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/IR/DiagnosticInfo.h" |
25 | #include "llvm/IR/MemoryModelRelaxationAnnotations.h" |
26 | #include "llvm/Support/AtomicOrdering.h" |
27 | #include "llvm/TargetParser/TargetParser.h" |
28 | |
29 | using namespace llvm; |
30 | using namespace llvm::AMDGPU; |
31 | |
32 | #define DEBUG_TYPE "si-memory-legalizer" |
33 | #define PASS_NAME "SI Memory Legalizer" |
34 | |
35 | static cl::opt<bool> AmdgcnSkipCacheInvalidations( |
36 | "amdgcn-skip-cache-invalidations" , cl::init(Val: false), cl::Hidden, |
37 | cl::desc("Use this to skip inserting cache invalidating instructions." )); |
38 | |
39 | namespace { |
40 | |
41 | LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE(); |
42 | |
43 | /// Memory operation flags. Can be ORed together. |
44 | enum class SIMemOp { |
45 | NONE = 0u, |
46 | LOAD = 1u << 0, |
47 | STORE = 1u << 1, |
48 | LLVM_MARK_AS_BITMASK_ENUM(/* LargestFlag = */ STORE) |
49 | }; |
50 | |
51 | /// Position to insert a new instruction relative to an existing |
52 | /// instruction. |
53 | enum class Position { |
54 | BEFORE, |
55 | AFTER |
56 | }; |
57 | |
58 | /// The atomic synchronization scopes supported by the AMDGPU target. |
59 | enum class SIAtomicScope { |
60 | NONE, |
61 | SINGLETHREAD, |
62 | WAVEFRONT, |
63 | WORKGROUP, |
64 | AGENT, |
65 | SYSTEM |
66 | }; |
67 | |
68 | /// The distinct address spaces supported by the AMDGPU target for |
69 | /// atomic memory operation. Can be ORed together. |
70 | enum class SIAtomicAddrSpace { |
71 | NONE = 0u, |
72 | GLOBAL = 1u << 0, |
73 | LDS = 1u << 1, |
74 | SCRATCH = 1u << 2, |
75 | GDS = 1u << 3, |
76 | OTHER = 1u << 4, |
77 | |
78 | /// The address spaces that can be accessed by a FLAT instruction. |
79 | FLAT = GLOBAL | LDS | SCRATCH, |
80 | |
81 | /// The address spaces that support atomic instructions. |
82 | ATOMIC = GLOBAL | LDS | SCRATCH | GDS, |
83 | |
84 | /// All address spaces. |
85 | ALL = GLOBAL | LDS | SCRATCH | GDS | OTHER, |
86 | |
87 | LLVM_MARK_AS_BITMASK_ENUM(/* LargestFlag = */ ALL) |
88 | }; |
89 | |
90 | class SIMemOpInfo final { |
91 | private: |
92 | |
93 | friend class SIMemOpAccess; |
94 | |
95 | AtomicOrdering Ordering = AtomicOrdering::NotAtomic; |
96 | AtomicOrdering FailureOrdering = AtomicOrdering::NotAtomic; |
97 | SIAtomicScope Scope = SIAtomicScope::SYSTEM; |
98 | SIAtomicAddrSpace OrderingAddrSpace = SIAtomicAddrSpace::NONE; |
99 | SIAtomicAddrSpace InstrAddrSpace = SIAtomicAddrSpace::NONE; |
100 | bool IsCrossAddressSpaceOrdering = false; |
101 | bool IsVolatile = false; |
102 | bool IsNonTemporal = false; |
103 | bool IsLastUse = false; |
104 | |
105 | SIMemOpInfo( |
106 | AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent, |
107 | SIAtomicScope Scope = SIAtomicScope::SYSTEM, |
108 | SIAtomicAddrSpace OrderingAddrSpace = SIAtomicAddrSpace::ATOMIC, |
109 | SIAtomicAddrSpace InstrAddrSpace = SIAtomicAddrSpace::ALL, |
110 | bool IsCrossAddressSpaceOrdering = true, |
111 | AtomicOrdering FailureOrdering = AtomicOrdering::SequentiallyConsistent, |
112 | bool IsVolatile = false, bool IsNonTemporal = false, |
113 | bool IsLastUse = false) |
114 | : Ordering(Ordering), FailureOrdering(FailureOrdering), Scope(Scope), |
115 | OrderingAddrSpace(OrderingAddrSpace), InstrAddrSpace(InstrAddrSpace), |
116 | IsCrossAddressSpaceOrdering(IsCrossAddressSpaceOrdering), |
117 | IsVolatile(IsVolatile), IsNonTemporal(IsNonTemporal), |
118 | IsLastUse(IsLastUse) { |
119 | |
120 | if (Ordering == AtomicOrdering::NotAtomic) { |
121 | assert(Scope == SIAtomicScope::NONE && |
122 | OrderingAddrSpace == SIAtomicAddrSpace::NONE && |
123 | !IsCrossAddressSpaceOrdering && |
124 | FailureOrdering == AtomicOrdering::NotAtomic); |
125 | return; |
126 | } |
127 | |
128 | assert(Scope != SIAtomicScope::NONE && |
129 | (OrderingAddrSpace & SIAtomicAddrSpace::ATOMIC) != |
130 | SIAtomicAddrSpace::NONE && |
131 | (InstrAddrSpace & SIAtomicAddrSpace::ATOMIC) != |
132 | SIAtomicAddrSpace::NONE); |
133 | |
134 | // There is also no cross address space ordering if the ordering |
135 | // address space is the same as the instruction address space and |
136 | // only contains a single address space. |
137 | if ((OrderingAddrSpace == InstrAddrSpace) && |
138 | isPowerOf2_32(Value: uint32_t(InstrAddrSpace))) |
139 | this->IsCrossAddressSpaceOrdering = false; |
140 | |
141 | // Limit the scope to the maximum supported by the instruction's address |
142 | // spaces. |
143 | if ((InstrAddrSpace & ~SIAtomicAddrSpace::SCRATCH) == |
144 | SIAtomicAddrSpace::NONE) { |
145 | this->Scope = std::min(a: Scope, b: SIAtomicScope::SINGLETHREAD); |
146 | } else if ((InstrAddrSpace & |
147 | ~(SIAtomicAddrSpace::SCRATCH | SIAtomicAddrSpace::LDS)) == |
148 | SIAtomicAddrSpace::NONE) { |
149 | this->Scope = std::min(a: Scope, b: SIAtomicScope::WORKGROUP); |
150 | } else if ((InstrAddrSpace & |
151 | ~(SIAtomicAddrSpace::SCRATCH | SIAtomicAddrSpace::LDS | |
152 | SIAtomicAddrSpace::GDS)) == SIAtomicAddrSpace::NONE) { |
153 | this->Scope = std::min(a: Scope, b: SIAtomicScope::AGENT); |
154 | } |
155 | } |
156 | |
157 | public: |
158 | /// \returns Atomic synchronization scope of the machine instruction used to |
159 | /// create this SIMemOpInfo. |
160 | SIAtomicScope getScope() const { |
161 | return Scope; |
162 | } |
163 | |
164 | /// \returns Ordering constraint of the machine instruction used to |
165 | /// create this SIMemOpInfo. |
166 | AtomicOrdering getOrdering() const { |
167 | return Ordering; |
168 | } |
169 | |
170 | /// \returns Failure ordering constraint of the machine instruction used to |
171 | /// create this SIMemOpInfo. |
172 | AtomicOrdering getFailureOrdering() const { |
173 | return FailureOrdering; |
174 | } |
175 | |
176 | /// \returns The address spaces be accessed by the machine |
177 | /// instruction used to create this SIMemOpInfo. |
178 | SIAtomicAddrSpace getInstrAddrSpace() const { |
179 | return InstrAddrSpace; |
180 | } |
181 | |
182 | /// \returns The address spaces that must be ordered by the machine |
183 | /// instruction used to create this SIMemOpInfo. |
184 | SIAtomicAddrSpace getOrderingAddrSpace() const { |
185 | return OrderingAddrSpace; |
186 | } |
187 | |
188 | /// \returns Return true iff memory ordering of operations on |
189 | /// different address spaces is required. |
190 | bool getIsCrossAddressSpaceOrdering() const { |
191 | return IsCrossAddressSpaceOrdering; |
192 | } |
193 | |
194 | /// \returns True if memory access of the machine instruction used to |
195 | /// create this SIMemOpInfo is volatile, false otherwise. |
196 | bool isVolatile() const { |
197 | return IsVolatile; |
198 | } |
199 | |
200 | /// \returns True if memory access of the machine instruction used to |
201 | /// create this SIMemOpInfo is nontemporal, false otherwise. |
202 | bool isNonTemporal() const { |
203 | return IsNonTemporal; |
204 | } |
205 | |
206 | /// \returns True if memory access of the machine instruction used to |
207 | /// create this SIMemOpInfo is last use, false otherwise. |
208 | bool isLastUse() const { return IsLastUse; } |
209 | |
210 | /// \returns True if ordering constraint of the machine instruction used to |
211 | /// create this SIMemOpInfo is unordered or higher, false otherwise. |
212 | bool isAtomic() const { |
213 | return Ordering != AtomicOrdering::NotAtomic; |
214 | } |
215 | |
216 | }; |
217 | |
218 | class SIMemOpAccess final { |
219 | private: |
220 | const AMDGPUMachineModuleInfo *MMI = nullptr; |
221 | |
222 | /// Reports unsupported message \p Msg for \p MI to LLVM context. |
223 | void reportUnsupported(const MachineBasicBlock::iterator &MI, |
224 | const char *Msg) const; |
225 | |
226 | /// Inspects the target synchronization scope \p SSID and determines |
227 | /// the SI atomic scope it corresponds to, the address spaces it |
228 | /// covers, and whether the memory ordering applies between address |
229 | /// spaces. |
230 | std::optional<std::tuple<SIAtomicScope, SIAtomicAddrSpace, bool>> |
231 | toSIAtomicScope(SyncScope::ID SSID, SIAtomicAddrSpace InstrAddrSpace) const; |
232 | |
233 | /// \return Return a bit set of the address spaces accessed by \p AS. |
234 | SIAtomicAddrSpace toSIAtomicAddrSpace(unsigned AS) const; |
235 | |
236 | /// \returns Info constructed from \p MI, which has at least machine memory |
237 | /// operand. |
238 | std::optional<SIMemOpInfo> |
239 | constructFromMIWithMMO(const MachineBasicBlock::iterator &MI) const; |
240 | |
241 | public: |
242 | /// Construct class to support accessing the machine memory operands |
243 | /// of instructions in the machine function \p MF. |
244 | SIMemOpAccess(const AMDGPUMachineModuleInfo &MMI); |
245 | |
246 | /// \returns Load info if \p MI is a load operation, "std::nullopt" otherwise. |
247 | std::optional<SIMemOpInfo> |
248 | getLoadInfo(const MachineBasicBlock::iterator &MI) const; |
249 | |
250 | /// \returns Store info if \p MI is a store operation, "std::nullopt" |
251 | /// otherwise. |
252 | std::optional<SIMemOpInfo> |
253 | getStoreInfo(const MachineBasicBlock::iterator &MI) const; |
254 | |
255 | /// \returns Atomic fence info if \p MI is an atomic fence operation, |
256 | /// "std::nullopt" otherwise. |
257 | std::optional<SIMemOpInfo> |
258 | getAtomicFenceInfo(const MachineBasicBlock::iterator &MI) const; |
259 | |
260 | /// \returns Atomic cmpxchg/rmw info if \p MI is an atomic cmpxchg or |
261 | /// rmw operation, "std::nullopt" otherwise. |
262 | std::optional<SIMemOpInfo> |
263 | getAtomicCmpxchgOrRmwInfo(const MachineBasicBlock::iterator &MI) const; |
264 | }; |
265 | |
266 | class SICacheControl { |
267 | protected: |
268 | |
269 | /// AMDGPU subtarget info. |
270 | const GCNSubtarget &ST; |
271 | |
272 | /// Instruction info. |
273 | const SIInstrInfo *TII = nullptr; |
274 | |
275 | IsaVersion IV; |
276 | |
277 | /// Whether to insert cache invalidating instructions. |
278 | bool InsertCacheInv; |
279 | |
280 | SICacheControl(const GCNSubtarget &ST); |
281 | |
282 | /// Sets named bit \p BitName to "true" if present in instruction \p MI. |
283 | /// \returns Returns true if \p MI is modified, false otherwise. |
284 | bool enableNamedBit(const MachineBasicBlock::iterator MI, |
285 | AMDGPU::CPol::CPol Bit) const; |
286 | |
287 | public: |
288 | |
289 | /// Create a cache control for the subtarget \p ST. |
290 | static std::unique_ptr<SICacheControl> create(const GCNSubtarget &ST); |
291 | |
292 | /// Update \p MI memory load instruction to bypass any caches up to |
293 | /// the \p Scope memory scope for address spaces \p |
294 | /// AddrSpace. Return true iff the instruction was modified. |
295 | virtual bool enableLoadCacheBypass(const MachineBasicBlock::iterator &MI, |
296 | SIAtomicScope Scope, |
297 | SIAtomicAddrSpace AddrSpace) const = 0; |
298 | |
299 | /// Update \p MI memory store instruction to bypass any caches up to |
300 | /// the \p Scope memory scope for address spaces \p |
301 | /// AddrSpace. Return true iff the instruction was modified. |
302 | virtual bool enableStoreCacheBypass(const MachineBasicBlock::iterator &MI, |
303 | SIAtomicScope Scope, |
304 | SIAtomicAddrSpace AddrSpace) const = 0; |
305 | |
306 | /// Update \p MI memory read-modify-write instruction to bypass any caches up |
307 | /// to the \p Scope memory scope for address spaces \p AddrSpace. Return true |
308 | /// iff the instruction was modified. |
309 | virtual bool enableRMWCacheBypass(const MachineBasicBlock::iterator &MI, |
310 | SIAtomicScope Scope, |
311 | SIAtomicAddrSpace AddrSpace) const = 0; |
312 | |
313 | /// Update \p MI memory instruction of kind \p Op associated with address |
314 | /// spaces \p AddrSpace to indicate it is volatile and/or |
315 | /// nontemporal/last-use. Return true iff the instruction was modified. |
316 | virtual bool enableVolatileAndOrNonTemporal(MachineBasicBlock::iterator &MI, |
317 | SIAtomicAddrSpace AddrSpace, |
318 | SIMemOp Op, bool IsVolatile, |
319 | bool IsNonTemporal, |
320 | bool IsLastUse = false) const = 0; |
321 | |
322 | virtual bool expandSystemScopeStore(MachineBasicBlock::iterator &MI) const { |
323 | return false; |
324 | }; |
325 | |
326 | /// Inserts any necessary instructions at position \p Pos relative |
327 | /// to instruction \p MI to ensure memory instructions before \p Pos of kind |
328 | /// \p Op associated with address spaces \p AddrSpace have completed. Used |
329 | /// between memory instructions to enforce the order they become visible as |
330 | /// observed by other memory instructions executing in memory scope \p Scope. |
331 | /// \p IsCrossAddrSpaceOrdering indicates if the memory ordering is between |
332 | /// address spaces. Returns true iff any instructions inserted. |
333 | virtual bool insertWait(MachineBasicBlock::iterator &MI, |
334 | SIAtomicScope Scope, |
335 | SIAtomicAddrSpace AddrSpace, |
336 | SIMemOp Op, |
337 | bool IsCrossAddrSpaceOrdering, |
338 | Position Pos) 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 | virtual bool tryForceStoreSC0SC1(const SIMemOpInfo &MOI, |
366 | MachineBasicBlock::iterator &MI) const { |
367 | return false; |
368 | } |
369 | }; |
370 | |
371 | class SIGfx6CacheControl : public SICacheControl { |
372 | protected: |
373 | |
374 | /// Sets GLC bit to "true" if present in \p MI. Returns true if \p MI |
375 | /// is modified, false otherwise. |
376 | bool enableGLCBit(const MachineBasicBlock::iterator &MI) const { |
377 | return enableNamedBit(MI, Bit: AMDGPU::CPol::GLC); |
378 | } |
379 | |
380 | /// Sets SLC bit to "true" if present in \p MI. Returns true if \p MI |
381 | /// is modified, false otherwise. |
382 | bool enableSLCBit(const MachineBasicBlock::iterator &MI) const { |
383 | return enableNamedBit(MI, Bit: AMDGPU::CPol::SLC); |
384 | } |
385 | |
386 | public: |
387 | |
388 | SIGfx6CacheControl(const GCNSubtarget &ST) : SICacheControl(ST) {} |
389 | |
390 | bool enableLoadCacheBypass(const MachineBasicBlock::iterator &MI, |
391 | SIAtomicScope Scope, |
392 | SIAtomicAddrSpace AddrSpace) const override; |
393 | |
394 | bool enableStoreCacheBypass(const MachineBasicBlock::iterator &MI, |
395 | SIAtomicScope Scope, |
396 | SIAtomicAddrSpace AddrSpace) const override; |
397 | |
398 | bool enableRMWCacheBypass(const MachineBasicBlock::iterator &MI, |
