1 //===- SIMemoryLegalizer.cpp ----------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// Memory legalizer - implements memory model. More information can be
13 /// http://llvm.org/docs/AMDGPUUsage.html#memory-model
15 //===----------------------------------------------------------------------===//
18 #include "AMDGPUMachineModuleInfo.h"
19 #include "AMDGPUSubtarget.h"
20 #include "SIDefines.h"
21 #include "SIInstrInfo.h"
22 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
23 #include "Utils/AMDGPUBaseInfo.h"
24 #include "llvm/ADT/BitmaskEnum.h"
25 #include "llvm/ADT/None.h"
26 #include "llvm/ADT/Optional.h"
27 #include "llvm/CodeGen/MachineBasicBlock.h"
28 #include "llvm/CodeGen/MachineFunction.h"
29 #include "llvm/CodeGen/MachineFunctionPass.h"
30 #include "llvm/CodeGen/MachineInstrBuilder.h"
31 #include "llvm/CodeGen/MachineMemOperand.h"
32 #include "llvm/CodeGen/MachineModuleInfo.h"
33 #include "llvm/CodeGen/MachineOperand.h"
34 #include "llvm/IR/DebugLoc.h"
35 #include "llvm/IR/DiagnosticInfo.h"
36 #include "llvm/IR/Function.h"
37 #include "llvm/IR/LLVMContext.h"
38 #include "llvm/MC/MCInstrDesc.h"
39 #include "llvm/Pass.h"
40 #include "llvm/Support/AtomicOrdering.h"
41 #include "llvm/Support/MathExtras.h"
46 using namespace llvm::AMDGPU;
48 #define DEBUG_TYPE "si-memory-legalizer"
49 #define PASS_NAME "SI Memory Legalizer"
53 LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE();
55 /// Memory operation flags. Can be ORed together.
60 LLVM_MARK_AS_BITMASK_ENUM(/* LargestFlag = */ STORE)
63 /// Position to insert a new instruction relative to an existing
70 /// The atomic synchronization scopes supported by the AMDGPU target.
71 enum class SIAtomicScope {
80 /// The distinct address spaces supported by the AMDGPU target for
81 /// atomic memory operation. Can be ORed toether.
82 enum class SIAtomicAddrSpace {
90 /// The address spaces that can be accessed by a FLAT instruction.
91 FLAT = GLOBAL | LDS | SCRATCH,
93 /// The address spaces that support atomic instructions.
94 ATOMIC = GLOBAL | LDS | SCRATCH | GDS,
96 /// All address spaces.
97 ALL = GLOBAL | LDS | SCRATCH | GDS | OTHER,
99 LLVM_MARK_AS_BITMASK_ENUM(/* LargestFlag = */ ALL)
102 /// Sets named bit \p BitName to "true" if present in instruction \p MI.
103 /// \returns Returns true if \p MI is modified, false otherwise.
104 template <uint16_t BitName>
105 bool enableNamedBit(const MachineBasicBlock::iterator &MI) {
106 int BitIdx = AMDGPU::getNamedOperandIdx(MI->getOpcode(), BitName);
110 MachineOperand &Bit = MI->getOperand(BitIdx);
111 if (Bit.getImm() != 0)
118 class SIMemOpInfo final {
121 friend class SIMemOpAccess;
123 AtomicOrdering Ordering = AtomicOrdering::NotAtomic;
124 AtomicOrdering FailureOrdering = AtomicOrdering::NotAtomic;
125 SIAtomicScope Scope = SIAtomicScope::SYSTEM;
126 SIAtomicAddrSpace OrderingAddrSpace = SIAtomicAddrSpace::NONE;
127 SIAtomicAddrSpace InstrAddrSpace = SIAtomicAddrSpace::NONE;
128 bool IsCrossAddressSpaceOrdering = false;
129 bool IsNonTemporal = false;
131 SIMemOpInfo(AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent,
132 SIAtomicScope Scope = SIAtomicScope::SYSTEM,
133 SIAtomicAddrSpace OrderingAddrSpace = SIAtomicAddrSpace::ATOMIC,
134 SIAtomicAddrSpace InstrAddrSpace = SIAtomicAddrSpace::ALL,
135 bool IsCrossAddressSpaceOrdering = true,
136 AtomicOrdering FailureOrdering =
137 AtomicOrdering::SequentiallyConsistent,
138 bool IsNonTemporal = false)
139 : Ordering(Ordering), FailureOrdering(FailureOrdering),
140 Scope(Scope), OrderingAddrSpace(OrderingAddrSpace),
141 InstrAddrSpace(InstrAddrSpace),
142 IsCrossAddressSpaceOrdering(IsCrossAddressSpaceOrdering),
143 IsNonTemporal(IsNonTemporal) {
144 // There is also no cross address space ordering if the ordering
145 // address space is the same as the instruction address space and
146 // only contains a single address space.
