1 | //===- InterleavedLoadCombine.cpp - Combine Interleaved Loads ---*- C++ -*-===// |
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 | // |
11 | // This file defines the interleaved-load-combine pass. The pass searches for |
12 | // ShuffleVectorInstruction that execute interleaving loads. If a matching |
13 | // pattern is found, it adds a combined load and further instructions in a |
14 | // pattern that is detectable by InterleavedAccesPass. The old instructions are |
15 | // left dead to be removed later. The pass is specifically designed to be |
16 | // executed just before InterleavedAccesPass to find any left-over instances |
17 | // that are not detected within former passes. |
18 | // |
19 | //===----------------------------------------------------------------------===// |
20 | |
21 | #include "llvm/ADT/Statistic.h" |
22 | #include "llvm/Analysis/MemorySSA.h" |
23 | #include "llvm/Analysis/MemorySSAUpdater.h" |
24 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
25 | #include "llvm/Analysis/TargetTransformInfo.h" |
26 | #include "llvm/CodeGen/InterleavedLoadCombine.h" |
27 | #include "llvm/CodeGen/Passes.h" |
28 | #include "llvm/CodeGen/TargetLowering.h" |
29 | #include "llvm/CodeGen/TargetPassConfig.h" |
30 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
31 | #include "llvm/IR/DataLayout.h" |
32 | #include "llvm/IR/Dominators.h" |
33 | #include "llvm/IR/Function.h" |
34 | #include "llvm/IR/IRBuilder.h" |
35 | #include "llvm/IR/Instructions.h" |
36 | #include "llvm/IR/Module.h" |
37 | #include "llvm/InitializePasses.h" |
38 | #include "llvm/Pass.h" |
39 | #include "llvm/Support/Debug.h" |
40 | #include "llvm/Support/ErrorHandling.h" |
41 | #include "llvm/Support/raw_ostream.h" |
42 | #include "llvm/Target/TargetMachine.h" |
43 | |
44 | #include <algorithm> |
45 | #include <cassert> |
46 | #include <list> |
47 | |
48 | using namespace llvm; |
49 | |
50 | #define DEBUG_TYPE "interleaved-load-combine" |
51 | |
52 | namespace { |
53 | |
54 | /// Statistic counter |
55 | STATISTIC(NumInterleavedLoadCombine, "Number of combined loads" ); |
56 | |
57 | /// Option to disable the pass |
58 | static cl::opt<bool> DisableInterleavedLoadCombine( |
59 | "disable-" DEBUG_TYPE, cl::init(Val: false), cl::Hidden, |
60 | cl::desc("Disable combining of interleaved loads" )); |
61 | |
62 | struct VectorInfo; |
63 | |
64 | struct InterleavedLoadCombineImpl { |
65 | public: |
66 | InterleavedLoadCombineImpl(Function &F, DominatorTree &DT, MemorySSA &MSSA, |
67 | const TargetTransformInfo &TTI, |
68 | const TargetMachine &TM) |
69 | : F(F), DT(DT), MSSA(MSSA), |
70 | TLI(*TM.getSubtargetImpl(F)->getTargetLowering()), TTI(TTI) {} |
71 | |
72 | /// Scan the function for interleaved load candidates and execute the |
73 | /// replacement if applicable. |
74 | bool run(); |
75 | |
76 | private: |
77 | /// Function this pass is working on |
78 | Function &F; |
79 | |
80 | /// Dominator Tree Analysis |
81 | DominatorTree &DT; |
82 | |
83 | /// Memory Alias Analyses |
84 | MemorySSA &MSSA; |
85 | |
86 | /// Target Lowering Information |
87 | const TargetLowering &TLI; |
88 | |
89 | /// Target Transform Information |
90 | const TargetTransformInfo &TTI; |
91 | |
92 | /// Find the instruction in sets LIs that dominates all others, return nullptr |
93 | /// if there is none. |
94 | LoadInst *findFirstLoad(const std::set<LoadInst *> &LIs); |
95 | |
96 | /// Replace interleaved load candidates. It does additional |
97 | /// analyses if this makes sense. Returns true on success and false |
98 | /// of nothing has been changed. |
99 | bool combine(std::list<VectorInfo> &InterleavedLoad, |
100 | OptimizationRemarkEmitter &ORE); |
101 | |
102 | /// Given a set of VectorInfo containing candidates for a given interleave |
103 | /// factor, find a set that represents a 'factor' interleaved load. |
104 | bool findPattern(std::list<VectorInfo> &Candidates, |
105 | std::list<VectorInfo> &InterleavedLoad, unsigned Factor, |
106 | const DataLayout &DL); |
107 | }; // InterleavedLoadCombine |
108 | |
109 | /// First Order Polynomial on an n-Bit Integer Value |
110 | /// |
111 | /// Polynomial(Value) = Value * B + A + E*2^(n-e) |
112 | /// |
113 | /// A and B are the coefficients. E*2^(n-e) is an error within 'e' most |
