1 | //===- SIMemoryLegalizer.cpp ----------------------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | /// \file |
10 | /// Memory legalizer - implements memory model. More information can be |
11 | /// found here: |
12 | /// http://llvm.org/docs/AMDGPUUsage.html#memory-model |
13 | // |
14 | //===----------------------------------------------------------------------===// |
15 | |
16 | #include "AMDGPU.h" |
17 | #include "AMDGPUMachineModuleInfo.h" |
18 | #include "GCNSubtarget.h" |
19 | #include "MCTargetDesc/AMDGPUMCTargetDesc.h" |
20 | #include "llvm/ADT/BitmaskEnum.h" |
21 | #include "llvm/ADT/StringExtras.h" |
22 | #include "llvm/CodeGen/MachineBasicBlock.h" |
23 | #include "llvm/CodeGen/MachineFunctionPass.h" |
24 | #include "llvm/CodeGen/MachinePassManager.h" |
25 | #include "llvm/IR/DiagnosticInfo.h" |
26 | #include "llvm/IR/MemoryModelRelaxationAnnotations.h" |
27 | #include "llvm/IR/PassManager.h" |
28 | #include "llvm/Support/AtomicOrdering.h" |
29 | #include "llvm/TargetParser/TargetParser.h" |
30 | |
31 | using namespace llvm; |
32 | using namespace llvm::AMDGPU; |
33 | |
34 | #define DEBUG_TYPE "si-memory-legalizer" |
35 | #define PASS_NAME "SI Memory Legalizer" |
36 | |
37 | static cl::opt<bool> AmdgcnSkipCacheInvalidations( |
38 | "amdgcn-skip-cache-invalidations" , cl::init(Val: false), cl::Hidden, |
39 | cl::desc("Use this to skip inserting cache invalidating instructions." )); |
40 | |
41 | namespace { |
42 | |
43 | LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE(); |
44 | |
45 | /// Memory operation flags. Can be ORed together. |
46 | enum class SIMemOp { |
47 | NONE = 0u, |
48 | LOAD = 1u << 0, |
49 | STORE = 1u << 1, |
50 | LLVM_MARK_AS_BITMASK_ENUM(/* LargestFlag = */ STORE) |
51 | }; |
52 | |
53 | /// Position to insert a new instruction relative to an existing |
54 | /// instruction. |
55 | enum class Position { |
56 | BEFORE, |
57 | AFTER |
58 | }; |
59 | |
60 | /// The atomic synchronization scopes supported by the AMDGPU target. |
61 | enum class SIAtomicScope { |
62 | NONE, |
63 | SINGLETHREAD, |
64 | WAVEFRONT, |
65 | WORKGROUP, |
66 | AGENT, |
67 | SYSTEM |
68 | }; |
69 | |
70 | /// The distinct address spaces supported by the AMDGPU target for |
71 | /// atomic memory operation. Can be ORed together. |
72 | enum class SIAtomicAddrSpace { |
73 | NONE = 0u, |
74 | GLOBAL = 1u << 0, |
75 | LDS = 1u << 1, |
76 | SCRATCH = 1u << 2, |
77 | GDS = 1u << 3, |
78 | OTHER = 1u << 4, |
79 | |
80 | /// The address spaces that can be accessed by a FLAT instruction. |
81 | FLAT = GLOBAL | LDS | SCRATCH, |
82 | |
83 | /// The address spaces that support atomic instructions. |
84 | ATOMIC = GLOBAL | LDS | SCRATCH | GDS, |
85 | |
86 | /// All address spaces. |
87 | ALL = GLOBAL | LDS | SCRATCH | GDS | OTHER, |
88 | |
89 | LLVM_MARK_AS_BITMASK_ENUM(/* LargestFlag = */ ALL) |
90 | }; |
91 | |
92 | class SIMemOpInfo final { |
93 | private: |
94 | |
95 | friend class SIMemOpAccess; |
96 | |
97 | AtomicOrdering Ordering = AtomicOrdering::NotAtomic; |
98 | AtomicOrdering FailureOrdering = AtomicOrdering::NotAtomic; |
99 | SIAtomicScope Scope = SIAtomicScope::SYSTEM; |
100 | SIAtomicAddrSpace OrderingAddrSpace = SIAtomicAddrSpace::NONE; |
101 | SIAtomicAddrSpace InstrAddrSpace = SIAtomicAddrSpace::NONE; |
102 | bool IsCrossAddressSpaceOrdering = false; |
103 | bool IsVolatile = false; |
104 | bool IsNonTemporal = false; |
105 | bool IsLastUse = false; |
106 | |
107 | SIMemOpInfo( |
108 | AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent, |
109 | SIAtomicScope Scope = SIAtomicScope::SYSTEM, |
110 | SIAtomicAddrSpace OrderingAddrSpace = SIAtomicAddrSpace::ATOMIC, |
111 | SIAtomicAddrSpace InstrAddrSpace = SIAtomicAddrSpace::ALL, |
112 | bool IsCrossAddressSpaceOrdering = true, |
113 | AtomicOrdering FailureOrdering = AtomicOrdering::SequentiallyConsistent, |
114 | bool IsVolatile = false, bool IsNonTemporal = false, |
115 | bool IsLastUse = false) |
116 | : Ordering(Ordering), FailureOrdering(FailureOrdering), Scope(Scope), |
117 | OrderingAddrSpace(OrderingAddrSpace), InstrAddrSpace(InstrAddrSpace), |
118 | IsCrossAddressSpaceOrdering(IsCrossAddressSpaceOrdering), |
119 | IsVolatile(IsVolatile), IsNonTemporal(IsNonTemporal), |
120 | IsLastUse(IsLastUse) { |
121 | |
122 | if (Ordering == AtomicOrdering::NotAtomic) { |
123 | assert(Scope == SIAtomicScope::NONE && |
124 | OrderingAddrSpace == SIAtomicAddrSpace::NONE && |
125 | !IsCrossAddressSpaceOrdering && |
126 | FailureOrdering == AtomicOrdering::NotAtomic); |
127 | return; |
128 | } |
129 | |
130 | assert(Scope != SIAtomicScope::NONE && |
131 | (OrderingAddrSpace & SIAtomicAddrSpace::ATOMIC) != |
132 | SIAtomicAddrSpace::NONE && |
133 | (InstrAddrSpace & SIAtomicAddrSpace::ATOMIC) != |
134 | SIAtomicAddrSpace::NONE); |
135 | |
136 | // There is also no cross address space ordering if the ordering |
137 | // address space is the same as the instruction address space and |
138 | // only contains a single address space. |
139 | if ((OrderingAddrSpace == InstrAddrSpace) && |
140 | isPowerOf2_32(Value: uint32_t(InstrAddrSpace))) |
141 | this->IsCrossAddressSpaceOrdering = false; |
142 | |
143 | // Limit the scope to the maximum supported by the instruction's address |
144 | // spaces. |
145 | if ((InstrAddrSpace & ~SIAtomicAddrSpace::SCRATCH) == |
146 | SIAtomicAddrSpace::NONE) { |
147 | this->Scope = std::min(a: Scope, b: SIAtomicScope::SINGLETHREAD); |
148 | } else if ((InstrAddrSpace & |
149 | ~(SIAtomicAddrSpace::SCRATCH | SIAtomicAddrSpace::LDS)) == |
150 | SIAtomicAddrSpace::NONE) { |
151 | this->Scope = std::min(a: Scope, b: SIAtomicScope::WORKGROUP); |
152 | } else if ((InstrAddrSpace & |
153 | ~(SIAtomicAddrSpace::SCRATCH | SIAtomicAddrSpace::LDS | |
154 | SIAtomicAddrSpace::GDS)) == SIAtomicAddrSpace::NONE) { |
155 | this->Scope = std::min(a: Scope, b: SIAtomicScope::AGENT); |
156 | } |
157 | } |
158 | |
159 | public: |
160 | /// \returns Atomic synchronization scope of the machine instruction used to |
161 | /// create this SIMemOpInfo. |
162 | SIAtomicScope getScope() const { |
163 | return Scope; |
164 | } |
165 | |
166 | /// \returns Ordering constraint of the machine instruction used to |
167 | /// create this SIMemOpInfo. |
168 | AtomicOrdering getOrdering() const { |
169 | return Ordering; |
170 | } |
171 | |
172 | /// \returns Failure ordering constraint of the machine instruction used to |
173 | /// create this SIMemOpInfo. |
174 | AtomicOrdering getFailureOrdering() const { |
175 | return FailureOrdering; |
176 | } |
177 | |
178 | /// \returns The address spaces be accessed by the machine |
179 | /// instruction used to create this SIMemOpInfo. |
180 | SIAtomicAddrSpace getInstrAddrSpace() const { |
181 | return InstrAddrSpace; |
182 | } |
183 | |
184 | /// \returns The address spaces that must be ordered by the machine |
185 | /// instruction used to create this SIMemOpInfo. |
186 | SIAtomicAddrSpace getOrderingAddrSpace() const { |
187 | return OrderingAddrSpace; |
188 | } |
189 | |
190 | /// \returns Return true iff memory ordering of operations on |
191 | /// different address spaces is required. |
192 | bool getIsCrossAddressSpaceOrdering() const { |
193 | return IsCrossAddressSpaceOrdering; |
194 | } |
195 | |
196 | /// \returns True if memory access of the machine instruction used to |
197 | /// create this SIMemOpInfo is volatile, false otherwise. |
198 | bool isVolatile() const { |
199 | return IsVolatile; |
200 | } |
201 | |
202 | /// \returns True if memory access of the machine instruction used to |
203 | /// create this SIMemOpInfo is nontemporal, false otherwise. |
204 | bool isNonTemporal() const { |
205 | return IsNonTemporal; |
206 | } |
207 | |
208 | /// \returns True if memory access of the machine instruction used to |
209 | /// create this SIMemOpInfo is last use, false otherwise. |
210 | bool isLastUse() const { return IsLastUse; } |
211 | |
212 | /// \returns True if ordering constraint of the machine instruction used to |
213 | /// create this SIMemOpInfo is unordered or higher, false otherwise. |
214 | bool isAtomic() const { |
215 | return Ordering != AtomicOrdering::NotAtomic; |
216 | } |
217 | |
218 | }; |
219 | |
220 | class SIMemOpAccess final { |
221 | private: |
222 | const AMDGPUMachineModuleInfo *MMI = nullptr; |
223 | |
224 | /// Reports unsupported message \p Msg for \p MI to LLVM context. |
225 | void reportUnsupported(const MachineBasicBlock::iterator &MI, |
226 | const char *Msg) const; |
227 | |
228 | /// Inspects the target synchronization scope \p SSID and determines |
229 | /// the SI atomic scope it corresponds to, the address spaces it |
230 | /// covers, and whether the memory ordering applies between address |
231 | /// spaces. |
232 | std::optional<std::tuple<SIAtomicScope, SIAtomicAddrSpace, bool>> |
233 | toSIAtomicScope(SyncScope::ID SSID, SIAtomicAddrSpace InstrAddrSpace) const; |
234 | |
235 | /// \return Return a bit set of the address spaces accessed by \p AS. |
236 | SIAtomicAddrSpace toSIAtomicAddrSpace(unsigned AS) const; |
237 | |
238 | /// \returns Info constructed from \p MI, which has at least machine memory |
239 | /// operand. |
240 | std::optional<SIMemOpInfo> |
241 | constructFromMIWithMMO(const MachineBasicBlock::iterator &MI) const; |
242 | |
243 | public: |
244 | /// Construct class to support accessing the machine memory operands |
245 | /// of instructions in the machine function \p MF. |
246 | SIMemOpAccess(const AMDGPUMachineModuleInfo &MMI); |
247 | |
248 | /// \returns Load info if \p MI is a load operation, "std::nullopt" otherwise. |
249 | std::optional<SIMemOpInfo> |
250 | getLoadInfo(const MachineBasicBlock::iterator &MI) const; |
251 | |
252 | /// \returns Store info if \p MI is a store operation, "std::nullopt" |
253 | /// otherwise. |
254 | std::optional<SIMemOpInfo> |
255 | getStoreInfo(const MachineBasicBlock::iterator &MI) const; |
256 | |
257 | /// \returns Atomic fence info if \p MI is an atomic fence operation, |
258 | /// "std::nullopt" otherwise. |
259 | std::optional<SIMemOpInfo> |
260 | getAtomicFenceInfo(const MachineBasicBlock::iterator &MI) const; |
261 | |
262 | /// \returns Atomic cmpxchg/rmw info if \p MI is an atomic cmpxchg or |
263 | /// rmw operation, "std::nullopt" otherwise. |
264 | std::optional<SIMemOpInfo> |
265 | getAtomicCmpxchgOrRmwInfo(const MachineBasicBlock::iterator &MI) const; |
266 | }; |
267 | |
268 | class SICacheControl { |
269 | protected: |
270 | |
271 | /// AMDGPU subtarget info. |
272 | const GCNSubtarget &ST; |
273 | |
274 | /// Instruction info. |
275 | const SIInstrInfo *TII = nullptr; |
276 | |
277 | IsaVersion IV; |
278 | |
279 | /// Whether to insert cache invalidating instructions. |
280 | bool InsertCacheInv; |
281 | |
282 | SICacheControl(const GCNSubtarget &ST); |
283 | |
284 | /// Sets named bit \p BitName to "true" if present in instruction \p MI. |
285 | /// \returns Returns true if \p MI is modified, false otherwise. |
286 | bool enableNamedBit(const MachineBasicBlock::iterator MI, |
287 | AMDGPU::CPol::CPol Bit) const; |
288 | |
289 | public: |
290 | |
291 | /// Create a cache control for the subtarget \p ST. |
292 | static std::unique_ptr<SICacheControl> create(const GCNSubtarget &ST); |
293 | |
294 | /// Update \p MI memory load instruction to bypass any caches up to |
295 | /// the \p Scope memory scope for address spaces \p |
296 | /// AddrSpace. Return true iff the instruction was modified. |
297 | virtual bool enableLoadCacheBypass(const MachineBasicBlock::iterator &MI, |
298 | SIAtomicScope Scope, |
299 | SIAtomicAddrSpace AddrSpace) const = 0; |
300 | |
301 | /// Update \p MI memory store instruction to bypass any caches up to |
302 | /// the \p Scope memory scope for address spaces \p |
303 | /// AddrSpace. Return true iff the instruction was modified. |
304 | virtual bool enableStoreCacheBypass(const MachineBasicBlock::iterator &MI, |
305 | SIAtomicScope Scope, |
306 | SIAtomicAddrSpace AddrSpace) const = 0; |
307 | |
308 | /// Update \p MI memory read-modify-write instruction to bypass any caches up |
309 | /// to the \p Scope memory scope for address spaces \p AddrSpace. Return true |
310 | /// iff the instruction was modified. |
311 | virtual bool enableRMWCacheBypass(const MachineBasicBlock::iterator &MI, |
312 | SIAtomicScope Scope, |
313 | SIAtomicAddrSpace AddrSpace) const = 0; |
314 | |
315 | /// Update \p MI memory instruction of kind \p Op associated with address |
316 | /// spaces \p AddrSpace to indicate it is volatile and/or |
317 | /// nontemporal/last-use. Return true iff the instruction was modified. |
318 | virtual bool enableVolatileAndOrNonTemporal(MachineBasicBlock::iterator &MI, |
319 | SIAtomicAddrSpace AddrSpace, |
320 | SIMemOp Op, bool IsVolatile, |
321 | bool IsNonTemporal, |
322 | bool IsLastUse = false) const = 0; |
323 | |
324 | virtual bool expandSystemScopeStore(MachineBasicBlock::iterator &MI) const { |
325 | return false; |
326 | }; |
327 | |
328 | /// Inserts any necessary instructions at position \p Pos relative |
329 | /// to instruction \p MI to ensure memory instructions before \p Pos of kind |
330 | /// \p Op associated with address spaces \p AddrSpace have completed. Used |
331 | /// between memory instructions to enforce the order they become visible as |
332 | /// observed by other memory instructions executing in memory scope \p Scope. |
333 | /// \p IsCrossAddrSpaceOrdering indicates if the memory ordering is between |
334 | /// address spaces. Returns true iff any instructions inserted. |
335 | virtual bool insertWait(MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
336 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
337 | bool IsCrossAddrSpaceOrdering, Position Pos, |
338 | AtomicOrdering Order) const = 0; |
339 | |
340 | /// Inserts any necessary instructions at position \p Pos relative to |
341 | /// instruction \p MI to ensure any subsequent memory instructions of this |
342 | /// thread with address spaces \p AddrSpace will observe the previous memory |
343 | /// operations by any thread for memory scopes up to memory scope \p Scope . |
344 | /// Returns true iff any instructions inserted. |
345 | virtual bool insertAcquire(MachineBasicBlock::iterator &MI, |
346 | SIAtomicScope Scope, |
347 | SIAtomicAddrSpace AddrSpace, |
348 | Position Pos) const = 0; |
349 | |
350 | /// Inserts any necessary instructions at position \p Pos relative to |
351 | /// instruction \p MI to ensure previous memory instructions by this thread |
352 | /// with address spaces \p AddrSpace have completed and can be observed by |
353 | /// subsequent memory instructions by any thread executing in memory scope \p |