147 if ((OrderingAddrSpace == InstrAddrSpace) &&
148 isPowerOf2_32(uint32_t(InstrAddrSpace)))
149 IsCrossAddressSpaceOrdering = false;
153 /// \returns Atomic synchronization scope of the machine instruction used to
154 /// create this SIMemOpInfo.
155 SIAtomicScope getScope() const {
159 /// \returns Ordering constraint of the machine instruction used to
160 /// create this SIMemOpInfo.
161 AtomicOrdering getOrdering() const {
165 /// \returns Failure ordering constraint of the machine instruction used to
166 /// create this SIMemOpInfo.
167 AtomicOrdering getFailureOrdering() const {
168 return FailureOrdering;
171 /// \returns The address spaces be accessed by the machine
172 /// instruction used to create this SiMemOpInfo.
173 SIAtomicAddrSpace getInstrAddrSpace() const {
174 return InstrAddrSpace;
177 /// \returns The address spaces that must be ordered by the machine
178 /// instruction used to create this SiMemOpInfo.
179 SIAtomicAddrSpace getOrderingAddrSpace() const {
180 return OrderingAddrSpace;
183 /// \returns Return true iff memory ordering of operations on
184 /// different address spaces is required.
185 bool getIsCrossAddressSpaceOrdering() const {
186 return IsCrossAddressSpaceOrdering;
189 /// \returns True if memory access of the machine instruction used to
190 /// create this SIMemOpInfo is non-temporal, false otherwise.
191 bool isNonTemporal() const {
192 return IsNonTemporal;
195 /// \returns True if ordering constraint of the machine instruction used to
196 /// create this SIMemOpInfo is unordered or higher, false otherwise.
197 bool isAtomic() const {
198 return Ordering != AtomicOrdering::NotAtomic;
203 class SIMemOpAccess final {
205 AMDGPUMachineModuleInfo *MMI = nullptr;
207 /// Reports unsupported message \p Msg for \p MI to LLVM context.
208 void reportUnsupported(const MachineBasicBlock::iterator &MI,
209 const char *Msg) const;
211 /// Inspects the target synchonization scope \p SSID and determines
212 /// the SI atomic scope it corresponds to, the address spaces it
213 /// covers, and whether the memory ordering applies between address
215 Optional<std::tuple<SIAtomicScope, SIAtomicAddrSpace, bool>>
216 toSIAtomicScope(SyncScope::ID SSID, SIAtomicAddrSpace InstrScope) const;
218 /// \return Return a bit set of the address spaces accessed by \p AS.
219 SIAtomicAddrSpace toSIAtomicAddrSpace(unsigned AS) const;
221 /// \returns Info constructed from \p MI, which has at least machine memory
223 Optional<SIMemOpInfo> constructFromMIWithMMO(
224 const MachineBasicBlock::iterator &MI) const;
227 /// Construct class to support accessing the machine memory operands
228 /// of instructions in the machine function \p MF.
229 SIMemOpAccess(MachineFunction &MF);
231 /// \returns Load info if \p MI is a load operation, "None" otherwise.
232 Optional<SIMemOpInfo> getLoadInfo(
233 const MachineBasicBlock::iterator &MI) const;
235 /// \returns Store info if \p MI is a store operation, "None" otherwise.
236 Optional<SIMemOpInfo> getStoreInfo(
237 const MachineBasicBlock::iterator &MI) const;
239 /// \returns Atomic fence info if \p MI is an atomic fence operation,
240 /// "None" otherwise.
241 Optional<SIMemOpInfo> getAtomicFenceInfo(
242 const MachineBasicBlock::iterator &MI) const;
244 /// \returns Atomic cmpxchg/rmw info if \p MI is an atomic cmpxchg or
245 /// rmw operation, "None" otherwise.
246 Optional<SIMemOpInfo> getAtomicCmpxchgOrRmwInfo(
247 const MachineBasicBlock::iterator &MI) const;
250 class SICacheControl {
253 /// Instruction info.
254 const SIInstrInfo *TII = nullptr;
258 SICacheControl(const GCNSubtarget &ST);
262 /// Create a cache control for the subtarget \p ST.
263 static std::unique_ptr<SICacheControl> create(const GCNSubtarget &ST);
265 /// Update \p MI memory load instruction to bypass any caches up to
266 /// the \p Scope memory scope for address spaces \p
267 /// AddrSpace. Return true iff the instruction was modified.
268 virtual bool enableLoadCacheBypass(const MachineBasicBlock::iterator &MI,
270 SIAtomicAddrSpace AddrSpace) const = 0;
272 /// Update \p MI memory instruction to indicate it is
273 /// nontemporal. Return true iff the instruction was modified.
