute("amdgpu-expert-scheduling-mode")
- .getValueAsBool());
- MaxCounter = IsExpertMode ? NUM_EXPERT_INST_CNTS : NUM_EXTENDED_INST_CNTS;
if (!WCG)
- WCG = std::make_unique<WaitcntGeneratorGFX12Plus>(MF, MaxCounter, &Limits,
- IsExpertMode);
+ WCG = std::make_unique<WaitcntGeneratorGFX12Plus>(MF, &Limits);
} else {
- MaxCounter = NUM_NORMAL_INST_CNTS;
if (!WCG)
- WCG = std::make_unique<WaitcntGeneratorPreGFX12>(MF, NUM_NORMAL_INST_CNTS,
- &Limits);
+ WCG = std::make_unique<WaitcntGeneratorPreGFX12>(&MF, &Limits);
}
for (auto T : inst_counter_types())
@@ -3617,7 +3628,7 @@ bool SIInsertWaitcnts::run(MachineFunction &MF) {
TII->get(instrsForExtendedCounterTypes[CT]))
.addImm(0);
}
- if (IsExpertMode) {
+ if (isExpertMode(MF, *ST)) {
unsigned Enc = AMDGPU::DepCtr::encodeFieldVaVdst(0, *ST);
Enc = AMDGPU::DepCtr::encodeFieldVmVsrc(Enc, 0);
BuildMI(EntryBB, I, DebugLoc(), TII->get(AMDGPU::S_WAITCNT_DEPCTR))
@@ -3756,7 +3767,7 @@ bool SIInsertWaitcnts::run(MachineFunction &MF) {
}
}
- if (IsExpertMode) {
+ if (isExpertMode(MF, *ST)) {
// Enable expert scheduling on function entry. To satisfy ABI requirements
// and to allow calls between function with different expert scheduling
// settings, disable it around calls and before returns.
>From 44ed8f17be08ecb20ac55487cc9dcf5bbd11cc1f Mon Sep 17 00:00:00 2001
From: Vasileios Porpodas <vasileios.porpodas at amd.com>
Date: Thu, 12 Feb 2026 17:31:53 +0000
Subject: [PATCH 2/4] fixup! [AMDGPU][SIInsertWaitcnts][NFC] Drop MaxCounter
member variable
---
llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp | 32 ++++++++++-----------
1 file changed, 15 insertions(+), 17 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
index 4251916f94c45..eea11929a6013 100644
--- a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
@@ -530,7 +530,7 @@ class SIInsertWaitcnts {
const SIRegisterInfo *TRI = nullptr;
const MachineRegisterInfo *MRI = nullptr;
InstCounterType SmemAccessCounter;
- const MachineFunction *MF = nullptr;
+ bool IsExpertMode = false;
private:
DenseMap<const Value *, MachineBasicBlock *> SLoadAddresses;
@@ -1081,8 +1081,7 @@ bool WaitcntBrackets::hasPointSamplePendingVmemTypes(const MachineInstr &MI,
void WaitcntBrackets::updateByEvent(WaitEventType E, MachineInstr &Inst) {
InstCounterType T = Context->getCounterFromEvent(E);
- assert(T < getMaxCounter(*Context->ST,
- isExpertMode(*Context->MF, *Context->ST)) &&
+ assert(T < getMaxCounter(*Context->ST, Context->IsExpertMode) &&
"T not supported in current mode!");
unsigned UB = getScoreUB(T);
@@ -1303,8 +1302,7 @@ void WaitcntBrackets::recordAsyncMark(MachineInstr &Inst) {
void WaitcntBrackets::print(raw_ostream &OS) const {
const GCNSubtarget *ST = Context->ST;
- for (auto T : inst_counter_types(
- getMaxCounter(*ST, isExpertMode(*Context->MF, *Context->ST)))) {
+ for (auto T : inst_counter_types(getMaxCounter(*ST, Context->IsExpertMode))) {
unsigned SR = getScoreRange(T);
switch (T) {
case LOAD_CNT:
@@ -2842,7 +2840,7 @@ void SIInsertWaitcnts::updateEventWaitcntAfter(MachineInstr &Inst,
bool IsVMEMAccess = false;
bool IsSMEMAccess = false;
- if (isExpertMode(*MF, *ST)) {