354 | /// Scope. \p IsCrossAddrSpaceOrdering indicates if the memory ordering is |
355 | /// between address spaces. Returns true iff any instructions inserted. |
356 | virtual bool insertRelease(MachineBasicBlock::iterator &MI, |
357 | SIAtomicScope Scope, |
358 | SIAtomicAddrSpace AddrSpace, |
359 | bool IsCrossAddrSpaceOrdering, |
360 | Position Pos) const = 0; |
361 | |
362 | /// Virtual destructor to allow derivations to be deleted. |
363 | virtual ~SICacheControl() = default; |
364 | }; |
365 | |
366 | class SIGfx6CacheControl : public SICacheControl { |
367 | protected: |
368 | |
369 | /// Sets GLC bit to "true" if present in \p MI. Returns true if \p MI |
370 | /// is modified, false otherwise. |
371 | bool enableGLCBit(const MachineBasicBlock::iterator &MI) const { |
372 | return enableNamedBit(MI, Bit: AMDGPU::CPol::GLC); |
373 | } |
374 | |
375 | /// Sets SLC bit to "true" if present in \p MI. Returns true if \p MI |
376 | /// is modified, false otherwise. |
377 | bool enableSLCBit(const MachineBasicBlock::iterator &MI) const { |
378 | return enableNamedBit(MI, Bit: AMDGPU::CPol::SLC); |
379 | } |
380 | |
381 | public: |
382 | |
383 | SIGfx6CacheControl(const GCNSubtarget &ST) : SICacheControl(ST) {} |
384 | |
385 | bool enableLoadCacheBypass(const MachineBasicBlock::iterator &MI, |
386 | SIAtomicScope Scope, |
387 | SIAtomicAddrSpace AddrSpace) const override; |
388 | |
389 | bool enableStoreCacheBypass(const MachineBasicBlock::iterator &MI, |
390 | SIAtomicScope Scope, |
391 | SIAtomicAddrSpace AddrSpace) const override; |
392 | |
393 | bool enableRMWCacheBypass(const MachineBasicBlock::iterator &MI, |
394 | SIAtomicScope Scope, |
395 | SIAtomicAddrSpace AddrSpace) const override; |
396 | |
397 | bool enableVolatileAndOrNonTemporal(MachineBasicBlock::iterator &MI, |
398 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
399 | bool IsVolatile, bool IsNonTemporal, |
400 | bool IsLastUse) const override; |
401 | |
402 | bool insertWait(MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
403 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
404 | bool IsCrossAddrSpaceOrdering, Position Pos, |
405 | AtomicOrdering Order) const override; |
406 | |
407 | bool insertAcquire(MachineBasicBlock::iterator &MI, |
408 | SIAtomicScope Scope, |
409 | SIAtomicAddrSpace AddrSpace, |
410 | Position Pos) const override; |
411 | |
412 | bool insertRelease(MachineBasicBlock::iterator &MI, |
413 | SIAtomicScope Scope, |
414 | SIAtomicAddrSpace AddrSpace, |
415 | bool IsCrossAddrSpaceOrdering, |
416 | Position Pos) const override; |
417 | }; |
418 | |
419 | class SIGfx7CacheControl : public SIGfx6CacheControl { |
420 | public: |
421 | |
422 | SIGfx7CacheControl(const GCNSubtarget &ST) : SIGfx6CacheControl(ST) {} |
423 | |
424 | bool insertAcquire(MachineBasicBlock::iterator &MI, |
425 | SIAtomicScope Scope, |
426 | SIAtomicAddrSpace AddrSpace, |
427 | Position Pos) const override; |
428 | |
429 | }; |
430 | |
431 | class SIGfx90ACacheControl : public SIGfx7CacheControl { |
432 | public: |
433 | |
434 | SIGfx90ACacheControl(const GCNSubtarget &ST) : SIGfx7CacheControl(ST) {} |
435 | |
436 | bool enableLoadCacheBypass(const MachineBasicBlock::iterator &MI, |
437 | SIAtomicScope Scope, |
438 | SIAtomicAddrSpace AddrSpace) const override; |
439 | |
440 | bool enableStoreCacheBypass(const MachineBasicBlock::iterator &MI, |
441 | SIAtomicScope Scope, |
442 | SIAtomicAddrSpace AddrSpace) const override; |
443 | |
444 | bool enableRMWCacheBypass(const MachineBasicBlock::iterator &MI, |
445 | SIAtomicScope Scope, |
446 | SIAtomicAddrSpace AddrSpace) const override; |
447 | |
448 | bool enableVolatileAndOrNonTemporal(MachineBasicBlock::iterator &MI, |
449 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
450 | bool IsVolatile, bool IsNonTemporal, |
451 | bool IsLastUse) const override; |
452 | |
453 | bool insertWait(MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
454 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
455 | bool IsCrossAddrSpaceOrdering, Position Pos, |
456 | AtomicOrdering Order) const override; |
457 | |
458 | bool insertAcquire(MachineBasicBlock::iterator &MI, |
459 | SIAtomicScope Scope, |
460 | SIAtomicAddrSpace AddrSpace, |
461 | Position Pos) const override; |
462 | |
463 | bool insertRelease(MachineBasicBlock::iterator &MI, |
464 | SIAtomicScope Scope, |
465 | SIAtomicAddrSpace AddrSpace, |
466 | bool IsCrossAddrSpaceOrdering, |
467 | Position Pos) const override; |
468 | }; |
469 | |
470 | class SIGfx940CacheControl : public SIGfx90ACacheControl { |
471 | protected: |
472 | |
473 | /// Sets SC0 bit to "true" if present in \p MI. Returns true if \p MI |
474 | /// is modified, false otherwise. |
475 | bool enableSC0Bit(const MachineBasicBlock::iterator &MI) const { |
476 | return enableNamedBit(MI, Bit: AMDGPU::CPol::SC0); |
477 | } |
478 | |
479 | /// Sets SC1 bit to "true" if present in \p MI. Returns true if \p MI |
480 | /// is modified, false otherwise. |
481 | bool enableSC1Bit(const MachineBasicBlock::iterator &MI) const { |
482 | return enableNamedBit(MI, Bit: AMDGPU::CPol::SC1); |
483 | } |
484 | |
485 | /// Sets NT bit to "true" if present in \p MI. Returns true if \p MI |
486 | /// is modified, false otherwise. |
487 | bool enableNTBit(const MachineBasicBlock::iterator &MI) const { |
488 | return enableNamedBit(MI, Bit: AMDGPU::CPol::NT); |
489 | } |
490 | |
491 | public: |
492 | SIGfx940CacheControl(const GCNSubtarget &ST) : SIGfx90ACacheControl(ST) {}; |
493 | |
494 | bool enableLoadCacheBypass(const MachineBasicBlock::iterator &MI, |
495 | SIAtomicScope Scope, |
496 | SIAtomicAddrSpace AddrSpace) const override; |
497 | |
498 | bool enableStoreCacheBypass(const MachineBasicBlock::iterator &MI, |
499 | SIAtomicScope Scope, |
500 | SIAtomicAddrSpace AddrSpace) const override; |
501 | |
502 | bool enableRMWCacheBypass(const MachineBasicBlock::iterator &MI, |
503 | SIAtomicScope Scope, |
504 | SIAtomicAddrSpace AddrSpace) const override; |
505 | |
506 | bool enableVolatileAndOrNonTemporal(MachineBasicBlock::iterator &MI, |
507 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
508 | bool IsVolatile, bool IsNonTemporal, |
509 | bool IsLastUse) const override; |
510 | |
511 | bool insertAcquire(MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
512 | SIAtomicAddrSpace AddrSpace, Position Pos) const override; |
513 | |
514 | bool insertRelease(MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
515 | SIAtomicAddrSpace AddrSpace, bool IsCrossAddrSpaceOrdering, |
516 | Position Pos) const override; |
517 | }; |
518 | |
519 | class SIGfx10CacheControl : public SIGfx7CacheControl { |
520 | protected: |
521 | |
522 | /// Sets DLC bit to "true" if present in \p MI. Returns true if \p MI |
523 | /// is modified, false otherwise. |
524 | bool enableDLCBit(const MachineBasicBlock::iterator &MI) const { |
525 | return enableNamedBit(MI, Bit: AMDGPU::CPol::DLC); |
526 | } |
527 | |
528 | public: |
529 | |
530 | SIGfx10CacheControl(const GCNSubtarget &ST) : SIGfx7CacheControl(ST) {} |
531 | |
532 | bool enableLoadCacheBypass(const MachineBasicBlock::iterator &MI, |
533 | SIAtomicScope Scope, |
534 | SIAtomicAddrSpace AddrSpace) const override; |
535 | |
536 | bool enableVolatileAndOrNonTemporal(MachineBasicBlock::iterator &MI, |
537 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
538 | bool IsVolatile, bool IsNonTemporal, |
539 | bool IsLastUse) const override; |
540 | |
541 | bool insertWait(MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
542 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
543 | bool IsCrossAddrSpaceOrdering, Position Pos, |
544 | AtomicOrdering Order) const override; |
545 | |
546 | bool insertAcquire(MachineBasicBlock::iterator &MI, |
547 | SIAtomicScope Scope, |
548 | SIAtomicAddrSpace AddrSpace, |
549 | Position Pos) const override; |
550 | }; |
551 | |
552 | class SIGfx11CacheControl : public SIGfx10CacheControl { |
553 | public: |
554 | SIGfx11CacheControl(const GCNSubtarget &ST) : SIGfx10CacheControl(ST) {} |
555 | |
556 | bool enableLoadCacheBypass(const MachineBasicBlock::iterator &MI, |
557 | SIAtomicScope Scope, |
558 | SIAtomicAddrSpace AddrSpace) const override; |
559 | |
560 | bool enableVolatileAndOrNonTemporal(MachineBasicBlock::iterator &MI, |
561 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
562 | bool IsVolatile, bool IsNonTemporal, |
563 | bool IsLastUse) const override; |
564 | }; |
565 | |
566 | class SIGfx12CacheControl : public SIGfx11CacheControl { |
567 | protected: |
568 | // Sets TH policy to \p Value if CPol operand is present in instruction \p MI. |
569 | // \returns Returns true if \p MI is modified, false otherwise. |
570 | bool setTH(const MachineBasicBlock::iterator MI, |
571 | AMDGPU::CPol::CPol Value) const; |
572 | // Sets Scope policy to \p Value if CPol operand is present in instruction \p |
573 | // MI. \returns Returns true if \p MI is modified, false otherwise. |
574 | bool setScope(const MachineBasicBlock::iterator MI, |
575 | AMDGPU::CPol::CPol Value) const; |
576 | |
577 | // Stores with system scope (SCOPE_SYS) need to wait for: |
578 | // - loads or atomics(returning) - wait for {LOAD|SAMPLE|BVH|KM}CNT==0 |
579 | // - non-returning-atomics - wait for STORECNT==0 |
580 | // TODO: SIInsertWaitcnts will not always be able to remove STORECNT waits |
581 | // since it does not distinguish atomics-with-return from regular stores. |
582 | // There is no need to wait if memory is cached (mtype != UC). |
583 | bool |
584 | insertWaitsBeforeSystemScopeStore(const MachineBasicBlock::iterator MI) const; |
585 | |
586 | bool setAtomicScope(const MachineBasicBlock::iterator &MI, |
587 | SIAtomicScope Scope, SIAtomicAddrSpace AddrSpace) const; |
588 | |
589 | public: |
590 | SIGfx12CacheControl(const GCNSubtarget &ST) : SIGfx11CacheControl(ST) {} |
591 | |
592 | bool insertWait(MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
593 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
594 | bool IsCrossAddrSpaceOrdering, Position Pos, |
595 | AtomicOrdering Order) const override; |
596 | |
597 | bool insertAcquire(MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
598 | SIAtomicAddrSpace AddrSpace, Position Pos) const override; |
599 | |
600 | bool enableVolatileAndOrNonTemporal(MachineBasicBlock::iterator &MI, |
601 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
602 | bool IsVolatile, bool IsNonTemporal, |
603 | bool IsLastUse) const override; |
604 | |
605 | bool expandSystemScopeStore(MachineBasicBlock::iterator &MI) const override; |
606 | |
607 | bool insertRelease(MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
608 | SIAtomicAddrSpace AddrSpace, bool IsCrossAddrSpaceOrdering, |
609 | Position Pos) const override; |
610 | |
611 | bool enableLoadCacheBypass(const MachineBasicBlock::iterator &MI, |
612 | SIAtomicScope Scope, |
613 | SIAtomicAddrSpace AddrSpace) const override { |
614 | return setAtomicScope(MI, Scope, AddrSpace); |
615 | } |
616 | |
617 | bool enableStoreCacheBypass(const MachineBasicBlock::iterator &MI, |
618 | SIAtomicScope Scope, |
619 | SIAtomicAddrSpace AddrSpace) const override { |
620 | return setAtomicScope(MI, Scope, AddrSpace); |
621 | } |
622 | |
623 | bool enableRMWCacheBypass(const MachineBasicBlock::iterator &MI, |
624 | SIAtomicScope Scope, |
625 | SIAtomicAddrSpace AddrSpace) const override { |
626 | return setAtomicScope(MI, Scope, AddrSpace); |
627 | } |
628 | }; |
629 | |
630 | class SIMemoryLegalizer final { |
631 | private: |
632 | const MachineModuleInfo &MMI; |
633 | /// Cache Control. |
634 | std::unique_ptr<SICacheControl> CC = nullptr; |
635 | |
636 | /// List of atomic pseudo instructions. |
637 | std::list<MachineBasicBlock::iterator> AtomicPseudoMIs; |
638 | |
639 | /// Return true iff instruction \p MI is a atomic instruction that |
640 | /// returns a result. |
641 | bool isAtomicRet(const MachineInstr &MI) const { |
642 | return SIInstrInfo::isAtomicRet(MI); |
643 | } |
644 | |
645 | /// Removes all processed atomic pseudo instructions from the current |
646 | /// function. Returns true if current function is modified, false otherwise. |
647 | bool removeAtomicPseudoMIs(); |
648 | |
649 | /// Expands load operation \p MI. Returns true if instructions are |
650 | /// added/deleted or \p MI is modified, false otherwise. |
651 | bool expandLoad(const SIMemOpInfo &MOI, |
652 | MachineBasicBlock::iterator &MI); |
653 | /// Expands store operation \p MI. Returns true if instructions are |
654 | /// added/deleted or \p MI is modified, false otherwise. |
655 | bool expandStore(const SIMemOpInfo &MOI, |
656 | MachineBasicBlock::iterator &MI); |
657 | /// Expands atomic fence operation \p MI. Returns true if |
658 | /// instructions are added/deleted or \p MI is modified, false otherwise. |
659 | bool expandAtomicFence(const SIMemOpInfo &MOI, |
660 | MachineBasicBlock::iterator &MI); |
661 | /// Expands atomic cmpxchg or rmw operation \p MI. Returns true if |
662 | /// instructions are added/deleted or \p MI is modified, false otherwise. |
663 | bool expandAtomicCmpxchgOrRmw(const SIMemOpInfo &MOI, |
664 | MachineBasicBlock::iterator &MI); |
665 | |
666 | public: |
667 | SIMemoryLegalizer(const MachineModuleInfo &MMI) : MMI(MMI) {}; |
668 | bool run(MachineFunction &MF); |
669 | }; |
670 | |
671 | class SIMemoryLegalizerLegacy final : public MachineFunctionPass { |
672 | public: |
673 | static char ID; |
674 | |
675 | SIMemoryLegalizerLegacy() : MachineFunctionPass(ID) {} |
676 | |
677 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
678 | AU.setPreservesCFG(); |
679 | MachineFunctionPass::getAnalysisUsage(AU); |
680 | } |
681 | |
682 | StringRef getPassName() const override { |
683 | return PASS_NAME; |
684 | } |
685 | |
686 | bool runOnMachineFunction(MachineFunction &MF) override; |
687 | }; |
688 | |
689 | static const StringMap<SIAtomicAddrSpace> ASNames = {{ |
690 | {"global" , SIAtomicAddrSpace::GLOBAL}, |
691 | {"local" , SIAtomicAddrSpace::LDS}, |
692 | }}; |
693 | |
694 | void diagnoseUnknownMMRAASName(const MachineInstr &MI, StringRef AS) { |
695 | const MachineFunction *MF = MI.getMF(); |
696 | const Function &Fn = MF->getFunction(); |
697 | SmallString<128> Str; |
698 | raw_svector_ostream OS(Str); |
699 | OS << "unknown address space '" << AS << "'; expected one of " ; |
700 | ListSeparator LS; |
701 | for (const auto &[Name, Val] : ASNames) |
702 | OS << LS << '\'' << Name << '\''; |
703 | Fn.getContext().diagnose( |
704 | DI: DiagnosticInfoUnsupported(Fn, Str.str(), MI.getDebugLoc(), DS_Warning)); |
705 | } |
706 | |
707 | /// Reads \p MI's MMRAs to parse the "amdgpu-as" MMRA. |
708 | /// If this tag isn't present, or if it has no meaningful values, returns \p |