114 | /// significant bits. It is introduced if an exact computation cannot be proven |
115 | /// (e.q. division by 2). |
116 | /// |
117 | /// As part of this optimization multiple loads will be combined. It necessary |
118 | /// to prove that loads are within some relative offset to each other. This |
119 | /// class is used to prove relative offsets of values loaded from memory. |
120 | /// |
121 | /// Representing an integer in this form is sound since addition in two's |
122 | /// complement is associative (trivial) and multiplication distributes over the |
123 | /// addition (see Proof(1) in Polynomial::mul). Further, both operations |
124 | /// commute. |
125 | // |
126 | // Example: |
127 | // declare @fn(i64 %IDX, <4 x float>* %PTR) { |
128 | // %Pa1 = add i64 %IDX, 2 |
129 | // %Pa2 = lshr i64 %Pa1, 1 |
130 | // %Pa3 = getelementptr inbounds <4 x float>, <4 x float>* %PTR, i64 %Pa2 |
131 | // %Va = load <4 x float>, <4 x float>* %Pa3 |
132 | // |
133 | // %Pb1 = add i64 %IDX, 4 |
134 | // %Pb2 = lshr i64 %Pb1, 1 |
135 | // %Pb3 = getelementptr inbounds <4 x float>, <4 x float>* %PTR, i64 %Pb2 |
136 | // %Vb = load <4 x float>, <4 x float>* %Pb3 |
137 | // ... } |
138 | // |
139 | // The goal is to prove that two loads load consecutive addresses. |
140 | // |
141 | // In this case the polynomials are constructed by the following |
142 | // steps. |
143 | // |
144 | // The number tag #e specifies the error bits. |
145 | // |
146 | // Pa_0 = %IDX #0 |
147 | // Pa_1 = %IDX + 2 #0 | add 2 |
148 | // Pa_2 = %IDX/2 + 1 #1 | lshr 1 |
149 | // Pa_3 = %IDX/2 + 1 #1 | GEP, step signext to i64 |
150 | // Pa_4 = (%IDX/2)*16 + 16 #0 | GEP, multiply index by sizeof(4) for floats |
151 | // Pa_5 = (%IDX/2)*16 + 16 #0 | GEP, add offset of leading components |
152 | // |
153 | // Pb_0 = %IDX #0 |
154 | // Pb_1 = %IDX + 4 #0 | add 2 |
155 | // Pb_2 = %IDX/2 + 2 #1 | lshr 1 |
156 | // Pb_3 = %IDX/2 + 2 #1 | GEP, step signext to i64 |
157 | // Pb_4 = (%IDX/2)*16 + 32 #0 | GEP, multiply index by sizeof(4) for floats |
158 | // Pb_5 = (%IDX/2)*16 + 16 #0 | GEP, add offset of leading components |
159 | // |
160 | // Pb_5 - Pa_5 = 16 #0 | subtract to get the offset |
161 | // |
162 | // Remark: %PTR is not maintained within this class. So in this instance the |
163 | // offset of 16 can only be assumed if the pointers are equal. |
164 | // |
165 | class Polynomial { |
166 | /// Operations on B |
167 | enum BOps { |
168 | LShr, |
169 | Mul, |
170 | SExt, |
171 | Trunc, |
172 | }; |
173 | |
174 | /// Number of Error Bits e |
175 | unsigned ErrorMSBs = (unsigned)-1; |
176 | |
177 | /// Value |
178 | Value *V = nullptr; |
179 | |
180 | /// Coefficient B |
181 | SmallVector<std::pair<BOps, APInt>, 4> B; |
182 | |
183 | /// Coefficient A |
184 | APInt A; |
185 | |
186 | public: |
187 | Polynomial(Value *V) : V(V) { |
188 | IntegerType *Ty = dyn_cast<IntegerType>(Val: V->getType()); |
189 | if (Ty) { |
190 | ErrorMSBs = 0; |
191 | this->V = V; |
192 | A = APInt(Ty->getBitWidth(), 0); |
193 | } |
194 | } |
195 | |
196 | Polynomial(const APInt &A, unsigned ErrorMSBs = 0) |
197 | : ErrorMSBs(ErrorMSBs), A(A) {} |
198 | |
199 | Polynomial(unsigned BitWidth, uint64_t A, unsigned ErrorMSBs = 0) |
200 | : ErrorMSBs(ErrorMSBs), A(BitWidth, A) {} |
201 | |
202 | Polynomial() = default; |
203 | |
204 | /// Increment and clamp the number of undefined bits. |
205 | void incErrorMSBs(unsigned amt) { |
206 | if (ErrorMSBs == (unsigned)-1) |
207 | return; |
208 | |
209 | ErrorMSBs += amt; |
210 | if (ErrorMSBs > A.getBitWidth()) |
211 | ErrorMSBs = A.getBitWidth(); |
212 | } |
213 | |
214 | /// Decrement and clamp the number of undefined bits. |
215 | void decErrorMSBs(unsigned amt) { |
216 | if (ErrorMSBs == (unsigned)-1) |
217 | return; |
218 | |
219 | if (ErrorMSBs > amt) |
220 | ErrorMSBs -= amt; |
221 | else |
222 | ErrorMSBs = 0; |
223 | } |
224 | |
225 | /// Apply an add on the polynomial |
226 | Polynomial &add(const APInt &C) { |
227 | // Note: Addition is associative in two's complement even when in case of |
228 | // signed overflow. |
229 | // |
230 | // Error bits can only propagate into higher significant bits. As these are |
231 | // already regarded as undefined, there is no change. |