274 virtual bool enableNonTemporal(const MachineBasicBlock::iterator &MI)
277 /// Inserts any necessary instructions at position \p Pos relative
278 /// to instruction \p MI to ensure any caches associated with
279 /// address spaces \p AddrSpace for memory scopes up to memory scope
280 /// \p Scope are invalidated. Returns true iff any instructions
282 virtual bool insertCacheInvalidate(MachineBasicBlock::iterator &MI,
284 SIAtomicAddrSpace AddrSpace,
285 Position Pos) const = 0;
287 /// Inserts any necessary instructions at position \p Pos relative
288 /// to instruction \p MI to ensure memory instructions of kind \p Op
289 /// associated with address spaces \p AddrSpace have completed as
290 /// observed by other memory instructions executing in memory scope
291 /// \p Scope. \p IsCrossAddrSpaceOrdering indicates if the memory
292 /// ordering is between address spaces. Returns true iff any
293 /// instructions inserted.
294 virtual bool insertWait(MachineBasicBlock::iterator &MI,
296 SIAtomicAddrSpace AddrSpace,
298 bool IsCrossAddrSpaceOrdering,
299 Position Pos) const = 0;
301 /// Virtual destructor to allow derivations to be deleted.
302 virtual ~SICacheControl() = default;
306 class SIGfx6CacheControl : public SICacheControl {
309 /// Sets GLC bit to "true" if present in \p MI. Returns true if \p MI
310 /// is modified, false otherwise.
311 bool enableGLCBit(const MachineBasicBlock::iterator &MI) const {
312 return enableNamedBit<AMDGPU::OpName::glc>(MI);
315 /// Sets SLC bit to "true" if present in \p MI. Returns true if \p MI
316 /// is modified, false otherwise.
317 bool enableSLCBit(const MachineBasicBlock::iterator &MI) const {
318 return enableNamedBit<AMDGPU::OpName::slc>(MI);
323 SIGfx6CacheControl(const GCNSubtarget &ST) : SICacheControl(ST) {};
325 bool enableLoadCacheBypass(const MachineBasicBlock::iterator &MI,
327 SIAtomicAddrSpace AddrSpace) const override;
329 bool enableNonTemporal(const MachineBasicBlock::iterator &MI) const override;
331 bool insertCacheInvalidate(MachineBasicBlock::iterator &MI,
333 SIAtomicAddrSpace AddrSpace,
334 Position Pos) const override;
336 bool insertWait(MachineBasicBlock::iterator &MI,
338 SIAtomicAddrSpace AddrSpace,
340 bool IsCrossAddrSpaceOrdering,
341 Position Pos) const override;
344 class SIGfx7CacheControl : public SIGfx6CacheControl {
347 SIGfx7CacheControl(const GCNSubtarget &ST) : SIGfx6CacheControl(ST) {};
349 bool insertCacheInvalidate(MachineBasicBlock::iterator &MI,
351 SIAtomicAddrSpace AddrSpace,
352 Position Pos) const override;
356 class SIMemoryLegalizer final : public MachineFunctionPass {
360 std::unique_ptr<SICacheControl> CC = nullptr;
362 /// List of atomic pseudo instructions.
363 std::list<MachineBasicBlock::iterator> AtomicPseudoMIs;
365 /// Return true iff instruction \p MI is a atomic instruction that
366 /// returns a result.
367 bool isAtomicRet(const MachineInstr &MI) const {
368 return AMDGPU::getAtomicNoRetOp(MI.getOpcode()) != -1;
371 /// Removes all processed atomic pseudo instructions from the current
372 /// function. Returns true if current function is modified, false otherwise.
373 bool removeAtomicPseudoMIs();
375 /// Expands load operation \p MI. Returns true if instructions are
376 /// added/deleted or \p MI is modified, false otherwise.
377 bool expandLoad(const SIMemOpInfo &MOI,
378 MachineBasicBlock::iterator &MI);
379 /// Expands store operation \p MI. Returns true if instructions are
380 /// added/deleted or \p MI is modified, false otherwise.
381 bool expandStore(const SIMemOpInfo &MOI,
382 MachineBasicBlock::iterator &MI);
383 /// Expands atomic fence operation \p MI. Returns true if
384 /// instructions are added/deleted or \p MI is modified, false otherwise.
385 bool expandAtomicFence(const SIMemOpInfo &MOI,
386 MachineBasicBlock::iterator &MI);
387 /// Expands atomic cmpxchg or rmw operation \p MI. Returns true if
388 /// instructions are added/deleted or \p MI is modified, false otherwise.