709 | /// Default. Otherwise returns all the address spaces concerned by the MMRA. |
710 | static SIAtomicAddrSpace getFenceAddrSpaceMMRA(const MachineInstr &MI, |
711 | SIAtomicAddrSpace Default) { |
712 | static constexpr StringLiteral FenceASPrefix = "amdgpu-as" ; |
713 | |
714 | auto MMRA = MMRAMetadata(MI.getMMRAMetadata()); |
715 | if (!MMRA) |
716 | return Default; |
717 | |
718 | SIAtomicAddrSpace Result = SIAtomicAddrSpace::NONE; |
719 | for (const auto &[Prefix, Suffix] : MMRA) { |
720 | if (Prefix != FenceASPrefix) |
721 | continue; |
722 | |
723 | if (auto It = ASNames.find(Key: Suffix); It != ASNames.end()) |
724 | Result |= It->second; |
725 | else |
726 | diagnoseUnknownMMRAASName(MI, AS: Suffix); |
727 | } |
728 | |
729 | return (Result != SIAtomicAddrSpace::NONE) ? Result : Default; |
730 | } |
731 | |
732 | } // end anonymous namespace |
733 | |
734 | void SIMemOpAccess::reportUnsupported(const MachineBasicBlock::iterator &MI, |
735 | const char *Msg) const { |
736 | const Function &Func = MI->getParent()->getParent()->getFunction(); |
737 | Func.getContext().diagnose( |
738 | DI: DiagnosticInfoUnsupported(Func, Msg, MI->getDebugLoc())); |
739 | } |
740 | |
741 | std::optional<std::tuple<SIAtomicScope, SIAtomicAddrSpace, bool>> |
742 | SIMemOpAccess::toSIAtomicScope(SyncScope::ID SSID, |
743 | SIAtomicAddrSpace InstrAddrSpace) const { |
744 | if (SSID == SyncScope::System) |
745 | return std::tuple(SIAtomicScope::SYSTEM, SIAtomicAddrSpace::ATOMIC, true); |
746 | if (SSID == MMI->getAgentSSID()) |
747 | return std::tuple(SIAtomicScope::AGENT, SIAtomicAddrSpace::ATOMIC, true); |
748 | if (SSID == MMI->getWorkgroupSSID()) |
749 | return std::tuple(SIAtomicScope::WORKGROUP, SIAtomicAddrSpace::ATOMIC, |
750 | true); |
751 | if (SSID == MMI->getWavefrontSSID()) |
752 | return std::tuple(SIAtomicScope::WAVEFRONT, SIAtomicAddrSpace::ATOMIC, |
753 | true); |
754 | if (SSID == SyncScope::SingleThread) |
755 | return std::tuple(SIAtomicScope::SINGLETHREAD, SIAtomicAddrSpace::ATOMIC, |
756 | true); |
757 | if (SSID == MMI->getSystemOneAddressSpaceSSID()) |
758 | return std::tuple(SIAtomicScope::SYSTEM, |
759 | SIAtomicAddrSpace::ATOMIC & InstrAddrSpace, false); |
760 | if (SSID == MMI->getAgentOneAddressSpaceSSID()) |
761 | return std::tuple(SIAtomicScope::AGENT, |
762 | SIAtomicAddrSpace::ATOMIC & InstrAddrSpace, false); |
763 | if (SSID == MMI->getWorkgroupOneAddressSpaceSSID()) |
764 | return std::tuple(SIAtomicScope::WORKGROUP, |
765 | SIAtomicAddrSpace::ATOMIC & InstrAddrSpace, false); |
766 | if (SSID == MMI->getWavefrontOneAddressSpaceSSID()) |
767 | return std::tuple(SIAtomicScope::WAVEFRONT, |
768 | SIAtomicAddrSpace::ATOMIC & InstrAddrSpace, false); |
769 | if (SSID == MMI->getSingleThreadOneAddressSpaceSSID()) |
770 | return std::tuple(SIAtomicScope::SINGLETHREAD, |
771 | SIAtomicAddrSpace::ATOMIC & InstrAddrSpace, false); |
772 | return std::nullopt; |
773 | } |
774 | |
775 | SIAtomicAddrSpace SIMemOpAccess::toSIAtomicAddrSpace(unsigned AS) const { |
776 | if (AS == AMDGPUAS::FLAT_ADDRESS) |
777 | return SIAtomicAddrSpace::FLAT; |
778 | if (AS == AMDGPUAS::GLOBAL_ADDRESS) |
779 | return SIAtomicAddrSpace::GLOBAL; |
780 | if (AS == AMDGPUAS::LOCAL_ADDRESS) |
781 | return SIAtomicAddrSpace::LDS; |
782 | if (AS == AMDGPUAS::PRIVATE_ADDRESS) |
783 | return SIAtomicAddrSpace::SCRATCH; |
784 | if (AS == AMDGPUAS::REGION_ADDRESS) |
785 | return SIAtomicAddrSpace::GDS; |
786 | |
787 | return SIAtomicAddrSpace::OTHER; |
788 | } |
789 | |
790 | SIMemOpAccess::SIMemOpAccess(const AMDGPUMachineModuleInfo &MMI_) |
791 | : MMI(&MMI_) {} |
792 | |
793 | std::optional<SIMemOpInfo> SIMemOpAccess::constructFromMIWithMMO( |
794 | const MachineBasicBlock::iterator &MI) const { |
795 | assert(MI->getNumMemOperands() > 0); |
796 | |
797 | SyncScope::ID SSID = SyncScope::SingleThread; |
798 | AtomicOrdering Ordering = AtomicOrdering::NotAtomic; |
799 | AtomicOrdering FailureOrdering = AtomicOrdering::NotAtomic; |
800 | SIAtomicAddrSpace InstrAddrSpace = SIAtomicAddrSpace::NONE; |
801 | bool IsNonTemporal = true; |
802 | bool IsVolatile = false; |
803 | bool IsLastUse = false; |
804 | |
805 | // Validator should check whether or not MMOs cover the entire set of |
806 | // locations accessed by the memory instruction. |
807 | for (const auto &MMO : MI->memoperands()) { |
808 | IsNonTemporal &= MMO->isNonTemporal(); |
809 | IsVolatile |= MMO->isVolatile(); |
810 | IsLastUse |= MMO->getFlags() & MOLastUse; |
811 | InstrAddrSpace |= |
812 | toSIAtomicAddrSpace(AS: MMO->getPointerInfo().getAddrSpace()); |
813 | AtomicOrdering OpOrdering = MMO->getSuccessOrdering(); |
814 | if (OpOrdering != AtomicOrdering::NotAtomic) { |
815 | const auto &IsSyncScopeInclusion = |
816 | MMI->isSyncScopeInclusion(A: SSID, B: MMO->getSyncScopeID()); |
817 | if (!IsSyncScopeInclusion) { |
818 | reportUnsupported(MI, |
819 | Msg: "Unsupported non-inclusive atomic synchronization scope" ); |
820 | return std::nullopt; |
821 | } |
822 | |
823 | SSID = *IsSyncScopeInclusion ? SSID : MMO->getSyncScopeID(); |
824 | Ordering = getMergedAtomicOrdering(AO: Ordering, Other: OpOrdering); |
825 | assert(MMO->getFailureOrdering() != AtomicOrdering::Release && |
826 | MMO->getFailureOrdering() != AtomicOrdering::AcquireRelease); |
827 | FailureOrdering = |
828 | getMergedAtomicOrdering(AO: FailureOrdering, Other: MMO->getFailureOrdering()); |
829 | } |
830 | } |
831 | |
832 | SIAtomicScope Scope = SIAtomicScope::NONE; |
833 | SIAtomicAddrSpace OrderingAddrSpace = SIAtomicAddrSpace::NONE; |
834 | bool IsCrossAddressSpaceOrdering = false; |
835 | if (Ordering != AtomicOrdering::NotAtomic) { |
836 | auto ScopeOrNone = toSIAtomicScope(SSID, InstrAddrSpace); |
837 | if (!ScopeOrNone) { |
838 | reportUnsupported(MI, Msg: "Unsupported atomic synchronization scope" ); |
839 | return std::nullopt; |
840 | } |
841 | std::tie(args&: Scope, args&: OrderingAddrSpace, args&: IsCrossAddressSpaceOrdering) = |
842 | *ScopeOrNone; |
843 | if ((OrderingAddrSpace == SIAtomicAddrSpace::NONE) || |
844 | ((OrderingAddrSpace & SIAtomicAddrSpace::ATOMIC) != OrderingAddrSpace) || |
845 | ((InstrAddrSpace & SIAtomicAddrSpace::ATOMIC) == SIAtomicAddrSpace::NONE)) { |
846 | reportUnsupported(MI, Msg: "Unsupported atomic address space" ); |
847 | return std::nullopt; |
848 | } |
849 | } |
850 | return SIMemOpInfo(Ordering, Scope, OrderingAddrSpace, InstrAddrSpace, |
851 | IsCrossAddressSpaceOrdering, FailureOrdering, IsVolatile, |
852 | IsNonTemporal, IsLastUse); |
853 | } |
854 | |
855 | std::optional<SIMemOpInfo> |
856 | SIMemOpAccess::getLoadInfo(const MachineBasicBlock::iterator &MI) const { |
857 | assert(MI->getDesc().TSFlags & SIInstrFlags::maybeAtomic); |
858 | |
859 | if (!(MI->mayLoad() && !MI->mayStore())) |
860 | return std::nullopt; |
861 | |
862 | // Be conservative if there are no memory operands. |
863 | if (MI->getNumMemOperands() == 0) |
864 | return SIMemOpInfo(); |
865 | |
866 | return constructFromMIWithMMO(MI); |
867 | } |
868 | |
869 | std::optional<SIMemOpInfo> |
870 | SIMemOpAccess::getStoreInfo(const MachineBasicBlock::iterator &MI) const { |
871 | assert(MI->getDesc().TSFlags & SIInstrFlags::maybeAtomic); |
872 | |
873 | if (!(!MI->mayLoad() && MI->mayStore())) |
874 | return std::nullopt; |
875 | |
876 | // Be conservative if there are no memory operands. |
877 | if (MI->getNumMemOperands() == 0) |
878 | return SIMemOpInfo(); |
879 | |
880 | return constructFromMIWithMMO(MI); |
881 | } |
882 | |
883 | std::optional<SIMemOpInfo> |
884 | SIMemOpAccess::getAtomicFenceInfo(const MachineBasicBlock::iterator &MI) const { |
885 | assert(MI->getDesc().TSFlags & SIInstrFlags::maybeAtomic); |
886 | |
887 | if (MI->getOpcode() != AMDGPU::ATOMIC_FENCE) |
888 | return std::nullopt; |
889 | |
890 | AtomicOrdering Ordering = |
891 | static_cast<AtomicOrdering>(MI->getOperand(i: 0).getImm()); |
892 | |
893 | SyncScope::ID SSID = static_cast<SyncScope::ID>(MI->getOperand(i: 1).getImm()); |
894 | auto ScopeOrNone = toSIAtomicScope(SSID, InstrAddrSpace: SIAtomicAddrSpace::ATOMIC); |
895 | if (!ScopeOrNone) { |
896 | reportUnsupported(MI, Msg: "Unsupported atomic synchronization scope" ); |
897 | return std::nullopt; |
898 | } |
899 | |
900 | SIAtomicScope Scope = SIAtomicScope::NONE; |
901 | SIAtomicAddrSpace OrderingAddrSpace = SIAtomicAddrSpace::NONE; |
902 | bool IsCrossAddressSpaceOrdering = false; |
903 | std::tie(args&: Scope, args&: OrderingAddrSpace, args&: IsCrossAddressSpaceOrdering) = |
904 | *ScopeOrNone; |
905 | |
906 | if ((OrderingAddrSpace == SIAtomicAddrSpace::NONE) || |
907 | ((OrderingAddrSpace & SIAtomicAddrSpace::ATOMIC) != OrderingAddrSpace)) { |
908 | reportUnsupported(MI, Msg: "Unsupported atomic address space" ); |
909 | return std::nullopt; |
910 | } |
911 | |
912 | return SIMemOpInfo(Ordering, Scope, OrderingAddrSpace, SIAtomicAddrSpace::ATOMIC, |
913 | IsCrossAddressSpaceOrdering, AtomicOrdering::NotAtomic); |
914 | } |
915 | |
916 | std::optional<SIMemOpInfo> SIMemOpAccess::getAtomicCmpxchgOrRmwInfo( |
917 | const MachineBasicBlock::iterator &MI) const { |
918 | assert(MI->getDesc().TSFlags & SIInstrFlags::maybeAtomic); |
919 | |
920 | if (!(MI->mayLoad() && MI->mayStore())) |
921 | return std::nullopt; |
922 | |
923 | // Be conservative if there are no memory operands. |
924 | if (MI->getNumMemOperands() == 0) |
925 | return SIMemOpInfo(); |
926 | |
927 | return constructFromMIWithMMO(MI); |
928 | } |
929 | |
930 | SICacheControl::SICacheControl(const GCNSubtarget &ST) : ST(ST) { |
931 | TII = ST.getInstrInfo(); |
932 | IV = getIsaVersion(GPU: ST.getCPU()); |
933 | InsertCacheInv = !AmdgcnSkipCacheInvalidations; |
934 | } |
935 | |
936 | bool SICacheControl::enableNamedBit(const MachineBasicBlock::iterator MI, |
937 | AMDGPU::CPol::CPol Bit) const { |
938 | MachineOperand *CPol = TII->getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::cpol); |
939 | if (!CPol) |
940 | return false; |
941 | |
942 | CPol->setImm(CPol->getImm() | Bit); |
943 | return true; |
944 | } |
945 | |
946 | /* static */ |
947 | std::unique_ptr<SICacheControl> SICacheControl::create(const GCNSubtarget &ST) { |
948 | GCNSubtarget::Generation Generation = ST.getGeneration(); |
949 | if (ST.hasGFX940Insts()) |
950 | return std::make_unique<SIGfx940CacheControl>(args: ST); |
951 | if (ST.hasGFX90AInsts()) |
952 | return std::make_unique<SIGfx90ACacheControl>(args: ST); |
953 | if (Generation <= AMDGPUSubtarget::SOUTHERN_ISLANDS) |
954 | return std::make_unique<SIGfx6CacheControl>(args: ST); |
955 | if (Generation < AMDGPUSubtarget::GFX10) |
956 | return std::make_unique<SIGfx7CacheControl>(args: ST); |
957 | if (Generation < AMDGPUSubtarget::GFX11) |
958 | return std::make_unique<SIGfx10CacheControl>(args: ST); |
959 | if (Generation < AMDGPUSubtarget::GFX12) |
960 | return std::make_unique<SIGfx11CacheControl>(args: ST); |
961 | return std::make_unique<SIGfx12CacheControl>(args: ST); |
962 | } |
963 | |
964 | bool SIGfx6CacheControl::enableLoadCacheBypass( |
965 | const MachineBasicBlock::iterator &MI, |
966 | SIAtomicScope Scope, |
967 | SIAtomicAddrSpace AddrSpace) const { |
968 | assert(MI->mayLoad() && !MI->mayStore()); |
969 | bool Changed = false; |
970 | |
971 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
972 | switch (Scope) { |
973 | case SIAtomicScope::SYSTEM: |
974 | case SIAtomicScope::AGENT: |
975 | // Set L1 cache policy to MISS_EVICT. |
976 | // Note: there is no L2 cache bypass policy at the ISA level. |
977 | Changed |= enableGLCBit(MI); |
978 | break; |
979 | case SIAtomicScope::WORKGROUP: |
980 | case SIAtomicScope::WAVEFRONT: |
981 | case SIAtomicScope::SINGLETHREAD: |
982 | // No cache to bypass. |
983 | break; |
984 | default: |
985 | llvm_unreachable("Unsupported synchronization scope" ); |
986 | } |
987 | } |
988 | |
989 | /// The scratch address space does not need the global memory caches |
990 | /// to be bypassed as all memory operations by the same thread are |
991 | /// sequentially consistent, and no other thread can access scratch |
992 | /// memory. |
993 | |
994 | /// Other address spaces do not have a cache. |
995 | |
996 | return Changed; |
997 | } |
998 | |
999 | bool SIGfx6CacheControl::enableStoreCacheBypass( |
1000 | const MachineBasicBlock::iterator &MI, |
1001 | SIAtomicScope Scope, |
1002 | SIAtomicAddrSpace AddrSpace) const { |
1003 | assert(!MI->mayLoad() && MI->mayStore()); |
1004 | bool Changed = false; |
1005 | |
1006 | /// The L1 cache is write through so does not need to be bypassed. There is no |
1007 | /// bypass control for the L2 cache at the isa level. |
1008 | |
1009 | return Changed; |
1010 | } |
1011 | |
1012 | bool SIGfx6CacheControl::enableRMWCacheBypass( |
1013 | const MachineBasicBlock::iterator &MI, |
1014 | SIAtomicScope Scope, |
1015 | SIAtomicAddrSpace AddrSpace) const { |
1016 | assert(MI->mayLoad() && MI->mayStore()); |
1017 | bool Changed = false; |
1018 | |
1019 | /// Do not set GLC for RMW atomic operations as L0/L1 cache is automatically |
1020 | /// bypassed, and the GLC bit is instead used to indicate if they are |
1021 | /// return or no-return. |
1022 | /// Note: there is no L2 cache coherent bypass control at the ISA level. |
1023 | |
1024 | return Changed; |
1025 | } |
1026 | |
1027 | bool SIGfx6CacheControl::enableVolatileAndOrNonTemporal( |
1028 | MachineBasicBlock::iterator &MI, SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
1029 | bool IsVolatile, bool IsNonTemporal, bool IsLastUse = false) const { |
1030 | // Only handle load and store, not atomic read-modify-write insructions. The |
1031 | // latter use glc to indicate if the atomic returns a result and so must not |
1032 | // be used for cache control. |
1033 | assert(MI->mayLoad() ^ MI->mayStore()); |
1034 | |
1035 | // Only update load and store, not LLVM IR atomic read-modify-write |
1036 | // instructions. The latter are always marked as volatile so cannot sensibly |
1037 | // handle it as do not want to pessimize all atomics. Also they do not support |
1038 | // the nontemporal attribute. |
1039 | assert(Op == SIMemOp::LOAD || Op == SIMemOp::STORE); |
1040 | |
1041 | bool Changed = false; |
1042 | |
1043 | if (IsVolatile) { |
1044 | // Set L1 cache policy to be MISS_EVICT for load instructions |
1045 | // and MISS_LRU for store instructions. |
1046 | // Note: there is no L2 cache bypass policy at the ISA level. |
1047 | if (Op == SIMemOp::LOAD) |
1048 | Changed |= enableGLCBit(MI); |
1049 | |
1050 | // Ensure operation has completed at system scope to cause all volatile |
1051 | // operations to be visible outside the program in a global order. Do not |