+ if (IsExpertMode) {
if (const auto ET = getExpertSchedulingEventType(Inst))
ScoreBrackets->updateByEvent(*ET, Inst);
}
@@ -3016,8 +3014,8 @@ bool WaitcntBrackets::mergeAsyncMarks(ArrayRef<MergeInfo> MergeInfos,
unsigned MergeCount = std::min(OtherSize, OurSize);
assert(OurSize == MaxSize);
for (unsigned Idx = 1; Idx <= MergeCount; ++Idx) {
- for (auto T : inst_counter_types(getMaxCounter(
- *Context->ST, isExpertMode(*Context->MF, *Context->ST)))) {
+ for (auto T : inst_counter_types(
+ getMaxCounter(*Context->ST, Context->IsExpertMode))) {
StrictDom |= mergeScore(MergeInfos[T], AsyncMarks[OurSize - Idx][T],
OtherMarks[OtherSize - Idx][T]);
}
@@ -3052,8 +3050,8 @@ bool WaitcntBrackets::merge(const WaitcntBrackets &Other) {
// Array to store MergeInfo for each counter type
MergeInfo MergeInfos[NUM_INST_CNTS];
- for (auto T : inst_counter_types(getMaxCounter(
- *Context->ST, isExpertMode(*Context->MF, *Context->ST)))) {
+ for (auto T :
+ inst_counter_types(getMaxCounter(*Context->ST, Context->IsExpertMode))) {
// Merge event flags for this counter
const WaitEventSet &EventsForT = Context->getWaitEvents(T);
const WaitEventSet OldEvents = PendingEvents & EventsForT;
@@ -3116,8 +3114,8 @@ bool WaitcntBrackets::merge(const WaitcntBrackets &Other) {
}
StrictDom |= mergeAsyncMarks(MergeInfos, Other.AsyncMarks);
- for (auto T : inst_counter_types(getMaxCounter(
- *Context->ST, isExpertMode(*Context->MF, *Context->ST))))
+ for (auto T :
+ inst_counter_types(getMaxCounter(*Context->ST, Context->IsExpertMode)))
StrictDom |= mergeScore(MergeInfos[T], AsyncScore[T], Other.AsyncScore[T]);
purgeEmptyTrackingData();
@@ -3222,8 +3220,8 @@ bool SIInsertWaitcnts::insertWaitcntInBlock(MachineFunction &MF,
// Track pre-existing waitcnts that were added in earlier iterations or by
// the memory legalizer.
- if (isWaitInstr(Inst) || (isExpertMode(MF, *ST) &&
- Inst.getOpcode() == AMDGPU::S_WAITCNT_DEPCTR)) {
+ if (isWaitInstr(Inst) ||
+ (IsExpertMode && Inst.getOpcode() == AMDGPU::S_WAITCNT_DEPCTR)) {
++Iter;
bool IsSoftXcnt = isSoftXcnt(Inst);
// The Memory Legalizer conservatively inserts a soft xcnt before each
@@ -3577,7 +3575,7 @@ bool SIInsertWaitcnts::run(MachineFunction &MF) {
TRI = &TII->getRegisterInfo();
MRI = &MF.getRegInfo();
const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
- this->MF = &MF;
+ IsExpertMode = isExpertMode(MF, *ST);
AMDGPU::IsaVersion IV = AMDGPU::getIsaVersion(ST->getCPU());
@@ -3628,7 +3626,7 @@ bool SIInsertWaitcnts::run(MachineFunction &MF) {
TII->get(instrsForExtendedCounterTypes[CT]))
.addImm(0);
}
- if (isExpertMode(MF, *ST)) {
+ if (IsExpertMode) {
unsigned Enc = AMDGPU::DepCtr::encodeFieldVaVdst(0, *ST);
Enc = AMDGPU::DepCtr::encodeFieldVmVsrc(Enc, 0);
BuildMI(EntryBB, I, DebugLoc(), TII->get(AMDGPU::S_WAITCNT_DEPCTR))
@@ -3767,7 +3765,7 @@ bool SIInsertWaitcnts::run(MachineFunction &MF) {
}
}
- if (isExpertMode(MF, *ST)) {
+ if (IsExpertMode) {
// Enable expert scheduling on function entry. To satisfy ABI requirements
// and to allow calls between function with different expert scheduling
// settings, disable it around calls and before returns.