399 | SIAtomicScope Scope, |
400 | SIAtomicAddrSpace AddrSpace) const override; |
401 | |
402 | bool enableVolatileAndOrNonTemporal(MachineBasicBlock::iterator &MI, |
403 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
404 | bool IsVolatile, bool IsNonTemporal, |
405 | bool IsLastUse) const override; |
406 | |
407 | bool insertWait(MachineBasicBlock::iterator &MI, |
408 | SIAtomicScope Scope, |
409 | SIAtomicAddrSpace AddrSpace, |
410 | SIMemOp Op, |
411 | bool IsCrossAddrSpaceOrdering, |
412 | Position Pos) const override; |
413 | |
414 | bool insertAcquire(MachineBasicBlock::iterator &MI, |
415 | SIAtomicScope Scope, |
416 | SIAtomicAddrSpace AddrSpace, |
417 | Position Pos) const override; |
418 | |
419 | bool insertRelease(MachineBasicBlock::iterator &MI, |
420 | SIAtomicScope Scope, |
421 | SIAtomicAddrSpace AddrSpace, |
422 | bool IsCrossAddrSpaceOrdering, |
423 | Position Pos) const override; |
424 | }; |
425 | |
426 | class SIGfx7CacheControl : public SIGfx6CacheControl { |
427 | public: |
428 | |
429 | SIGfx7CacheControl(const GCNSubtarget &ST) : SIGfx6CacheControl(ST) {} |
430 | |
431 | bool insertAcquire(MachineBasicBlock::iterator &MI, |
432 | SIAtomicScope Scope, |
433 | SIAtomicAddrSpace AddrSpace, |
434 | Position Pos) const override; |
435 | |
436 | }; |
437 | |
438 | class SIGfx90ACacheControl : public SIGfx7CacheControl { |
439 | public: |
440 | |
441 | SIGfx90ACacheControl(const GCNSubtarget &ST) : SIGfx7CacheControl(ST) {} |
442 | |
443 | bool enableLoadCacheBypass(const MachineBasicBlock::iterator &MI, |
444 | SIAtomicScope Scope, |
445 | SIAtomicAddrSpace AddrSpace) const override; |
446 | |
447 | bool enableStoreCacheBypass(const MachineBasicBlock::iterator &MI, |
448 | SIAtomicScope Scope, |
449 | SIAtomicAddrSpace AddrSpace) const override; |
450 | |
451 | bool enableRMWCacheBypass(const MachineBasicBlock::iterator &MI, |
452 | SIAtomicScope Scope, |
453 | SIAtomicAddrSpace AddrSpace) const override; |
454 | |
455 | bool enableVolatileAndOrNonTemporal(MachineBasicBlock::iterator &MI, |
456 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
457 | bool IsVolatile, bool IsNonTemporal, |
458 | bool IsLastUse) const override; |
459 | |
460 | bool insertWait(MachineBasicBlock::iterator &MI, |
461 | SIAtomicScope Scope, |
462 | SIAtomicAddrSpace AddrSpace, |
463 | SIMemOp Op, |
464 | bool IsCrossAddrSpaceOrdering, |
465 | Position Pos) const override; |
466 | |
467 | bool insertAcquire(MachineBasicBlock::iterator &MI, |
468 | SIAtomicScope Scope, |
469 | SIAtomicAddrSpace AddrSpace, |
470 | Position Pos) const override; |
471 | |
472 | bool insertRelease(MachineBasicBlock::iterator &MI, |
473 | SIAtomicScope Scope, |
474 | SIAtomicAddrSpace AddrSpace, |
475 | bool IsCrossAddrSpaceOrdering, |
476 | Position Pos) const override; |
477 | }; |
478 | |
479 | class SIGfx940CacheControl : public SIGfx90ACacheControl { |
480 | protected: |
481 | |
482 | /// Sets SC0 bit to "true" if present in \p MI. Returns true if \p MI |
483 | /// is modified, false otherwise. |
484 | bool enableSC0Bit(const MachineBasicBlock::iterator &MI) const { |
485 | return enableNamedBit(MI, Bit: AMDGPU::CPol::SC0); |
486 | } |
487 | |
488 | /// Sets SC1 bit to "true" if present in \p MI. Returns true if \p MI |
489 | /// is modified, false otherwise. |
490 | bool enableSC1Bit(const MachineBasicBlock::iterator &MI) const { |
491 | return enableNamedBit(MI, Bit: AMDGPU::CPol::SC1); |
492 | } |
493 | |
494 | /// Sets NT bit to "true" if present in \p MI. Returns true if \p MI |
495 | /// is modified, false otherwise. |
496 | bool enableNTBit(const MachineBasicBlock::iterator &MI) const { |
497 | return enableNamedBit(MI, Bit: AMDGPU::CPol::NT); |
498 | } |
499 | |
500 | public: |
501 | |
502 | SIGfx940CacheControl(const GCNSubtarget &ST) : SIGfx90ACacheControl(ST) {}; |
503 | |
504 | bool enableLoadCacheBypass(const MachineBasicBlock::iterator &MI, |
505 | SIAtomicScope Scope, |
506 | SIAtomicAddrSpace AddrSpace) const override; |
507 | |
508 | bool enableStoreCacheBypass(const MachineBasicBlock::iterator &MI, |
509 | SIAtomicScope Scope, |
510 | SIAtomicAddrSpace AddrSpace) const override; |
511 | |
512 | bool enableRMWCacheBypass(const MachineBasicBlock::iterator &MI, |
513 | SIAtomicScope Scope, |
514 | SIAtomicAddrSpace AddrSpace) const override; |
515 | |
516 | bool enableVolatileAndOrNonTemporal(MachineBasicBlock::iterator &MI, |
517 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
518 | bool IsVolatile, bool IsNonTemporal, |
519 | bool IsLastUse) const override; |
520 | |
521 | bool insertAcquire(MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
522 | SIAtomicAddrSpace AddrSpace, Position Pos) const override; |
523 | |
524 | bool insertRelease(MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
525 | SIAtomicAddrSpace AddrSpace, bool IsCrossAddrSpaceOrdering, |
526 | Position Pos) const override; |
527 | |
528 | bool tryForceStoreSC0SC1(const SIMemOpInfo &MOI, |
529 | MachineBasicBlock::iterator &MI) const override { |
530 | bool Changed = false; |
531 | if (ST.hasForceStoreSC0SC1() && |
532 | (MOI.getInstrAddrSpace() & (SIAtomicAddrSpace::SCRATCH | |
533 | SIAtomicAddrSpace::GLOBAL | |
534 | SIAtomicAddrSpace::OTHER)) != |
535 | SIAtomicAddrSpace::NONE) { |
536 | Changed |= enableSC0Bit(MI); |
537 | Changed |= enableSC1Bit(MI); |
538 | } |
539 | return Changed; |
540 | } |
541 | }; |
542 | |
543 | class SIGfx10CacheControl : public SIGfx7CacheControl { |
544 | protected: |
545 | |
546 | /// Sets DLC bit to "true" if present in \p MI. Returns true if \p MI |
547 | /// is modified, false otherwise. |
548 | bool enableDLCBit(const MachineBasicBlock::iterator &MI) const { |
549 | return enableNamedBit(MI, Bit: AMDGPU::CPol::DLC); |
550 | } |
551 | |
552 | public: |
553 | |
554 | SIGfx10CacheControl(const GCNSubtarget &ST) : SIGfx7CacheControl(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 | bool insertWait(MachineBasicBlock::iterator &MI, |
566 | SIAtomicScope Scope, |
567 | SIAtomicAddrSpace AddrSpace, |
568 | SIMemOp Op, |
569 | bool IsCrossAddrSpaceOrdering, |
570 | Position Pos) const override; |
571 | |
572 | bool insertAcquire(MachineBasicBlock::iterator &MI, |
573 | SIAtomicScope Scope, |
574 | SIAtomicAddrSpace AddrSpace, |
575 | Position Pos) const override; |
576 | }; |
577 | |
578 | class SIGfx11CacheControl : public SIGfx10CacheControl { |
579 | public: |
580 | SIGfx11CacheControl(const GCNSubtarget &ST) : SIGfx10CacheControl(ST) {} |
581 | |
582 | bool enableLoadCacheBypass(const MachineBasicBlock::iterator &MI, |
583 | SIAtomicScope Scope, |
584 | SIAtomicAddrSpace AddrSpace) const override; |
585 | |
586 | bool enableVolatileAndOrNonTemporal(MachineBasicBlock::iterator &MI, |
587 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
588 | bool IsVolatile, bool IsNonTemporal, |
589 | bool IsLastUse) const override; |
590 | }; |
591 | |
592 | class SIGfx12CacheControl : public SIGfx11CacheControl { |
593 | protected: |
594 | // Sets TH policy to \p Value if CPol operand is present in instruction \p MI. |
595 | // \returns Returns true if \p MI is modified, false otherwise. |
596 | bool setTH(const MachineBasicBlock::iterator MI, |
597 | AMDGPU::CPol::CPol Value) const; |
598 | // Sets Scope policy to \p Value if CPol operand is present in instruction \p |
599 | // MI. \returns Returns true if \p MI is modified, false otherwise. |
600 | bool setScope(const MachineBasicBlock::iterator MI, |
601 | AMDGPU::CPol::CPol Value) const; |
602 | |
603 | // Stores with system scope (SCOPE_SYS) need to wait for: |
604 | // - loads or atomics(returning) - wait for {LOAD|SAMPLE|BVH|KM}CNT==0 |
605 | // - non-returning-atomics - wait for STORECNT==0 |
606 | // TODO: SIInsertWaitcnts will not always be able to remove STORECNT waits |
607 | // since it does not distinguish atomics-with-return from regular stores. |
608 | // There is no need to wait if memory is cached (mtype != UC). |
609 | bool |
610 | insertWaitsBeforeSystemScopeStore(const MachineBasicBlock::iterator MI) const; |
611 | |
612 | bool setAtomicScope(const MachineBasicBlock::iterator &MI, |
613 | SIAtomicScope Scope, SIAtomicAddrSpace AddrSpace) const; |
614 | |
615 | public: |
616 | SIGfx12CacheControl(const GCNSubtarget &ST) : SIGfx11CacheControl(ST) {} |
617 | |
618 | bool insertWait(MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
619 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
620 | bool IsCrossAddrSpaceOrdering, Position Pos) const override; |
621 | |
622 | bool insertAcquire(MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
623 | SIAtomicAddrSpace AddrSpace, Position Pos) const override; |
624 | |
625 | bool enableVolatileAndOrNonTemporal(MachineBasicBlock::iterator &MI, |
626 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
627 | bool IsVolatile, bool IsNonTemporal, |
628 | bool IsLastUse) const override; |
629 | |
630 | bool expandSystemScopeStore(MachineBasicBlock::iterator &MI) const override; |
631 | |
632 | bool insertRelease(MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
633 | SIAtomicAddrSpace AddrSpace, bool IsCrossAddrSpaceOrdering, |
634 | Position Pos) const override; |
635 | |
636 | bool enableLoadCacheBypass(const MachineBasicBlock::iterator &MI, |
637 | SIAtomicScope Scope, |
638 | SIAtomicAddrSpace AddrSpace) const override { |
639 | return setAtomicScope(MI, Scope, AddrSpace); |
640 | } |
641 | |
642 | bool enableStoreCacheBypass(const MachineBasicBlock::iterator &MI, |
643 | SIAtomicScope Scope, |
644 | SIAtomicAddrSpace AddrSpace) const override { |
645 | return setAtomicScope(MI, Scope, AddrSpace); |
646 | } |
647 | |
648 | bool enableRMWCacheBypass(const MachineBasicBlock::iterator &MI, |
649 | SIAtomicScope Scope, |
650 | SIAtomicAddrSpace AddrSpace) const override { |
651 | return setAtomicScope(MI, Scope, AddrSpace); |
652 | } |
653 | }; |
654 | |
655 | class SIMemoryLegalizer final : public MachineFunctionPass { |
656 | private: |
657 | |
658 | /// Cache Control. |
659 | std::unique_ptr<SICacheControl> CC = nullptr; |
660 | |
661 | /// List of atomic pseudo instructions. |
662 | std::list<MachineBasicBlock::iterator> AtomicPseudoMIs; |
663 | |
664 | /// Return true iff instruction \p MI is a atomic instruction that |
665 | /// returns a result. |
666 | bool isAtomicRet(const MachineInstr &MI) const { |
667 | return SIInstrInfo::isAtomicRet(MI); |
668 | } |
669 | |
670 | /// Removes all processed atomic pseudo instructions from the current |
671 | /// function. Returns true if current function is modified, false otherwise. |
672 | bool removeAtomicPseudoMIs(); |
673 | |
674 | /// Expands load operation \p MI. Returns true if instructions are |
675 | /// added/deleted or \p MI is modified, false otherwise. |
676 | bool expandLoad(const SIMemOpInfo &MOI, |
677 | MachineBasicBlock::iterator &MI); |
678 | /// Expands store operation \p MI. Returns true if instructions are |
679 | /// added/deleted or \p MI is modified, false otherwise. |
680 | bool expandStore(const SIMemOpInfo &MOI, |
681 | MachineBasicBlock::iterator &MI); |
682 | /// Expands atomic fence operation \p MI. Returns true if |
683 | /// instructions are added/deleted or \p MI is modified, false otherwise. |
684 | bool expandAtomicFence(const SIMemOpInfo &MOI, |
685 | MachineBasicBlock::iterator &MI); |
686 | /// Expands atomic cmpxchg or rmw operation \p MI. Returns true if |
687 | /// instructions are added/deleted or \p MI is modified, false otherwise. |
688 | bool expandAtomicCmpxchgOrRmw(const SIMemOpInfo &MOI, |
689 | MachineBasicBlock::iterator &MI); |
690 | |
691 | public: |
692 | static char ID; |
693 | |
694 | SIMemoryLegalizer() : MachineFunctionPass(ID) {} |
695 | |
696 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
697 | AU.setPreservesCFG(); |
698 | MachineFunctionPass::getAnalysisUsage(AU); |
699 | } |
700 | |
701 | StringRef getPassName() const override { |
702 | return PASS_NAME; |
703 | } |
704 | |
705 | bool runOnMachineFunction(MachineFunction &MF) override; |
706 | }; |
707 | |
708 | static const StringMap<SIAtomicAddrSpace> ASNames = {{ |
709 | {"global" , SIAtomicAddrSpace::GLOBAL}, |
710 | {"local" , SIAtomicAddrSpace::LDS}, |
711 | }}; |
712 | |
713 | void diagnoseUnknownMMRAASName(const MachineInstr &MI, StringRef AS) { |
714 | const MachineFunction *MF = MI.getMF(); |
715 | const Function &Fn = MF->getFunction(); |
716 | SmallString<128> Str; |
717 | raw_svector_ostream OS(Str); |
718 | OS << "unknown address space '" << AS << "'; expected one of " ; |
719 | ListSeparator LS; |
720 | for (const auto &[Name, Val] : ASNames) |
721 | OS << LS << '\'' << Name << '\''; |
722 | DiagnosticInfoUnsupported BadTag(Fn, Str.str(), MI.getDebugLoc(), DS_Warning); |
723 | Fn.getContext().diagnose(DI: BadTag); |
724 | } |
725 | |
726 | /// Reads \p MI's MMRAs to parse the "amdgpu-as" MMRA. |
727 | /// If this tag isn't present, or if it has no meaningful values, returns \p |
728 | /// Default. Otherwise returns all the address spaces concerned by the MMRA. |
729 | static SIAtomicAddrSpace getFenceAddrSpaceMMRA(const MachineInstr &MI, |
730 | SIAtomicAddrSpace Default) { |
731 | static constexpr StringLiteral FenceASPrefix = "amdgpu-as" ; |
732 | |
733 | auto MMRA = MMRAMetadata(MI.getMMRAMetadata()); |
734 | if (!MMRA) |
735 | return Default; |
736 | |
737 | SIAtomicAddrSpace Result = SIAtomicAddrSpace::NONE; |
738 | for (const auto &[Prefix, Suffix] : MMRA) { |
739 | if (Prefix != FenceASPrefix) |
740 | continue; |
741 | |
742 | if (auto It = ASNames.find(Key: Suffix); It != ASNames.end()) |
743 | Result |= It->second; |
744 | else |
745 | diagnoseUnknownMMRAASName(MI, AS: Suffix); |
746 | } |
747 | |
748 | return (Result != SIAtomicAddrSpace::NONE) ? Result : Default; |
749 | } |
750 | |
751 | } // end anonymous namespace |
752 | |
753 | void SIMemOpAccess::reportUnsupported(const MachineBasicBlock::iterator &MI, |
754 | const char *Msg) const { |
755 | const Function &Func = MI->getParent()->getParent()->getFunction(); |
756 | DiagnosticInfoUnsupported Diag(Func, Msg, MI->getDebugLoc()); |
757 | Func.getContext().diagnose(DI: Diag); |
758 | } |
759 | |
760 | std::optional<std::tuple<SIAtomicScope, SIAtomicAddrSpace, bool>> |
761 | SIMemOpAccess::toSIAtomicScope(SyncScope::ID SSID, |