1052 | // request cross address space as only the global address space can be |
1053 | // observable outside the program, so no need to cause a waitcnt for LDS |
1054 | // address space operations. |
1055 | Changed |= insertWait(MI, Scope: SIAtomicScope::SYSTEM, AddrSpace, Op, IsCrossAddrSpaceOrdering: false, |
1056 | Pos: Position::AFTER, Order: AtomicOrdering::Unordered); |
1057 | |
1058 | return Changed; |
1059 | } |
1060 | |
1061 | if (IsNonTemporal) { |
1062 | // Setting both GLC and SLC configures L1 cache policy to MISS_EVICT |
1063 | // for both loads and stores, and the L2 cache policy to STREAM. |
1064 | Changed |= enableGLCBit(MI); |
1065 | Changed |= enableSLCBit(MI); |
1066 | return Changed; |
1067 | } |
1068 | |
1069 | return Changed; |
1070 | } |
1071 | |
1072 | bool SIGfx6CacheControl::insertWait(MachineBasicBlock::iterator &MI, |
1073 | SIAtomicScope Scope, |
1074 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
1075 | bool IsCrossAddrSpaceOrdering, Position Pos, |
1076 | AtomicOrdering Order) const { |
1077 | bool Changed = false; |
1078 | |
1079 | MachineBasicBlock &MBB = *MI->getParent(); |
1080 | DebugLoc DL = MI->getDebugLoc(); |
1081 | |
1082 | if (Pos == Position::AFTER) |
1083 | ++MI; |
1084 | |
1085 | bool VMCnt = false; |
1086 | bool LGKMCnt = false; |
1087 | |
1088 | if ((AddrSpace & (SIAtomicAddrSpace::GLOBAL | SIAtomicAddrSpace::SCRATCH)) != |
1089 | SIAtomicAddrSpace::NONE) { |
1090 | switch (Scope) { |
1091 | case SIAtomicScope::SYSTEM: |
1092 | case SIAtomicScope::AGENT: |
1093 | VMCnt |= true; |
1094 | break; |
1095 | case SIAtomicScope::WORKGROUP: |
1096 | case SIAtomicScope::WAVEFRONT: |
1097 | case SIAtomicScope::SINGLETHREAD: |
1098 | // The L1 cache keeps all memory operations in order for |
1099 | // wavefronts in the same work-group. |
1100 | break; |
1101 | default: |
1102 | llvm_unreachable("Unsupported synchronization scope" ); |
1103 | } |
1104 | } |
1105 | |
1106 | if ((AddrSpace & SIAtomicAddrSpace::LDS) != SIAtomicAddrSpace::NONE) { |
1107 | switch (Scope) { |
1108 | case SIAtomicScope::SYSTEM: |
1109 | case SIAtomicScope::AGENT: |
1110 | case SIAtomicScope::WORKGROUP: |
1111 | // If no cross address space ordering then an "S_WAITCNT lgkmcnt(0)" is |
1112 | // not needed as LDS operations for all waves are executed in a total |
1113 | // global ordering as observed by all waves. Required if also |
1114 | // synchronizing with global/GDS memory as LDS operations could be |
1115 | // reordered with respect to later global/GDS memory operations of the |
1116 | // same wave. |
1117 | LGKMCnt |= IsCrossAddrSpaceOrdering; |
1118 | break; |
1119 | case SIAtomicScope::WAVEFRONT: |
1120 | case SIAtomicScope::SINGLETHREAD: |
1121 | // The LDS keeps all memory operations in order for |
1122 | // the same wavefront. |
1123 | break; |
1124 | default: |
1125 | llvm_unreachable("Unsupported synchronization scope" ); |
1126 | } |
1127 | } |
1128 | |
1129 | if ((AddrSpace & SIAtomicAddrSpace::GDS) != SIAtomicAddrSpace::NONE) { |
1130 | switch (Scope) { |
1131 | case SIAtomicScope::SYSTEM: |
1132 | case SIAtomicScope::AGENT: |
1133 | // If no cross address space ordering then an GDS "S_WAITCNT lgkmcnt(0)" |
1134 | // is not needed as GDS operations for all waves are executed in a total |
1135 | // global ordering as observed by all waves. Required if also |
1136 | // synchronizing with global/LDS memory as GDS operations could be |
1137 | // reordered with respect to later global/LDS memory operations of the |
1138 | // same wave. |
1139 | LGKMCnt |= IsCrossAddrSpaceOrdering; |
1140 | break; |
1141 | case SIAtomicScope::WORKGROUP: |
1142 | case SIAtomicScope::WAVEFRONT: |
1143 | case SIAtomicScope::SINGLETHREAD: |
1144 | // The GDS keeps all memory operations in order for |
1145 | // the same work-group. |
1146 | break; |
1147 | default: |
1148 | llvm_unreachable("Unsupported synchronization scope" ); |
1149 | } |
1150 | } |
1151 | |
1152 | if (VMCnt || LGKMCnt) { |
1153 | unsigned WaitCntImmediate = |
1154 | AMDGPU::encodeWaitcnt(Version: IV, |
1155 | Vmcnt: VMCnt ? 0 : getVmcntBitMask(Version: IV), |
1156 | Expcnt: getExpcntBitMask(Version: IV), |
1157 | Lgkmcnt: LGKMCnt ? 0 : getLgkmcntBitMask(Version: IV)); |
1158 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_WAITCNT_soft)) |
1159 | .addImm(Val: WaitCntImmediate); |
1160 | Changed = true; |
1161 | } |
1162 | |
1163 | if (Pos == Position::AFTER) |
1164 | --MI; |
1165 | |
1166 | return Changed; |
1167 | } |
1168 | |
1169 | bool SIGfx6CacheControl::insertAcquire(MachineBasicBlock::iterator &MI, |
1170 | SIAtomicScope Scope, |
1171 | SIAtomicAddrSpace AddrSpace, |
1172 | Position Pos) const { |
1173 | if (!InsertCacheInv) |
1174 | return false; |
1175 | |
1176 | bool Changed = false; |
1177 | |
1178 | MachineBasicBlock &MBB = *MI->getParent(); |
1179 | DebugLoc DL = MI->getDebugLoc(); |
1180 | |
1181 | if (Pos == Position::AFTER) |
1182 | ++MI; |
1183 | |
1184 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1185 | switch (Scope) { |
1186 | case SIAtomicScope::SYSTEM: |
1187 | case SIAtomicScope::AGENT: |
1188 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_WBINVL1)); |
1189 | Changed = true; |
1190 | break; |
1191 | case SIAtomicScope::WORKGROUP: |
1192 | case SIAtomicScope::WAVEFRONT: |
1193 | case SIAtomicScope::SINGLETHREAD: |
1194 | // No cache to invalidate. |
1195 | break; |
1196 | default: |
1197 | llvm_unreachable("Unsupported synchronization scope" ); |
1198 | } |
1199 | } |
1200 | |
1201 | /// The scratch address space does not need the global memory cache |
1202 | /// to be flushed as all memory operations by the same thread are |
1203 | /// sequentially consistent, and no other thread can access scratch |
1204 | /// memory. |
1205 | |
1206 | /// Other address spaces do not have a cache. |
1207 | |
1208 | if (Pos == Position::AFTER) |
1209 | --MI; |
1210 | |
1211 | return Changed; |
1212 | } |
1213 | |
1214 | bool SIGfx6CacheControl::insertRelease(MachineBasicBlock::iterator &MI, |
1215 | SIAtomicScope Scope, |
1216 | SIAtomicAddrSpace AddrSpace, |
1217 | bool IsCrossAddrSpaceOrdering, |
1218 | Position Pos) const { |
1219 | return insertWait(MI, Scope, AddrSpace, Op: SIMemOp::LOAD | SIMemOp::STORE, |
1220 | IsCrossAddrSpaceOrdering, Pos, Order: AtomicOrdering::Release); |
1221 | } |
1222 | |
1223 | bool SIGfx7CacheControl::insertAcquire(MachineBasicBlock::iterator &MI, |
1224 | SIAtomicScope Scope, |
1225 | SIAtomicAddrSpace AddrSpace, |
1226 | Position Pos) const { |
1227 | if (!InsertCacheInv) |
1228 | return false; |
1229 | |
1230 | bool Changed = false; |
1231 | |
1232 | MachineBasicBlock &MBB = *MI->getParent(); |
1233 | DebugLoc DL = MI->getDebugLoc(); |
1234 | |
1235 | const GCNSubtarget &STM = MBB.getParent()->getSubtarget<GCNSubtarget>(); |
1236 | |
1237 | const unsigned InvalidateL1 = STM.isAmdPalOS() || STM.isMesa3DOS() |
1238 | ? AMDGPU::BUFFER_WBINVL1 |
1239 | : AMDGPU::BUFFER_WBINVL1_VOL; |
1240 | |
1241 | if (Pos == Position::AFTER) |
1242 | ++MI; |
1243 | |
1244 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1245 | switch (Scope) { |
1246 | case SIAtomicScope::SYSTEM: |
1247 | case SIAtomicScope::AGENT: |
1248 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: InvalidateL1)); |
1249 | Changed = true; |
1250 | break; |
1251 | case SIAtomicScope::WORKGROUP: |
1252 | case SIAtomicScope::WAVEFRONT: |
1253 | case SIAtomicScope::SINGLETHREAD: |
1254 | // No cache to invalidate. |
1255 | break; |
1256 | default: |
1257 | llvm_unreachable("Unsupported synchronization scope" ); |
1258 | } |
1259 | } |
1260 | |
1261 | /// The scratch address space does not need the global memory cache |
1262 | /// to be flushed as all memory operations by the same thread are |
1263 | /// sequentially consistent, and no other thread can access scratch |
1264 | /// memory. |
1265 | |
1266 | /// Other address spaces do not have a cache. |
1267 | |
1268 | if (Pos == Position::AFTER) |
1269 | --MI; |
1270 | |
1271 | return Changed; |
1272 | } |
1273 | |
1274 | bool SIGfx90ACacheControl::enableLoadCacheBypass( |
1275 | const MachineBasicBlock::iterator &MI, |
1276 | SIAtomicScope Scope, |
1277 | SIAtomicAddrSpace AddrSpace) const { |
1278 | assert(MI->mayLoad() && !MI->mayStore()); |
1279 | bool Changed = false; |
1280 | |
1281 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1282 | switch (Scope) { |
1283 | case SIAtomicScope::SYSTEM: |
1284 | case SIAtomicScope::AGENT: |
1285 | // Set the L1 cache policy to MISS_LRU. |
1286 | // Note: there is no L2 cache bypass policy at the ISA level. |
1287 | Changed |= enableGLCBit(MI); |
1288 | break; |
1289 | case SIAtomicScope::WORKGROUP: |
1290 | // In threadgroup split mode the waves of a work-group can be executing on |
1291 | // different CUs. Therefore need to bypass the L1 which is per CU. |
1292 | // Otherwise in non-threadgroup split mode all waves of a work-group are |
1293 | // on the same CU, and so the L1 does not need to be bypassed. |
1294 | if (ST.isTgSplitEnabled()) |
1295 | Changed |= enableGLCBit(MI); |
1296 | break; |
1297 | case SIAtomicScope::WAVEFRONT: |
1298 | case SIAtomicScope::SINGLETHREAD: |
1299 | // No cache to bypass. |
1300 | break; |
1301 | default: |
1302 | llvm_unreachable("Unsupported synchronization scope" ); |
1303 | } |
1304 | } |
1305 | |
1306 | /// The scratch address space does not need the global memory caches |
1307 | /// to be bypassed as all memory operations by the same thread are |
1308 | /// sequentially consistent, and no other thread can access scratch |
1309 | /// memory. |
1310 | |
1311 | /// Other address spaces do not have a cache. |
1312 | |
1313 | return Changed; |
1314 | } |
1315 | |
1316 | bool SIGfx90ACacheControl::enableStoreCacheBypass( |
1317 | const MachineBasicBlock::iterator &MI, |
1318 | SIAtomicScope Scope, |
1319 | SIAtomicAddrSpace AddrSpace) const { |
1320 | assert(!MI->mayLoad() && MI->mayStore()); |
1321 | bool Changed = false; |
1322 | |
1323 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1324 | switch (Scope) { |
1325 | case SIAtomicScope::SYSTEM: |
1326 | case SIAtomicScope::AGENT: |
1327 | /// Do not set glc for store atomic operations as they implicitly write |
1328 | /// through the L1 cache. |
1329 | break; |
1330 | case SIAtomicScope::WORKGROUP: |
1331 | case SIAtomicScope::WAVEFRONT: |
1332 | case SIAtomicScope::SINGLETHREAD: |
1333 | // No cache to bypass. Store atomics implicitly write through the L1 |
1334 | // cache. |
1335 | break; |
1336 | default: |
1337 | llvm_unreachable("Unsupported synchronization scope" ); |
1338 | } |
1339 | } |
1340 | |
1341 | /// The scratch address space does not need the global memory caches |
1342 | /// to be bypassed as all memory operations by the same thread are |
1343 | /// sequentially consistent, and no other thread can access scratch |
1344 | /// memory. |
1345 | |
1346 | /// Other address spaces do not have a cache. |
1347 | |
1348 | return Changed; |
1349 | } |
1350 | |
1351 | bool SIGfx90ACacheControl::enableRMWCacheBypass( |
1352 | const MachineBasicBlock::iterator &MI, |
1353 | SIAtomicScope Scope, |
1354 | SIAtomicAddrSpace AddrSpace) const { |
1355 | assert(MI->mayLoad() && MI->mayStore()); |
1356 | bool Changed = false; |
1357 | |
1358 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1359 | switch (Scope) { |
1360 | case SIAtomicScope::SYSTEM: |
1361 | case SIAtomicScope::AGENT: |
1362 | /// Do not set glc for RMW atomic operations as they implicitly bypass |
1363 | /// the L1 cache, and the glc bit is instead used to indicate if they are |
1364 | /// return or no-return. |
1365 | break; |
1366 | case SIAtomicScope::WORKGROUP: |
1367 | case SIAtomicScope::WAVEFRONT: |
1368 | case SIAtomicScope::SINGLETHREAD: |
1369 | // No cache to bypass. RMW atomics implicitly bypass the L1 cache. |
1370 | break; |
1371 | default: |
1372 | llvm_unreachable("Unsupported synchronization scope" ); |
1373 | } |
1374 | } |
1375 | |
1376 | return Changed; |
1377 | } |
1378 | |
1379 | bool SIGfx90ACacheControl::enableVolatileAndOrNonTemporal( |
1380 | MachineBasicBlock::iterator &MI, SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
1381 | bool IsVolatile, bool IsNonTemporal, bool IsLastUse = false) const { |
1382 | // Only handle load and store, not atomic read-modify-write insructions. The |
1383 | // latter use glc to indicate if the atomic returns a result and so must not |
1384 | // be used for cache control. |
1385 | assert(MI->mayLoad() ^ MI->mayStore()); |
1386 | |
1387 | // Only update load and store, not LLVM IR atomic read-modify-write |
1388 | // instructions. The latter are always marked as volatile so cannot sensibly |
1389 | // handle it as do not want to pessimize all atomics. Also they do not support |
1390 | // the nontemporal attribute. |
1391 | assert(Op == SIMemOp::LOAD || Op == SIMemOp::STORE); |
1392 | |
1393 | bool Changed = false; |
1394 | |
1395 | if (IsVolatile) { |
1396 | // Set L1 cache policy to be MISS_EVICT for load instructions |
1397 | // and MISS_LRU for store instructions. |
1398 | // Note: there is no L2 cache bypass policy at the ISA level. |
1399 | if (Op == SIMemOp::LOAD) |
1400 | Changed |= enableGLCBit(MI); |
1401 | |
1402 | // Ensure operation has completed at system scope to cause all volatile |
1403 | // operations to be visible outside the program in a global order. Do not |
1404 | // request cross address space as only the global address space can be |
1405 | // observable outside the program, so no need to cause a waitcnt for LDS |
1406 | // address space operations. |
1407 | Changed |= insertWait(MI, Scope: SIAtomicScope::SYSTEM, AddrSpace, Op, IsCrossAddrSpaceOrdering: false, |
1408 | Pos: Position::AFTER, Order: AtomicOrdering::Unordered); |
1409 | |
1410 | return Changed; |
1411 | } |
1412 | |
1413 | if (IsNonTemporal) { |
1414 | // Setting both GLC and SLC configures L1 cache policy to MISS_EVICT |
1415 | // for both loads and stores, and the L2 cache policy to STREAM. |
1416 | Changed |= enableGLCBit(MI); |
1417 | Changed |= enableSLCBit(MI); |
1418 | return Changed; |
1419 | } |
1420 | |
1421 | return Changed; |
1422 | } |
1423 | |
1424 | bool SIGfx90ACacheControl::insertWait(MachineBasicBlock::iterator &MI, |
1425 | SIAtomicScope Scope, |
1426 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
1427 | bool IsCrossAddrSpaceOrdering, |
1428 | Position Pos, |
1429 | AtomicOrdering Order) const { |
1430 | if (ST.isTgSplitEnabled()) { |
1431 | // In threadgroup split mode the waves of a work-group can be executing on |
1432 | // different CUs. Therefore need to wait for global or GDS memory operations |
1433 | // to complete to ensure they are visible to waves in the other CUs. |