>From 011c3a36e7acf4aef52e3b1a9e6216bfe6f6ecc1 Mon Sep 17 00:00:00 2001
From: Vasileios Porpodas <vasileios.porpodas at amd.com>
Date: Thu, 12 Feb 2026 19:10:20 +0000
Subject: [PATCH 3/4] fixup! fixup! [AMDGPU][SIInsertWaitcnts][NFC] Drop
MaxCounter member variable
---
llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
index eea11929a6013..5e99c132b3270 100644
--- a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
@@ -3013,9 +3013,10 @@ bool WaitcntBrackets::mergeAsyncMarks(ArrayRef<MergeInfo> MergeInfos,
unsigned OurSize = AsyncMarks.size();
unsigned MergeCount = std::min(OtherSize, OurSize);
assert(OurSize == MaxSize);
+ const InstCounterType MaxCounter =
+ getMaxCounter(*Context->ST, Context->IsExpertMode);
for (unsigned Idx = 1; Idx <= MergeCount; ++Idx) {
- for (auto T : inst_counter_types(
- getMaxCounter(*Context->ST, Context->IsExpertMode))) {
+ for (auto T : inst_counter_types(MaxCounter)) {
StrictDom |= mergeScore(MergeInfos[T], AsyncMarks[OurSize - Idx][T],
OtherMarks[OtherSize - Idx][T]);
}
>From d5409e2f0781208b4f9826c2e9d062eb794758de Mon Sep 17 00:00:00 2001
From: Vasileios Porpodas <vasileios.porpodas at amd.com>
Date: Mon, 16 Feb 2026 19:14:00 +0000
Subject: [PATCH 4/4] fixup! fixup! fixup! [AMDGPU][SIInsertWaitcnts][NFC] Drop
MaxCounter member variable
---
llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp | 40 ++++++++++-----------
1 file changed, 19 insertions(+), 21 deletions(-)
diff --git a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
index 5e99c132b3270..9bfcf05e6b113 100644
--- a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
@@ -194,16 +194,6 @@ template <> struct enum_iteration_traits<WaitEventType> {
namespace {
-bool isExpertMode(const MachineFunction &MF, const GCNSubtarget &ST) {
- return ST.hasExpertSchedulingMode() &&
- (ExpertSchedulingModeFlag.getNumOccurrences()
- ? ExpertSchedulingModeFlag
- : MF.getFunction()
- .getFnAttribute("amdgpu-expert-scheduling-mode")
- .getValueAsBool());
-
-}
-
InstCounterType getMaxCounter(const GCNSubtarget &ST, bool IsExpertMode) {
if (ST.hasExtendedWaitCounts())
return IsExpertMode ? NUM_EXPERT_INST_CNTS : NUM_EXTENDED_INST_CNTS;
@@ -370,23 +360,24 @@ class WaitcntGenerator {
const GCNSubtarget &ST;
const SIInstrInfo &TII;
AMDGPU::IsaVersion IV;
+ // TODO: Shouldn't this be a reference, because IsExpertMode can change per MF
+ bool IsExpertMode;
bool OptNone;
bool ExpandWaitcntProfiling = false;
const AMDGPU::HardwareLimits *Limits = nullptr;
- bool IsExpertMode = false;
public:
WaitcntGenerator() = delete;
WaitcntGenerator(const WaitcntGenerator &) = delete;
- WaitcntGenerator(const MachineFunction &MF,
+ WaitcntGenerator(const MachineFunction &MF, const bool &IsExpertMode,
const AMDGPU::HardwareLimits *Limits)
: ST(MF.getSubtarget<GCNSubtarget>()), TII(*ST.getInstrInfo()),
- IV(AMDGPU::getIsaVersion(ST.getCPU())),
+ IV(AMDGPU::getIsaVersion(ST.getCPU())), IsExpertMode(IsExpertMode),
OptNone(MF.getFunction().hasOptNone() ||
MF.getTarget().getOptLevel() == CodeGenOptLevel::None),
ExpandWaitcntProfiling(
MF.getFunction().hasFnAttribute("amdgpu-expand-waitcnt-profiling")),
- Limits(Limits), IsExpertMode(isExpertMode(MF, ST)) {}
+ Limits(Limits) {}
// Return true if the current function should be compiled with no
// optimization.