762 | SIAtomicAddrSpace InstrAddrSpace) const { |
763 | if (SSID == SyncScope::System) |
764 | return std::tuple(SIAtomicScope::SYSTEM, SIAtomicAddrSpace::ATOMIC, true); |
765 | if (SSID == MMI->getAgentSSID()) |
766 | return std::tuple(SIAtomicScope::AGENT, SIAtomicAddrSpace::ATOMIC, true); |
767 | if (SSID == MMI->getWorkgroupSSID()) |
768 | return std::tuple(SIAtomicScope::WORKGROUP, SIAtomicAddrSpace::ATOMIC, |
769 | true); |
770 | if (SSID == MMI->getWavefrontSSID()) |
771 | return std::tuple(SIAtomicScope::WAVEFRONT, SIAtomicAddrSpace::ATOMIC, |
772 | true); |
773 | if (SSID == SyncScope::SingleThread) |
774 | return std::tuple(SIAtomicScope::SINGLETHREAD, SIAtomicAddrSpace::ATOMIC, |
775 | true); |
776 | if (SSID == MMI->getSystemOneAddressSpaceSSID()) |
777 | return std::tuple(SIAtomicScope::SYSTEM, |
778 | SIAtomicAddrSpace::ATOMIC & InstrAddrSpace, false); |
779 | if (SSID == MMI->getAgentOneAddressSpaceSSID()) |
780 | return std::tuple(SIAtomicScope::AGENT, |
781 | SIAtomicAddrSpace::ATOMIC & InstrAddrSpace, false); |
782 | if (SSID == MMI->getWorkgroupOneAddressSpaceSSID()) |
783 | return std::tuple(SIAtomicScope::WORKGROUP, |
784 | SIAtomicAddrSpace::ATOMIC & InstrAddrSpace, false); |
785 | if (SSID == MMI->getWavefrontOneAddressSpaceSSID()) |
786 | return std::tuple(SIAtomicScope::WAVEFRONT, |
787 | SIAtomicAddrSpace::ATOMIC & InstrAddrSpace, false); |
788 | if (SSID == MMI->getSingleThreadOneAddressSpaceSSID()) |
789 | return std::tuple(SIAtomicScope::SINGLETHREAD, |
790 | SIAtomicAddrSpace::ATOMIC & InstrAddrSpace, false); |
791 | return std::nullopt; |
792 | } |
793 | |
794 | SIAtomicAddrSpace SIMemOpAccess::toSIAtomicAddrSpace(unsigned AS) const { |
795 | if (AS == AMDGPUAS::FLAT_ADDRESS) |
796 | return SIAtomicAddrSpace::FLAT; |
797 | if (AS == AMDGPUAS::GLOBAL_ADDRESS) |
798 | return SIAtomicAddrSpace::GLOBAL; |
799 | if (AS == AMDGPUAS::LOCAL_ADDRESS) |
800 | return SIAtomicAddrSpace::LDS; |
801 | if (AS == AMDGPUAS::PRIVATE_ADDRESS) |
802 | return SIAtomicAddrSpace::SCRATCH; |
803 | if (AS == AMDGPUAS::REGION_ADDRESS) |
804 | return SIAtomicAddrSpace::GDS; |
805 | |
806 | return SIAtomicAddrSpace::OTHER; |
807 | } |
808 | |
809 | SIMemOpAccess::SIMemOpAccess(const AMDGPUMachineModuleInfo &MMI_) |
810 | : MMI(&MMI_) {} |
811 | |
812 | std::optional<SIMemOpInfo> SIMemOpAccess::constructFromMIWithMMO( |
813 | const MachineBasicBlock::iterator &MI) const { |
814 | assert(MI->getNumMemOperands() > 0); |
815 | |
816 | SyncScope::ID SSID = SyncScope::SingleThread; |
817 | AtomicOrdering Ordering = AtomicOrdering::NotAtomic; |
818 | AtomicOrdering FailureOrdering = AtomicOrdering::NotAtomic; |
819 | SIAtomicAddrSpace InstrAddrSpace = SIAtomicAddrSpace::NONE; |
820 | bool IsNonTemporal = true; |
821 | bool IsVolatile = false; |
822 | bool IsLastUse = false; |
823 | |
824 | // Validator should check whether or not MMOs cover the entire set of |
825 | // locations accessed by the memory instruction. |
826 | for (const auto &MMO : MI->memoperands()) { |
827 | IsNonTemporal &= MMO->isNonTemporal(); |
828 | IsVolatile |= MMO->isVolatile(); |
829 | IsLastUse |= MMO->getFlags() & MOLastUse; |
830 | InstrAddrSpace |= |
831 | toSIAtomicAddrSpace(AS: MMO->getPointerInfo().getAddrSpace()); |
832 | AtomicOrdering OpOrdering = MMO->getSuccessOrdering(); |
833 | if (OpOrdering != AtomicOrdering::NotAtomic) { |
834 | const auto &IsSyncScopeInclusion = |
835 | MMI->isSyncScopeInclusion(A: SSID, B: MMO->getSyncScopeID()); |
836 | if (!IsSyncScopeInclusion) { |
837 | reportUnsupported(MI, |
838 | Msg: "Unsupported non-inclusive atomic synchronization scope" ); |
839 | return std::nullopt; |
840 | } |
841 | |
842 | SSID = *IsSyncScopeInclusion ? SSID : MMO->getSyncScopeID(); |
843 | Ordering = getMergedAtomicOrdering(AO: Ordering, Other: OpOrdering); |
844 | assert(MMO->getFailureOrdering() != AtomicOrdering::Release && |
845 | MMO->getFailureOrdering() != AtomicOrdering::AcquireRelease); |
846 | FailureOrdering = |
847 | getMergedAtomicOrdering(AO: FailureOrdering, Other: MMO->getFailureOrdering()); |
848 | } |
849 | } |
850 | |
851 | SIAtomicScope Scope = SIAtomicScope::NONE; |
852 | SIAtomicAddrSpace OrderingAddrSpace = SIAtomicAddrSpace::NONE; |
853 | bool IsCrossAddressSpaceOrdering = false; |
854 | if (Ordering != AtomicOrdering::NotAtomic) { |
855 | auto ScopeOrNone = toSIAtomicScope(SSID, InstrAddrSpace); |
856 | if (!ScopeOrNone) { |
857 | reportUnsupported(MI, Msg: "Unsupported atomic synchronization scope" ); |
858 | return std::nullopt; |
859 | } |
860 | std::tie(args&: Scope, args&: OrderingAddrSpace, args&: IsCrossAddressSpaceOrdering) = |
861 | *ScopeOrNone; |
862 | if ((OrderingAddrSpace == SIAtomicAddrSpace::NONE) || |
863 | ((OrderingAddrSpace & SIAtomicAddrSpace::ATOMIC) != OrderingAddrSpace) || |
864 | ((InstrAddrSpace & SIAtomicAddrSpace::ATOMIC) == SIAtomicAddrSpace::NONE)) { |
865 | reportUnsupported(MI, Msg: "Unsupported atomic address space" ); |
866 | return std::nullopt; |
867 | } |
868 | } |
869 | return SIMemOpInfo(Ordering, Scope, OrderingAddrSpace, InstrAddrSpace, |
870 | IsCrossAddressSpaceOrdering, FailureOrdering, IsVolatile, |
871 | IsNonTemporal, IsLastUse); |
872 | } |
873 | |
874 | std::optional<SIMemOpInfo> |
875 | SIMemOpAccess::getLoadInfo(const MachineBasicBlock::iterator &MI) const { |
876 | assert(MI->getDesc().TSFlags & SIInstrFlags::maybeAtomic); |
877 | |
878 | if (!(MI->mayLoad() && !MI->mayStore())) |
879 | return std::nullopt; |
880 | |
881 | // Be conservative if there are no memory operands. |
882 | if (MI->getNumMemOperands() == 0) |
883 | return SIMemOpInfo(); |
884 | |
885 | return constructFromMIWithMMO(MI); |
886 | } |
887 | |
888 | std::optional<SIMemOpInfo> |
889 | SIMemOpAccess::getStoreInfo(const MachineBasicBlock::iterator &MI) const { |
890 | assert(MI->getDesc().TSFlags & SIInstrFlags::maybeAtomic); |
891 | |
892 | if (!(!MI->mayLoad() && MI->mayStore())) |
893 | return std::nullopt; |
894 | |
895 | // Be conservative if there are no memory operands. |
896 | if (MI->getNumMemOperands() == 0) |
897 | return SIMemOpInfo(); |
898 | |
899 | return constructFromMIWithMMO(MI); |
900 | } |
901 | |
902 | std::optional<SIMemOpInfo> |
903 | SIMemOpAccess::getAtomicFenceInfo(const MachineBasicBlock::iterator &MI) const { |
904 | assert(MI->getDesc().TSFlags & SIInstrFlags::maybeAtomic); |
905 | |
906 | if (MI->getOpcode() != AMDGPU::ATOMIC_FENCE) |
907 | return std::nullopt; |
908 | |
909 | AtomicOrdering Ordering = |
910 | static_cast<AtomicOrdering>(MI->getOperand(i: 0).getImm()); |
911 | |
912 | SyncScope::ID SSID = static_cast<SyncScope::ID>(MI->getOperand(i: 1).getImm()); |
913 | auto ScopeOrNone = toSIAtomicScope(SSID, InstrAddrSpace: SIAtomicAddrSpace::ATOMIC); |
914 | if (!ScopeOrNone) { |
915 | reportUnsupported(MI, Msg: "Unsupported atomic synchronization scope" ); |
916 | return std::nullopt; |
917 | } |
918 | |
919 | SIAtomicScope Scope = SIAtomicScope::NONE; |
920 | SIAtomicAddrSpace OrderingAddrSpace = SIAtomicAddrSpace::NONE; |
921 | bool IsCrossAddressSpaceOrdering = false; |
922 | std::tie(args&: Scope, args&: OrderingAddrSpace, args&: IsCrossAddressSpaceOrdering) = |
923 | *ScopeOrNone; |
924 | |
925 | if ((OrderingAddrSpace == SIAtomicAddrSpace::NONE) || |
926 | ((OrderingAddrSpace & SIAtomicAddrSpace::ATOMIC) != OrderingAddrSpace)) { |
927 | reportUnsupported(MI, Msg: "Unsupported atomic address space" ); |
928 | return std::nullopt; |
929 | } |
930 | |
931 | return SIMemOpInfo(Ordering, Scope, OrderingAddrSpace, SIAtomicAddrSpace::ATOMIC, |
932 | IsCrossAddressSpaceOrdering, AtomicOrdering::NotAtomic); |
933 | } |
934 | |
935 | std::optional<SIMemOpInfo> SIMemOpAccess::getAtomicCmpxchgOrRmwInfo( |
936 | const MachineBasicBlock::iterator &MI) const { |
937 | assert(MI->getDesc().TSFlags & SIInstrFlags::maybeAtomic); |
938 | |
939 | if (!(MI->mayLoad() && MI->mayStore())) |
940 | return std::nullopt; |
941 | |
942 | // Be conservative if there are no memory operands. |
943 | if (MI->getNumMemOperands() == 0) |
944 | return SIMemOpInfo(); |
945 | |
946 | return constructFromMIWithMMO(MI); |
947 | } |
948 | |
949 | SICacheControl::SICacheControl(const GCNSubtarget &ST) : ST(ST) { |
950 | TII = ST.getInstrInfo(); |
951 | IV = getIsaVersion(GPU: ST.getCPU()); |
952 | InsertCacheInv = !AmdgcnSkipCacheInvalidations; |
953 | } |
954 | |
955 | bool SICacheControl::enableNamedBit(const MachineBasicBlock::iterator MI, |
956 | AMDGPU::CPol::CPol Bit) const { |
957 | MachineOperand *CPol = TII->getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::cpol); |
958 | if (!CPol) |
959 | return false; |
960 | |
961 | CPol->setImm(CPol->getImm() | Bit); |
962 | return true; |
963 | } |
964 | |
965 | /* static */ |
966 | std::unique_ptr<SICacheControl> SICacheControl::create(const GCNSubtarget &ST) { |
967 | GCNSubtarget::Generation Generation = ST.getGeneration(); |
968 | if (ST.hasGFX940Insts()) |
969 | return std::make_unique<SIGfx940CacheControl>(args: ST); |
970 | if (ST.hasGFX90AInsts()) |
971 | return std::make_unique<SIGfx90ACacheControl>(args: ST); |
972 | if (Generation <= AMDGPUSubtarget::SOUTHERN_ISLANDS) |
973 | return std::make_unique<SIGfx6CacheControl>(args: ST); |
974 | if (Generation < AMDGPUSubtarget::GFX10) |
975 | return std::make_unique<SIGfx7CacheControl>(args: ST); |
976 | if (Generation < AMDGPUSubtarget::GFX11) |
977 | return std::make_unique<SIGfx10CacheControl>(args: ST); |
978 | if (Generation < AMDGPUSubtarget::GFX12) |
979 | return std::make_unique<SIGfx11CacheControl>(args: ST); |
980 | return std::make_unique<SIGfx12CacheControl>(args: ST); |
981 | } |
982 | |
983 | bool SIGfx6CacheControl::enableLoadCacheBypass( |
984 | const MachineBasicBlock::iterator &MI, |
985 | SIAtomicScope Scope, |
986 | SIAtomicAddrSpace AddrSpace) const { |
987 | assert(MI->mayLoad() && !MI->mayStore()); |
988 | bool Changed = false; |
989 | |
990 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
991 | switch (Scope) { |
992 | case SIAtomicScope::SYSTEM: |
993 | case SIAtomicScope::AGENT: |
994 | // Set L1 cache policy to MISS_EVICT. |
995 | // Note: there is no L2 cache bypass policy at the ISA level. |
996 | Changed |= enableGLCBit(MI); |
997 | break; |
998 | case SIAtomicScope::WORKGROUP: |
999 | case SIAtomicScope::WAVEFRONT: |
1000 | case SIAtomicScope::SINGLETHREAD: |
1001 | // No cache to bypass. |
1002 | break; |
1003 | default: |
1004 | llvm_unreachable("Unsupported synchronization scope" ); |
1005 | } |
1006 | } |
1007 | |
1008 | /// The scratch address space does not need the global memory caches |
1009 | /// to be bypassed as all memory operations by the same thread are |
1010 | /// sequentially consistent, and no other thread can access scratch |
1011 | /// memory. |
1012 | |
1013 | /// Other address spaces do not have a cache. |
1014 | |
1015 | return Changed; |
1016 | } |
1017 | |
1018 | bool SIGfx6CacheControl::enableStoreCacheBypass( |
1019 | const MachineBasicBlock::iterator &MI, |
1020 | SIAtomicScope Scope, |
1021 | SIAtomicAddrSpace AddrSpace) const { |
1022 | assert(!MI->mayLoad() && MI->mayStore()); |
1023 | bool Changed = false; |
1024 | |
1025 | /// The L1 cache is write through so does not need to be bypassed. There is no |
1026 | /// bypass control for the L2 cache at the isa level. |
1027 | |
1028 | return Changed; |
1029 | } |
1030 | |
1031 | bool SIGfx6CacheControl::enableRMWCacheBypass( |
1032 | const MachineBasicBlock::iterator &MI, |
1033 | SIAtomicScope Scope, |
1034 | SIAtomicAddrSpace AddrSpace) const { |
1035 | assert(MI->mayLoad() && MI->mayStore()); |
1036 | bool Changed = false; |
1037 | |
1038 | /// Do not set GLC for RMW atomic operations as L0/L1 cache is automatically |
1039 | /// bypassed, and the GLC bit is instead used to indicate if they are |
1040 | /// return or no-return. |
1041 | /// Note: there is no L2 cache coherent bypass control at the ISA level. |
1042 | |
1043 | return Changed; |
1044 | } |
1045 | |
1046 | bool SIGfx6CacheControl::enableVolatileAndOrNonTemporal( |
1047 | MachineBasicBlock::iterator &MI, SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
1048 | bool IsVolatile, bool IsNonTemporal, bool IsLastUse = false) const { |
1049 | // Only handle load and store, not atomic read-modify-write insructions. The |
1050 | // latter use glc to indicate if the atomic returns a result and so must not |
1051 | // be used for cache control. |
1052 | assert(MI->mayLoad() ^ MI->mayStore()); |
1053 | |
1054 | // Only update load and store, not LLVM IR atomic read-modify-write |
1055 | // instructions. The latter are always marked as volatile so cannot sensibly |
1056 | // handle it as do not want to pessimize all atomics. Also they do not support |
1057 | // the nontemporal attribute. |
1058 | assert(Op == SIMemOp::LOAD || Op == SIMemOp::STORE); |
1059 | |
1060 | bool Changed = false; |
1061 | |
1062 | if (IsVolatile) { |
1063 | // Set L1 cache policy to be MISS_EVICT for load instructions |
1064 | // and MISS_LRU for store instructions. |
1065 | // Note: there is no L2 cache bypass policy at the ISA level. |
1066 | if (Op == SIMemOp::LOAD) |
1067 | Changed |= enableGLCBit(MI); |
1068 | |
1069 | // Ensure operation has completed at system scope to cause all volatile |
1070 | // operations to be visible outside the program in a global order. Do not |
1071 | // request cross address space as only the global address space can be |
1072 | // observable outside the program, so no need to cause a waitcnt for LDS |
1073 | // address space operations. |
1074 | Changed |= insertWait(MI, Scope: SIAtomicScope::SYSTEM, AddrSpace, Op, IsCrossAddrSpaceOrdering: false, |
1075 | Pos: Position::AFTER); |
1076 | |
1077 | return Changed; |
1078 | } |
1079 | |
1080 | if (IsNonTemporal) { |
1081 | // Setting both GLC and SLC configures L1 cache policy to MISS_EVICT |
1082 | // for both loads and stores, and the L2 cache policy to STREAM. |
1083 | Changed |= enableGLCBit(MI); |
1084 | Changed |= enableSLCBit(MI); |
1085 | return Changed; |
1086 | } |
1087 | |
1088 | return Changed; |
1089 | } |
1090 | |
1091 | bool SIGfx6CacheControl::insertWait(MachineBasicBlock::iterator &MI, |
1092 | SIAtomicScope Scope, |
1093 | SIAtomicAddrSpace AddrSpace, |
1094 | SIMemOp Op, |
1095 | bool IsCrossAddrSpaceOrdering, |
1096 | Position Pos) const { |
1097 | bool Changed = false; |