1434 | // Otherwise in non-threadgroup split mode all waves of a work-group are on |
1435 | // the same CU, so no need to wait for global memory as all waves in the |
1436 | // work-group access the same the L1, nor wait for GDS as access are ordered |
1437 | // on a CU. |
1438 | if (((AddrSpace & (SIAtomicAddrSpace::GLOBAL | SIAtomicAddrSpace::SCRATCH | |
1439 | SIAtomicAddrSpace::GDS)) != SIAtomicAddrSpace::NONE) && |
1440 | (Scope == SIAtomicScope::WORKGROUP)) { |
1441 | // Same as GFX7 using agent scope. |
1442 | Scope = SIAtomicScope::AGENT; |
1443 | } |
1444 | // In threadgroup split mode LDS cannot be allocated so no need to wait for |
1445 | // LDS memory operations. |
1446 | AddrSpace &= ~SIAtomicAddrSpace::LDS; |
1447 | } |
1448 | return SIGfx7CacheControl::insertWait(MI, Scope, AddrSpace, Op, |
1449 | IsCrossAddrSpaceOrdering, Pos, Order); |
1450 | } |
1451 | |
1452 | bool SIGfx90ACacheControl::insertAcquire(MachineBasicBlock::iterator &MI, |
1453 | SIAtomicScope Scope, |
1454 | SIAtomicAddrSpace AddrSpace, |
1455 | Position Pos) const { |
1456 | if (!InsertCacheInv) |
1457 | return false; |
1458 | |
1459 | bool Changed = false; |
1460 | |
1461 | MachineBasicBlock &MBB = *MI->getParent(); |
1462 | DebugLoc DL = MI->getDebugLoc(); |
1463 | |
1464 | if (Pos == Position::AFTER) |
1465 | ++MI; |
1466 | |
1467 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1468 | switch (Scope) { |
1469 | case SIAtomicScope::SYSTEM: |
1470 | // Ensures that following loads will not see stale remote VMEM data or |
1471 | // stale local VMEM data with MTYPE NC. Local VMEM data with MTYPE RW and |
1472 | // CC will never be stale due to the local memory probes. |
1473 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_INVL2)); |
1474 | // Inserting a "S_WAITCNT vmcnt(0)" after is not required because the |
1475 | // hardware does not reorder memory operations by the same wave with |
1476 | // respect to a preceding "BUFFER_INVL2". The invalidate is guaranteed to |
1477 | // remove any cache lines of earlier writes by the same wave and ensures |
1478 | // later reads by the same wave will refetch the cache lines. |
1479 | Changed = true; |
1480 | break; |
1481 | case SIAtomicScope::AGENT: |
1482 | // Same as GFX7. |
1483 | break; |
1484 | case SIAtomicScope::WORKGROUP: |
1485 | // In threadgroup split mode the waves of a work-group can be executing on |
1486 | // different CUs. Therefore need to invalidate the L1 which is per CU. |
1487 | // Otherwise in non-threadgroup split mode all waves of a work-group are |
1488 | // on the same CU, and so the L1 does not need to be invalidated. |
1489 | if (ST.isTgSplitEnabled()) { |
1490 | // Same as GFX7 using agent scope. |
1491 | Scope = SIAtomicScope::AGENT; |
1492 | } |
1493 | break; |
1494 | case SIAtomicScope::WAVEFRONT: |
1495 | case SIAtomicScope::SINGLETHREAD: |
1496 | // Same as GFX7. |
1497 | break; |
1498 | default: |
1499 | llvm_unreachable("Unsupported synchronization scope" ); |
1500 | } |
1501 | } |
1502 | |
1503 | /// The scratch address space does not need the global memory cache |
1504 | /// to be flushed as all memory operations by the same thread are |
1505 | /// sequentially consistent, and no other thread can access scratch |
1506 | /// memory. |
1507 | |
1508 | /// Other address spaces do not have a cache. |
1509 | |
1510 | if (Pos == Position::AFTER) |
1511 | --MI; |
1512 | |
1513 | Changed |= SIGfx7CacheControl::insertAcquire(MI, Scope, AddrSpace, Pos); |
1514 | |
1515 | return Changed; |
1516 | } |
1517 | |
1518 | bool SIGfx90ACacheControl::insertRelease(MachineBasicBlock::iterator &MI, |
1519 | SIAtomicScope Scope, |
1520 | SIAtomicAddrSpace AddrSpace, |
1521 | bool IsCrossAddrSpaceOrdering, |
1522 | Position Pos) const { |
1523 | bool Changed = false; |
1524 | |
1525 | MachineBasicBlock &MBB = *MI->getParent(); |
1526 | const DebugLoc &DL = MI->getDebugLoc(); |
1527 | |
1528 | if (Pos == Position::AFTER) |
1529 | ++MI; |
1530 | |
1531 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1532 | switch (Scope) { |
1533 | case SIAtomicScope::SYSTEM: |
1534 | // Inserting a "S_WAITCNT vmcnt(0)" before is not required because the |
1535 | // hardware does not reorder memory operations by the same wave with |
1536 | // respect to a following "BUFFER_WBL2". The "BUFFER_WBL2" is guaranteed |
1537 | // to initiate writeback of any dirty cache lines of earlier writes by the |
1538 | // same wave. A "S_WAITCNT vmcnt(0)" is needed after to ensure the |
1539 | // writeback has completed. |
1540 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_WBL2)) |
1541 | // Set SC bits to indicate system scope. |
1542 | .addImm(Val: AMDGPU::CPol::SC0 | AMDGPU::CPol::SC1); |
1543 | // Followed by same as GFX7, which will ensure the necessary "S_WAITCNT |
1544 | // vmcnt(0)" needed by the "BUFFER_WBL2". |
1545 | Changed = true; |
1546 | break; |
1547 | case SIAtomicScope::AGENT: |
1548 | case SIAtomicScope::WORKGROUP: |
1549 | case SIAtomicScope::WAVEFRONT: |
1550 | case SIAtomicScope::SINGLETHREAD: |
1551 | // Same as GFX7. |
1552 | break; |
1553 | default: |
1554 | llvm_unreachable("Unsupported synchronization scope" ); |
1555 | } |
1556 | } |
1557 | |
1558 | if (Pos == Position::AFTER) |
1559 | --MI; |
1560 | |
1561 | Changed |= |
1562 | SIGfx7CacheControl::insertRelease(MI, Scope, AddrSpace, |
1563 | IsCrossAddrSpaceOrdering, Pos); |
1564 | |
1565 | return Changed; |
1566 | } |
1567 | |
1568 | bool SIGfx940CacheControl::enableLoadCacheBypass( |
1569 | const MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
1570 | SIAtomicAddrSpace AddrSpace) const { |
1571 | assert(MI->mayLoad() && !MI->mayStore()); |
1572 | bool Changed = false; |
1573 | |
1574 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1575 | switch (Scope) { |
1576 | case SIAtomicScope::SYSTEM: |
1577 | // Set SC bits to indicate system scope. |
1578 | Changed |= enableSC0Bit(MI); |
1579 | Changed |= enableSC1Bit(MI); |
1580 | break; |
1581 | case SIAtomicScope::AGENT: |
1582 | // Set SC bits to indicate agent scope. |
1583 | Changed |= enableSC1Bit(MI); |
1584 | break; |
1585 | case SIAtomicScope::WORKGROUP: |
1586 | // In threadgroup split mode the waves of a work-group can be executing on |
1587 | // different CUs. Therefore need to bypass the L1 which is per CU. |
1588 | // Otherwise in non-threadgroup split mode all waves of a work-group are |
1589 | // on the same CU, and so the L1 does not need to be bypassed. Setting SC |
1590 | // bits to indicate work-group scope will do this automatically. |
1591 | Changed |= enableSC0Bit(MI); |
1592 | break; |
1593 | case SIAtomicScope::WAVEFRONT: |
1594 | case SIAtomicScope::SINGLETHREAD: |
1595 | // Leave SC bits unset to indicate wavefront scope. |
1596 | break; |
1597 | default: |
1598 | llvm_unreachable("Unsupported synchronization scope" ); |
1599 | } |
1600 | } |
1601 | |
1602 | /// The scratch address space does not need the global memory caches |
1603 | /// to be bypassed as all memory operations by the same thread are |
1604 | /// sequentially consistent, and no other thread can access scratch |
1605 | /// memory. |
1606 | |
1607 | /// Other address spaces do not have a cache. |
1608 | |
1609 | return Changed; |
1610 | } |
1611 | |
1612 | bool SIGfx940CacheControl::enableStoreCacheBypass( |
1613 | const MachineBasicBlock::iterator &MI, |
1614 | SIAtomicScope Scope, SIAtomicAddrSpace AddrSpace) const { |
1615 | assert(!MI->mayLoad() && MI->mayStore()); |
1616 | bool Changed = false; |
1617 | |
1618 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1619 | switch (Scope) { |
1620 | case SIAtomicScope::SYSTEM: |
1621 | // Set SC bits to indicate system scope. |
1622 | Changed |= enableSC0Bit(MI); |
1623 | Changed |= enableSC1Bit(MI); |
1624 | break; |
1625 | case SIAtomicScope::AGENT: |
1626 | // Set SC bits to indicate agent scope. |
1627 | Changed |= enableSC1Bit(MI); |
1628 | break; |
1629 | case SIAtomicScope::WORKGROUP: |
1630 | // Set SC bits to indicate workgroup scope. |
1631 | Changed |= enableSC0Bit(MI); |
1632 | break; |
1633 | case SIAtomicScope::WAVEFRONT: |
1634 | case SIAtomicScope::SINGLETHREAD: |
1635 | // Leave SC bits unset to indicate wavefront scope. |
1636 | break; |
1637 | default: |
1638 | llvm_unreachable("Unsupported synchronization scope" ); |
1639 | } |
1640 | } |
1641 | |
1642 | /// The scratch address space does not need the global memory caches |
1643 | /// to be bypassed as all memory operations by the same thread are |
1644 | /// sequentially consistent, and no other thread can access scratch |
1645 | /// memory. |
1646 | |
1647 | /// Other address spaces do not have a cache. |
1648 | |
1649 | return Changed; |
1650 | } |
1651 | |
1652 | bool SIGfx940CacheControl::enableRMWCacheBypass( |
1653 | const MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
1654 | SIAtomicAddrSpace AddrSpace) const { |
1655 | assert(MI->mayLoad() && MI->mayStore()); |
1656 | bool Changed = false; |
1657 | |
1658 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1659 | switch (Scope) { |
1660 | case SIAtomicScope::SYSTEM: |
1661 | // Set SC1 bit to indicate system scope. |
1662 | Changed |= enableSC1Bit(MI); |
1663 | break; |
1664 | case SIAtomicScope::AGENT: |
1665 | case SIAtomicScope::WORKGROUP: |
1666 | case SIAtomicScope::WAVEFRONT: |
1667 | case SIAtomicScope::SINGLETHREAD: |
1668 | // RMW atomic operations implicitly bypass the L1 cache and only use SC1 |
1669 | // to indicate system or agent scope. The SC0 bit is used to indicate if |
1670 | // they are return or no-return. Leave SC1 bit unset to indicate agent |
1671 | // scope. |
1672 | break; |
1673 | default: |
1674 | llvm_unreachable("Unsupported synchronization scope" ); |
1675 | } |
1676 | } |
1677 | |
1678 | return Changed; |
1679 | } |
1680 | |
1681 | bool SIGfx940CacheControl::enableVolatileAndOrNonTemporal( |
1682 | MachineBasicBlock::iterator &MI, SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
1683 | bool IsVolatile, bool IsNonTemporal, bool IsLastUse = false) const { |
1684 | // Only handle load and store, not atomic read-modify-write insructions. The |
1685 | // latter use glc to indicate if the atomic returns a result and so must not |
1686 | // be used for cache control. |
1687 | assert(MI->mayLoad() ^ MI->mayStore()); |
1688 | |
1689 | // Only update load and store, not LLVM IR atomic read-modify-write |
1690 | // instructions. The latter are always marked as volatile so cannot sensibly |
1691 | // handle it as do not want to pessimize all atomics. Also they do not support |
1692 | // the nontemporal attribute. |
1693 | assert(Op == SIMemOp::LOAD || Op == SIMemOp::STORE); |
1694 | |
1695 | bool Changed = false; |
1696 | |
1697 | if (IsVolatile) { |
1698 | // Set SC bits to indicate system scope. |
1699 | Changed |= enableSC0Bit(MI); |
1700 | Changed |= enableSC1Bit(MI); |
1701 | |
1702 | // Ensure operation has completed at system scope to cause all volatile |
1703 | // operations to be visible outside the program in a global order. Do not |
1704 | // request cross address space as only the global address space can be |
1705 | // observable outside the program, so no need to cause a waitcnt for LDS |
1706 | // address space operations. |
1707 | Changed |= insertWait(MI, Scope: SIAtomicScope::SYSTEM, AddrSpace, Op, IsCrossAddrSpaceOrdering: false, |
1708 | Pos: Position::AFTER, Order: AtomicOrdering::Unordered); |
1709 | |
1710 | return Changed; |
1711 | } |
1712 | |
1713 | if (IsNonTemporal) { |
1714 | Changed |= enableNTBit(MI); |
1715 | return Changed; |
1716 | } |
1717 | |
1718 | return Changed; |
1719 | } |
1720 | |
1721 | bool SIGfx940CacheControl::insertAcquire(MachineBasicBlock::iterator &MI, |
1722 | SIAtomicScope Scope, |
1723 | SIAtomicAddrSpace AddrSpace, |
1724 | Position Pos) const { |
1725 | if (!InsertCacheInv) |
1726 | return false; |
1727 | |
1728 | bool Changed = false; |
1729 | |
1730 | MachineBasicBlock &MBB = *MI->getParent(); |
1731 | DebugLoc DL = MI->getDebugLoc(); |
1732 | |
1733 | if (Pos == Position::AFTER) |
1734 | ++MI; |
1735 | |
1736 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1737 | switch (Scope) { |
1738 | case SIAtomicScope::SYSTEM: |
1739 | // Ensures that following loads will not see stale remote VMEM data or |
1740 | // stale local VMEM data with MTYPE NC. Local VMEM data with MTYPE RW and |
1741 | // CC will never be stale due to the local memory probes. |
1742 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_INV)) |
1743 | // Set SC bits to indicate system scope. |
1744 | .addImm(Val: AMDGPU::CPol::SC0 | AMDGPU::CPol::SC1); |
1745 | // Inserting a "S_WAITCNT vmcnt(0)" after is not required because the |
1746 | // hardware does not reorder memory operations by the same wave with |
1747 | // respect to a preceding "BUFFER_INV". The invalidate is guaranteed to |
1748 | // remove any cache lines of earlier writes by the same wave and ensures |
1749 | // later reads by the same wave will refetch the cache lines. |
1750 | Changed = true; |
1751 | break; |
1752 | case SIAtomicScope::AGENT: |
1753 | // Ensures that following loads will not see stale remote date or local |
1754 | // MTYPE NC global data. Local MTYPE RW and CC memory will never be stale |
1755 | // due to the memory probes. |
1756 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_INV)) |
1757 | // Set SC bits to indicate agent scope. |
1758 | .addImm(Val: AMDGPU::CPol::SC1); |
1759 | // Inserting "S_WAITCNT vmcnt(0)" is not required because the hardware |
1760 | // does not reorder memory operations with respect to preceeding buffer |
1761 | // invalidate. The invalidate is guaranteed to remove any cache lines of |
1762 | // earlier writes and ensures later writes will refetch the cache lines. |
1763 | Changed = true; |
1764 | break; |
1765 | case SIAtomicScope::WORKGROUP: |
1766 | // In threadgroup split mode the waves of a work-group can be executing on |
1767 | // different CUs. Therefore need to invalidate the L1 which is per CU. |
1768 | // Otherwise in non-threadgroup split mode all waves of a work-group are |
1769 | // on the same CU, and so the L1 does not need to be invalidated. |
1770 | if (ST.isTgSplitEnabled()) { |
1771 | // Ensures L1 is invalidated if in threadgroup split mode. In |
1772 | // non-threadgroup split mode it is a NOP, but no point generating it in |
1773 | // that case if know not in that mode. |
1774 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_INV)) |
1775 | // Set SC bits to indicate work-group scope. |
1776 | .addImm(Val: AMDGPU::CPol::SC0); |
1777 | // Inserting "S_WAITCNT vmcnt(0)" is not required because the hardware |
1778 | // does not reorder memory operations with respect to preceeding buffer |