389 bool expandAtomicCmpxchgOrRmw(const SIMemOpInfo &MOI,
390 MachineBasicBlock::iterator &MI);
395 SIMemoryLegalizer() : MachineFunctionPass(ID) {}
397 void getAnalysisUsage(AnalysisUsage &AU) const override {
398 AU.setPreservesCFG();
399 MachineFunctionPass::getAnalysisUsage(AU);
402 StringRef getPassName() const override {
406 bool runOnMachineFunction(MachineFunction &MF) override;
409 } // end namespace anonymous
411 void SIMemOpAccess::reportUnsupported(const MachineBasicBlock::iterator &MI,
412 const char *Msg) const {
413 const Function &Func = MI->getParent()->getParent()->getFunction();
414 DiagnosticInfoUnsupported Diag(Func, Msg, MI->getDebugLoc());
415 Func.getContext().diagnose(Diag);
418 Optional<std::tuple<SIAtomicScope, SIAtomicAddrSpace, bool>>
419 SIMemOpAccess::toSIAtomicScope(SyncScope::ID SSID,
420 SIAtomicAddrSpace InstrScope) const {
421 /// TODO: For now assume OpenCL memory model which treats each
422 /// address space as having a separate happens-before relation, and
423 /// so an instruction only has ordering with respect to the address
424 /// space it accesses, and if it accesses multiple address spaces it
425 /// does not require ordering of operations in different address
427 if (SSID == SyncScope::System)
428 return std::make_tuple(SIAtomicScope::SYSTEM,
429 SIAtomicAddrSpace::ATOMIC & InstrScope,
431 if (SSID == MMI->getAgentSSID())
432 return std::make_tuple(SIAtomicScope::AGENT,
433 SIAtomicAddrSpace::ATOMIC & InstrScope,
435 if (SSID == MMI->getWorkgroupSSID())
436 return std::make_tuple(SIAtomicScope::WORKGROUP,
437 SIAtomicAddrSpace::ATOMIC & InstrScope,
439 if (SSID == MMI->getWavefrontSSID())
440 return std::make_tuple(SIAtomicScope::WAVEFRONT,
441 SIAtomicAddrSpace::ATOMIC & InstrScope,
443 if (SSID == SyncScope::SingleThread)
444 return std::make_tuple(SIAtomicScope::SINGLETHREAD,
445 SIAtomicAddrSpace::ATOMIC & InstrScope,
447 /// TODO: To support HSA Memory Model need to add additional memory
448 /// scopes that specify that do require cross address space
453 SIAtomicAddrSpace SIMemOpAccess::toSIAtomicAddrSpace(unsigned AS) const {
454 if (AS == AMDGPUAS::FLAT_ADDRESS)
455 return SIAtomicAddrSpace::FLAT;
456 if (AS == AMDGPUAS::GLOBAL_ADDRESS)
457 return SIAtomicAddrSpace::GLOBAL;
458 if (AS == AMDGPUAS::LOCAL_ADDRESS)
459 return SIAtomicAddrSpace::LDS;
460 if (AS == AMDGPUAS::PRIVATE_ADDRESS)
461 return SIAtomicAddrSpace::SCRATCH;
462 if (AS == AMDGPUAS::REGION_ADDRESS)
463 return SIAtomicAddrSpace::GDS;
465 return SIAtomicAddrSpace::OTHER;
468 SIMemOpAccess::SIMemOpAccess(MachineFunction &MF) {
469 MMI = &MF.getMMI().getObjFileInfo<AMDGPUMachineModuleInfo>();
472 Optional<SIMemOpInfo> SIMemOpAccess::constructFromMIWithMMO(
473 const MachineBasicBlock::iterator &MI) const {
474 assert(MI->getNumMemOperands() > 0);
476 SyncScope::ID SSID = SyncScope::SingleThread;
477 AtomicOrdering Ordering = AtomicOrdering::NotAtomic;
478 AtomicOrdering FailureOrdering = AtomicOrdering::NotAtomic;
479 SIAtomicAddrSpace InstrAddrSpace = SIAtomicAddrSpace::NONE;
480 bool IsNonTemporal = true;
482 // Validator should check whether or not MMOs cover the entire set of
483 // locations accessed by the memory instruction.
484 for (const auto &MMO : MI->memoperands()) {
485 IsNonTemporal &= MMO->isNonTemporal();
487 toSIAtomicAddrSpace(MMO->getPointerInfo().getAddrSpace());
488 AtomicOrdering OpOrdering = MMO->getOrdering();
489 if (OpOrdering != AtomicOrdering::NotAtomic) {
490 const auto &IsSyncScopeInclusion =
491 MMI->isSyncScopeInclusion(SSID, MMO->getSyncScopeID());
492 if (!IsSyncScopeInclusion) {
493 reportUnsupported(MI,
494 "Unsupported non-inclusive atomic synchronization scope");
498 SSID = IsSyncScopeInclusion.getValue() ? SSID : MMO->getSyncScopeID();
500 isStrongerThan(Ordering, OpOrdering) ?
501 Ordering : MMO->getOrdering();
502 assert(MMO->getFailureOrdering() != AtomicOrdering::Release &&
503 MMO->getFailureOrdering() != AtomicOrdering::AcquireRelease);
505 isStrongerThan(FailureOrdering, MMO->getFailureOrdering()) ?