1098 | |
1099 | MachineBasicBlock &MBB = *MI->getParent(); |
1100 | DebugLoc DL = MI->getDebugLoc(); |
1101 | |
1102 | if (Pos == Position::AFTER) |
1103 | ++MI; |
1104 | |
1105 | bool VMCnt = false; |
1106 | bool LGKMCnt = false; |
1107 | |
1108 | if ((AddrSpace & (SIAtomicAddrSpace::GLOBAL | SIAtomicAddrSpace::SCRATCH)) != |
1109 | SIAtomicAddrSpace::NONE) { |
1110 | switch (Scope) { |
1111 | case SIAtomicScope::SYSTEM: |
1112 | case SIAtomicScope::AGENT: |
1113 | VMCnt |= true; |
1114 | break; |
1115 | case SIAtomicScope::WORKGROUP: |
1116 | case SIAtomicScope::WAVEFRONT: |
1117 | case SIAtomicScope::SINGLETHREAD: |
1118 | // The L1 cache keeps all memory operations in order for |
1119 | // wavefronts in the same work-group. |
1120 | break; |
1121 | default: |
1122 | llvm_unreachable("Unsupported synchronization scope" ); |
1123 | } |
1124 | } |
1125 | |
1126 | if ((AddrSpace & SIAtomicAddrSpace::LDS) != SIAtomicAddrSpace::NONE) { |
1127 | switch (Scope) { |
1128 | case SIAtomicScope::SYSTEM: |
1129 | case SIAtomicScope::AGENT: |
1130 | case SIAtomicScope::WORKGROUP: |
1131 | // If no cross address space ordering then an "S_WAITCNT lgkmcnt(0)" is |
1132 | // not needed as LDS operations for all waves are executed in a total |
1133 | // global ordering as observed by all waves. Required if also |
1134 | // synchronizing with global/GDS memory as LDS operations could be |
1135 | // reordered with respect to later global/GDS memory operations of the |
1136 | // same wave. |
1137 | LGKMCnt |= IsCrossAddrSpaceOrdering; |
1138 | break; |
1139 | case SIAtomicScope::WAVEFRONT: |
1140 | case SIAtomicScope::SINGLETHREAD: |
1141 | // The LDS keeps all memory operations in order for |
1142 | // the same wavefront. |
1143 | break; |
1144 | default: |
1145 | llvm_unreachable("Unsupported synchronization scope" ); |
1146 | } |
1147 | } |
1148 | |
1149 | if ((AddrSpace & SIAtomicAddrSpace::GDS) != SIAtomicAddrSpace::NONE) { |
1150 | switch (Scope) { |
1151 | case SIAtomicScope::SYSTEM: |
1152 | case SIAtomicScope::AGENT: |
1153 | // If no cross address space ordering then an GDS "S_WAITCNT lgkmcnt(0)" |
1154 | // is not needed as GDS operations for all waves are executed in a total |
1155 | // global ordering as observed by all waves. Required if also |
1156 | // synchronizing with global/LDS memory as GDS operations could be |
1157 | // reordered with respect to later global/LDS memory operations of the |
1158 | // same wave. |
1159 | LGKMCnt |= IsCrossAddrSpaceOrdering; |
1160 | break; |
1161 | case SIAtomicScope::WORKGROUP: |
1162 | case SIAtomicScope::WAVEFRONT: |
1163 | case SIAtomicScope::SINGLETHREAD: |
1164 | // The GDS keeps all memory operations in order for |
1165 | // the same work-group. |
1166 | break; |
1167 | default: |
1168 | llvm_unreachable("Unsupported synchronization scope" ); |
1169 | } |
1170 | } |
1171 | |
1172 | if (VMCnt || LGKMCnt) { |
1173 | unsigned WaitCntImmediate = |
1174 | AMDGPU::encodeWaitcnt(Version: IV, |
1175 | Vmcnt: VMCnt ? 0 : getVmcntBitMask(Version: IV), |
1176 | Expcnt: getExpcntBitMask(Version: IV), |
1177 | Lgkmcnt: LGKMCnt ? 0 : getLgkmcntBitMask(Version: IV)); |
1178 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_WAITCNT_soft)) |
1179 | .addImm(Val: WaitCntImmediate); |
1180 | Changed = true; |
1181 | } |
1182 | |
1183 | if (Pos == Position::AFTER) |
1184 | --MI; |
1185 | |
1186 | return Changed; |
1187 | } |
1188 | |
1189 | bool SIGfx6CacheControl::insertAcquire(MachineBasicBlock::iterator &MI, |
1190 | SIAtomicScope Scope, |
1191 | SIAtomicAddrSpace AddrSpace, |
1192 | Position Pos) const { |
1193 | if (!InsertCacheInv) |
1194 | return false; |
1195 | |
1196 | bool Changed = false; |
1197 | |
1198 | MachineBasicBlock &MBB = *MI->getParent(); |
1199 | DebugLoc DL = MI->getDebugLoc(); |
1200 | |
1201 | if (Pos == Position::AFTER) |
1202 | ++MI; |
1203 | |
1204 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1205 | switch (Scope) { |
1206 | case SIAtomicScope::SYSTEM: |
1207 | case SIAtomicScope::AGENT: |
1208 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_WBINVL1)); |
1209 | Changed = true; |
1210 | break; |
1211 | case SIAtomicScope::WORKGROUP: |
1212 | case SIAtomicScope::WAVEFRONT: |
1213 | case SIAtomicScope::SINGLETHREAD: |
1214 | // No cache to invalidate. |
1215 | break; |
1216 | default: |
1217 | llvm_unreachable("Unsupported synchronization scope" ); |
1218 | } |
1219 | } |
1220 | |
1221 | /// The scratch address space does not need the global memory cache |
1222 | /// to be flushed as all memory operations by the same thread are |
1223 | /// sequentially consistent, and no other thread can access scratch |
1224 | /// memory. |
1225 | |
1226 | /// Other address spaces do not have a cache. |
1227 | |
1228 | if (Pos == Position::AFTER) |
1229 | --MI; |
1230 | |
1231 | return Changed; |
1232 | } |
1233 | |
1234 | bool SIGfx6CacheControl::insertRelease(MachineBasicBlock::iterator &MI, |
1235 | SIAtomicScope Scope, |
1236 | SIAtomicAddrSpace AddrSpace, |
1237 | bool IsCrossAddrSpaceOrdering, |
1238 | Position Pos) const { |
1239 | return insertWait(MI, Scope, AddrSpace, Op: SIMemOp::LOAD | SIMemOp::STORE, |
1240 | IsCrossAddrSpaceOrdering, Pos); |
1241 | } |
1242 | |
1243 | bool SIGfx7CacheControl::insertAcquire(MachineBasicBlock::iterator &MI, |
1244 | SIAtomicScope Scope, |
1245 | SIAtomicAddrSpace AddrSpace, |
1246 | Position Pos) const { |
1247 | if (!InsertCacheInv) |
1248 | return false; |
1249 | |
1250 | bool Changed = false; |
1251 | |
1252 | MachineBasicBlock &MBB = *MI->getParent(); |
1253 | DebugLoc DL = MI->getDebugLoc(); |
1254 | |
1255 | const GCNSubtarget &STM = MBB.getParent()->getSubtarget<GCNSubtarget>(); |
1256 | |
1257 | const unsigned InvalidateL1 = STM.isAmdPalOS() || STM.isMesa3DOS() |
1258 | ? AMDGPU::BUFFER_WBINVL1 |
1259 | : AMDGPU::BUFFER_WBINVL1_VOL; |
1260 | |
1261 | if (Pos == Position::AFTER) |
1262 | ++MI; |
1263 | |
1264 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1265 | switch (Scope) { |
1266 | case SIAtomicScope::SYSTEM: |
1267 | case SIAtomicScope::AGENT: |
1268 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: InvalidateL1)); |
1269 | Changed = true; |
1270 | break; |
1271 | case SIAtomicScope::WORKGROUP: |
1272 | case SIAtomicScope::WAVEFRONT: |
1273 | case SIAtomicScope::SINGLETHREAD: |
1274 | // No cache to invalidate. |
1275 | break; |
1276 | default: |
1277 | llvm_unreachable("Unsupported synchronization scope" ); |
1278 | } |
1279 | } |
1280 | |
1281 | /// The scratch address space does not need the global memory cache |
1282 | /// to be flushed as all memory operations by the same thread are |
1283 | /// sequentially consistent, and no other thread can access scratch |
1284 | /// memory. |
1285 | |
1286 | /// Other address spaces do not have a cache. |
1287 | |
1288 | if (Pos == Position::AFTER) |
1289 | --MI; |
1290 | |
1291 | return Changed; |
1292 | } |
1293 | |
1294 | bool SIGfx90ACacheControl::enableLoadCacheBypass( |
1295 | const MachineBasicBlock::iterator &MI, |
1296 | SIAtomicScope Scope, |
1297 | SIAtomicAddrSpace AddrSpace) const { |
1298 | assert(MI->mayLoad() && !MI->mayStore()); |
1299 | bool Changed = false; |
1300 | |
1301 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1302 | switch (Scope) { |
1303 | case SIAtomicScope::SYSTEM: |
1304 | case SIAtomicScope::AGENT: |
1305 | // Set the L1 cache policy to MISS_LRU. |
1306 | // Note: there is no L2 cache bypass policy at the ISA level. |
1307 | Changed |= enableGLCBit(MI); |
1308 | break; |
1309 | case SIAtomicScope::WORKGROUP: |
1310 | // In threadgroup split mode the waves of a work-group can be executing on |
1311 | // different CUs. Therefore need to bypass the L1 which is per CU. |
1312 | // Otherwise in non-threadgroup split mode all waves of a work-group are |
1313 | // on the same CU, and so the L1 does not need to be bypassed. |
1314 | if (ST.isTgSplitEnabled()) |
1315 | Changed |= enableGLCBit(MI); |
1316 | break; |
1317 | case SIAtomicScope::WAVEFRONT: |
1318 | case SIAtomicScope::SINGLETHREAD: |
1319 | // No cache to bypass. |
1320 | break; |
1321 | default: |
1322 | llvm_unreachable("Unsupported synchronization scope" ); |
1323 | } |
1324 | } |
1325 | |
1326 | /// The scratch address space does not need the global memory caches |
1327 | /// to be bypassed as all memory operations by the same thread are |
1328 | /// sequentially consistent, and no other thread can access scratch |
1329 | /// memory. |
1330 | |
1331 | /// Other address spaces do not have a cache. |
1332 | |
1333 | return Changed; |
1334 | } |
1335 | |
1336 | bool SIGfx90ACacheControl::enableStoreCacheBypass( |
1337 | const MachineBasicBlock::iterator &MI, |
1338 | SIAtomicScope Scope, |
1339 | SIAtomicAddrSpace AddrSpace) const { |
1340 | assert(!MI->mayLoad() && MI->mayStore()); |
1341 | bool Changed = false; |
1342 | |
1343 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1344 | switch (Scope) { |
1345 | case SIAtomicScope::SYSTEM: |
1346 | case SIAtomicScope::AGENT: |
1347 | /// Do not set glc for store atomic operations as they implicitly write |
1348 | /// through the L1 cache. |
1349 | break; |
1350 | case SIAtomicScope::WORKGROUP: |
1351 | case SIAtomicScope::WAVEFRONT: |
1352 | case SIAtomicScope::SINGLETHREAD: |
1353 | // No cache to bypass. Store atomics implicitly write through the L1 |
1354 | // cache. |
1355 | break; |
1356 | default: |
1357 | llvm_unreachable("Unsupported synchronization scope" ); |
1358 | } |
1359 | } |
1360 | |
1361 | /// The scratch address space does not need the global memory caches |
1362 | /// to be bypassed as all memory operations by the same thread are |
1363 | /// sequentially consistent, and no other thread can access scratch |
1364 | /// memory. |
1365 | |
1366 | /// Other address spaces do not have a cache. |
1367 | |
1368 | return Changed; |
1369 | } |
1370 | |
1371 | bool SIGfx90ACacheControl::enableRMWCacheBypass( |
1372 | const MachineBasicBlock::iterator &MI, |
1373 | SIAtomicScope Scope, |
1374 | SIAtomicAddrSpace AddrSpace) const { |
1375 | assert(MI->mayLoad() && MI->mayStore()); |
1376 | bool Changed = false; |
1377 | |
1378 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1379 | switch (Scope) { |
1380 | case SIAtomicScope::SYSTEM: |
1381 | case SIAtomicScope::AGENT: |
1382 | /// Do not set glc for RMW atomic operations as they implicitly bypass |
1383 | /// the L1 cache, and the glc bit is instead used to indicate if they are |
1384 | /// return or no-return. |
1385 | break; |
1386 | case SIAtomicScope::WORKGROUP: |
1387 | case SIAtomicScope::WAVEFRONT: |
1388 | case SIAtomicScope::SINGLETHREAD: |
1389 | // No cache to bypass. RMW atomics implicitly bypass the L1 cache. |
1390 | break; |
1391 | default: |
1392 | llvm_unreachable("Unsupported synchronization scope" ); |
1393 | } |
1394 | } |
1395 | |
1396 | return Changed; |
1397 | } |
1398 | |
1399 | bool SIGfx90ACacheControl::enableVolatileAndOrNonTemporal( |
1400 | MachineBasicBlock::iterator &MI, SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
1401 | bool IsVolatile, bool IsNonTemporal, bool IsLastUse = false) const { |
1402 | // Only handle load and store, not atomic read-modify-write insructions. The |
1403 | // latter use glc to indicate if the atomic returns a result and so must not |
1404 | // be used for cache control. |
1405 | assert(MI->mayLoad() ^ MI->mayStore()); |
1406 | |
1407 | // Only update load and store, not LLVM IR atomic read-modify-write |
1408 | // instructions. The latter are always marked as volatile so cannot sensibly |
1409 | // handle it as do not want to pessimize all atomics. Also they do not support |
1410 | // the nontemporal attribute. |
1411 | assert(Op == SIMemOp::LOAD || Op == SIMemOp::STORE); |
1412 | |
1413 | bool Changed = false; |
1414 | |
1415 | if (IsVolatile) { |
1416 | // Set L1 cache policy to be MISS_EVICT for load instructions |
1417 | // and MISS_LRU for store instructions. |
1418 | // Note: there is no L2 cache bypass policy at the ISA level. |
1419 | if (Op == SIMemOp::LOAD) |
1420 | Changed |= enableGLCBit(MI); |
1421 | |
1422 | // Ensure operation has completed at system scope to cause all volatile |
1423 | // operations to be visible outside the program in a global order. Do not |
1424 | // request cross address space as only the global address space can be |
1425 | // observable outside the program, so no need to cause a waitcnt for LDS |
1426 | // address space operations. |
1427 | Changed |= insertWait(MI, Scope: SIAtomicScope::SYSTEM, AddrSpace, Op, IsCrossAddrSpaceOrdering: false, |
1428 | Pos: Position::AFTER); |
1429 | |
1430 | return Changed; |
1431 | } |
1432 | |
1433 | if (IsNonTemporal) { |
1434 | // Setting both GLC and SLC configures L1 cache policy to MISS_EVICT |
1435 | // for both loads and stores, and the L2 cache policy to STREAM. |
1436 | Changed |= enableGLCBit(MI); |
1437 | Changed |= enableSLCBit(MI); |
1438 | return Changed; |
1439 | } |
1440 | |
1441 | return Changed; |
1442 | } |
1443 | |
1444 | bool SIGfx90ACacheControl::insertWait(MachineBasicBlock::iterator &MI, |
1445 | SIAtomicScope Scope, |
1446 | SIAtomicAddrSpace AddrSpace, |
1447 | SIMemOp Op, |
1448 | bool IsCrossAddrSpaceOrdering, |
1449 | Position Pos) const { |
1450 | if (ST.isTgSplitEnabled()) { |
1451 | // In threadgroup split mode the waves of a work-group can be executing on |
1452 | // different CUs. Therefore need to wait for global or GDS memory operations |
1453 | // to complete to ensure they are visible to waves in the other CUs. |
1454 | // Otherwise in non-threadgroup split mode all waves of a work-group are on |
1455 | // the same CU, so no need to wait for global memory as all waves in the |
1456 | // work-group access the same the L1, nor wait for GDS as access are ordered |
1457 | // on a CU. |
1458 | if (((AddrSpace & (SIAtomicAddrSpace::GLOBAL | SIAtomicAddrSpace::SCRATCH | |
1459 | SIAtomicAddrSpace::GDS)) != SIAtomicAddrSpace::NONE) && |
1460 | (Scope == SIAtomicScope::WORKGROUP)) { |
1461 | // Same as GFX7 using agent scope. |
1462 | Scope = SIAtomicScope::AGENT; |
1463 | } |
1464 | // In threadgroup split mode LDS cannot be allocated so no need to wait for |
1465 | // LDS memory operations. |
1466 | AddrSpace &= ~SIAtomicAddrSpace::LDS; |
1467 | } |