232 | // |
233 | // Theorem: Adding a constant to a polynomial does not change the error |
234 | // term. |
235 | // |
236 | // Proof: |
237 | // |
238 | // Since the addition is associative and commutes: |
239 | // |
240 | // (B + A + E*2^(n-e)) + C = B + (A + C) + E*2^(n-e) |
241 | // [qed] |
242 | |
243 | if (C.getBitWidth() != A.getBitWidth()) { |
244 | ErrorMSBs = (unsigned)-1; |
245 | return *this; |
246 | } |
247 | |
248 | A += C; |
249 | return *this; |
250 | } |
251 | |
252 | /// Apply a multiplication onto the polynomial. |
253 | Polynomial &mul(const APInt &C) { |
254 | // Note: Multiplication distributes over the addition |
255 | // |
256 | // Theorem: Multiplication distributes over the addition |
257 | // |
258 | // Proof(1): |
259 | // |
260 | // (B+A)*C =- |
261 | // = (B + A) + (B + A) + .. {C Times} |
262 | // addition is associative and commutes, hence |
263 | // = B + B + .. {C Times} .. + A + A + .. {C times} |
264 | // = B*C + A*C |
265 | // (see (function add) for signed values and overflows) |
266 | // [qed] |
267 | // |
268 | // Theorem: If C has c trailing zeros, errors bits in A or B are shifted out |
269 | // to the left. |
270 | // |
271 | // Proof(2): |
272 | // |
273 | // Let B' and A' be the n-Bit inputs with some unknown errors EA, |
274 | // EB at e leading bits. B' and A' can be written down as: |
275 | // |
276 | // B' = B + 2^(n-e)*EB |
277 | // A' = A + 2^(n-e)*EA |
278 | // |
279 | // Let C' be an input with c trailing zero bits. C' can be written as |
280 | // |
281 | // C' = C*2^c |
282 | // |
283 | // Therefore we can compute the result by using distributivity and |
284 | // commutativity. |
285 | // |
286 | // (B'*C' + A'*C') = [B + 2^(n-e)*EB] * C' + [A + 2^(n-e)*EA] * C' = |
287 | // = [B + 2^(n-e)*EB + A + 2^(n-e)*EA] * C' = |
288 | // = (B'+A') * C' = |
289 | // = [B + 2^(n-e)*EB + A + 2^(n-e)*EA] * C' = |
290 | // = [B + A + 2^(n-e)*EB + 2^(n-e)*EA] * C' = |
291 | // = (B + A) * C' + [2^(n-e)*EB + 2^(n-e)*EA)] * C' = |
292 | // = (B + A) * C' + [2^(n-e)*EB + 2^(n-e)*EA)] * C*2^c = |
293 | // = (B + A) * C' + C*(EB + EA)*2^(n-e)*2^c = |
294 | // |
295 | // Let EC be the final error with EC = C*(EB + EA) |
296 | // |
297 | // = (B + A)*C' + EC*2^(n-e)*2^c = |
298 | // = (B + A)*C' + EC*2^(n-(e-c)) |
299 | // |
300 | // Since EC is multiplied by 2^(n-(e-c)) the resulting error contains c |
301 | // less error bits than the input. c bits are shifted out to the left. |
302 | // [qed] |
303 | |
304 | if (C.getBitWidth() != A.getBitWidth()) { |
305 | ErrorMSBs = (unsigned)-1; |
306 | return *this; |
307 | } |
308 | |
309 | // Multiplying by one is a no-op. |
310 | if (C.isOne()) { |
311 | return *this; |
312 | } |
313 | |
314 | // Multiplying by zero removes the coefficient B and defines all bits. |
315 | if (C.isZero()) { |
316 | ErrorMSBs = 0; |
317 | deleteB(); |
318 | } |
319 | |
320 | // See Proof(2): Trailing zero bits indicate a left shift. This removes |
321 | // leading bits from the result even if they are undefined. |
322 | decErrorMSBs(amt: C.countr_zero()); |
323 | |
324 | A *= C; |
325 | pushBOperation(Op: Mul, C); |
326 | return *this; |
327 | } |
328 | |
329 | /// Apply a logical shift right on the polynomial |
330 | Polynomial &lshr(const APInt &C) { |
331 | // Theorem(1): (B + A + E*2^(n-e)) >> 1 => (B >> 1) + (A >> 1) + E'*2^(n-e') |
332 | // where |
333 | // e' = e + 1, |
334 | // E is a e-bit number, |
335 | // E' is a e'-bit number, |
336 | // holds under the following precondition: |
337 | // pre(1): A % 2 = 0 |
338 | // pre(2): e < n, (see Theorem(2) for the trivial case with e=n) |
339 | // where >> expresses a logical shift to the right, with adding zeros. |
340 | // |
341 | // We need to show that for every, E there is a E' |
342 | // |
343 | // B = b_h * 2^(n-1) + b_m * 2 + b_l |
344 | // A = a_h * 2^(n-1) + a_m * 2 (pre(1)) |
345 | // |
346 | // where a_h, b_h, b_l are single bits, and a_m, b_m are (n-2) bit numbers |
347 | // |
348 | // Let X = (B + A + E*2^(n-e)) >> 1 |
349 | // Let Y = (B >> 1) + (A >> 1) + E*2^(n-e) >> 1 |
350 | // |
351 | // X = [B + A + E*2^(n-e)] >> 1 = |
352 | // = [ b_h * 2^(n-1) + b_m * 2 + b_l + |
353 | // + a_h * 2^(n-1) + a_m * 2 + |
354 | // + E * 2^(n-e) ] >> 1 = |