506 FailureOrdering : MMO->getFailureOrdering();
510 SIAtomicScope Scope = SIAtomicScope::NONE;
511 SIAtomicAddrSpace OrderingAddrSpace = SIAtomicAddrSpace::NONE;
512 bool IsCrossAddressSpaceOrdering = false;
513 if (Ordering != AtomicOrdering::NotAtomic) {
514 auto ScopeOrNone = toSIAtomicScope(SSID, InstrAddrSpace);
516 reportUnsupported(MI, "Unsupported atomic synchronization scope");
519 std::tie(Scope, OrderingAddrSpace, IsCrossAddressSpaceOrdering) =
520 ScopeOrNone.getValue();
521 if ((OrderingAddrSpace == SIAtomicAddrSpace::NONE) ||
522 ((OrderingAddrSpace & SIAtomicAddrSpace::ATOMIC) != OrderingAddrSpace)) {
523 reportUnsupported(MI, "Unsupported atomic address space");
527 return SIMemOpInfo(Ordering, Scope, OrderingAddrSpace, InstrAddrSpace,
528 IsCrossAddressSpaceOrdering, FailureOrdering, IsNonTemporal);
531 Optional<SIMemOpInfo> SIMemOpAccess::getLoadInfo(
532 const MachineBasicBlock::iterator &MI) const {
533 assert(MI->getDesc().TSFlags & SIInstrFlags::maybeAtomic);
535 if (!(MI->mayLoad() && !MI->mayStore()))
538 // Be conservative if there are no memory operands.
539 if (MI->getNumMemOperands() == 0)
540 return SIMemOpInfo();
542 return constructFromMIWithMMO(MI);
545 Optional<SIMemOpInfo> SIMemOpAccess::getStoreInfo(
546 const MachineBasicBlock::iterator &MI) const {
547 assert(MI->getDesc().TSFlags & SIInstrFlags::maybeAtomic);
549 if (!(!MI->mayLoad() && MI->mayStore()))
552 // Be conservative if there are no memory operands.
553 if (MI->getNumMemOperands() == 0)
554 return SIMemOpInfo();
556 return constructFromMIWithMMO(MI);
559 Optional<SIMemOpInfo> SIMemOpAccess::getAtomicFenceInfo(
560 const MachineBasicBlock::iterator &MI) const {
561 assert(MI->getDesc().TSFlags & SIInstrFlags::maybeAtomic);
563 if (MI->getOpcode() != AMDGPU::ATOMIC_FENCE)
566 AtomicOrdering Ordering =
567 static_cast<AtomicOrdering>(MI->getOperand(0).getImm());
569 SyncScope::ID SSID = static_cast<SyncScope::ID>(MI->getOperand(1).getImm());
570 auto ScopeOrNone = toSIAtomicScope(SSID, SIAtomicAddrSpace::ATOMIC);
572 reportUnsupported(MI, "Unsupported atomic synchronization scope");
576 SIAtomicScope Scope = SIAtomicScope::NONE;
577 SIAtomicAddrSpace OrderingAddrSpace = SIAtomicAddrSpace::NONE;
578 bool IsCrossAddressSpaceOrdering = false;
579 std::tie(Scope, OrderingAddrSpace, IsCrossAddressSpaceOrdering) =
580 ScopeOrNone.getValue();
582 if ((OrderingAddrSpace == SIAtomicAddrSpace::NONE) ||
583 ((OrderingAddrSpace & SIAtomicAddrSpace::ATOMIC) != OrderingAddrSpace)) {
584 reportUnsupported(MI, "Unsupported atomic address space");
588 return SIMemOpInfo(Ordering, Scope, OrderingAddrSpace, SIAtomicAddrSpace::ATOMIC,
589 IsCrossAddressSpaceOrdering);
592 Optional<SIMemOpInfo> SIMemOpAccess::getAtomicCmpxchgOrRmwInfo(
593 const MachineBasicBlock::iterator &MI) const {
594 assert(MI->getDesc().TSFlags & SIInstrFlags::maybeAtomic);
596 if (!(MI->mayLoad() && MI->mayStore()))
599 // Be conservative if there are no memory operands.
600 if (MI->getNumMemOperands() == 0)
601 return SIMemOpInfo();
603 return constructFromMIWithMMO(MI);
606 SICacheControl::SICacheControl(const GCNSubtarget &ST) {
607 TII = ST.getInstrInfo();
608 IV = getIsaVersion(ST.getCPU());
612 std::unique_ptr<SICacheControl> SICacheControl::create(const GCNSubtarget &ST) {
613 GCNSubtarget::Generation Generation = ST.getGeneration();
614 if (Generation <= AMDGPUSubtarget::SOUTHERN_ISLANDS)
615 return make_unique<SIGfx6CacheControl>(ST);
616 return make_unique<SIGfx7CacheControl>(ST);
619 bool SIGfx6CacheControl::enableLoadCacheBypass(
620 const MachineBasicBlock::iterator &MI,
622 SIAtomicAddrSpace AddrSpace) const {
623 assert(MI->mayLoad() && !MI->mayStore());
624 bool Changed = false;
626 if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) {
627 /// TODO: Do not set glc for rmw atomic operations as they
628 /// implicitly bypass the L1 cache.