1468 | return SIGfx7CacheControl::insertWait(MI, Scope, AddrSpace, Op, |
1469 | IsCrossAddrSpaceOrdering, Pos); |
1470 | } |
1471 | |
1472 | bool SIGfx90ACacheControl::insertAcquire(MachineBasicBlock::iterator &MI, |
1473 | SIAtomicScope Scope, |
1474 | SIAtomicAddrSpace AddrSpace, |
1475 | Position Pos) const { |
1476 | if (!InsertCacheInv) |
1477 | return false; |
1478 | |
1479 | bool Changed = false; |
1480 | |
1481 | MachineBasicBlock &MBB = *MI->getParent(); |
1482 | DebugLoc DL = MI->getDebugLoc(); |
1483 | |
1484 | if (Pos == Position::AFTER) |
1485 | ++MI; |
1486 | |
1487 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1488 | switch (Scope) { |
1489 | case SIAtomicScope::SYSTEM: |
1490 | // Ensures that following loads will not see stale remote VMEM data or |
1491 | // stale local VMEM data with MTYPE NC. Local VMEM data with MTYPE RW and |
1492 | // CC will never be stale due to the local memory probes. |
1493 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_INVL2)); |
1494 | // Inserting a "S_WAITCNT vmcnt(0)" after is not required because the |
1495 | // hardware does not reorder memory operations by the same wave with |
1496 | // respect to a preceding "BUFFER_INVL2". The invalidate is guaranteed to |
1497 | // remove any cache lines of earlier writes by the same wave and ensures |
1498 | // later reads by the same wave will refetch the cache lines. |
1499 | Changed = true; |
1500 | break; |
1501 | case SIAtomicScope::AGENT: |
1502 | // Same as GFX7. |
1503 | break; |
1504 | case SIAtomicScope::WORKGROUP: |
1505 | // In threadgroup split mode the waves of a work-group can be executing on |
1506 | // different CUs. Therefore need to invalidate the L1 which is per CU. |
1507 | // Otherwise in non-threadgroup split mode all waves of a work-group are |
1508 | // on the same CU, and so the L1 does not need to be invalidated. |
1509 | if (ST.isTgSplitEnabled()) { |
1510 | // Same as GFX7 using agent scope. |
1511 | Scope = SIAtomicScope::AGENT; |
1512 | } |
1513 | break; |
1514 | case SIAtomicScope::WAVEFRONT: |
1515 | case SIAtomicScope::SINGLETHREAD: |
1516 | // Same as GFX7. |
1517 | break; |
1518 | default: |
1519 | llvm_unreachable("Unsupported synchronization scope" ); |
1520 | } |
1521 | } |
1522 | |
1523 | /// The scratch address space does not need the global memory cache |
1524 | /// to be flushed as all memory operations by the same thread are |
1525 | /// sequentially consistent, and no other thread can access scratch |
1526 | /// memory. |
1527 | |
1528 | /// Other address spaces do not have a cache. |
1529 | |
1530 | if (Pos == Position::AFTER) |
1531 | --MI; |
1532 | |
1533 | Changed |= SIGfx7CacheControl::insertAcquire(MI, Scope, AddrSpace, Pos); |
1534 | |
1535 | return Changed; |
1536 | } |
1537 | |
1538 | bool SIGfx90ACacheControl::insertRelease(MachineBasicBlock::iterator &MI, |
1539 | SIAtomicScope Scope, |
1540 | SIAtomicAddrSpace AddrSpace, |
1541 | bool IsCrossAddrSpaceOrdering, |
1542 | Position Pos) const { |
1543 | bool Changed = false; |
1544 | |
1545 | MachineBasicBlock &MBB = *MI->getParent(); |
1546 | const DebugLoc &DL = MI->getDebugLoc(); |
1547 | |
1548 | if (Pos == Position::AFTER) |
1549 | ++MI; |
1550 | |
1551 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1552 | switch (Scope) { |
1553 | case SIAtomicScope::SYSTEM: |
1554 | // Inserting a "S_WAITCNT vmcnt(0)" before is not required because the |
1555 | // hardware does not reorder memory operations by the same wave with |
1556 | // respect to a following "BUFFER_WBL2". The "BUFFER_WBL2" is guaranteed |
1557 | // to initiate writeback of any dirty cache lines of earlier writes by the |
1558 | // same wave. A "S_WAITCNT vmcnt(0)" is needed after to ensure the |
1559 | // writeback has completed. |
1560 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_WBL2)) |
1561 | // Set SC bits to indicate system scope. |
1562 | .addImm(Val: AMDGPU::CPol::SC0 | AMDGPU::CPol::SC1); |
1563 | // Followed by same as GFX7, which will ensure the necessary "S_WAITCNT |
1564 | // vmcnt(0)" needed by the "BUFFER_WBL2". |
1565 | Changed = true; |
1566 | break; |
1567 | case SIAtomicScope::AGENT: |
1568 | case SIAtomicScope::WORKGROUP: |
1569 | case SIAtomicScope::WAVEFRONT: |
1570 | case SIAtomicScope::SINGLETHREAD: |
1571 | // Same as GFX7. |
1572 | break; |
1573 | default: |
1574 | llvm_unreachable("Unsupported synchronization scope" ); |
1575 | } |
1576 | } |
1577 | |
1578 | if (Pos == Position::AFTER) |
1579 | --MI; |
1580 | |
1581 | Changed |= |
1582 | SIGfx7CacheControl::insertRelease(MI, Scope, AddrSpace, |
1583 | IsCrossAddrSpaceOrdering, Pos); |
1584 | |
1585 | return Changed; |
1586 | } |
1587 | |
1588 | bool SIGfx940CacheControl::enableLoadCacheBypass( |
1589 | const MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
1590 | SIAtomicAddrSpace AddrSpace) const { |
1591 | assert(MI->mayLoad() && !MI->mayStore()); |
1592 | bool Changed = false; |
1593 | |
1594 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1595 | switch (Scope) { |
1596 | case SIAtomicScope::SYSTEM: |
1597 | // Set SC bits to indicate system scope. |
1598 | Changed |= enableSC0Bit(MI); |
1599 | Changed |= enableSC1Bit(MI); |
1600 | break; |
1601 | case SIAtomicScope::AGENT: |
1602 | // Set SC bits to indicate agent scope. |
1603 | Changed |= enableSC1Bit(MI); |
1604 | break; |
1605 | case SIAtomicScope::WORKGROUP: |
1606 | // In threadgroup split mode the waves of a work-group can be executing on |
1607 | // different CUs. Therefore need to bypass the L1 which is per CU. |
1608 | // Otherwise in non-threadgroup split mode all waves of a work-group are |
1609 | // on the same CU, and so the L1 does not need to be bypassed. Setting SC |
1610 | // bits to indicate work-group scope will do this automatically. |
1611 | Changed |= enableSC0Bit(MI); |
1612 | break; |
1613 | case SIAtomicScope::WAVEFRONT: |
1614 | case SIAtomicScope::SINGLETHREAD: |
1615 | // Leave SC bits unset to indicate wavefront scope. |
1616 | break; |
1617 | default: |
1618 | llvm_unreachable("Unsupported synchronization scope" ); |
1619 | } |
1620 | } |
1621 | |
1622 | /// The scratch address space does not need the global memory caches |
1623 | /// to be bypassed as all memory operations by the same thread are |
1624 | /// sequentially consistent, and no other thread can access scratch |
1625 | /// memory. |
1626 | |
1627 | /// Other address spaces do not have a cache. |
1628 | |
1629 | return Changed; |
1630 | } |
1631 | |
1632 | bool SIGfx940CacheControl::enableStoreCacheBypass( |
1633 | const MachineBasicBlock::iterator &MI, |
1634 | SIAtomicScope Scope, SIAtomicAddrSpace AddrSpace) const { |
1635 | assert(!MI->mayLoad() && MI->mayStore()); |
1636 | bool Changed = false; |
1637 | |
1638 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1639 | switch (Scope) { |
1640 | case SIAtomicScope::SYSTEM: |
1641 | // Set SC bits to indicate system scope. |
1642 | Changed |= enableSC0Bit(MI); |
1643 | Changed |= enableSC1Bit(MI); |
1644 | break; |
1645 | case SIAtomicScope::AGENT: |
1646 | // Set SC bits to indicate agent scope. |
1647 | Changed |= enableSC1Bit(MI); |
1648 | break; |
1649 | case SIAtomicScope::WORKGROUP: |
1650 | // Set SC bits to indicate workgroup scope. |
1651 | Changed |= enableSC0Bit(MI); |
1652 | break; |
1653 | case SIAtomicScope::WAVEFRONT: |
1654 | case SIAtomicScope::SINGLETHREAD: |
1655 | // Leave SC bits unset to indicate wavefront scope. |
1656 | break; |
1657 | default: |
1658 | llvm_unreachable("Unsupported synchronization scope" ); |
1659 | } |
1660 | } |
1661 | |
1662 | /// The scratch address space does not need the global memory caches |
1663 | /// to be bypassed as all memory operations by the same thread are |
1664 | /// sequentially consistent, and no other thread can access scratch |
1665 | /// memory. |
1666 | |
1667 | /// Other address spaces do not have a cache. |
1668 | |
1669 | return Changed; |
1670 | } |
1671 | |
1672 | bool SIGfx940CacheControl::enableRMWCacheBypass( |
1673 | const MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
1674 | SIAtomicAddrSpace AddrSpace) const { |
1675 | assert(MI->mayLoad() && MI->mayStore()); |
1676 | bool Changed = false; |
1677 | |
1678 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1679 | switch (Scope) { |
1680 | case SIAtomicScope::SYSTEM: |
1681 | // Set SC1 bit to indicate system scope. |
1682 | Changed |= enableSC1Bit(MI); |
1683 | break; |
1684 | case SIAtomicScope::AGENT: |
1685 | case SIAtomicScope::WORKGROUP: |
1686 | case SIAtomicScope::WAVEFRONT: |
1687 | case SIAtomicScope::SINGLETHREAD: |
1688 | // RMW atomic operations implicitly bypass the L1 cache and only use SC1 |
1689 | // to indicate system or agent scope. The SC0 bit is used to indicate if |
1690 | // they are return or no-return. Leave SC1 bit unset to indicate agent |
1691 | // scope. |
1692 | break; |
1693 | default: |
1694 | llvm_unreachable("Unsupported synchronization scope" ); |
1695 | } |
1696 | } |
1697 | |
1698 | return Changed; |
1699 | } |
1700 | |
1701 | bool SIGfx940CacheControl::enableVolatileAndOrNonTemporal( |
1702 | MachineBasicBlock::iterator &MI, SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
1703 | bool IsVolatile, bool IsNonTemporal, bool IsLastUse = false) const { |
1704 | // Only handle load and store, not atomic read-modify-write insructions. The |
1705 | // latter use glc to indicate if the atomic returns a result and so must not |
1706 | // be used for cache control. |
1707 | assert(MI->mayLoad() ^ MI->mayStore()); |
1708 | |
1709 | // Only update load and store, not LLVM IR atomic read-modify-write |
1710 | // instructions. The latter are always marked as volatile so cannot sensibly |
1711 | // handle it as do not want to pessimize all atomics. Also they do not support |
1712 | // the nontemporal attribute. |
1713 | assert(Op == SIMemOp::LOAD || Op == SIMemOp::STORE); |
1714 | |
1715 | bool Changed = false; |
1716 | |
1717 | if (IsVolatile) { |
1718 | // Set SC bits to indicate system scope. |
1719 | Changed |= enableSC0Bit(MI); |
1720 | Changed |= enableSC1Bit(MI); |
1721 | |
1722 | // Ensure operation has completed at system scope to cause all volatile |
1723 | // operations to be visible outside the program in a global order. Do not |
1724 | // request cross address space as only the global address space can be |
1725 | // observable outside the program, so no need to cause a waitcnt for LDS |
1726 | // address space operations. |
1727 | Changed |= insertWait(MI, Scope: SIAtomicScope::SYSTEM, AddrSpace, Op, IsCrossAddrSpaceOrdering: false, |
1728 | Pos: Position::AFTER); |
1729 | |
1730 | return Changed; |
1731 | } |
1732 | |
1733 | if (IsNonTemporal) { |
1734 | Changed |= enableNTBit(MI); |
1735 | return Changed; |
1736 | } |
1737 | |
1738 | return Changed; |
1739 | } |
1740 | |
1741 | bool SIGfx940CacheControl::insertAcquire(MachineBasicBlock::iterator &MI, |
1742 | SIAtomicScope Scope, |
1743 | SIAtomicAddrSpace AddrSpace, |
1744 | Position Pos) const { |
1745 | if (!InsertCacheInv) |
1746 | return false; |
1747 | |
1748 | bool Changed = false; |
1749 | |
1750 | MachineBasicBlock &MBB = *MI->getParent(); |
1751 | DebugLoc DL = MI->getDebugLoc(); |
1752 | |
1753 | if (Pos == Position::AFTER) |
1754 | ++MI; |
1755 | |
1756 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1757 | switch (Scope) { |
1758 | case SIAtomicScope::SYSTEM: |
1759 | // Ensures that following loads will not see stale remote VMEM data or |
1760 | // stale local VMEM data with MTYPE NC. Local VMEM data with MTYPE RW and |
1761 | // CC will never be stale due to the local memory probes. |
1762 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_INV)) |
1763 | // Set SC bits to indicate system scope. |
1764 | .addImm(Val: AMDGPU::CPol::SC0 | AMDGPU::CPol::SC1); |
1765 | // Inserting a "S_WAITCNT vmcnt(0)" after is not required because the |
1766 | // hardware does not reorder memory operations by the same wave with |
1767 | // respect to a preceding "BUFFER_INV". The invalidate is guaranteed to |
1768 | // remove any cache lines of earlier writes by the same wave and ensures |
1769 | // later reads by the same wave will refetch the cache lines. |
1770 | Changed = true; |
1771 | break; |
1772 | case SIAtomicScope::AGENT: |
1773 | // Ensures that following loads will not see stale remote date or local |
1774 | // MTYPE NC global data. Local MTYPE RW and CC memory will never be stale |
1775 | // due to the memory probes. |
1776 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_INV)) |
1777 | // Set SC bits to indicate agent scope. |
1778 | .addImm(Val: AMDGPU::CPol::SC1); |
1779 | // Inserting "S_WAITCNT vmcnt(0)" is not required because the hardware |
1780 | // does not reorder memory operations with respect to preceeding buffer |
1781 | // invalidate. The invalidate is guaranteed to remove any cache lines of |
1782 | // earlier writes and ensures later writes will refetch the cache lines. |
1783 | Changed = true; |
1784 | break; |
1785 | case SIAtomicScope::WORKGROUP: |
1786 | // In threadgroup split mode the waves of a work-group can be executing on |
1787 | // different CUs. Therefore need to invalidate the L1 which is per CU. |
1788 | // Otherwise in non-threadgroup split mode all waves of a work-group are |
1789 | // on the same CU, and so the L1 does not need to be invalidated. |
1790 | if (ST.isTgSplitEnabled()) { |
1791 | // Ensures L1 is invalidated if in threadgroup split mode. In |
1792 | // non-threadgroup split mode it is a NOP, but no point generating it in |
1793 | // that case if know not in that mode. |
1794 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_INV)) |
1795 | // Set SC bits to indicate work-group scope. |
1796 | .addImm(Val: AMDGPU::CPol::SC0); |
1797 | // Inserting "S_WAITCNT vmcnt(0)" is not required because the hardware |
1798 | // does not reorder memory operations with respect to preceeding buffer |
1799 | // invalidate. The invalidate is guaranteed to remove any cache lines of |
1800 | // earlier writes and ensures later writes will refetch the cache lines. |
1801 | Changed = true; |
1802 | } |
1803 | break; |
1804 | case SIAtomicScope::WAVEFRONT: |
1805 | case SIAtomicScope::SINGLETHREAD: |
1806 | // Could generate "BUFFER_INV" but it would do nothing as there are no |
1807 | // caches to invalidate. |