355 | // |
356 | // The sum is built by putting the overflow of [a_m + b+n] into the term |
357 | // 2^(n-1). As there are no more bits beyond 2^(n-1) the overflow within |
358 | // this bit is discarded. This is expressed by % 2. |
359 | // |
360 | // The bit in position 0 cannot overflow into the term (b_m + a_m). |
361 | // |
362 | // = [ ([b_h + a_h + (b_m + a_m) >> (n-2)] % 2) * 2^(n-1) + |
363 | // + ((b_m + a_m) % 2^(n-2)) * 2 + |
364 | // + b_l + E * 2^(n-e) ] >> 1 = |
365 | // |
366 | // The shift is computed by dividing the terms by 2 and by cutting off |
367 | // b_l. |
368 | // |
369 | // = ([b_h + a_h + (b_m + a_m) >> (n-2)] % 2) * 2^(n-2) + |
370 | // + ((b_m + a_m) % 2^(n-2)) + |
371 | // + E * 2^(n-(e+1)) = |
372 | // |
373 | // by the definition in the Theorem e+1 = e' |
374 | // |
375 | // = ([b_h + a_h + (b_m + a_m) >> (n-2)] % 2) * 2^(n-2) + |
376 | // + ((b_m + a_m) % 2^(n-2)) + |
377 | // + E * 2^(n-e') = |
378 | // |
379 | // Compute Y by applying distributivity first |
380 | // |
381 | // Y = (B >> 1) + (A >> 1) + E*2^(n-e') = |
382 | // = (b_h * 2^(n-1) + b_m * 2 + b_l) >> 1 + |
383 | // + (a_h * 2^(n-1) + a_m * 2) >> 1 + |
384 | // + E * 2^(n-e) >> 1 = |
385 | // |
386 | // Again, the shift is computed by dividing the terms by 2 and by cutting |
387 | // off b_l. |
388 | // |
389 | // = b_h * 2^(n-2) + b_m + |
390 | // + a_h * 2^(n-2) + a_m + |
391 | // + E * 2^(n-(e+1)) = |
392 | // |
393 | // Again, the sum is built by putting the overflow of [a_m + b+n] into |
394 | // the term 2^(n-1). But this time there is room for a second bit in the |
395 | // term 2^(n-2) we add this bit to a new term and denote it o_h in a |
396 | // second step. |
397 | // |
398 | // = ([b_h + a_h + (b_m + a_m) >> (n-2)] >> 1) * 2^(n-1) + |
399 | // + ([b_h + a_h + (b_m + a_m) >> (n-2)] % 2) * 2^(n-2) + |
400 | // + ((b_m + a_m) % 2^(n-2)) + |
401 | // + E * 2^(n-(e+1)) = |
402 | // |
403 | // Let o_h = [b_h + a_h + (b_m + a_m) >> (n-2)] >> 1 |
404 | // Further replace e+1 by e'. |
405 | // |
406 | // = o_h * 2^(n-1) + |
407 | // + ([b_h + a_h + (b_m + a_m) >> (n-2)] % 2) * 2^(n-2) + |
408 | // + ((b_m + a_m) % 2^(n-2)) + |
409 | // + E * 2^(n-e') = |
410 | // |
411 | // Move o_h into the error term and construct E'. To ensure that there is |
412 | // no 2^x with negative x, this step requires pre(2) (e < n). |
413 | // |
414 | // = ([b_h + a_h + (b_m + a_m) >> (n-2)] % 2) * 2^(n-2) + |
415 | // + ((b_m + a_m) % 2^(n-2)) + |
416 | // + o_h * 2^(e'-1) * 2^(n-e') + | pre(2), move 2^(e'-1) |
417 | // | out of the old exponent |
418 | // + E * 2^(n-e') = |
419 | // = ([b_h + a_h + (b_m + a_m) >> (n-2)] % 2) * 2^(n-2) + |
420 | // + ((b_m + a_m) % 2^(n-2)) + |
421 | // + [o_h * 2^(e'-1) + E] * 2^(n-e') + | move 2^(e'-1) out of |
422 | // | the old exponent |
423 | // |
424 | // Let E' = o_h * 2^(e'-1) + E |
425 | // |
426 | // = ([b_h + a_h + (b_m + a_m) >> (n-2)] % 2) * 2^(n-2) + |
427 | // + ((b_m + a_m) % 2^(n-2)) + |
428 | // + E' * 2^(n-e') |
429 | // |
430 | // Because X and Y are distinct only in there error terms and E' can be |
431 | // constructed as shown the theorem holds. |
432 | // [qed] |
433 | // |
434 | // For completeness in case of the case e=n it is also required to show that |
435 | // distributivity can be applied. |
436 | // |
437 | // In this case Theorem(1) transforms to (the pre-condition on A can also be |
438 | // dropped) |
439 | // |
440 | // Theorem(2): (B + A + E) >> 1 => (B >> 1) + (A >> 1) + E' |
441 | // where |
442 | // A, B, E, E' are two's complement numbers with the same bit |
443 | // width |
444 | // |
445 | // Let A + B + E = X |
446 | // Let (B >> 1) + (A >> 1) = Y |
447 | // |
448 | // Therefore we need to show that for every X and Y there is an E' which |
449 | // makes the equation |
450 | // |
451 | // X = Y + E' |
452 | // |
453 | // hold. This is trivially the case for E' = X - Y. |
454 | // |
455 | // [qed] |
456 | // |
457 | // Remark: Distributing lshr with and arbitrary number n can be expressed as |
458 | // ((((B + A) lshr 1) lshr 1) ... ) {n times}. |
459 | // This construction induces n additional error bits at the left. |
460 | |
461 | if (C.getBitWidth() != A.getBitWidth()) { |
462 | ErrorMSBs = (unsigned)-1; |