631 case SIAtomicScope::SYSTEM:
632 case SIAtomicScope::AGENT:
633 Changed |= enableGLCBit(MI);
635 case SIAtomicScope::WORKGROUP:
636 case SIAtomicScope::WAVEFRONT:
637 case SIAtomicScope::SINGLETHREAD:
638 // No cache to bypass.
641 llvm_unreachable("Unsupported synchronization scope");
645 /// The scratch address space does not need the global memory caches
646 /// to be bypassed as all memory operations by the same thread are
647 /// sequentially consistent, and no other thread can access scratch
650 /// Other address spaces do not hava a cache.
655 bool SIGfx6CacheControl::enableNonTemporal(
656 const MachineBasicBlock::iterator &MI) const {
657 assert(MI->mayLoad() ^ MI->mayStore());
658 bool Changed = false;
660 /// TODO: Do not enableGLCBit if rmw atomic.
661 Changed |= enableGLCBit(MI);
662 Changed |= enableSLCBit(MI);
667 bool SIGfx6CacheControl::insertCacheInvalidate(MachineBasicBlock::iterator &MI,
669 SIAtomicAddrSpace AddrSpace,
670 Position Pos) const {
671 bool Changed = false;
673 MachineBasicBlock &MBB = *MI->getParent();
674 DebugLoc DL = MI->getDebugLoc();
676 if (Pos == Position::AFTER)
679 if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) {
681 case SIAtomicScope::SYSTEM:
682 case SIAtomicScope::AGENT:
683 BuildMI(MBB, MI, DL, TII->get(AMDGPU::BUFFER_WBINVL1));
686 case SIAtomicScope::WORKGROUP:
687 case SIAtomicScope::WAVEFRONT:
688 case SIAtomicScope::SINGLETHREAD:
689 // No cache to invalidate.
692 llvm_unreachable("Unsupported synchronization scope");
696 /// The scratch address space does not need the global memory cache
697 /// to be flushed as all memory operations by the same thread are
698 /// sequentially consistent, and no other thread can access scratch
701 /// Other address spaces do not hava a cache.
703 if (Pos == Position::AFTER)
709 bool SIGfx6CacheControl::insertWait(MachineBasicBlock::iterator &MI,
711 SIAtomicAddrSpace AddrSpace,
713 bool IsCrossAddrSpaceOrdering,
714 Position Pos) const {
715 bool Changed = false;
717 MachineBasicBlock &MBB = *MI->getParent();
718 DebugLoc DL = MI->getDebugLoc();
720 if (Pos == Position::AFTER)
724 bool LGKMCnt = false;
727 if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) {
729 case SIAtomicScope::SYSTEM:
730 case SIAtomicScope::AGENT:
733 case SIAtomicScope::WORKGROUP:
734 case SIAtomicScope::WAVEFRONT:
735 case SIAtomicScope::SINGLETHREAD:
736 // The L1 cache keeps all memory operations in order for
737 // wavefronts in the same work-group.
740 llvm_unreachable("Unsupported synchronization scope");
744 if ((AddrSpace & SIAtomicAddrSpace::LDS) != SIAtomicAddrSpace::NONE) {
746 case SIAtomicScope::SYSTEM:
747 case SIAtomicScope::AGENT:
748 case SIAtomicScope::WORKGROUP:
749 // If no cross address space ordering then an LDS waitcnt is not
750 // needed as LDS operations for all waves are executed in a
751 // total global ordering as observed by all waves. Required if
752 // also synchronizing with global/GDS memory as LDS operations
753 // could be reordered with respect to later global/GDS memory
754 // operations of the same wave.
755 LGKMCnt = IsCrossAddrSpaceOrdering;
757 case SIAtomicScope::WAVEFRONT:
758 case SIAtomicScope::SINGLETHREAD:
759 // The LDS keeps all memory operations in order for
760 // the same wavesfront.
763 llvm_unreachable("Unsupported synchronization scope");
767 if ((AddrSpace & SIAtomicAddrSpace::GDS) != SIAtomicAddrSpace::NONE) {
769 case SIAtomicScope::SYSTEM:
770 case SIAtomicScope::AGENT:
771 // If no cross address space ordering then an GDS waitcnt is not
772 // needed as GDS operations for all waves are executed in a
773 // total global ordering as observed by all waves. Required if
774 // also synchronizing with global/LDS memory as GDS operations
775 // could be reordered with respect to later global/LDS memory
776 // operations of the same wave.