1808 | break; |
1809 | default: |
1810 | llvm_unreachable("Unsupported synchronization scope" ); |
1811 | } |
1812 | } |
1813 | |
1814 | /// The scratch address space does not need the global memory cache |
1815 | /// to be flushed as all memory operations by the same thread are |
1816 | /// sequentially consistent, and no other thread can access scratch |
1817 | /// memory. |
1818 | |
1819 | /// Other address spaces do not have a cache. |
1820 | |
1821 | if (Pos == Position::AFTER) |
1822 | --MI; |
1823 | |
1824 | return Changed; |
1825 | } |
1826 | |
1827 | bool SIGfx940CacheControl::insertRelease(MachineBasicBlock::iterator &MI, |
1828 | SIAtomicScope Scope, |
1829 | SIAtomicAddrSpace AddrSpace, |
1830 | bool IsCrossAddrSpaceOrdering, |
1831 | Position Pos) const { |
1832 | bool Changed = false; |
1833 | |
1834 | MachineBasicBlock &MBB = *MI->getParent(); |
1835 | DebugLoc DL = MI->getDebugLoc(); |
1836 | |
1837 | if (Pos == Position::AFTER) |
1838 | ++MI; |
1839 | |
1840 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1841 | switch (Scope) { |
1842 | case SIAtomicScope::SYSTEM: |
1843 | // Inserting a "S_WAITCNT vmcnt(0)" before is not required because the |
1844 | // hardware does not reorder memory operations by the same wave with |
1845 | // respect to a following "BUFFER_WBL2". The "BUFFER_WBL2" is guaranteed |
1846 | // to initiate writeback of any dirty cache lines of earlier writes by the |
1847 | // same wave. A "S_WAITCNT vmcnt(0)" is needed after to ensure the |
1848 | // writeback has completed. |
1849 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_WBL2)) |
1850 | // Set SC bits to indicate system scope. |
1851 | .addImm(Val: AMDGPU::CPol::SC0 | AMDGPU::CPol::SC1); |
1852 | // Since AddrSpace contains SIAtomicAddrSpace::GLOBAL and Scope is |
1853 | // SIAtomicScope::SYSTEM, the following insertWait will generate the |
1854 | // required "S_WAITCNT vmcnt(0)" needed by the "BUFFER_WBL2". |
1855 | Changed = true; |
1856 | break; |
1857 | case SIAtomicScope::AGENT: |
1858 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_WBL2)) |
1859 | // Set SC bits to indicate agent scope. |
1860 | .addImm(Val: AMDGPU::CPol::SC1); |
1861 | |
1862 | // Since AddrSpace contains SIAtomicAddrSpace::GLOBAL and Scope is |
1863 | // SIAtomicScope::AGENT, the following insertWait will generate the |
1864 | // required "S_WAITCNT vmcnt(0)". |
1865 | Changed = true; |
1866 | break; |
1867 | case SIAtomicScope::WORKGROUP: |
1868 | case SIAtomicScope::WAVEFRONT: |
1869 | case SIAtomicScope::SINGLETHREAD: |
1870 | // Do not generate "BUFFER_WBL2" as there are no caches it would |
1871 | // writeback, and would require an otherwise unnecessary |
1872 | // "S_WAITCNT vmcnt(0)". |
1873 | break; |
1874 | default: |
1875 | llvm_unreachable("Unsupported synchronization scope" ); |
1876 | } |
1877 | } |
1878 | |
1879 | if (Pos == Position::AFTER) |
1880 | --MI; |
1881 | |
1882 | // Ensure the necessary S_WAITCNT needed by any "BUFFER_WBL2" as well as other |
1883 | // S_WAITCNT needed. |
1884 | Changed |= insertWait(MI, Scope, AddrSpace, Op: SIMemOp::LOAD | SIMemOp::STORE, |
1885 | IsCrossAddrSpaceOrdering, Pos); |
1886 | |
1887 | return Changed; |
1888 | } |
1889 | |
1890 | bool SIGfx10CacheControl::enableLoadCacheBypass( |
1891 | const MachineBasicBlock::iterator &MI, |
1892 | SIAtomicScope Scope, |
1893 | SIAtomicAddrSpace AddrSpace) const { |
1894 | assert(MI->mayLoad() && !MI->mayStore()); |
1895 | bool Changed = false; |
1896 | |
1897 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
1898 | switch (Scope) { |
1899 | case SIAtomicScope::SYSTEM: |
1900 | case SIAtomicScope::AGENT: |
1901 | // Set the L0 and L1 cache policies to MISS_EVICT. |
1902 | // Note: there is no L2 cache coherent bypass control at the ISA level. |
1903 | Changed |= enableGLCBit(MI); |
1904 | Changed |= enableDLCBit(MI); |
1905 | break; |
1906 | case SIAtomicScope::WORKGROUP: |
1907 | // In WGP mode the waves of a work-group can be executing on either CU of |
1908 | // the WGP. Therefore need to bypass the L0 which is per CU. Otherwise in |
1909 | // CU mode all waves of a work-group are on the same CU, and so the L0 |
1910 | // does not need to be bypassed. |
1911 | if (!ST.isCuModeEnabled()) |
1912 | Changed |= enableGLCBit(MI); |
1913 | break; |
1914 | case SIAtomicScope::WAVEFRONT: |
1915 | case SIAtomicScope::SINGLETHREAD: |
1916 | // No cache to bypass. |
1917 | break; |
1918 | default: |
1919 | llvm_unreachable("Unsupported synchronization scope" ); |
1920 | } |
1921 | } |
1922 | |
1923 | /// The scratch address space does not need the global memory caches |
1924 | /// to be bypassed as all memory operations by the same thread are |
1925 | /// sequentially consistent, and no other thread can access scratch |
1926 | /// memory. |
1927 | |
1928 | /// Other address spaces do not have a cache. |
1929 | |
1930 | return Changed; |
1931 | } |
1932 | |
1933 | bool SIGfx10CacheControl::enableVolatileAndOrNonTemporal( |
1934 | MachineBasicBlock::iterator &MI, SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
1935 | bool IsVolatile, bool IsNonTemporal, bool IsLastUse = false) const { |
1936 | |
1937 | // Only handle load and store, not atomic read-modify-write insructions. The |
1938 | // latter use glc to indicate if the atomic returns a result and so must not |
1939 | // be used for cache control. |
1940 | assert(MI->mayLoad() ^ MI->mayStore()); |
1941 | |
1942 | // Only update load and store, not LLVM IR atomic read-modify-write |
1943 | // instructions. The latter are always marked as volatile so cannot sensibly |
1944 | // handle it as do not want to pessimize all atomics. Also they do not support |
1945 | // the nontemporal attribute. |
1946 | assert(Op == SIMemOp::LOAD || Op == SIMemOp::STORE); |
1947 | |
1948 | bool Changed = false; |
1949 | |
1950 | if (IsVolatile) { |
1951 | // Set L0 and L1 cache policy to be MISS_EVICT for load instructions |
1952 | // and MISS_LRU for store instructions. |
1953 | // Note: there is no L2 cache coherent bypass control at the ISA level. |
1954 | if (Op == SIMemOp::LOAD) { |
1955 | Changed |= enableGLCBit(MI); |
1956 | Changed |= enableDLCBit(MI); |
1957 | } |
1958 | |
1959 | // Ensure operation has completed at system scope to cause all volatile |
1960 | // operations to be visible outside the program in a global order. Do not |
1961 | // request cross address space as only the global address space can be |
1962 | // observable outside the program, so no need to cause a waitcnt for LDS |
1963 | // address space operations. |
1964 | Changed |= insertWait(MI, Scope: SIAtomicScope::SYSTEM, AddrSpace, Op, IsCrossAddrSpaceOrdering: false, |
1965 | Pos: Position::AFTER); |
1966 | return Changed; |
1967 | } |
1968 | |
1969 | if (IsNonTemporal) { |
1970 | // For loads setting SLC configures L0 and L1 cache policy to HIT_EVICT |
1971 | // and L2 cache policy to STREAM. |
1972 | // For stores setting both GLC and SLC configures L0 and L1 cache policy |
1973 | // to MISS_EVICT and the L2 cache policy to STREAM. |
1974 | if (Op == SIMemOp::STORE) |
1975 | Changed |= enableGLCBit(MI); |
1976 | Changed |= enableSLCBit(MI); |
1977 | |
1978 | return Changed; |
1979 | } |
1980 | |
1981 | return Changed; |
1982 | } |
1983 | |
1984 | bool SIGfx10CacheControl::insertWait(MachineBasicBlock::iterator &MI, |
1985 | SIAtomicScope Scope, |
1986 | SIAtomicAddrSpace AddrSpace, |
1987 | SIMemOp Op, |
1988 | bool IsCrossAddrSpaceOrdering, |
1989 | Position Pos) const { |
1990 | bool Changed = false; |
1991 | |
1992 | MachineBasicBlock &MBB = *MI->getParent(); |
1993 | DebugLoc DL = MI->getDebugLoc(); |
1994 | |
1995 | if (Pos == Position::AFTER) |
1996 | ++MI; |
1997 | |
1998 | bool VMCnt = false; |
1999 | bool VSCnt = false; |
2000 | bool LGKMCnt = false; |
2001 | |
2002 | if ((AddrSpace & (SIAtomicAddrSpace::GLOBAL | SIAtomicAddrSpace::SCRATCH)) != |
2003 | SIAtomicAddrSpace::NONE) { |
2004 | switch (Scope) { |
2005 | case SIAtomicScope::SYSTEM: |
2006 | case SIAtomicScope::AGENT: |
2007 | if ((Op & SIMemOp::LOAD) != SIMemOp::NONE) |
2008 | VMCnt |= true; |
2009 | if ((Op & SIMemOp::STORE) != SIMemOp::NONE) |
2010 | VSCnt |= true; |
2011 | break; |
2012 | case SIAtomicScope::WORKGROUP: |
2013 | // In WGP mode the waves of a work-group can be executing on either CU of |
2014 | // the WGP. Therefore need to wait for operations to complete to ensure |
2015 | // they are visible to waves in the other CU as the L0 is per CU. |
2016 | // Otherwise in CU mode and all waves of a work-group are on the same CU |
2017 | // which shares the same L0. |
2018 | if (!ST.isCuModeEnabled()) { |
2019 | if ((Op & SIMemOp::LOAD) != SIMemOp::NONE) |
2020 | VMCnt |= true; |
2021 | if ((Op & SIMemOp::STORE) != SIMemOp::NONE) |
2022 | VSCnt |= true; |
2023 | } |
2024 | break; |
2025 | case SIAtomicScope::WAVEFRONT: |
2026 | case SIAtomicScope::SINGLETHREAD: |
2027 | // The L0 cache keeps all memory operations in order for |
2028 | // work-items in the same wavefront. |
2029 | break; |
2030 | default: |
2031 | llvm_unreachable("Unsupported synchronization scope" ); |
2032 | } |
2033 | } |
2034 | |
2035 | if ((AddrSpace & SIAtomicAddrSpace::LDS) != SIAtomicAddrSpace::NONE) { |
2036 | switch (Scope) { |
2037 | case SIAtomicScope::SYSTEM: |
2038 | case SIAtomicScope::AGENT: |
2039 | case SIAtomicScope::WORKGROUP: |
2040 | // If no cross address space ordering then an "S_WAITCNT lgkmcnt(0)" is |
2041 | // not needed as LDS operations for all waves are executed in a total |
2042 | // global ordering as observed by all waves. Required if also |
2043 | // synchronizing with global/GDS memory as LDS operations could be |
2044 | // reordered with respect to later global/GDS memory operations of the |
2045 | // same wave. |
2046 | LGKMCnt |= IsCrossAddrSpaceOrdering; |
2047 | break; |
2048 | case SIAtomicScope::WAVEFRONT: |
2049 | case SIAtomicScope::SINGLETHREAD: |
2050 | // The LDS keeps all memory operations in order for |
2051 | // the same wavefront. |
2052 | break; |
2053 | default: |
2054 | llvm_unreachable("Unsupported synchronization scope" ); |
2055 | } |
2056 | } |
2057 | |
2058 | if ((AddrSpace & SIAtomicAddrSpace::GDS) != SIAtomicAddrSpace::NONE) { |
2059 | switch (Scope) { |
2060 | case SIAtomicScope::SYSTEM: |
2061 | case SIAtomicScope::AGENT: |
2062 | // If no cross address space ordering then an GDS "S_WAITCNT lgkmcnt(0)" |
2063 | // is not needed as GDS operations for all waves are executed in a total |
2064 | // global ordering as observed by all waves. Required if also |
2065 | // synchronizing with global/LDS memory as GDS operations could be |
2066 | // reordered with respect to later global/LDS memory operations of the |
2067 | // same wave. |
2068 | LGKMCnt |= IsCrossAddrSpaceOrdering; |
2069 | break; |
2070 | case SIAtomicScope::WORKGROUP: |
2071 | case SIAtomicScope::WAVEFRONT: |
2072 | case SIAtomicScope::SINGLETHREAD: |
2073 | // The GDS keeps all memory operations in order for |
2074 | // the same work-group. |
2075 | break; |
2076 | default: |
2077 | llvm_unreachable("Unsupported synchronization scope" ); |
2078 | } |
2079 | } |
2080 | |
2081 | if (VMCnt || LGKMCnt) { |
2082 | unsigned WaitCntImmediate = |
2083 | AMDGPU::encodeWaitcnt(Version: IV, |
2084 | Vmcnt: VMCnt ? 0 : getVmcntBitMask(Version: IV), |
2085 | Expcnt: getExpcntBitMask(Version: IV), |
2086 | Lgkmcnt: LGKMCnt ? 0 : getLgkmcntBitMask(Version: IV)); |
2087 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_WAITCNT_soft)) |
2088 | .addImm(Val: WaitCntImmediate); |
2089 | Changed = true; |
2090 | } |
2091 | |
2092 | if (VSCnt) { |
2093 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_WAITCNT_VSCNT_soft)) |
2094 | .addReg(RegNo: AMDGPU::SGPR_NULL, flags: RegState::Undef) |
2095 | .addImm(Val: 0); |
2096 | Changed = true; |
2097 | } |
2098 | |
2099 | if (Pos == Position::AFTER) |
2100 | --MI; |
2101 | |
2102 | return Changed; |
2103 | } |
2104 | |
2105 | bool SIGfx10CacheControl::insertAcquire(MachineBasicBlock::iterator &MI, |
2106 | SIAtomicScope Scope, |
2107 | SIAtomicAddrSpace AddrSpace, |
2108 | Position Pos) const { |
2109 | if (!InsertCacheInv) |
2110 | return false; |
2111 | |
2112 | bool Changed = false; |
2113 | |
2114 | MachineBasicBlock &MBB = *MI->getParent(); |
2115 | DebugLoc DL = MI->getDebugLoc(); |
2116 | |
2117 | if (Pos == Position::AFTER) |
2118 | ++MI; |
2119 | |
2120 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
2121 | switch (Scope) { |
2122 | case SIAtomicScope::SYSTEM: |
2123 | case SIAtomicScope::AGENT: |
2124 | // The order of invalidates matter here. We must invalidate "outer in" |
2125 | // so L1 -> L0 to avoid L0 pulling in stale data from L1 when it is |
2126 | // invalidated. |
2127 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_GL1_INV)); |
2128 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_GL0_INV)); |
2129 | Changed = true; |
2130 | break; |
2131 | case SIAtomicScope::WORKGROUP: |
2132 | // In WGP mode the waves of a work-group can be executing on either CU of |
2133 | // the WGP. Therefore need to invalidate the L0 which is per CU. Otherwise |
2134 | // in CU mode and all waves of a work-group are on the same CU, and so the |
2135 | // L0 does not need to be invalidated. |
2136 | if (!ST.isCuModeEnabled()) { |
2137 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::BUFFER_GL0_INV)); |
2138 | Changed = true; |
2139 | } |
2140 | break; |
2141 | case SIAtomicScope::WAVEFRONT: |
2142 | case SIAtomicScope::SINGLETHREAD: |
2143 | // No cache to invalidate. |
2144 | break; |
2145 | default: |
2146 | llvm_unreachable("Unsupported synchronization scope" ); |
2147 | } |
2148 | } |
2149 | |
2150 | /// The scratch address space does not need the global memory cache |
2151 | /// to be flushed as all memory operations by the same thread are |
2152 | /// sequentially consistent, and no other thread can access scratch |
2153 | /// memory. |
2154 | |
2155 | /// Other address spaces do not have a cache. |
2156 | |
2157 | if (Pos == Position::AFTER) |
2158 | --MI; |
2159 | |
2160 | return Changed; |
2161 | } |
2162 | |
2163 | bool SIGfx11CacheControl::enableLoadCacheBypass( |
2164 | const MachineBasicBlock::iterator &MI, SIAtomicScope Scope, |
2165 | SIAtomicAddrSpace AddrSpace) const { |
2166 | assert(MI->mayLoad() && !MI->mayStore()); |