1779 | // invalidate. The invalidate is guaranteed to remove any cache lines of |
1780 | // earlier writes and ensures later writes will refetch the cache lines. |
1781 | Changed = true; |
1782 | } |
1783 | break; |
1784 | case SIAtomicScope::WAVEFRONT: |
1785 | case SIAtomicScope::SINGLETHREAD: |
1786 | // Could generate "BUFFER_INV" but it would do nothing as there are no |
1787 | // caches to invalidate. |
1788 | break; |
1789 | default: |
1790 | llvm_unreachable("Unsupported synchronization scope" ); |
1791 | } |
1792 | } |
1793 | |
1794 | /// The scratch address space does not need the global memory cache |
1795 | /// to be flushed as all memory operations by the same thread are |
1796 | /// sequentially consistent, and no other thread can access scratch |
1797 | /// memory. |
1798 | |
1799 | /// Other address spaces do not have a cache. |
1800 | |
1801 | if (Pos == Position::AFTER) |
1802 | --MI; |
1803 | |
1804 | return Changed; |
1805 | } |
1806 | |
1807 | bool SIGfx940CacheControl::insertRelease(MachineBasicBlock::iterator &MI, |
1808 | SIAtomicScope Scope, |
1809 | SIAtomicAddrSpace AddrSpace, |
1810 | bool IsCrossAddrSpaceOrdering, |
1811 | Position Pos) const { |
1812 | bool Changed = false; |
1813 | |
1814 | MachineBasicBlock &MBB = *MI->getParent(); |
1815 | DebugLoc DL = MI->getDebugLoc(); |
1816 | |
1817 | if (Pos == Position::AFTER) |
1818 | ++MI; |
1819 | |
1820 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1821 | switch (Scope) { |
1822 | case SIAtomicScope::SYSTEM: |
1823 | // Inserting a "S_WAITCNT vmcnt(0)" before is not required because the |
1824 | // hardware does not reorder memory operations by the same wave with |
1825 | // respect to a following "BUFFER_WBL2". The "BUFFER_WBL2" is guaranteed |
1826 | // to initiate writeback of any dirty cache lines of earlier writes by the |
1827 | // same wave. A "S_WAITCNT vmcnt(0)" is needed after to ensure the |
1828 | // writeback has completed. |
1829 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_WBL2)) |
1830 | // Set SC bits to indicate system scope. |
1831 | .addImm(Val: AMDGPU::CPol::SC0 | AMDGPU::CPol::SC1); |
1832 | // Since AddrSpace contains SIAtomicAddrSpace::GLOBAL and Scope is |
1833 | // SIAtomicScope::SYSTEM, the following insertWait will generate the |
1834 | // required "S_WAITCNT vmcnt(0)" needed by the "BUFFER_WBL2". |
1835 | Changed = true; |
1836 | break; |
1837 | case SIAtomicScope::AGENT: |
1838 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_WBL2)) |
1839 | // Set SC bits to indicate agent scope. |
1840 | .addImm(Val: AMDGPU::CPol::SC1); |
1841 | |
1842 | // Since AddrSpace contains SIAtomicAddrSpace::GLOBAL and Scope is |
1843 | // SIAtomicScope::AGENT, the following insertWait will generate the |
1844 | // required "S_WAITCNT vmcnt(0)". |
1845 | Changed = true; |
1846 | break; |
1847 | case SIAtomicScope::WORKGROUP: |
1848 | case SIAtomicScope::WAVEFRONT: |
1849 | case SIAtomicScope::SINGLETHREAD: |
1850 | // Do not generate "BUFFER_WBL2" as there are no caches it would |
1851 | // writeback, and would require an otherwise unnecessary |
1852 | // "S_WAITCNT vmcnt(0)". |
1853 | break; |
1854 | default: |
1855 | llvm_unreachable("Unsupported synchronization scope" ); |
1856 | } |
1857 | } |
1858 | |
1859 | if (Pos == Position::AFTER) |
1860 | --MI; |
1861 | |
1862 | // Ensure the necessary S_WAITCNT needed by any "BUFFER_WBL2" as well as other |
1863 | // S_WAITCNT needed. |
1864 | Changed |= insertWait(MI, Scope, AddrSpace, Op: SIMemOp::LOAD | SIMemOp::STORE, |
1865 | IsCrossAddrSpaceOrdering, Pos, Order: AtomicOrdering::Release); |
1866 | |
1867 | return Changed; |
1868 | } |
1869 | |
1870 | bool SIGfx10CacheControl::enableLoadCacheBypass( |
1871 | const MachineBasicBlock::iterator &MI, |
1872 | SIAtomicScope Scope, |
1873 | SIAtomicAddrSpace AddrSpace) const { |
1874 | assert(MI->mayLoad() && !MI->mayStore()); |
1875 | bool Changed = false; |
1876 | |
1877 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1878 | switch (Scope) { |
1879 | case SIAtomicScope::SYSTEM: |
1880 | case SIAtomicScope::AGENT: |
1881 | // Set the L0 and L1 cache policies to MISS_EVICT. |
1882 | // Note: there is no L2 cache coherent bypass control at the ISA level. |
1883 | Changed |= enableGLCBit(MI); |
1884 | Changed |= enableDLCBit(MI); |
1885 | break; |
1886 | case SIAtomicScope::WORKGROUP: |
1887 | // In WGP mode the waves of a work-group can be executing on either CU of |
1888 | // the WGP. Therefore need to bypass the L0 which is per CU. Otherwise in |
1889 | // CU mode all waves of a work-group are on the same CU, and so the L0 |
1890 | // does not need to be bypassed. |
1891 | if (!ST.isCuModeEnabled()) |
1892 | Changed |= enableGLCBit(MI); |
1893 | break; |
1894 | case SIAtomicScope::WAVEFRONT: |
1895 | case SIAtomicScope::SINGLETHREAD: |
1896 | // No cache to bypass. |
1897 | break; |
1898 | default: |
1899 | llvm_unreachable("Unsupported synchronization scope" ); |
1900 | } |
1901 | } |
1902 | |
1903 | /// The scratch address space does not need the global memory caches |
1904 | /// to be bypassed as all memory operations by the same thread are |
1905 | /// sequentially consistent, and no other thread can access scratch |
1906 | /// memory. |
1907 | |
1908 | /// Other address spaces do not have a cache. |
1909 | |
1910 | return Changed; |
1911 | } |
1912 | |
1913 | bool SIGfx10CacheControl::enableVolatileAndOrNonTemporal( |
1914 | MachineBasicBlock::iterator &MI, SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
1915 | bool IsVolatile, bool IsNonTemporal, bool IsLastUse = false) const { |
1916 | |
1917 | // Only handle load and store, not atomic read-modify-write insructions. The |
1918 | // latter use glc to indicate if the atomic returns a result and so must not |
1919 | // be used for cache control. |
1920 | assert(MI->mayLoad() ^ MI->mayStore()); |
1921 | |
1922 | // Only update load and store, not LLVM IR atomic read-modify-write |
1923 | // instructions. The latter are always marked as volatile so cannot sensibly |
1924 | // handle it as do not want to pessimize all atomics. Also they do not support |
1925 | // the nontemporal attribute. |
1926 | assert(Op == SIMemOp::LOAD || Op == SIMemOp::STORE); |
1927 | |
1928 | bool Changed = false; |
1929 | |
1930 | if (IsVolatile) { |
1931 | // Set L0 and L1 cache policy to be MISS_EVICT for load instructions |
1932 | // and MISS_LRU for store instructions. |
1933 | // Note: there is no L2 cache coherent bypass control at the ISA level. |
1934 | if (Op == SIMemOp::LOAD) { |
1935 | Changed |= enableGLCBit(MI); |
1936 | Changed |= enableDLCBit(MI); |
1937 | } |
1938 | |
1939 | // Ensure operation has completed at system scope to cause all volatile |
1940 | // operations to be visible outside the program in a global order. Do not |
1941 | // request cross address space as only the global address space can be |
1942 | // observable outside the program, so no need to cause a waitcnt for LDS |
1943 | // address space operations. |
1944 | Changed |= insertWait(MI, Scope: SIAtomicScope::SYSTEM, AddrSpace, Op, IsCrossAddrSpaceOrdering: false, |
1945 | Pos: Position::AFTER, Order: AtomicOrdering::Unordered); |
1946 | return Changed; |
1947 | } |
1948 | |
1949 | if (IsNonTemporal) { |
1950 | // For loads setting SLC configures L0 and L1 cache policy to HIT_EVICT |
1951 | // and L2 cache policy to STREAM. |
1952 | // For stores setting both GLC and SLC configures L0 and L1 cache policy |
1953 | // to MISS_EVICT and the L2 cache policy to STREAM. |
1954 | if (Op == SIMemOp::STORE) |
1955 | Changed |= enableGLCBit(MI); |
1956 | Changed |= enableSLCBit(MI); |
1957 | |
1958 | return Changed; |
1959 | } |
1960 | |
1961 | return Changed; |
1962 | } |
1963 | |
1964 | bool SIGfx10CacheControl::insertWait(MachineBasicBlock::iterator &MI, |
1965 | SIAtomicScope Scope, |
1966 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
1967 | bool IsCrossAddrSpaceOrdering, |
1968 | Position Pos, AtomicOrdering Order) const { |
1969 | bool Changed = false; |
1970 | |
1971 | MachineBasicBlock &MBB = *MI->getParent(); |
1972 | DebugLoc DL = MI->getDebugLoc(); |
1973 | |
1974 | if (Pos == Position::AFTER) |
1975 | ++MI; |
1976 | |
1977 | bool VMCnt = false; |
1978 | bool VSCnt = false; |
1979 | bool LGKMCnt = false; |
1980 | |
1981 | if ((AddrSpace & (SIAtomicAddrSpace::GLOBAL | SIAtomicAddrSpace::SCRATCH)) != |
1982 | SIAtomicAddrSpace::NONE) { |
1983 | switch (Scope) { |
1984 | case SIAtomicScope::SYSTEM: |
1985 | case SIAtomicScope::AGENT: |
1986 | if ((Op & SIMemOp::LOAD) != SIMemOp::NONE) |
1987 | VMCnt |= true; |
1988 | if ((Op & SIMemOp::STORE) != SIMemOp::NONE) |
1989 | VSCnt |= true; |
1990 | break; |
1991 | case SIAtomicScope::WORKGROUP: |
1992 | // In WGP mode the waves of a work-group can be executing on either CU of |
1993 | // the WGP. Therefore need to wait for operations to complete to ensure |
1994 | // they are visible to waves in the other CU as the L0 is per CU. |
1995 | // Otherwise in CU mode and all waves of a work-group are on the same CU |
1996 | // which shares the same L0. |
1997 | if (!ST.isCuModeEnabled()) { |
1998 | if ((Op & SIMemOp::LOAD) != SIMemOp::NONE) |
1999 | VMCnt |= true; |
2000 | if ((Op & SIMemOp::STORE) != SIMemOp::NONE) |
2001 | VSCnt |= true; |
2002 | } |
2003 | break; |
2004 | case SIAtomicScope::WAVEFRONT: |
2005 | case SIAtomicScope::SINGLETHREAD: |
2006 | // The L0 cache keeps all memory operations in order for |
2007 | // work-items in the same wavefront. |
2008 | break; |
2009 | default: |
2010 | llvm_unreachable("Unsupported synchronization scope" ); |
2011 | } |
2012 | } |
2013 | |
2014 | if ((AddrSpace & SIAtomicAddrSpace::LDS) != SIAtomicAddrSpace::NONE) { |
2015 | switch (Scope) { |
2016 | case SIAtomicScope::SYSTEM: |
2017 | case SIAtomicScope::AGENT: |
2018 | case SIAtomicScope::WORKGROUP: |
2019 | // If no cross address space ordering then an "S_WAITCNT lgkmcnt(0)" is |
2020 | // not needed as LDS operations for all waves are executed in a total |
2021 | // global ordering as observed by all waves. Required if also |
2022 | // synchronizing with global/GDS memory as LDS operations could be |
2023 | // reordered with respect to later global/GDS memory operations of the |
2024 | // same wave. |
2025 | LGKMCnt |= IsCrossAddrSpaceOrdering; |
2026 | break; |
2027 | case SIAtomicScope::WAVEFRONT: |
2028 | case SIAtomicScope::SINGLETHREAD: |
2029 | // The LDS keeps all memory operations in order for |
2030 | // the same wavefront. |
2031 | break; |
2032 | default: |
2033 | llvm_unreachable("Unsupported synchronization scope" ); |
2034 | } |
2035 | } |
2036 | |
2037 | if ((AddrSpace & SIAtomicAddrSpace::GDS) != SIAtomicAddrSpace::NONE) { |
2038 | switch (Scope) { |
2039 | case SIAtomicScope::SYSTEM: |
2040 | case SIAtomicScope::AGENT: |
2041 | // If no cross address space ordering then an GDS "S_WAITCNT lgkmcnt(0)" |
2042 | // is not needed as GDS operations for all waves are executed in a total |
2043 | // global ordering as observed by all waves. Required if also |
2044 | // synchronizing with global/LDS memory as GDS operations could be |
2045 | // reordered with respect to later global/LDS memory operations of the |
2046 | // same wave. |
2047 | LGKMCnt |= IsCrossAddrSpaceOrdering; |
2048 | break; |
2049 | case SIAtomicScope::WORKGROUP: |
2050 | case SIAtomicScope::WAVEFRONT: |
2051 | case SIAtomicScope::SINGLETHREAD: |
2052 | // The GDS keeps all memory operations in order for |
2053 | // the same work-group. |
2054 | break; |
2055 | default: |
2056 | llvm_unreachable("Unsupported synchronization scope" ); |
2057 | } |
2058 | } |
2059 | |
2060 | if (VMCnt || LGKMCnt) { |
2061 | unsigned WaitCntImmediate = |
2062 | AMDGPU::encodeWaitcnt(Version: IV, |
2063 | Vmcnt: VMCnt ? 0 : getVmcntBitMask(Version: IV), |
2064 | Expcnt: getExpcntBitMask(Version: IV), |
2065 | Lgkmcnt: LGKMCnt ? 0 : getLgkmcntBitMask(Version: IV)); |
2066 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_WAITCNT_soft)) |
2067 | .addImm(Val: WaitCntImmediate); |
2068 | Changed = true; |
2069 | } |
2070 | |
2071 | if (VSCnt) { |
2072 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_WAITCNT_VSCNT_soft)) |
2073 | .addReg(RegNo: AMDGPU::SGPR_NULL, flags: RegState::Undef) |
2074 | .addImm(Val: 0); |
2075 | Changed = true; |
2076 | } |
2077 | |
2078 | if (Pos == Position::AFTER) |
2079 | --MI; |
2080 | |
2081 | return Changed; |
2082 | } |
2083 | |
2084 | bool SIGfx10CacheControl::insertAcquire(MachineBasicBlock::iterator &MI, |
2085 | SIAtomicScope Scope, |
2086 | SIAtomicAddrSpace AddrSpace, |
2087 | Position Pos) const { |
2088 | if (!InsertCacheInv) |
2089 | return false; |
2090 | |
2091 | bool Changed = false; |
2092 | |
2093 | MachineBasicBlock &MBB = *MI->getParent(); |
2094 | DebugLoc DL = MI->getDebugLoc(); |
2095 | |
2096 | if (Pos == Position::AFTER) |
2097 | ++MI; |
2098 | |
2099 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
2100 | switch (Scope) { |
2101 | case SIAtomicScope::SYSTEM: |
2102 | case SIAtomicScope::AGENT: |
2103 | // The order of invalidates matter here. We must invalidate "outer in" |
2104 | // so L1 -> L0 to avoid L0 pulling in stale data from L1 when it is |
2105 | // invalidated. |
2106 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_GL1_INV)); |
2107 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_GL0_INV)); |
2108 | Changed = true; |
2109 | break; |
2110 | case SIAtomicScope::WORKGROUP: |
2111 | // In WGP mode the waves of a work-group can be executing on either CU of |
2112 | // the WGP. Therefore need to invalidate the L0 which is per CU. Otherwise |
2113 | // in CU mode and all waves of a work-group are on the same CU, and so the |
2114 | // L0 does not need to be invalidated. |
2115 | if (!ST.isCuModeEnabled()) { |
2116 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_GL0_INV)); |
2117 | Changed = true; |
2118 | } |
2119 | break; |
2120 | case SIAtomicScope::WAVEFRONT: |
2121 | case SIAtomicScope::SINGLETHREAD: |
2122 | // No cache to invalidate. |
2123 | break; |
2124 | default: |
2125 | llvm_unreachable("Unsupported synchronization scope" ); |
2126 | } |
2127 | } |
2128 | |
2129 | /// The scratch address space does not need the global memory cache |
2130 | /// to be flushed as all memory operations by the same thread are |
2131 | /// sequentially consistent, and no other thread can access scratch |
2132 | /// memory. |
2133 | |
2134 | /// Other address spaces do not have a cache. |
2135 | |
2136 | if (Pos == Position::AFTER) |
2137 | --MI; |
2138 | |
2139 | return Changed; |
2140 | } |
2141 | |
2142 | bool SIGfx11CacheControl::enableLoadCacheBypass( |
2143 | const MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