777 EXPCnt = IsCrossAddrSpaceOrdering;
779 case SIAtomicScope::WORKGROUP:
780 case SIAtomicScope::WAVEFRONT:
781 case SIAtomicScope::SINGLETHREAD:
782 // The GDS keeps all memory operations in order for
783 // the same work-group.
786 llvm_unreachable("Unsupported synchronization scope");
790 if (VMCnt || LGKMCnt || EXPCnt) {
791 unsigned WaitCntImmediate =
792 AMDGPU::encodeWaitcnt(IV,
793 VMCnt ? 0 : getVmcntBitMask(IV),
794 EXPCnt ? 0 : getExpcntBitMask(IV),
795 LGKMCnt ? 0 : getLgkmcntBitMask(IV));
796 BuildMI(MBB, MI, DL, TII->get(AMDGPU::S_WAITCNT)).addImm(WaitCntImmediate);
800 if (Pos == Position::AFTER)
806 bool SIGfx7CacheControl::insertCacheInvalidate(MachineBasicBlock::iterator &MI,
808 SIAtomicAddrSpace AddrSpace,
809 Position Pos) const {
810 bool Changed = false;
812 MachineBasicBlock &MBB = *MI->getParent();
813 DebugLoc DL = MI->getDebugLoc();
815 const GCNSubtarget &STM = MBB.getParent()->getSubtarget<GCNSubtarget>();
817 const unsigned Flush = STM.isAmdPalOS() || STM.isMesa3DOS()
818 ? AMDGPU::BUFFER_WBINVL1
819 : AMDGPU::BUFFER_WBINVL1_VOL;
821 if (Pos == Position::AFTER)
824 if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) {
826 case SIAtomicScope::SYSTEM:
827 case SIAtomicScope::AGENT:
828 BuildMI(MBB, MI, DL, TII->get(Flush));
831 case SIAtomicScope::WORKGROUP:
832 case SIAtomicScope::WAVEFRONT:
833 case SIAtomicScope::SINGLETHREAD:
834 // No cache to invalidate.
837 llvm_unreachable("Unsupported synchronization scope");
841 /// The scratch address space does not need the global memory cache
842 /// to be flushed as all memory operations by the same thread are
843 /// sequentially consistent, and no other thread can access scratch
846 /// Other address spaces do not hava a cache.
848 if (Pos == Position::AFTER)
854 bool SIMemoryLegalizer::removeAtomicPseudoMIs() {
855 if (AtomicPseudoMIs.empty())
858 for (auto &MI : AtomicPseudoMIs)
859 MI->eraseFromParent();
861 AtomicPseudoMIs.clear();
865 bool SIMemoryLegalizer::expandLoad(const SIMemOpInfo &MOI,
866 MachineBasicBlock::iterator &MI) {
867 assert(MI->mayLoad() && !MI->mayStore());
869 bool Changed = false;
871 if (MOI.isAtomic()) {
872 if (MOI.getOrdering() == AtomicOrdering::Monotonic ||
873 MOI.getOrdering() == AtomicOrdering::Acquire ||
874 MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent) {
875 Changed |= CC->enableLoadCacheBypass(MI, MOI.getScope(),
876 MOI.getOrderingAddrSpace());
879 if (MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent)
880 Changed |= CC->insertWait(MI, MOI.getScope(),
881 MOI.getOrderingAddrSpace(),
882 SIMemOp::LOAD | SIMemOp::STORE,
883 MOI.getIsCrossAddressSpaceOrdering(),
886 if (MOI.getOrdering() == AtomicOrdering::Acquire ||
887 MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent) {
888 Changed |= CC->insertWait(MI, MOI.getScope(),
889 MOI.getInstrAddrSpace(),
891 MOI.getIsCrossAddressSpaceOrdering(),
893 Changed |= CC->insertCacheInvalidate(MI, MOI.getScope(),
894 MOI.getOrderingAddrSpace(),
901 // Atomic instructions do not have the nontemporal attribute.
902 if (MOI.isNonTemporal()) {
903 Changed |= CC->enableNonTemporal(MI);
910 bool SIMemoryLegalizer::expandStore(const SIMemOpInfo &MOI,
911 MachineBasicBlock::iterator &MI) {
912 assert(!MI->mayLoad() && MI->mayStore());
914 bool Changed = false;
916 if (MOI.isAtomic()) {
917 if (MOI.getOrdering() == AtomicOrdering::Release ||
918 MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent)
919 Changed |= CC->insertWait(MI, MOI.getScope(),
920 MOI.getOrderingAddrSpace(),
921 SIMemOp::LOAD | SIMemOp::STORE,
922 MOI.getIsCrossAddressSpaceOrdering(),
928 // Atomic instructions do not have the nontemporal attribute.