463 | return *this; |
464 | } |
465 | |
466 | if (C.isZero()) |
467 | return *this; |
468 | |
469 | // Test if the result will be zero |
470 | unsigned shiftAmt = C.getZExtValue(); |
471 | if (shiftAmt >= C.getBitWidth()) |
472 | return mul(C: APInt(C.getBitWidth(), 0)); |
473 | |
474 | // The proof that shiftAmt LSBs are zero for at least one summand is only |
475 | // possible for the constant number. |
476 | // |
477 | // If this can be proven add shiftAmt to the error counter |
478 | // `ErrorMSBs`. Otherwise set all bits as undefined. |
479 | if (A.countr_zero() < shiftAmt) |
480 | ErrorMSBs = A.getBitWidth(); |
481 | else |
482 | incErrorMSBs(amt: shiftAmt); |
483 | |
484 | // Apply the operation. |
485 | pushBOperation(Op: LShr, C); |
486 | A = A.lshr(shiftAmt); |
487 | |
488 | return *this; |
489 | } |
490 | |
491 | /// Apply a sign-extend or truncate operation on the polynomial. |
492 | Polynomial &sextOrTrunc(unsigned n) { |
493 | if (n < A.getBitWidth()) { |
494 | // Truncate: Clearly undefined Bits on the MSB side are removed |
495 | // if there are any. |
496 | decErrorMSBs(amt: A.getBitWidth() - n); |
497 | A = A.trunc(width: n); |
498 | pushBOperation(Op: Trunc, C: APInt(sizeof(n) * 8, n)); |
499 | } |
500 | if (n > A.getBitWidth()) { |
501 | // Extend: Clearly extending first and adding later is different |
502 | // to adding first and extending later in all extended bits. |
503 | incErrorMSBs(amt: n - A.getBitWidth()); |
504 | A = A.sext(width: n); |
505 | pushBOperation(Op: SExt, C: APInt(sizeof(n) * 8, n)); |
506 | } |
507 | |
508 | return *this; |
509 | } |
510 | |
511 | /// Test if there is a coefficient B. |
512 | bool isFirstOrder() const { return V != nullptr; } |
513 | |
514 | /// Test coefficient B of two Polynomials are equal. |
515 | bool isCompatibleTo(const Polynomial &o) const { |
516 | // The polynomial use different bit width. |
517 | if (A.getBitWidth() != o.A.getBitWidth()) |
518 | return false; |
519 | |
520 | // If neither Polynomial has the Coefficient B. |
521 | if (!isFirstOrder() && !o.isFirstOrder()) |
522 | return true; |
523 | |
524 | // The index variable is different. |
525 | if (V != o.V) |
526 | return false; |
527 | |
528 | // Check the operations. |
529 | if (B.size() != o.B.size()) |
530 | return false; |
531 | |
532 | auto *ob = o.B.begin(); |
533 | for (const auto &b : B) { |
534 | if (b != *ob) |
535 | return false; |
536 | ob++; |
537 | } |
538 | |
539 | return true; |
540 | } |
541 | |
542 | /// Subtract two polynomials, return an undefined polynomial if |
543 | /// subtraction is not possible. |
544 | Polynomial operator-(const Polynomial &o) const { |
545 | // Return an undefined polynomial if incompatible. |
546 | if (!isCompatibleTo(o)) |
547 | return Polynomial(); |
548 | |
549 | // If the polynomials are compatible (meaning they have the same |
550 | // coefficient on B), B is eliminated. Thus a polynomial solely |
551 | // containing A is returned |
552 | return Polynomial(A - o.A, std::max(a: ErrorMSBs, b: o.ErrorMSBs)); |
553 | } |
554 | |
555 | /// Subtract a constant from a polynomial, |
556 | Polynomial operator-(uint64_t C) const { |
557 | Polynomial Result(*this); |
558 | Result.A -= C; |
559 | return Result; |
560 | } |
561 | |
562 | /// Add a constant to a polynomial, |
563 | Polynomial operator+(uint64_t C) const { |
564 | Polynomial Result(*this); |
565 | Result.A += C; |
566 | return Result; |
567 | } |
568 | |
569 | /// Returns true if it can be proven that two Polynomials are equal. |
570 | bool isProvenEqualTo(const Polynomial &o) { |
571 | // Subtract both polynomials and test if it is fully defined and zero. |
572 | Polynomial r = *this - o; |
573 | return (r.ErrorMSBs == 0) && (!r.isFirstOrder()) && (r.A.isZero()); |
574 | } |
575 | |
576 | /// Print the polynomial into a stream. |
577 | void print(raw_ostream &OS) const { |
578 | OS << "[{#ErrBits:" << ErrorMSBs << "} " ; |
579 | |
580 | if (V) { |
581 | for (auto b : B) |
582 | OS << "(" ; |
583 | OS << "(" << *V << ") " ; |
584 | |
585 | for (auto b : B) { |
586 | switch (b.first) { |
587 | case LShr: |
588 | OS << "LShr " ; |
589 | break; |
590 | case Mul: |
591 | OS << "Mul " ; |
592 | break; |
593 | case SExt: |
594 | OS << "SExt " ; |