2144 | SIAtomicAddrSpace AddrSpace) const { |
2145 | assert(MI->mayLoad() && !MI->mayStore()); |
2146 | bool Changed = false; |
2147 | |
2148 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
2149 | switch (Scope) { |
2150 | case SIAtomicScope::SYSTEM: |
2151 | case SIAtomicScope::AGENT: |
2152 | // Set the L0 and L1 cache policies to MISS_EVICT. |
2153 | // Note: there is no L2 cache coherent bypass control at the ISA level. |
2154 | Changed |= enableGLCBit(MI); |
2155 | break; |
2156 | case SIAtomicScope::WORKGROUP: |
2157 | // In WGP mode the waves of a work-group can be executing on either CU of |
2158 | // the WGP. Therefore need to bypass the L0 which is per CU. Otherwise in |
2159 | // CU mode all waves of a work-group are on the same CU, and so the L0 |
2160 | // does not need to be bypassed. |
2161 | if (!ST.isCuModeEnabled()) |
2162 | Changed |= enableGLCBit(MI); |
2163 | break; |
2164 | case SIAtomicScope::WAVEFRONT: |
2165 | case SIAtomicScope::SINGLETHREAD: |
2166 | // No cache to bypass. |
2167 | break; |
2168 | default: |
2169 | llvm_unreachable("Unsupported synchronization scope" ); |
2170 | } |
2171 | } |
2172 | |
2173 | /// The scratch address space does not need the global memory caches |
2174 | /// to be bypassed as all memory operations by the same thread are |
2175 | /// sequentially consistent, and no other thread can access scratch |
2176 | /// memory. |
2177 | |
2178 | /// Other address spaces do not have a cache. |
2179 | |
2180 | return Changed; |
2181 | } |
2182 | |
2183 | bool SIGfx11CacheControl::enableVolatileAndOrNonTemporal( |
2184 | MachineBasicBlock::iterator &MI, SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
2185 | bool IsVolatile, bool IsNonTemporal, bool IsLastUse = false) const { |
2186 | |
2187 | // Only handle load and store, not atomic read-modify-write insructions. The |
2188 | // latter use glc to indicate if the atomic returns a result and so must not |
2189 | // be used for cache control. |
2190 | assert(MI->mayLoad() ^ MI->mayStore()); |
2191 | |
2192 | // Only update load and store, not LLVM IR atomic read-modify-write |
2193 | // instructions. The latter are always marked as volatile so cannot sensibly |
2194 | // handle it as do not want to pessimize all atomics. Also they do not support |
2195 | // the nontemporal attribute. |
2196 | assert(Op == SIMemOp::LOAD || Op == SIMemOp::STORE); |
2197 | |
2198 | bool Changed = false; |
2199 | |
2200 | if (IsVolatile) { |
2201 | // Set L0 and L1 cache policy to be MISS_EVICT for load instructions |
2202 | // and MISS_LRU for store instructions. |
2203 | // Note: there is no L2 cache coherent bypass control at the ISA level. |
2204 | if (Op == SIMemOp::LOAD) |
2205 | Changed |= enableGLCBit(MI); |
2206 | |
2207 | // Set MALL NOALLOC for load and store instructions. |
2208 | Changed |= enableDLCBit(MI); |
2209 | |
2210 | // Ensure operation has completed at system scope to cause all volatile |
2211 | // operations to be visible outside the program in a global order. Do not |
2212 | // request cross address space as only the global address space can be |
2213 | // observable outside the program, so no need to cause a waitcnt for LDS |
2214 | // address space operations. |
2215 | Changed |= insertWait(MI, Scope: SIAtomicScope::SYSTEM, AddrSpace, Op, IsCrossAddrSpaceOrdering: false, |
2216 | Pos: Position::AFTER, Order: AtomicOrdering::Unordered); |
2217 | return Changed; |
2218 | } |
2219 | |
2220 | if (IsNonTemporal) { |
2221 | // For loads setting SLC configures L0 and L1 cache policy to HIT_EVICT |
2222 | // and L2 cache policy to STREAM. |
2223 | // For stores setting both GLC and SLC configures L0 and L1 cache policy |
2224 | // to MISS_EVICT and the L2 cache policy to STREAM. |
2225 | if (Op == SIMemOp::STORE) |
2226 | Changed |= enableGLCBit(MI); |
2227 | Changed |= enableSLCBit(MI); |
2228 | |
2229 | // Set MALL NOALLOC for load and store instructions. |
2230 | Changed |= enableDLCBit(MI); |
2231 | return Changed; |
2232 | } |
2233 | |
2234 | return Changed; |
2235 | } |
2236 | |
2237 | bool SIGfx12CacheControl::setTH(const MachineBasicBlock::iterator MI, |
2238 | AMDGPU::CPol::CPol Value) const { |
2239 | MachineOperand *CPol = TII->getNamedOperand(MI&: *MI, OperandName: OpName::cpol); |
2240 | if (!CPol) |
2241 | return false; |
2242 | |
2243 | uint64_t NewTH = Value & AMDGPU::CPol::TH; |
2244 | if ((CPol->getImm() & AMDGPU::CPol::TH) != NewTH) { |
2245 | CPol->setImm((CPol->getImm() & ~AMDGPU::CPol::TH) | NewTH); |
2246 | return true; |
2247 | } |
2248 | |
2249 | return false; |
2250 | } |
2251 | |
2252 | bool SIGfx12CacheControl::setScope(const MachineBasicBlock::iterator MI, |
2253 | AMDGPU::CPol::CPol Value) const { |
2254 | MachineOperand *CPol = TII->getNamedOperand(MI&: *MI, OperandName: OpName::cpol); |
2255 | if (!CPol) |
2256 | return false; |
2257 | |
2258 | uint64_t NewScope = Value & AMDGPU::CPol::SCOPE; |
2259 | if ((CPol->getImm() & AMDGPU::CPol::SCOPE) != NewScope) { |
2260 | CPol->setImm((CPol->getImm() & ~AMDGPU::CPol::SCOPE) | NewScope); |
2261 | return true; |
2262 | } |
2263 | |
2264 | return false; |
2265 | } |
2266 | |
2267 | bool SIGfx12CacheControl::insertWaitsBeforeSystemScopeStore( |
2268 | const MachineBasicBlock::iterator MI) const { |
2269 | // TODO: implement flag for frontend to give us a hint not to insert waits. |
2270 | |
2271 | MachineBasicBlock &MBB = *MI->getParent(); |
2272 | const DebugLoc &DL = MI->getDebugLoc(); |
2273 | |
2274 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: S_WAIT_LOADCNT_soft)).addImm(Val: 0); |
2275 | if (ST.hasImageInsts()) { |
2276 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: S_WAIT_SAMPLECNT_soft)).addImm(Val: 0); |
2277 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: S_WAIT_BVHCNT_soft)).addImm(Val: 0); |
2278 | } |
2279 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: S_WAIT_KMCNT_soft)).addImm(Val: 0); |
2280 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: S_WAIT_STORECNT_soft)).addImm(Val: 0); |
2281 | |
2282 | return true; |
2283 | } |
2284 | |
2285 | bool SIGfx12CacheControl::insertWait(MachineBasicBlock::iterator &MI, |
2286 | SIAtomicScope Scope, |
2287 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
2288 | bool IsCrossAddrSpaceOrdering, |
2289 | Position Pos, AtomicOrdering Order) const { |
2290 | bool Changed = false; |
2291 | |
2292 | MachineBasicBlock &MBB = *MI->getParent(); |
2293 | DebugLoc DL = MI->getDebugLoc(); |
2294 | |
2295 | bool LOADCnt = false; |
2296 | bool DSCnt = false; |
2297 | bool STORECnt = false; |
2298 | |
2299 | if (Pos == Position::AFTER) |
2300 | ++MI; |
2301 | |
2302 | if ((AddrSpace & (SIAtomicAddrSpace::GLOBAL | SIAtomicAddrSpace::SCRATCH)) != |
2303 | SIAtomicAddrSpace::NONE) { |
2304 | switch (Scope) { |
2305 | case SIAtomicScope::SYSTEM: |
2306 | case SIAtomicScope::AGENT: |
2307 | if ((Op & SIMemOp::LOAD) != SIMemOp::NONE) |
2308 | LOADCnt |= true; |
2309 | if ((Op & SIMemOp::STORE) != SIMemOp::NONE) |
2310 | STORECnt |= true; |
2311 | break; |
2312 | case SIAtomicScope::WORKGROUP: |
2313 | // In WGP mode the waves of a work-group can be executing on either CU of |
2314 | // the WGP. Therefore need to wait for operations to complete to ensure |
2315 | // they are visible to waves in the other CU as the L0 is per CU. |
2316 | // Otherwise in CU mode and all waves of a work-group are on the same CU |
2317 | // which shares the same L0. |
2318 | if (!ST.isCuModeEnabled()) { |
2319 | if ((Op & SIMemOp::LOAD) != SIMemOp::NONE) |
2320 | LOADCnt |= true; |
2321 | if ((Op & SIMemOp::STORE) != SIMemOp::NONE) |
2322 | STORECnt |= true; |
2323 | } |
2324 | break; |
2325 | case SIAtomicScope::WAVEFRONT: |
2326 | case SIAtomicScope::SINGLETHREAD: |
2327 | // The L0 cache keeps all memory operations in order for |
2328 | // work-items in the same wavefront. |
2329 | break; |
2330 | default: |
2331 | llvm_unreachable("Unsupported synchronization scope" ); |
2332 | } |
2333 | } |
2334 | |
2335 | if ((AddrSpace & SIAtomicAddrSpace::LDS) != SIAtomicAddrSpace::NONE) { |
2336 | switch (Scope) { |
2337 | case SIAtomicScope::SYSTEM: |
2338 | case SIAtomicScope::AGENT: |
2339 | case SIAtomicScope::WORKGROUP: |
2340 | // If no cross address space ordering then an "S_WAITCNT lgkmcnt(0)" is |
2341 | // not needed as LDS operations for all waves are executed in a total |
2342 | // global ordering as observed by all waves. Required if also |
2343 | // synchronizing with global/GDS memory as LDS operations could be |
2344 | // reordered with respect to later global/GDS memory operations of the |
2345 | // same wave. |
2346 | DSCnt |= IsCrossAddrSpaceOrdering; |
2347 | break; |
2348 | case SIAtomicScope::WAVEFRONT: |
2349 | case SIAtomicScope::SINGLETHREAD: |
2350 | // The LDS keeps all memory operations in order for |
2351 | // the same wavefront. |
2352 | break; |
2353 | default: |
2354 | llvm_unreachable("Unsupported synchronization scope" ); |
2355 | } |
2356 | } |
2357 | |
2358 | if (LOADCnt) { |
2359 | // Acquire sequences only need to wait on the previous atomic operation. |
2360 | // e.g. a typical sequence looks like |
2361 | // atomic load |
2362 | // (wait) |
2363 | // global_inv |
2364 | // |
2365 | // We do not have BVH or SAMPLE atomics, so the atomic load is always going |
2366 | // to be tracked using loadcnt. |
2367 | // |
2368 | // This also applies to fences. Fences cannot pair with an instruction |
2369 | // tracked with bvh/samplecnt as we don't have any atomics that do that. |
2370 | if (Order != AtomicOrdering::Acquire) { |
2371 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_WAIT_BVHCNT_soft)).addImm(Val: 0); |
2372 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_WAIT_SAMPLECNT_soft)).addImm(Val: 0); |
2373 | } |
2374 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_WAIT_LOADCNT_soft)).addImm(Val: 0); |
2375 | Changed = true; |
2376 | } |
2377 | |
2378 | if (STORECnt) { |
2379 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_WAIT_STORECNT_soft)).addImm(Val: 0); |
2380 | Changed = true; |
2381 | } |
2382 | |
2383 | if (DSCnt) { |
2384 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_WAIT_DSCNT_soft)).addImm(Val: 0); |
2385 | Changed = true; |
2386 | } |
2387 | |
2388 | if (Pos == Position::AFTER) |
2389 | --MI; |
2390 | |
2391 | return Changed; |
2392 | } |
2393 | |
2394 | bool SIGfx12CacheControl::insertAcquire(MachineBasicBlock::iterator &MI, |
2395 | SIAtomicScope Scope, |
2396 | SIAtomicAddrSpace AddrSpace, |
2397 | Position Pos) const { |
2398 | if (!InsertCacheInv) |
2399 | return false; |
2400 | |
2401 | MachineBasicBlock &MBB = *MI->getParent(); |
2402 | DebugLoc DL = MI->getDebugLoc(); |
2403 | |
2404 | /// The scratch address space does not need the global memory cache |
2405 | /// to be flushed as all memory operations by the same thread are |
2406 | /// sequentially consistent, and no other thread can access scratch |
2407 | /// memory. |
2408 | |
2409 | /// Other address spaces do not have a cache. |
2410 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) == SIAtomicAddrSpace::NONE) |
2411 | return false; |
2412 | |
2413 | AMDGPU::CPol::CPol ScopeImm = AMDGPU::CPol::SCOPE_DEV; |
2414 | switch (Scope) { |
2415 | case SIAtomicScope::SYSTEM: |
2416 | ScopeImm = AMDGPU::CPol::SCOPE_SYS; |
2417 | break; |
2418 | case SIAtomicScope::AGENT: |
2419 | ScopeImm = AMDGPU::CPol::SCOPE_DEV; |
2420 | break; |
2421 | case SIAtomicScope::WORKGROUP: |
2422 | // In WGP mode the waves of a work-group can be executing on either CU of |
2423 | // the WGP. Therefore we need to invalidate the L0 which is per CU. |
2424 | // Otherwise in CU mode all waves of a work-group are on the same CU, and so |
2425 | // the L0 does not need to be invalidated. |
2426 | if (ST.isCuModeEnabled()) |
2427 | return false; |
2428 | |
2429 | ScopeImm = AMDGPU::CPol::SCOPE_SE; |
2430 | break; |
2431 | case SIAtomicScope::WAVEFRONT: |
2432 | case SIAtomicScope::SINGLETHREAD: |
2433 | // No cache to invalidate. |
2434 | return false; |
2435 | default: |
2436 | llvm_unreachable("Unsupported synchronization scope" ); |
2437 | } |
2438 | |
2439 | if (Pos == Position::AFTER) |
2440 | ++MI; |
2441 | |
2442 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::GLOBAL_INV)).addImm(Val: ScopeImm); |
2443 | |
2444 | if (Pos == Position::AFTER) |
2445 | --MI; |
2446 | |
2447 | return true; |
2448 | } |
2449 | |
2450 | bool SIGfx12CacheControl::insertRelease(MachineBasicBlock::iterator &MI, |
2451 | SIAtomicScope Scope, |
2452 | SIAtomicAddrSpace AddrSpace, |
2453 | bool IsCrossAddrSpaceOrdering, |
2454 | Position Pos) const { |
2455 | MachineBasicBlock &MBB = *MI->getParent(); |
2456 | DebugLoc DL = MI->getDebugLoc(); |
2457 | |
2458 | // The scratch address space does not need the global memory cache |
2459 | // writeback as all memory operations by the same thread are |
2460 | // sequentially consistent, and no other thread can access scratch |
2461 | // memory. |
2462 | |
2463 | // Other address spaces do not have a cache. |
2464 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) == SIAtomicAddrSpace::NONE) |
2465 | return false; |
2466 | |
2467 | if (Pos == Position::AFTER) |
2468 | ++MI; |
2469 | |
2470 | // global_wb is only necessary at system scope for gfx120x targets. |
2471 | // |
2472 | // Emitting it for lower scopes is a slow no-op, so we omit it |
2473 | // for performance. |
2474 | switch (Scope) { |
2475 | case SIAtomicScope::SYSTEM: |
2476 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::GLOBAL_WB)) |
2477 | .addImm(Val: AMDGPU::CPol::SCOPE_SYS); |
2478 | break; |
2479 | case SIAtomicScope::AGENT: |
2480 | case SIAtomicScope::WORKGROUP: |
2481 | // No WB necessary, but we still have to wait. |
2482 | break; |
2483 | case SIAtomicScope::WAVEFRONT: |
2484 | case SIAtomicScope::SINGLETHREAD: |
2485 | // No WB or wait necessary here. |
2486 | return false; |
2487 | default: |
2488 | llvm_unreachable("Unsupported synchronization scope" ); |
2489 | } |
2490 | |
2491 | if (Pos == Position::AFTER) |
2492 | --MI; |
2493 | |
2494 | // We always have to wait for previous memory operations (load/store) to |
2495 | // complete, whether we inserted a WB or not. If we inserted a WB (storecnt), |
2496 | // we of course need to wait for that as well. |
2497 | insertWait(MI, Scope, AddrSpace, Op: SIMemOp::LOAD | SIMemOp::STORE, |
2498 | IsCrossAddrSpaceOrdering, Pos, Order: AtomicOrdering::Release); |
2499 | |
2500 | return true; |
2501 | } |
2502 | |
2503 | bool SIGfx12CacheControl::enableVolatileAndOrNonTemporal( |
2504 | MachineBasicBlock::iterator &MI, SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