2167 | bool Changed = false; |
2168 | |
2169 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
2170 | switch (Scope) { |
2171 | case SIAtomicScope::SYSTEM: |
2172 | case SIAtomicScope::AGENT: |
2173 | // Set the L0 and L1 cache policies to MISS_EVICT. |
2174 | // Note: there is no L2 cache coherent bypass control at the ISA level. |
2175 | Changed |= enableGLCBit(MI); |
2176 | break; |
2177 | case SIAtomicScope::WORKGROUP: |
2178 | // In WGP mode the waves of a work-group can be executing on either CU of |
2179 | // the WGP. Therefore need to bypass the L0 which is per CU. Otherwise in |
2180 | // CU mode all waves of a work-group are on the same CU, and so the L0 |
2181 | // does not need to be bypassed. |
2182 | if (!ST.isCuModeEnabled()) |
2183 | Changed |= enableGLCBit(MI); |
2184 | break; |
2185 | case SIAtomicScope::WAVEFRONT: |
2186 | case SIAtomicScope::SINGLETHREAD: |
2187 | // No cache to bypass. |
2188 | break; |
2189 | default: |
2190 | llvm_unreachable("Unsupported synchronization scope" ); |
2191 | } |
2192 | } |
2193 | |
2194 | /// The scratch address space does not need the global memory caches |
2195 | /// to be bypassed as all memory operations by the same thread are |
2196 | /// sequentially consistent, and no other thread can access scratch |
2197 | /// memory. |
2198 | |
2199 | /// Other address spaces do not have a cache. |
2200 | |
2201 | return Changed; |
2202 | } |
2203 | |
2204 | bool SIGfx11CacheControl::enableVolatileAndOrNonTemporal( |
2205 | MachineBasicBlock::iterator &MI, SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
2206 | bool IsVolatile, bool IsNonTemporal, bool IsLastUse = false) const { |
2207 | |
2208 | // Only handle load and store, not atomic read-modify-write insructions. The |
2209 | // latter use glc to indicate if the atomic returns a result and so must not |
2210 | // be used for cache control. |
2211 | assert(MI->mayLoad() ^ MI->mayStore()); |
2212 | |
2213 | // Only update load and store, not LLVM IR atomic read-modify-write |
2214 | // instructions. The latter are always marked as volatile so cannot sensibly |
2215 | // handle it as do not want to pessimize all atomics. Also they do not support |
2216 | // the nontemporal attribute. |
2217 | assert(Op == SIMemOp::LOAD || Op == SIMemOp::STORE); |
2218 | |
2219 | bool Changed = false; |
2220 | |
2221 | if (IsVolatile) { |
2222 | // Set L0 and L1 cache policy to be MISS_EVICT for load instructions |
2223 | // and MISS_LRU for store instructions. |
2224 | // Note: there is no L2 cache coherent bypass control at the ISA level. |
2225 | if (Op == SIMemOp::LOAD) |
2226 | Changed |= enableGLCBit(MI); |
2227 | |
2228 | // Set MALL NOALLOC for load and store instructions. |
2229 | Changed |= enableDLCBit(MI); |
2230 | |
2231 | // Ensure operation has completed at system scope to cause all volatile |
2232 | // operations to be visible outside the program in a global order. Do not |
2233 | // request cross address space as only the global address space can be |
2234 | // observable outside the program, so no need to cause a waitcnt for LDS |
2235 | // address space operations. |
2236 | Changed |= insertWait(MI, Scope: SIAtomicScope::SYSTEM, AddrSpace, Op, IsCrossAddrSpaceOrdering: false, |
2237 | Pos: Position::AFTER); |
2238 | return Changed; |
2239 | } |
2240 | |
2241 | if (IsNonTemporal) { |
2242 | // For loads setting SLC configures L0 and L1 cache policy to HIT_EVICT |
2243 | // and L2 cache policy to STREAM. |
2244 | // For stores setting both GLC and SLC configures L0 and L1 cache policy |
2245 | // to MISS_EVICT and the L2 cache policy to STREAM. |
2246 | if (Op == SIMemOp::STORE) |
2247 | Changed |= enableGLCBit(MI); |
2248 | Changed |= enableSLCBit(MI); |
2249 | |
2250 | // Set MALL NOALLOC for load and store instructions. |
2251 | Changed |= enableDLCBit(MI); |
2252 | return Changed; |
2253 | } |
2254 | |
2255 | return Changed; |
2256 | } |
2257 | |
2258 | bool SIGfx12CacheControl::setTH(const MachineBasicBlock::iterator MI, |
2259 | AMDGPU::CPol::CPol Value) const { |
2260 | MachineOperand *CPol = TII->getNamedOperand(MI&: *MI, OperandName: OpName::cpol); |
2261 | if (!CPol) |
2262 | return false; |
2263 | |
2264 | uint64_t NewTH = Value & AMDGPU::CPol::TH; |
2265 | if ((CPol->getImm() & AMDGPU::CPol::TH) != NewTH) { |
2266 | CPol->setImm((CPol->getImm() & ~AMDGPU::CPol::TH) | NewTH); |
2267 | return true; |
2268 | } |
2269 | |
2270 | return false; |
2271 | } |
2272 | |
2273 | bool SIGfx12CacheControl::setScope(const MachineBasicBlock::iterator MI, |
2274 | AMDGPU::CPol::CPol Value) const { |
2275 | MachineOperand *CPol = TII->getNamedOperand(MI&: *MI, OperandName: OpName::cpol); |
2276 | if (!CPol) |
2277 | return false; |
2278 | |
2279 | uint64_t NewScope = Value & AMDGPU::CPol::SCOPE; |
2280 | if ((CPol->getImm() & AMDGPU::CPol::SCOPE) != NewScope) { |
2281 | CPol->setImm((CPol->getImm() & ~AMDGPU::CPol::SCOPE) | NewScope); |
2282 | return true; |
2283 | } |
2284 | |
2285 | return false; |
2286 | } |
2287 | |
2288 | bool SIGfx12CacheControl::insertWaitsBeforeSystemScopeStore( |
2289 | const MachineBasicBlock::iterator MI) const { |
2290 | // TODO: implement flag for frontend to give us a hint not to insert waits. |
2291 | |
2292 | MachineBasicBlock &MBB = *MI->getParent(); |
2293 | const DebugLoc &DL = MI->getDebugLoc(); |
2294 | |
2295 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: S_WAIT_LOADCNT_soft)).addImm(Val: 0); |
2296 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: S_WAIT_SAMPLECNT_soft)).addImm(Val: 0); |
2297 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: S_WAIT_BVHCNT_soft)).addImm(Val: 0); |
2298 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: S_WAIT_KMCNT_soft)).addImm(Val: 0); |
2299 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: S_WAIT_STORECNT_soft)).addImm(Val: 0); |
2300 | |
2301 | return true; |
2302 | } |
2303 | |
2304 | bool SIGfx12CacheControl::insertWait(MachineBasicBlock::iterator &MI, |
2305 | SIAtomicScope Scope, |
2306 | SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
2307 | bool IsCrossAddrSpaceOrdering, |
2308 | Position Pos) const { |
2309 | bool Changed = false; |
2310 | |
2311 | MachineBasicBlock &MBB = *MI->getParent(); |
2312 | DebugLoc DL = MI->getDebugLoc(); |
2313 | |
2314 | bool LOADCnt = false; |
2315 | bool DSCnt = false; |
2316 | bool STORECnt = false; |
2317 | |
2318 | if (Pos == Position::AFTER) |
2319 | ++MI; |
2320 | |
2321 | if ((AddrSpace & (SIAtomicAddrSpace::GLOBAL | SIAtomicAddrSpace::SCRATCH)) != |
2322 | SIAtomicAddrSpace::NONE) { |
2323 | switch (Scope) { |
2324 | case SIAtomicScope::SYSTEM: |
2325 | case SIAtomicScope::AGENT: |
2326 | if ((Op & SIMemOp::LOAD) != SIMemOp::NONE) |
2327 | LOADCnt |= true; |
2328 | if ((Op & SIMemOp::STORE) != SIMemOp::NONE) |
2329 | STORECnt |= true; |
2330 | break; |
2331 | case SIAtomicScope::WORKGROUP: |
2332 | // In WGP mode the waves of a work-group can be executing on either CU of |
2333 | // the WGP. Therefore need to wait for operations to complete to ensure |
2334 | // they are visible to waves in the other CU as the L0 is per CU. |
2335 | // Otherwise in CU mode and all waves of a work-group are on the same CU |
2336 | // which shares the same L0. |
2337 | if (!ST.isCuModeEnabled()) { |
2338 | if ((Op & SIMemOp::LOAD) != SIMemOp::NONE) |
2339 | LOADCnt |= true; |
2340 | if ((Op & SIMemOp::STORE) != SIMemOp::NONE) |
2341 | STORECnt |= true; |
2342 | } |
2343 | break; |
2344 | case SIAtomicScope::WAVEFRONT: |
2345 | case SIAtomicScope::SINGLETHREAD: |
2346 | // The L0 cache keeps all memory operations in order for |
2347 | // work-items in the same wavefront. |
2348 | break; |
2349 | default: |
2350 | llvm_unreachable("Unsupported synchronization scope" ); |
2351 | } |
2352 | } |
2353 | |
2354 | if ((AddrSpace & SIAtomicAddrSpace::LDS) != SIAtomicAddrSpace::NONE) { |
2355 | switch (Scope) { |
2356 | case SIAtomicScope::SYSTEM: |
2357 | case SIAtomicScope::AGENT: |
2358 | case SIAtomicScope::WORKGROUP: |
2359 | // If no cross address space ordering then an "S_WAITCNT lgkmcnt(0)" is |
2360 | // not needed as LDS operations for all waves are executed in a total |
2361 | // global ordering as observed by all waves. Required if also |
2362 | // synchronizing with global/GDS memory as LDS operations could be |
2363 | // reordered with respect to later global/GDS memory operations of the |
2364 | // same wave. |
2365 | DSCnt |= IsCrossAddrSpaceOrdering; |
2366 | break; |
2367 | case SIAtomicScope::WAVEFRONT: |
2368 | case SIAtomicScope::SINGLETHREAD: |
2369 | // The LDS keeps all memory operations in order for |
2370 | // the same wavefront. |
2371 | break; |
2372 | default: |
2373 | llvm_unreachable("Unsupported synchronization scope" ); |
2374 | } |
2375 | } |
2376 | |
2377 | if (LOADCnt) { |
2378 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_WAIT_BVHCNT_soft)).addImm(Val: 0); |
2379 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_WAIT_SAMPLECNT_soft)).addImm(Val: 0); |
2380 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_WAIT_LOADCNT_soft)).addImm(Val: 0); |
2381 | Changed = true; |
2382 | } |
2383 | |
2384 | if (STORECnt) { |
2385 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_WAIT_STORECNT_soft)).addImm(Val: 0); |
2386 | Changed = true; |
2387 | } |
2388 | |
2389 | if (DSCnt) { |
2390 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_WAIT_DSCNT_soft)).addImm(Val: 0); |
2391 | Changed = true; |
2392 | } |
2393 | |
2394 | if (Pos == Position::AFTER) |
2395 | --MI; |
2396 | |
2397 | return Changed; |
2398 | } |
2399 | |
2400 | bool SIGfx12CacheControl::insertAcquire(MachineBasicBlock::iterator &MI, |
2401 | SIAtomicScope Scope, |
2402 | SIAtomicAddrSpace AddrSpace, |
2403 | Position Pos) const { |
2404 | if (!InsertCacheInv) |
2405 | return false; |
2406 | |
2407 | MachineBasicBlock &MBB = *MI->getParent(); |
2408 | DebugLoc DL = MI->getDebugLoc(); |
2409 | |
2410 | /// The scratch address space does not need the global memory cache |
2411 | /// to be flushed as all memory operations by the same thread are |
2412 | /// sequentially consistent, and no other thread can access scratch |
2413 | /// memory. |
2414 | |
2415 | /// Other address spaces do not have a cache. |
2416 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) == SIAtomicAddrSpace::NONE) |
2417 | return false; |
2418 | |
2419 | AMDGPU::CPol::CPol ScopeImm = AMDGPU::CPol::SCOPE_DEV; |
2420 | switch (Scope) { |
2421 | case SIAtomicScope::SYSTEM: |
2422 | ScopeImm = AMDGPU::CPol::SCOPE_SYS; |
2423 | break; |
2424 | case SIAtomicScope::AGENT: |
2425 | ScopeImm = AMDGPU::CPol::SCOPE_DEV; |
2426 | break; |
2427 | case SIAtomicScope::WORKGROUP: |
2428 | // In WGP mode the waves of a work-group can be executing on either CU of |
2429 | // the WGP. Therefore we need to invalidate the L0 which is per CU. |
2430 | // Otherwise in CU mode all waves of a work-group are on the same CU, and so |
2431 | // the L0 does not need to be invalidated. |
2432 | if (ST.isCuModeEnabled()) |
2433 | return false; |
2434 | |
2435 | ScopeImm = AMDGPU::CPol::SCOPE_SE; |
2436 | break; |
2437 | case SIAtomicScope::WAVEFRONT: |
2438 | case SIAtomicScope::SINGLETHREAD: |
2439 | // No cache to invalidate. |
2440 | return false; |
2441 | default: |
2442 | llvm_unreachable("Unsupported synchronization scope" ); |
2443 | } |
2444 | |
2445 | if (Pos == Position::AFTER) |
2446 | ++MI; |
2447 | |
2448 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::GLOBAL_INV)).addImm(Val: ScopeImm); |
2449 | |
2450 | if (Pos == Position::AFTER) |
2451 | --MI; |
2452 | |
2453 | return true; |
2454 | } |
2455 | |
2456 | bool SIGfx12CacheControl::insertRelease(MachineBasicBlock::iterator &MI, |
2457 | SIAtomicScope Scope, |
2458 | SIAtomicAddrSpace AddrSpace, |
2459 | bool IsCrossAddrSpaceOrdering, |
2460 | Position Pos) const { |
2461 | MachineBasicBlock &MBB = *MI->getParent(); |
2462 | DebugLoc DL = MI->getDebugLoc(); |
2463 | |
2464 | // The scratch address space does not need the global memory cache |
2465 | // writeback as all memory operations by the same thread are |
2466 | // sequentially consistent, and no other thread can access scratch |
2467 | // memory. |
2468 | |
2469 | // Other address spaces do not have a cache. |
2470 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) == SIAtomicAddrSpace::NONE) |
2471 | return false; |
2472 | |
2473 | if (Pos == Position::AFTER) |
2474 | ++MI; |
2475 | |
2476 | // GLOBAL_WB is always needed, even for write-through caches, as it |
2477 | // additionally ensures all operations have reached the desired cache level. |
2478 | bool SkipWB = false; |
2479 | AMDGPU::CPol::CPol ScopeImm = AMDGPU::CPol::SCOPE_DEV; |
2480 | switch (Scope) { |
2481 | case SIAtomicScope::SYSTEM: |
2482 | ScopeImm = AMDGPU::CPol::SCOPE_SYS; |
2483 | break; |
2484 | case SIAtomicScope::AGENT: |
2485 | ScopeImm = AMDGPU::CPol::SCOPE_DEV; |
2486 | break; |
2487 | case SIAtomicScope::WORKGROUP: |
2488 | // In WGP mode the waves of a work-group can be executing on either CU of |
2489 | // the WGP. Therefore we need to ensure all operations have reached L1, |
2490 | // hence the SCOPE_SE WB. |
2491 | // For CU mode, we need operations to reach L0, so the wait is enough - |
2492 | // there are no ways for an operation to report completion without reaching |
2493 | // at least L0. |
2494 | if (ST.isCuModeEnabled()) |
2495 | SkipWB = true; |
2496 | else |
2497 | ScopeImm = AMDGPU::CPol::SCOPE_SE; |
2498 | break; |
2499 | case SIAtomicScope::WAVEFRONT: |
2500 | case SIAtomicScope::SINGLETHREAD: |
2501 | // No cache to invalidate. |
2502 | return false; |
2503 | default: |
2504 | llvm_unreachable("Unsupported synchronization scope" ); |
2505 | } |
2506 | |
2507 | if (!SkipWB) |
2508 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::GLOBAL_WB)).addImm(Val: ScopeImm); |
2509 | |
2510 | if (Pos == Position::AFTER) |
2511 | --MI; |
2512 | |
2513 | // We always have to wait for previous memory operations (load/store) to |
2514 | // complete, whether we inserted a WB or not. If we inserted a WB (storecnt), |
2515 | // we of course need to wait for that as well. |
2516 | insertWait(MI, Scope, AddrSpace, Op: SIMemOp::LOAD | SIMemOp::STORE, |
2517 | IsCrossAddrSpaceOrdering, Pos); |
2518 | |
2519 | return true; |