@@ -457,9 +448,9 @@ class WaitcntGeneratorPreGFX12 final : public WaitcntGenerator {
WaitEventSet()};
public:
- WaitcntGeneratorPreGFX12(MachineFunction *MF,
+ WaitcntGeneratorPreGFX12(MachineFunction *MF, const bool &IsExpertMode,
const AMDGPU::HardwareLimits *Limits)
- : WaitcntGenerator(*MF, Limits) {}
+ : WaitcntGenerator(*MF, IsExpertMode, Limits) {}
bool
applyPreexistingWaitcnt(WaitcntBrackets &ScoreBrackets,
MachineInstr &OldWaitcntInstr, AMDGPU::Waitcnt &Wait,
@@ -496,9 +487,9 @@ class WaitcntGeneratorGFX12Plus final : public WaitcntGenerator {
public:
WaitcntGeneratorGFX12Plus() = delete;
- WaitcntGeneratorGFX12Plus(const MachineFunction &MF,
+ WaitcntGeneratorGFX12Plus(const MachineFunction &MF, const bool &IsExpertMode,
const AMDGPU::HardwareLimits *Limits)
- : WaitcntGenerator(MF, Limits) {}
+ : WaitcntGenerator(MF, IsExpertMode, Limits) {}
bool
applyPreexistingWaitcnt(WaitcntBrackets &ScoreBrackets,
@@ -3576,7 +3567,6 @@ bool SIInsertWaitcnts::run(MachineFunction &MF) {
TRI = &TII->getRegisterInfo();
MRI = &MF.getRegInfo();
const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
- IsExpertMode = isExpertMode(MF, *ST);
AMDGPU::IsaVersion IV = AMDGPU::getIsaVersion(ST->getCPU());
@@ -3584,11 +3574,19 @@ bool SIInsertWaitcnts::run(MachineFunction &MF) {
Limits = AMDGPU::HardwareLimits(IV);
if (ST->hasExtendedWaitCounts()) {
+ IsExpertMode = ST->hasExpertSchedulingMode() &&
+ (ExpertSchedulingModeFlag.getNumOccurrences()
+ ? ExpertSchedulingModeFlag
+ : MF.getFunction()
+ .getFnAttribute("amdgpu-expert-scheduling-mode")
+ .getValueAsBool());
if (!WCG)
- WCG = std::make_unique<WaitcntGeneratorGFX12Plus>(MF, &Limits);
+ WCG = std::make_unique<WaitcntGeneratorGFX12Plus>(MF, IsExpertMode,
+ &Limits);
} else {
if (!WCG)
- WCG = std::make_unique<WaitcntGeneratorPreGFX12>(&MF, &Limits);
+ WCG = std::make_unique<WaitcntGeneratorPreGFX12>(&MF, IsExpertMode,
+ &Limits);
}
for (auto T : inst_counter_types())
More information about the llvm-commits
mailing list