2505 | bool IsVolatile, bool IsNonTemporal, bool IsLastUse = false) const { |
2506 | |
2507 | // Only handle load and store, not atomic read-modify-write instructions. |
2508 | assert(MI->mayLoad() ^ MI->mayStore()); |
2509 | |
2510 | // Only update load and store, not LLVM IR atomic read-modify-write |
2511 | // instructions. The latter are always marked as volatile so cannot sensibly |
2512 | // handle it as do not want to pessimize all atomics. Also they do not support |
2513 | // the nontemporal attribute. |
2514 | assert(Op == SIMemOp::LOAD || Op == SIMemOp::STORE); |
2515 | |
2516 | bool Changed = false; |
2517 | |
2518 | if (IsLastUse) { |
2519 | // Set last-use hint. |
2520 | Changed |= setTH(MI, Value: AMDGPU::CPol::TH_LU); |
2521 | } else if (IsNonTemporal) { |
2522 | // Set non-temporal hint for all cache levels. |
2523 | Changed |= setTH(MI, Value: AMDGPU::CPol::TH_NT); |
2524 | } |
2525 | |
2526 | if (IsVolatile) { |
2527 | Changed |= setScope(MI, Value: AMDGPU::CPol::SCOPE_SYS); |
2528 | |
2529 | if (Op == SIMemOp::STORE) |
2530 | Changed |= insertWaitsBeforeSystemScopeStore(MI); |
2531 | |
2532 | // Ensure operation has completed at system scope to cause all volatile |
2533 | // operations to be visible outside the program in a global order. Do not |
2534 | // request cross address space as only the global address space can be |
2535 | // observable outside the program, so no need to cause a waitcnt for LDS |
2536 | // address space operations. |
2537 | Changed |= insertWait(MI, Scope: SIAtomicScope::SYSTEM, AddrSpace, Op, IsCrossAddrSpaceOrdering: false, |
2538 | Pos: Position::AFTER, Order: AtomicOrdering::Unordered); |
2539 | } |
2540 | |
2541 | return Changed; |
2542 | } |
2543 | |
2544 | bool SIGfx12CacheControl::expandSystemScopeStore( |
2545 | MachineBasicBlock::iterator &MI) const { |
2546 | MachineOperand *CPol = TII->getNamedOperand(MI&: *MI, OperandName: OpName::cpol); |
2547 | if (CPol && ((CPol->getImm() & CPol::SCOPE) == CPol::SCOPE_SYS)) |
2548 | return insertWaitsBeforeSystemScopeStore(MI); |
2549 | |
2550 | return false; |
2551 | } |
2552 | |
2553 | bool SIGfx12CacheControl::setAtomicScope(const MachineBasicBlock::iterator &MI, |
2554 | SIAtomicScope Scope, |
2555 | SIAtomicAddrSpace AddrSpace) const { |
2556 | bool Changed = false; |
2557 | |
2558 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
2559 | switch (Scope) { |
2560 | case SIAtomicScope::SYSTEM: |
2561 | Changed |= setScope(MI, Value: AMDGPU::CPol::SCOPE_SYS); |
2562 | break; |
2563 | case SIAtomicScope::AGENT: |
2564 | Changed |= setScope(MI, Value: AMDGPU::CPol::SCOPE_DEV); |
2565 | break; |
2566 | case SIAtomicScope::WORKGROUP: |
2567 | // In workgroup mode, SCOPE_SE is needed as waves can executes on |
2568 | // different CUs that access different L0s. |
2569 | if (!ST.isCuModeEnabled()) |
2570 | Changed |= setScope(MI, Value: AMDGPU::CPol::SCOPE_SE); |
2571 | break; |
2572 | case SIAtomicScope::WAVEFRONT: |
2573 | case SIAtomicScope::SINGLETHREAD: |
2574 | // No cache to bypass. |
2575 | break; |
2576 | default: |
2577 | llvm_unreachable("Unsupported synchronization scope" ); |
2578 | } |
2579 | } |
2580 | |
2581 | // The scratch address space does not need the global memory caches |
2582 | // to be bypassed as all memory operations by the same thread are |
2583 | // sequentially consistent, and no other thread can access scratch |
2584 | // memory. |
2585 | |
2586 | // Other address spaces do not have a cache. |
2587 | |
2588 | return Changed; |
2589 | } |
2590 | |
2591 | bool SIMemoryLegalizer::removeAtomicPseudoMIs() { |
2592 | if (AtomicPseudoMIs.empty()) |
2593 | return false; |
2594 | |
2595 | for (auto &MI : AtomicPseudoMIs) |
2596 | MI->eraseFromParent(); |
2597 | |
2598 | AtomicPseudoMIs.clear(); |
2599 | return true; |
2600 | } |
2601 | |
2602 | bool SIMemoryLegalizer::expandLoad(const SIMemOpInfo &MOI, |
2603 | MachineBasicBlock::iterator &MI) { |
2604 | assert(MI->mayLoad() && !MI->mayStore()); |
2605 | |
2606 | bool Changed = false; |
2607 | |
2608 | if (MOI.isAtomic()) { |
2609 | const AtomicOrdering Order = MOI.getOrdering(); |
2610 | if (Order == AtomicOrdering::Monotonic || |
2611 | Order == AtomicOrdering::Acquire || |
2612 | Order == AtomicOrdering::SequentiallyConsistent) { |
2613 | Changed |= CC->enableLoadCacheBypass(MI, Scope: MOI.getScope(), |
2614 | AddrSpace: MOI.getOrderingAddrSpace()); |
2615 | } |
2616 | |
2617 | if (Order == AtomicOrdering::SequentiallyConsistent) |
2618 | Changed |= CC->insertWait(MI, Scope: MOI.getScope(), AddrSpace: MOI.getOrderingAddrSpace(), |
2619 | Op: SIMemOp::LOAD | SIMemOp::STORE, |
2620 | IsCrossAddrSpaceOrdering: MOI.getIsCrossAddressSpaceOrdering(), |
2621 | Pos: Position::BEFORE, Order); |
2622 | |
2623 | if (Order == AtomicOrdering::Acquire || |
2624 | Order == AtomicOrdering::SequentiallyConsistent) { |
2625 | Changed |= CC->insertWait( |
2626 | MI, Scope: MOI.getScope(), AddrSpace: MOI.getInstrAddrSpace(), Op: SIMemOp::LOAD, |
2627 | IsCrossAddrSpaceOrdering: MOI.getIsCrossAddressSpaceOrdering(), Pos: Position::AFTER, Order); |
2628 | Changed |= CC->insertAcquire(MI, Scope: MOI.getScope(), |
2629 | AddrSpace: MOI.getOrderingAddrSpace(), |
2630 | Pos: Position::AFTER); |
2631 | } |
2632 | |
2633 | return Changed; |
2634 | } |
2635 | |
2636 | // Atomic instructions already bypass caches to the scope specified by the |
2637 | // SyncScope operand. Only non-atomic volatile and nontemporal/last-use |
2638 | // instructions need additional treatment. |
2639 | Changed |= CC->enableVolatileAndOrNonTemporal( |
2640 | MI, AddrSpace: MOI.getInstrAddrSpace(), Op: SIMemOp::LOAD, IsVolatile: MOI.isVolatile(), |
2641 | IsNonTemporal: MOI.isNonTemporal(), IsLastUse: MOI.isLastUse()); |
2642 | |
2643 | return Changed; |
2644 | } |
2645 | |
2646 | bool SIMemoryLegalizer::expandStore(const SIMemOpInfo &MOI, |
2647 | MachineBasicBlock::iterator &MI) { |
2648 | assert(!MI->mayLoad() && MI->mayStore()); |
2649 | |
2650 | bool Changed = false; |
2651 | |
2652 | if (MOI.isAtomic()) { |
2653 | if (MOI.getOrdering() == AtomicOrdering::Monotonic || |
2654 | MOI.getOrdering() == AtomicOrdering::Release || |
2655 | MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent) { |
2656 | Changed |= CC->enableStoreCacheBypass(MI, Scope: MOI.getScope(), |
2657 | AddrSpace: MOI.getOrderingAddrSpace()); |
2658 | } |
2659 | |
2660 | if (MOI.getOrdering() == AtomicOrdering::Release || |
2661 | MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent) |
2662 | Changed |= CC->insertRelease(MI, Scope: MOI.getScope(), |
2663 | AddrSpace: MOI.getOrderingAddrSpace(), |
2664 | IsCrossAddrSpaceOrdering: MOI.getIsCrossAddressSpaceOrdering(), |
2665 | Pos: Position::BEFORE); |
2666 | |
2667 | return Changed; |
2668 | } |
2669 | |
2670 | // Atomic instructions already bypass caches to the scope specified by the |
2671 | // SyncScope operand. Only non-atomic volatile and nontemporal instructions |
2672 | // need additional treatment. |
2673 | Changed |= CC->enableVolatileAndOrNonTemporal( |
2674 | MI, AddrSpace: MOI.getInstrAddrSpace(), Op: SIMemOp::STORE, IsVolatile: MOI.isVolatile(), |
2675 | IsNonTemporal: MOI.isNonTemporal()); |
2676 | |
2677 | // GFX12 specific, scope(desired coherence domain in cache hierarchy) is |
2678 | // instruction field, do not confuse it with atomic scope. |
2679 | Changed |= CC->expandSystemScopeStore(MI); |
2680 | return Changed; |
2681 | } |
2682 | |
2683 | bool SIMemoryLegalizer::expandAtomicFence(const SIMemOpInfo &MOI, |
2684 | MachineBasicBlock::iterator &MI) { |
2685 | assert(MI->getOpcode() == AMDGPU::ATOMIC_FENCE); |
2686 | |
2687 | AtomicPseudoMIs.push_back(x: MI); |
2688 | bool Changed = false; |
2689 | |
2690 | // Refine fenced address space based on MMRAs. |
2691 | // |
2692 | // TODO: Should we support this MMRA on other atomic operations? |
2693 | auto OrderingAddrSpace = |
2694 | getFenceAddrSpaceMMRA(MI: *MI, Default: MOI.getOrderingAddrSpace()); |
2695 | |
2696 | if (MOI.isAtomic()) { |
2697 | const AtomicOrdering Order = MOI.getOrdering(); |
2698 | if (Order == AtomicOrdering::Acquire) { |
2699 | Changed |= CC->insertWait( |
2700 | MI, Scope: MOI.getScope(), AddrSpace: OrderingAddrSpace, Op: SIMemOp::LOAD | SIMemOp::STORE, |
2701 | IsCrossAddrSpaceOrdering: MOI.getIsCrossAddressSpaceOrdering(), Pos: Position::BEFORE, Order); |
2702 | } |
2703 | |
2704 | if (Order == AtomicOrdering::Release || |
2705 | Order == AtomicOrdering::AcquireRelease || |
2706 | Order == AtomicOrdering::SequentiallyConsistent) |
2707 | /// TODO: This relies on a barrier always generating a waitcnt |
2708 | /// for LDS to ensure it is not reordered with the completion of |
2709 | /// the proceeding LDS operations. If barrier had a memory |
2710 | /// ordering and memory scope, then library does not need to |
2711 | /// generate a fence. Could add support in this file for |
2712 | /// barrier. SIInsertWaitcnt.cpp could then stop unconditionally |
2713 | /// adding S_WAITCNT before a S_BARRIER. |
2714 | Changed |= CC->insertRelease(MI, Scope: MOI.getScope(), AddrSpace: OrderingAddrSpace, |
2715 | IsCrossAddrSpaceOrdering: MOI.getIsCrossAddressSpaceOrdering(), |
2716 | Pos: Position::BEFORE); |
2717 | |
2718 | // TODO: If both release and invalidate are happening they could be combined |
2719 | // to use the single "BUFFER_WBINV*" instruction. This could be done by |
2720 | // reorganizing this code or as part of optimizing SIInsertWaitcnt pass to |
2721 | // track cache invalidate and write back instructions. |
2722 | |
2723 | if (Order == AtomicOrdering::Acquire || |
2724 | Order == AtomicOrdering::AcquireRelease || |
2725 | Order == AtomicOrdering::SequentiallyConsistent) |
2726 | Changed |= CC->insertAcquire(MI, Scope: MOI.getScope(), AddrSpace: OrderingAddrSpace, |
2727 | Pos: Position::BEFORE); |
2728 | |
2729 | return Changed; |
2730 | } |
2731 | |
2732 | return Changed; |
2733 | } |
2734 | |
2735 | bool SIMemoryLegalizer::expandAtomicCmpxchgOrRmw(const SIMemOpInfo &MOI, |
2736 | MachineBasicBlock::iterator &MI) { |
2737 | assert(MI->mayLoad() && MI->mayStore()); |
2738 | |
2739 | bool Changed = false; |
2740 | |
2741 | if (MOI.isAtomic()) { |
2742 | const AtomicOrdering Order = MOI.getOrdering(); |
2743 | if (Order == AtomicOrdering::Monotonic || |
2744 | Order == AtomicOrdering::Acquire || Order == AtomicOrdering::Release || |
2745 | Order == AtomicOrdering::AcquireRelease || |
2746 | Order == AtomicOrdering::SequentiallyConsistent) { |
2747 | Changed |= CC->enableRMWCacheBypass(MI, Scope: MOI.getScope(), |
2748 | AddrSpace: MOI.getInstrAddrSpace()); |
2749 | } |
2750 | |
2751 | if (Order == AtomicOrdering::Release || |
2752 | Order == AtomicOrdering::AcquireRelease || |
2753 | Order == AtomicOrdering::SequentiallyConsistent || |
2754 | MOI.getFailureOrdering() == AtomicOrdering::SequentiallyConsistent) |
2755 | Changed |= CC->insertRelease(MI, Scope: MOI.getScope(), |
2756 | AddrSpace: MOI.getOrderingAddrSpace(), |
2757 | IsCrossAddrSpaceOrdering: MOI.getIsCrossAddressSpaceOrdering(), |
2758 | Pos: Position::BEFORE); |
2759 | |
2760 | if (Order == AtomicOrdering::Acquire || |
2761 | Order == AtomicOrdering::AcquireRelease || |
2762 | Order == AtomicOrdering::SequentiallyConsistent || |
2763 | MOI.getFailureOrdering() == AtomicOrdering::Acquire || |
2764 | MOI.getFailureOrdering() == AtomicOrdering::SequentiallyConsistent) { |
2765 | Changed |= CC->insertWait( |
2766 | MI, Scope: MOI.getScope(), AddrSpace: MOI.getInstrAddrSpace(), |
2767 | Op: isAtomicRet(MI: *MI) ? SIMemOp::LOAD : SIMemOp::STORE, |
2768 | IsCrossAddrSpaceOrdering: MOI.getIsCrossAddressSpaceOrdering(), Pos: Position::AFTER, Order); |
2769 | Changed |= CC->insertAcquire(MI, Scope: MOI.getScope(), |
2770 | AddrSpace: MOI.getOrderingAddrSpace(), |
2771 | Pos: Position::AFTER); |
2772 | } |
2773 | |
2774 | return Changed; |
2775 | } |
2776 | |
2777 | return Changed; |
2778 | } |
2779 | |
2780 | bool SIMemoryLegalizerLegacy::runOnMachineFunction(MachineFunction &MF) { |
2781 | const MachineModuleInfo &MMI = |
2782 | getAnalysis<MachineModuleInfoWrapperPass>().getMMI(); |
2783 | return SIMemoryLegalizer(MMI).run(MF); |
2784 | } |
2785 | |
2786 | PreservedAnalyses |
2787 | SIMemoryLegalizerPass::run(MachineFunction &MF, |
2788 | MachineFunctionAnalysisManager &MFAM) { |
2789 | auto *MMI = MFAM.getResult<ModuleAnalysisManagerMachineFunctionProxy>(IR&: MF) |
2790 | .getCachedResult<MachineModuleAnalysis>( |
2791 | IR&: *MF.getFunction().getParent()); |
2792 | assert(MMI && "MachineModuleAnalysis must be available" ); |
2793 | if (!SIMemoryLegalizer(MMI->getMMI()).run(MF)) |
2794 | return PreservedAnalyses::all(); |
2795 | return getMachineFunctionPassPreservedAnalyses().preserveSet<CFGAnalyses>(); |
2796 | } |
2797 | |
2798 | bool SIMemoryLegalizer::run(MachineFunction &MF) { |
2799 | bool Changed = false; |
2800 | |
2801 | SIMemOpAccess MOA(MMI.getObjFileInfo<AMDGPUMachineModuleInfo>()); |
2802 | CC = SICacheControl::create(ST: MF.getSubtarget<GCNSubtarget>()); |
2803 | |
2804 | for (auto &MBB : MF) { |
2805 | for (auto MI = MBB.begin(); MI != MBB.end(); ++MI) { |
2806 | |
2807 | // Unbundle instructions after the post-RA scheduler. |
2808 | if (MI->isBundle() && MI->mayLoadOrStore()) { |
2809 | MachineBasicBlock::instr_iterator II(MI->getIterator()); |
2810 | for (MachineBasicBlock::instr_iterator I = ++II, E = MBB.instr_end(); |
2811 | I != E && I->isBundledWithPred(); ++I) { |
2812 | I->unbundleFromPred(); |
2813 | for (MachineOperand &MO : I->operands()) |
2814 | if (MO.isReg()) |
2815 | MO.setIsInternalRead(false); |
2816 | } |
2817 | |
2818 | MI->eraseFromParent(); |
2819 | MI = II->getIterator(); |
2820 | } |
2821 | |
2822 | if (!(MI->getDesc().TSFlags & SIInstrFlags::maybeAtomic)) |
2823 | continue; |
2824 | |
2825 | if (const auto &MOI = MOA.getLoadInfo(MI)) |
2826 | Changed |= expandLoad(MOI: *MOI, MI); |
2827 | else if (const auto &MOI = MOA.getStoreInfo(MI)) { |
2828 | Changed |= expandStore(MOI: *MOI, MI); |
2829 | } else if (const auto &MOI = MOA.getAtomicFenceInfo(MI)) |
2830 | Changed |= expandAtomicFence(MOI: *MOI, MI); |
2831 | else if (const auto &MOI = MOA.getAtomicCmpxchgOrRmwInfo(MI)) |
2832 | Changed |= expandAtomicCmpxchgOrRmw(MOI: *MOI, MI); |
2833 | } |
2834 | } |
2835 | |
2836 | Changed |= removeAtomicPseudoMIs(); |
2837 | return Changed; |
2838 | } |
2839 | |
2840 | INITIALIZE_PASS(SIMemoryLegalizerLegacy, DEBUG_TYPE, PASS_NAME, false, false) |
2841 | |
2842 | char SIMemoryLegalizerLegacy::ID = 0; |
2843 | char &llvm::SIMemoryLegalizerID = SIMemoryLegalizerLegacy::ID; |
2844 | |
2845 | FunctionPass *llvm::createSIMemoryLegalizerPass() { |
2846 | return new SIMemoryLegalizerLegacy(); |
2847 | } |
2848 | |