595 | break; |
596 | case Trunc: |
597 | OS << "Trunc " ; |
598 | break; |
599 | } |
600 | |
601 | OS << b.second << ") " ; |
602 | } |
603 | } |
604 | |
605 | OS << "+ " << A << "]" ; |
606 | } |
607 | |
608 | private: |
609 | void deleteB() { |
610 | V = nullptr; |
611 | B.clear(); |
612 | } |
613 | |
614 | void pushBOperation(const BOps Op, const APInt &C) { |
615 | if (isFirstOrder()) { |
616 | B.push_back(Elt: std::make_pair(x: Op, y: C)); |
617 | return; |
618 | } |
619 | } |
620 | }; |
621 | |
622 | #ifndef NDEBUG |
623 | static raw_ostream &operator<<(raw_ostream &OS, const Polynomial &S) { |
624 | S.print(OS); |
625 | return OS; |
626 | } |
627 | #endif |
628 | |
629 | /// VectorInfo stores abstract the following information for each vector |
630 | /// element: |
631 | /// |
632 | /// 1) The memory address loaded into the element as Polynomial |
633 | /// 2) a set of load instruction necessary to construct the vector, |
634 | /// 3) a set of all other instructions that are necessary to create the vector and |
635 | /// 4) a pointer value that can be used as relative base for all elements. |
636 | struct VectorInfo { |
637 | private: |
638 | VectorInfo(const VectorInfo &c) : VTy(c.VTy) { |
639 | llvm_unreachable( |
640 | "Copying VectorInfo is neither implemented nor necessary," ); |
641 | } |
642 | |
643 | public: |
644 | /// Information of a Vector Element |
645 | struct ElementInfo { |
646 | /// Offset Polynomial. |
647 | Polynomial Ofs; |
648 | |
649 | /// The Load Instruction used to Load the entry. LI is null if the pointer |
650 | /// of the load instruction does not point on to the entry |
651 | LoadInst *LI; |
652 | |
653 | ElementInfo(Polynomial Offset = Polynomial(), LoadInst *LI = nullptr) |
654 | : Ofs(Offset), LI(LI) {} |
655 | }; |
656 | |
657 | /// Basic-block the load instructions are within |
658 | BasicBlock *BB = nullptr; |
659 | |
660 | /// Pointer value of all participation load instructions |
661 | Value *PV = nullptr; |
662 | |
663 | /// Participating load instructions |
664 | std::set<LoadInst *> LIs; |
665 | |
666 | /// Participating instructions |
667 | std::set<Instruction *> Is; |
668 | |
669 | /// Final shuffle-vector instruction |
670 | ShuffleVectorInst *SVI = nullptr; |
671 | |
672 | /// Information of the offset for each vector element |
673 | ElementInfo *EI; |
674 | |
675 | /// Vector Type |
676 | FixedVectorType *const VTy; |
677 | |
678 | VectorInfo(FixedVectorType *VTy) : VTy(VTy) { |
679 | EI = new ElementInfo[VTy->getNumElements()]; |
680 | } |
681 | |
682 | VectorInfo &operator=(const VectorInfo &other) = delete; |
683 | |
684 | virtual ~VectorInfo() { delete[] EI; } |
685 | |
686 | unsigned getDimension() const { return VTy->getNumElements(); } |
687 | |
688 | /// Test if the VectorInfo can be part of an interleaved load with the |
689 | /// specified factor. |
690 | /// |
691 | /// \param Factor of the interleave |
692 | /// \param DL Targets Datalayout |
693 | /// |
694 | /// \returns true if this is possible and false if not |
695 | bool isInterleaved(unsigned Factor, const DataLayout &DL) const { |
696 | unsigned Size = DL.getTypeAllocSize(Ty: VTy->getElementType()); |
697 | for (unsigned i = 1; i < getDimension(); i++) { |
698 | if (!EI[i].Ofs.isProvenEqualTo(o: EI[0].Ofs + i * Factor * Size)) { |
699 | return false; |
700 | } |
701 | } |
702 | return true; |
703 | } |
704 | |
705 | /// Recursively computes the vector information stored in V. |
706 | /// |
707 | /// This function delegates the work to specialized implementations |
708 | /// |
709 | /// \param V Value to operate on |
710 | /// \param Result Result of the computation |
711 | /// |
712 | /// \returns false if no sensible information can be gathered. |
713 | static bool compute(Value *V, VectorInfo &Result, const DataLayout &DL) { |
714 | ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Val: V); |
715 | if (SVI) |
716 | return computeFromSVI(SVI, Result, DL); |
717 | LoadInst *LI = dyn_cast<LoadInst>(Val: V); |
718 | if (LI) |
719 | return computeFromLI(LI, Result, DL); |
720 | BitCastInst *BCI = dyn_cast<BitCastInst>(Val: V); |
721 | if (BCI) |
722 | return computeFromBCI(BCI, Result, DL); |
723 | return false; |
724 | } |
725 | |