929 if (MOI.isNonTemporal()) {
930 Changed |= CC->enableNonTemporal(MI);
937 bool SIMemoryLegalizer::expandAtomicFence(const SIMemOpInfo &MOI,
938 MachineBasicBlock::iterator &MI) {
939 assert(MI->getOpcode() == AMDGPU::ATOMIC_FENCE);
941 AtomicPseudoMIs.push_back(MI);
942 bool Changed = false;
944 if (MOI.isAtomic()) {
945 if (MOI.getOrdering() == AtomicOrdering::Acquire ||
946 MOI.getOrdering() == AtomicOrdering::Release ||
947 MOI.getOrdering() == AtomicOrdering::AcquireRelease ||
948 MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent)
949 /// TODO: This relies on a barrier always generating a waitcnt
950 /// for LDS to ensure it is not reordered with the completion of
951 /// the proceeding LDS operations. If barrier had a memory
952 /// ordering and memory scope, then library does not need to
953 /// generate a fence. Could add support in this file for
954 /// barrier. SIInsertWaitcnt.cpp could then stop unconditionally
955 /// adding waitcnt before a S_BARRIER.
956 Changed |= CC->insertWait(MI, MOI.getScope(),
957 MOI.getOrderingAddrSpace(),
958 SIMemOp::LOAD | SIMemOp::STORE,
959 MOI.getIsCrossAddressSpaceOrdering(),
962 if (MOI.getOrdering() == AtomicOrdering::Acquire ||
963 MOI.getOrdering() == AtomicOrdering::AcquireRelease ||
964 MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent)
965 Changed |= CC->insertCacheInvalidate(MI, MOI.getScope(),
966 MOI.getOrderingAddrSpace(),
975 bool SIMemoryLegalizer::expandAtomicCmpxchgOrRmw(const SIMemOpInfo &MOI,
976 MachineBasicBlock::iterator &MI) {
977 assert(MI->mayLoad() && MI->mayStore());
979 bool Changed = false;
981 if (MOI.isAtomic()) {
982 if (MOI.getOrdering() == AtomicOrdering::Release ||
983 MOI.getOrdering() == AtomicOrdering::AcquireRelease ||
984 MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent ||
985 MOI.getFailureOrdering() == AtomicOrdering::SequentiallyConsistent)
986 Changed |= CC->insertWait(MI, MOI.getScope(),
987 MOI.getOrderingAddrSpace(),
988 SIMemOp::LOAD | SIMemOp::STORE,
989 MOI.getIsCrossAddressSpaceOrdering(),
992 if (MOI.getOrdering() == AtomicOrdering::Acquire ||
993 MOI.getOrdering() == AtomicOrdering::AcquireRelease ||
994 MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent ||
995 MOI.getFailureOrdering() == AtomicOrdering::Acquire ||
996 MOI.getFailureOrdering() == AtomicOrdering::SequentiallyConsistent) {
997 Changed |= CC->insertWait(MI, MOI.getScope(),
998 MOI.getOrderingAddrSpace(),
999 isAtomicRet(*MI) ? SIMemOp::LOAD :
1001 MOI.getIsCrossAddressSpaceOrdering(),
1003 Changed |= CC->insertCacheInvalidate(MI, MOI.getScope(),
1004 MOI.getOrderingAddrSpace(),
1014 bool SIMemoryLegalizer::runOnMachineFunction(MachineFunction &MF) {
1015 bool Changed = false;
1017 SIMemOpAccess MOA(MF);
1018 CC = SICacheControl::create(MF.getSubtarget<GCNSubtarget>());
1020 for (auto &MBB : MF) {
1021 for (auto MI = MBB.begin(); MI != MBB.end(); ++MI) {
1022 if (!(MI->getDesc().TSFlags & SIInstrFlags::maybeAtomic))
1025 if (const auto &MOI = MOA.getLoadInfo(MI))
1026 Changed |= expandLoad(MOI.getValue(), MI);
1027 else if (const auto &MOI = MOA.getStoreInfo(MI))
1028 Changed |= expandStore(MOI.getValue(), MI);
1029 else if (const auto &MOI = MOA.getAtomicFenceInfo(MI))
1030 Changed |= expandAtomicFence(MOI.getValue(), MI);
1031 else if (const auto &MOI = MOA.getAtomicCmpxchgOrRmwInfo(MI))
1032 Changed |= expandAtomicCmpxchgOrRmw(MOI.getValue(), MI);
1036 Changed |= removeAtomicPseudoMIs();
1040 INITIALIZE_PASS(SIMemoryLegalizer, DEBUG_TYPE, PASS_NAME, false, false)
1042 char SIMemoryLegalizer::ID = 0;
1043 char &llvm::SIMemoryLegalizerID = SIMemoryLegalizer::ID;
1045 FunctionPass *llvm::createSIMemoryLegalizerPass() {
1046 return new SIMemoryLegalizer();