2520 | } |
2521 | |
2522 | bool SIGfx12CacheControl::enableVolatileAndOrNonTemporal( |
2523 | MachineBasicBlock::iterator &MI, SIAtomicAddrSpace AddrSpace, SIMemOp Op, |
2524 | bool IsVolatile, bool IsNonTemporal, bool IsLastUse = false) const { |
2525 | |
2526 | // Only handle load and store, not atomic read-modify-write instructions. |
2527 | assert(MI->mayLoad() ^ MI->mayStore()); |
2528 | |
2529 | // Only update load and store, not LLVM IR atomic read-modify-write |
2530 | // instructions. The latter are always marked as volatile so cannot sensibly |
2531 | // handle it as do not want to pessimize all atomics. Also they do not support |
2532 | // the nontemporal attribute. |
2533 | assert(Op == SIMemOp::LOAD || Op == SIMemOp::STORE); |
2534 | |
2535 | bool Changed = false; |
2536 | |
2537 | if (IsLastUse) { |
2538 | // Set last-use hint. |
2539 | Changed |= setTH(MI, Value: AMDGPU::CPol::TH_LU); |
2540 | } else if (IsNonTemporal) { |
2541 | // Set non-temporal hint for all cache levels. |
2542 | Changed |= setTH(MI, Value: AMDGPU::CPol::TH_NT); |
2543 | } |
2544 | |
2545 | if (IsVolatile) { |
2546 | Changed |= setScope(MI, Value: AMDGPU::CPol::SCOPE_SYS); |
2547 | |
2548 | if (Op == SIMemOp::STORE) |
2549 | Changed |= insertWaitsBeforeSystemScopeStore(MI); |
2550 | |
2551 | // Ensure operation has completed at system scope to cause all volatile |
2552 | // operations to be visible outside the program in a global order. Do not |
2553 | // request cross address space as only the global address space can be |
2554 | // observable outside the program, so no need to cause a waitcnt for LDS |
2555 | // address space operations. |
2556 | Changed |= insertWait(MI, Scope: SIAtomicScope::SYSTEM, AddrSpace, Op, IsCrossAddrSpaceOrdering: false, |
2557 | Pos: Position::AFTER); |
2558 | } |
2559 | |
2560 | return Changed; |
2561 | } |
2562 | |
2563 | bool SIGfx12CacheControl::expandSystemScopeStore( |
2564 | MachineBasicBlock::iterator &MI) const { |
2565 | MachineOperand *CPol = TII->getNamedOperand(MI&: *MI, OperandName: OpName::cpol); |
2566 | if (CPol && ((CPol->getImm() & CPol::SCOPE) == CPol::SCOPE_SYS)) |
2567 | return insertWaitsBeforeSystemScopeStore(MI); |
2568 | |
2569 | return false; |
2570 | } |
2571 | |
2572 | bool SIGfx12CacheControl::setAtomicScope(const MachineBasicBlock::iterator &MI, |
2573 | SIAtomicScope Scope, |
2574 | SIAtomicAddrSpace AddrSpace) const { |
2575 | bool Changed = false; |
2576 | |
2577 | if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) { |
2578 | switch (Scope) { |
2579 | case SIAtomicScope::SYSTEM: |
2580 | Changed |= setScope(MI, Value: AMDGPU::CPol::SCOPE_SYS); |
2581 | break; |
2582 | case SIAtomicScope::AGENT: |
2583 | Changed |= setScope(MI, Value: AMDGPU::CPol::SCOPE_DEV); |
2584 | break; |
2585 | case SIAtomicScope::WORKGROUP: |
2586 | // In workgroup mode, SCOPE_SE is needed as waves can executes on |
2587 | // different CUs that access different L0s. |
2588 | if (!ST.isCuModeEnabled()) |
2589 | Changed |= setScope(MI, Value: AMDGPU::CPol::SCOPE_SE); |
2590 | break; |
2591 | case SIAtomicScope::WAVEFRONT: |
2592 | case SIAtomicScope::SINGLETHREAD: |
2593 | // No cache to bypass. |
2594 | break; |
2595 | default: |
2596 | llvm_unreachable("Unsupported synchronization scope" ); |
2597 | } |
2598 | } |
2599 | |
2600 | // The scratch address space does not need the global memory caches |
2601 | // to be bypassed as all memory operations by the same thread are |
2602 | // sequentially consistent, and no other thread can access scratch |
2603 | // memory. |
2604 | |
2605 | // Other address spaces do not have a cache. |
2606 | |
2607 | return Changed; |
2608 | } |
2609 | |
2610 | bool SIMemoryLegalizer::removeAtomicPseudoMIs() { |
2611 | if (AtomicPseudoMIs.empty()) |
2612 | return false; |
2613 | |
2614 | for (auto &MI : AtomicPseudoMIs) |
2615 | MI->eraseFromParent(); |
2616 | |
2617 | AtomicPseudoMIs.clear(); |
2618 | return true; |
2619 | } |
2620 | |
2621 | bool SIMemoryLegalizer::expandLoad(const SIMemOpInfo &MOI, |
2622 | MachineBasicBlock::iterator &MI) { |
2623 | assert(MI->mayLoad() && !MI->mayStore()); |
2624 | |
2625 | bool Changed = false; |
2626 | |
2627 | if (MOI.isAtomic()) { |
2628 | if (MOI.getOrdering() == AtomicOrdering::Monotonic || |
2629 | MOI.getOrdering() == AtomicOrdering::Acquire || |
2630 | MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent) { |
2631 | Changed |= CC->enableLoadCacheBypass(MI, Scope: MOI.getScope(), |
2632 | AddrSpace: MOI.getOrderingAddrSpace()); |
2633 | } |
2634 | |
2635 | if (MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent) |
2636 | Changed |= CC->insertWait(MI, Scope: MOI.getScope(), |
2637 | AddrSpace: MOI.getOrderingAddrSpace(), |
2638 | Op: SIMemOp::LOAD | SIMemOp::STORE, |
2639 | IsCrossAddrSpaceOrdering: MOI.getIsCrossAddressSpaceOrdering(), |
2640 | Pos: Position::BEFORE); |
2641 | |
2642 | if (MOI.getOrdering() == AtomicOrdering::Acquire || |
2643 | MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent) { |
2644 | Changed |= CC->insertWait(MI, Scope: MOI.getScope(), |
2645 | AddrSpace: MOI.getInstrAddrSpace(), |
2646 | Op: SIMemOp::LOAD, |
2647 | IsCrossAddrSpaceOrdering: MOI.getIsCrossAddressSpaceOrdering(), |
2648 | Pos: Position::AFTER); |
2649 | Changed |= CC->insertAcquire(MI, Scope: MOI.getScope(), |
2650 | AddrSpace: MOI.getOrderingAddrSpace(), |
2651 | Pos: Position::AFTER); |
2652 | } |
2653 | |
2654 | return Changed; |
2655 | } |
2656 | |
2657 | // Atomic instructions already bypass caches to the scope specified by the |
2658 | // SyncScope operand. Only non-atomic volatile and nontemporal/last-use |
2659 | // instructions need additional treatment. |
2660 | Changed |= CC->enableVolatileAndOrNonTemporal( |
2661 | MI, AddrSpace: MOI.getInstrAddrSpace(), Op: SIMemOp::LOAD, IsVolatile: MOI.isVolatile(), |
2662 | IsNonTemporal: MOI.isNonTemporal(), IsLastUse: MOI.isLastUse()); |
2663 | |
2664 | return Changed; |
2665 | } |
2666 | |
2667 | bool SIMemoryLegalizer::expandStore(const SIMemOpInfo &MOI, |
2668 | MachineBasicBlock::iterator &MI) { |
2669 | assert(!MI->mayLoad() && MI->mayStore()); |
2670 | |
2671 | bool Changed = false; |
2672 | |
2673 | if (MOI.isAtomic()) { |
2674 | if (MOI.getOrdering() == AtomicOrdering::Monotonic || |
2675 | MOI.getOrdering() == AtomicOrdering::Release || |
2676 | MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent) { |
2677 | Changed |= CC->enableStoreCacheBypass(MI, Scope: MOI.getScope(), |
2678 | AddrSpace: MOI.getOrderingAddrSpace()); |
2679 | } |
2680 | |
2681 | if (MOI.getOrdering() == AtomicOrdering::Release || |
2682 | MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent) |
2683 | Changed |= CC->insertRelease(MI, Scope: MOI.getScope(), |
2684 | AddrSpace: MOI.getOrderingAddrSpace(), |
2685 | IsCrossAddrSpaceOrdering: MOI.getIsCrossAddressSpaceOrdering(), |
2686 | Pos: Position::BEFORE); |
2687 | |
2688 | return Changed; |
2689 | } |
2690 | |
2691 | // Atomic instructions already bypass caches to the scope specified by the |
2692 | // SyncScope operand. Only non-atomic volatile and nontemporal instructions |
2693 | // need additional treatment. |
2694 | Changed |= CC->enableVolatileAndOrNonTemporal( |
2695 | MI, AddrSpace: MOI.getInstrAddrSpace(), Op: SIMemOp::STORE, IsVolatile: MOI.isVolatile(), |
2696 | IsNonTemporal: MOI.isNonTemporal()); |
2697 | |
2698 | // GFX12 specific, scope(desired coherence domain in cache hierarchy) is |
2699 | // instruction field, do not confuse it with atomic scope. |
2700 | Changed |= CC->expandSystemScopeStore(MI); |
2701 | return Changed; |
2702 | } |
2703 | |
2704 | bool SIMemoryLegalizer::expandAtomicFence(const SIMemOpInfo &MOI, |
2705 | MachineBasicBlock::iterator &MI) { |
2706 | assert(MI->getOpcode() == AMDGPU::ATOMIC_FENCE); |
2707 | |
2708 | AtomicPseudoMIs.push_back(x: MI); |
2709 | bool Changed = false; |
2710 | |
2711 | // Refine fenced address space based on MMRAs. |
2712 | // |
2713 | // TODO: Should we support this MMRA on other atomic operations? |
2714 | auto OrderingAddrSpace = |
2715 | getFenceAddrSpaceMMRA(MI: *MI, Default: MOI.getOrderingAddrSpace()); |
2716 | |
2717 | if (MOI.isAtomic()) { |
2718 | if (MOI.getOrdering() == AtomicOrdering::Acquire) |
2719 | Changed |= CC->insertWait( |
2720 | MI, Scope: MOI.getScope(), AddrSpace: OrderingAddrSpace, Op: SIMemOp::LOAD | SIMemOp::STORE, |
2721 | IsCrossAddrSpaceOrdering: MOI.getIsCrossAddressSpaceOrdering(), Pos: Position::BEFORE); |
2722 | |
2723 | if (MOI.getOrdering() == AtomicOrdering::Release || |
2724 | MOI.getOrdering() == AtomicOrdering::AcquireRelease || |
2725 | MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent) |
2726 | /// TODO: This relies on a barrier always generating a waitcnt |
2727 | /// for LDS to ensure it is not reordered with the completion of |
2728 | /// the proceeding LDS operations. If barrier had a memory |
2729 | /// ordering and memory scope, then library does not need to |
2730 | /// generate a fence. Could add support in this file for |
2731 | /// barrier. SIInsertWaitcnt.cpp could then stop unconditionally |
2732 | /// adding S_WAITCNT before a S_BARRIER. |
2733 | Changed |= CC->insertRelease(MI, Scope: MOI.getScope(), AddrSpace: OrderingAddrSpace, |
2734 | IsCrossAddrSpaceOrdering: MOI.getIsCrossAddressSpaceOrdering(), |
2735 | Pos: Position::BEFORE); |
2736 | |
2737 | // TODO: If both release and invalidate are happening they could be combined |
2738 | // to use the single "BUFFER_WBINV*" instruction. This could be done by |
2739 | // reorganizing this code or as part of optimizing SIInsertWaitcnt pass to |
2740 | // track cache invalidate and write back instructions. |
2741 | |
2742 | if (MOI.getOrdering() == AtomicOrdering::Acquire || |
2743 | MOI.getOrdering() == AtomicOrdering::AcquireRelease || |
2744 | MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent) |
2745 | Changed |= CC->insertAcquire(MI, Scope: MOI.getScope(), AddrSpace: OrderingAddrSpace, |
2746 | Pos: Position::BEFORE); |
2747 | |
2748 | return Changed; |
2749 | } |
2750 | |
2751 | return Changed; |
2752 | } |
2753 | |
2754 | bool SIMemoryLegalizer::expandAtomicCmpxchgOrRmw(const SIMemOpInfo &MOI, |
2755 | MachineBasicBlock::iterator &MI) { |
2756 | assert(MI->mayLoad() && MI->mayStore()); |
2757 | |
2758 | bool Changed = false; |
2759 | |
2760 | if (MOI.isAtomic()) { |
2761 | if (MOI.getOrdering() == AtomicOrdering::Monotonic || |
2762 | MOI.getOrdering() == AtomicOrdering::Acquire || |
2763 | MOI.getOrdering() == AtomicOrdering::Release || |
2764 | MOI.getOrdering() == AtomicOrdering::AcquireRelease || |
2765 | MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent) { |
2766 | Changed |= CC->enableRMWCacheBypass(MI, Scope: MOI.getScope(), |
2767 | AddrSpace: MOI.getInstrAddrSpace()); |
2768 | } |
2769 | |
2770 | if (MOI.getOrdering() == AtomicOrdering::Release || |
2771 | MOI.getOrdering() == AtomicOrdering::AcquireRelease || |
2772 | MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent || |
2773 | MOI.getFailureOrdering() == AtomicOrdering::SequentiallyConsistent) |
2774 | Changed |= CC->insertRelease(MI, Scope: MOI.getScope(), |
2775 | AddrSpace: MOI.getOrderingAddrSpace(), |
2776 | IsCrossAddrSpaceOrdering: MOI.getIsCrossAddressSpaceOrdering(), |
2777 | Pos: Position::BEFORE); |
2778 | |
2779 | if (MOI.getOrdering() == AtomicOrdering::Acquire || |
2780 | MOI.getOrdering() == AtomicOrdering::AcquireRelease || |
2781 | MOI.getOrdering() == AtomicOrdering::SequentiallyConsistent || |
2782 | MOI.getFailureOrdering() == AtomicOrdering::Acquire || |
2783 | MOI.getFailureOrdering() == AtomicOrdering::SequentiallyConsistent) { |
2784 | Changed |= CC->insertWait(MI, Scope: MOI.getScope(), |
2785 | AddrSpace: MOI.getInstrAddrSpace(), |
2786 | Op: isAtomicRet(MI: *MI) ? SIMemOp::LOAD : |
2787 | SIMemOp::STORE, |
2788 | IsCrossAddrSpaceOrdering: MOI.getIsCrossAddressSpaceOrdering(), |
2789 | Pos: Position::AFTER); |
2790 | Changed |= CC->insertAcquire(MI, Scope: MOI.getScope(), |
2791 | AddrSpace: MOI.getOrderingAddrSpace(), |
2792 | Pos: Position::AFTER); |
2793 | } |
2794 | |
2795 | return Changed; |
2796 | } |
2797 | |
2798 | return Changed; |
2799 | } |
2800 | |
2801 | bool SIMemoryLegalizer::runOnMachineFunction(MachineFunction &MF) { |
2802 | bool Changed = false; |
2803 | |
2804 | const MachineModuleInfo &MMI = |
2805 | getAnalysis<MachineModuleInfoWrapperPass>().getMMI(); |
2806 | |
2807 | SIMemOpAccess MOA(MMI.getObjFileInfo<AMDGPUMachineModuleInfo>()); |
2808 | CC = SICacheControl::create(ST: MF.getSubtarget<GCNSubtarget>()); |
2809 | |
2810 | for (auto &MBB : MF) { |
2811 | for (auto MI = MBB.begin(); MI != MBB.end(); ++MI) { |
2812 | |
2813 | // Unbundle instructions after the post-RA scheduler. |
2814 | if (MI->isBundle() && MI->mayLoadOrStore()) { |
2815 | MachineBasicBlock::instr_iterator II(MI->getIterator()); |
2816 | for (MachineBasicBlock::instr_iterator I = ++II, E = MBB.instr_end(); |
2817 | I != E && I->isBundledWithPred(); ++I) { |
2818 | I->unbundleFromPred(); |
2819 | for (MachineOperand &MO : I->operands()) |
2820 | if (MO.isReg()) |
2821 | MO.setIsInternalRead(false); |
2822 | } |
2823 | |
2824 | MI->eraseFromParent(); |
2825 | MI = II->getIterator(); |
2826 | } |
2827 | |
2828 | if (!(MI->getDesc().TSFlags & SIInstrFlags::maybeAtomic)) |
2829 | continue; |
2830 | |
2831 | if (const auto &MOI = MOA.getLoadInfo(MI)) |
2832 | Changed |= expandLoad(MOI: *MOI, MI); |
2833 | else if (const auto &MOI = MOA.getStoreInfo(MI)) { |
2834 | Changed |= expandStore(MOI: *MOI, MI); |
2835 | Changed |= CC->tryForceStoreSC0SC1(MOI: *MOI, MI); |
2836 | } else if (const auto &MOI = MOA.getAtomicFenceInfo(MI)) |
2837 | Changed |= expandAtomicFence(MOI: *MOI, MI); |
2838 | else if (const auto &MOI = MOA.getAtomicCmpxchgOrRmwInfo(MI)) |
2839 | Changed |= expandAtomicCmpxchgOrRmw(MOI: *MOI, MI); |
2840 | } |
2841 | } |
2842 | |
2843 | Changed |= removeAtomicPseudoMIs(); |
2844 | return Changed; |
2845 | } |
2846 | |
2847 | INITIALIZE_PASS(SIMemoryLegalizer, DEBUG_TYPE, PASS_NAME, false, false) |
2848 | |
2849 | char SIMemoryLegalizer::ID = 0; |
2850 | char &llvm::SIMemoryLegalizerID = SIMemoryLegalizer::ID; |
2851 | |
2852 | FunctionPass *llvm::createSIMemoryLegalizerPass() { |
2853 | return new SIMemoryLegalizer(); |
2854 | } |
2855 | |