726 | /// BitCastInst specialization to compute the vector information. |
727 | /// |
728 | /// \param BCI BitCastInst to operate on |
729 | /// \param Result Result of the computation |
730 | /// |
731 | /// \returns false if no sensible information can be gathered. |
732 | static bool computeFromBCI(BitCastInst *BCI, VectorInfo &Result, |
733 | const DataLayout &DL) { |
734 | Instruction *Op = dyn_cast<Instruction>(Val: BCI->getOperand(i_nocapture: 0)); |
735 | |
736 | if (!Op) |
737 | return false; |
738 | |
739 | FixedVectorType *VTy = dyn_cast<FixedVectorType>(Val: Op->getType()); |
740 | if (!VTy) |
741 | return false; |
742 | |
743 | // We can only cast from large to smaller vectors |
744 | if (Result.VTy->getNumElements() % VTy->getNumElements()) |
745 | return false; |
746 | |
747 | unsigned Factor = Result.VTy->getNumElements() / VTy->getNumElements(); |
748 | unsigned NewSize = DL.getTypeAllocSize(Ty: Result.VTy->getElementType()); |
749 | unsigned OldSize = DL.getTypeAllocSize(Ty: VTy->getElementType()); |
750 | |
751 | if (NewSize * Factor != OldSize) |
752 | return false; |
753 | |
754 | VectorInfo Old(VTy); |
755 | if (!compute(V: Op, Result&: Old, DL)) |
756 | return false; |
757 | |
758 | for (unsigned i = 0; i < Result.VTy->getNumElements(); i += Factor) { |
759 | for (unsigned j = 0; j < Factor; j++) { |
760 | Result.EI[i + j] = |
761 | ElementInfo(Old.EI[i / Factor].Ofs + j * NewSize, |
762 | j == 0 ? Old.EI[i / Factor].LI : nullptr); |
763 | } |
764 | } |
765 | |
766 | Result.BB = Old.BB; |
767 | Result.PV = Old.PV; |
768 | Result.LIs.insert(first: Old.LIs.begin(), last: Old.LIs.end()); |
769 | Result.Is.insert(first: Old.Is.begin(), last: Old.Is.end()); |
770 | Result.Is.insert(x: BCI); |
771 | Result.SVI = nullptr; |
772 | |
773 | return true; |
774 | } |
775 | |
776 | /// ShuffleVectorInst specialization to compute vector information. |
777 | /// |
778 | /// \param SVI ShuffleVectorInst to operate on |
779 | /// \param Result Result of the computation |
780 | /// |
781 | /// Compute the left and the right side vector information and merge them by |
782 | /// applying the shuffle operation. This function also ensures that the left |
783 | /// and right side have compatible loads. This means that all loads are with |
784 | /// in the same basic block and are based on the same pointer. |
785 | /// |
786 | /// \returns false if no sensible information can be gathered. |
787 | static bool computeFromSVI(ShuffleVectorInst *SVI, VectorInfo &Result, |
788 | const DataLayout &DL) { |
789 | FixedVectorType *ArgTy = |
790 | cast<FixedVectorType>(Val: SVI->getOperand(i_nocapture: 0)->getType()); |
791 | |
792 | // Compute the left hand vector information. |
793 | VectorInfo LHS(ArgTy); |
794 | if (!compute(V: SVI->getOperand(i_nocapture: 0), Result&: LHS, DL)) |
795 | LHS.BB = nullptr; |
796 | |
797 | // Compute the right hand vector information. |
798 | VectorInfo RHS(ArgTy); |
799 | if (!compute(V: SVI->getOperand(i_nocapture: 1), Result&: RHS, DL)) |
800 | RHS.BB = nullptr; |
801 | |
802 | // Neither operand produced sensible results? |
803 | if (!LHS.BB && !RHS.BB) |
804 | return false; |
805 | // Only RHS produced sensible results? |
806 | else if (!LHS.BB) { |
807 | Result.BB = RHS.BB; |
808 | Result.PV = RHS.PV; |
809 | } |
810 | // Only LHS produced sensible results? |
811 | else if (!RHS.BB) { |
812 | Result.BB = LHS.BB; |
813 | Result.PV = LHS.PV; |
814 | } |
815 | // Both operands produced sensible results? |
816 | else if ((LHS.BB == RHS.BB) && (LHS.PV == RHS.PV)) { |
817 | Result.BB = LHS.BB; |
818 | Result.PV = LHS.PV; |
819 | } |
820 | // Both operands produced sensible results but they are incompatible. |
821 | else { |
822 | return false; |
823 | } |
824 | |
825 | // Merge and apply the operation on the offset information. |
826 | if (LHS.BB) { |
827 | Result.LIs.insert(first: LHS.LIs.begin(), last: LHS.LIs.end()); |
828 | Result.Is.insert(first: LHS.Is.begin(), last: LHS.Is.end()); |
829 | } |
830 | if (RHS.BB) { |
831 | Result.LIs.insert(first: RHS.LIs.begin(), last: RHS.LIs.end()); |
832 | Result.Is.insert(first: RHS.Is.begin(), last: RHS.Is.end()); |
833 | } |
834 | Result.Is.insert(x: SVI); |
835 | Result.SVI = SVI; |
836 | |
837 | int j = 0; |
838 | for (int i : SVI->getShuffleMask()) { |
839 | assert((i < 2 * (signed)ArgTy->getNumElements()) && |
840 | "Invalid ShuffleVectorInst (index out of bounds)" ); |
841 | |
842 | if (i < 0) |
843 | Result.EI[j] = ElementInfo(); |
844 | else if (i < (signed)ArgTy->getNumElements()) { |
845 | if (LHS.BB) |
846 | Result.EI[j] = LHS.EI[i]; |
847 | else |
848 | Result.EI[j] = ElementInfo(); |
849 | } else { |
850 | if (RHS.BB) |
851 | Result.EI[j] = RHS.EI[i - ArgTy->getNumElements()]; |
852 | else |
853 | Result.EI[j] = ElementInfo(); |
854 | } |
855 | j++; |
856 | } |
857 | |
858 | return true; |
859 | } |
860 | |
861 | /// LoadInst specialization to compute vector information. |
862 | /// |
863 | /// This function also acts as abort condition to the recursion. |
864 | /// |
865 | /// \param LI LoadInst to operate on |
866 | /// \param Result Result of the computation |
867 | /// |
868 | /// \returns false if no sensible information can be gathered. |
869 | static bool computeFromLI(LoadInst *LI, VectorInfo &Result, |
870 | const DataLayout &DL) { |
871 | Value *BasePtr; |
872 | Polynomial Offset; |
873 | |
874 | if (LI->isVolatile()) |
875 | return false; |
876 | |
877 | if (LI->isAtomic()) |
878 | return false; |
879 | |
880 | if (!DL.typeSizeEqualsStoreSize(Ty: Result.VTy->getElementType())) |
881 | return false; |
882 | |
883 | // Get the base polynomial |
884 | computePolynomialFromPointer(Ptr&: *LI->getPointerOperand(), Result&: Offset, BasePtr, DL); |
885 | |
886 | Result.BB = LI->getParent(); |
887 | Result.PV = BasePtr; |
888 | Result.LIs.insert(x: LI); |
889 | Result.Is.insert(x: LI); |
890 | |
891 | for (unsigned i = 0; i < Result.getDimension(); i++) { |
892 | Value *Idx[2] = { |
893 | ConstantInt::get(Ty: Type::getInt32Ty(C&: LI->getContext()), V: 0), |
894 | ConstantInt::get(Ty: Type::getInt32Ty(C&: LI->getContext()), V: i), |
895 | }; |
896 | int64_t Ofs = DL.getIndexedOffsetInType(ElemTy: Result.VTy, Indices: Idx); |
897 | Result.EI[i] = ElementInfo(Offset + Ofs, i == 0 ? LI : nullptr); |
898 | } |
899 | |
900 | return true; |
901 | } |
902 | |
903 | /// Recursively compute polynomial of a value. |
904 | /// |
905 | /// \param BO Input binary operation |
906 | /// \param Result Result polynomial |
907 | static void computePolynomialBinOp(BinaryOperator &BO, Polynomial &Result) { |
908 | Value *LHS = BO.getOperand(i_nocapture: 0); |
909 | Value *RHS = BO.getOperand(i_nocapture: 1); |
910 | |
911 | // Find the RHS Constant if any |
912 | ConstantInt *C = dyn_cast<ConstantInt>(Val: RHS); |
913 | if ((!C) && BO.isCommutative()) { |
914 | C = dyn_cast<ConstantInt>(Val: LHS); |
915 | if (C) |
916 | std::swap(a&: LHS, b&: RHS); |
917 | } |
918 | |
919 | switch (BO.getOpcode()) { |
920 | case Instruction::Add: |
921 | if (!C) |
922 | break; |
923 | |
924 | computePolynomial(V&: *LHS, Result); |
925 | Result.add(C: C->getValue()); |
926 | return; |
927 | |
928 | case Instruction::LShr: |
929 | if (!C) |
930 | break; |
931 | |
932 | computePolynomial(V&: *LHS, Result); |
933 | Result.lshr(C: C->getValue()); |
934 | return; |
935 | |
936 | default: |
937 | break; |
938 | } |
939 | |
940 | Result = Polynomial(&BO); |
941 | } |
942 | |
943 | /// Recursively compute polynomial of a value |
944 | /// |
945 | /// \param V input value |
946 | /// \param Result result polynomial |
947 | static void computePolynomial(Value &V, Polynomial &Result) { |
948 | if (auto *BO = dyn_cast<BinaryOperator>(Val: &V)) |
949 | computePolynomialBinOp(BO&: *BO, Result); |
950 | else |
951 | Result = Polynomial(&V); |
952 | } |
953 | |
954 | /// Compute the Polynomial representation of a Pointer type. |
955 | /// |
956 | /// \param Ptr input pointer value |
957 | /// \param Result result polynomial |
958 | /// \param BasePtr pointer the polynomial is based on |
959 | /// \param DL Datalayout of the target machine |
960 | static void computePolynomialFromPointer(Value &Ptr, Polynomial &Result, |
961 | Value *&BasePtr, |
962 | const DataLayout &DL) { |
963 | // Not a pointer type? Return an undefined polynomial |
964 | PointerType *PtrTy = dyn_cast<PointerType>(Val: Ptr.getType()); |
965 | if (!PtrTy) { |
966 | Result = Polynomial(); |
967 | BasePtr = nullptr; |
968 | return; |
969 | } |
970 | unsigned PointerBits = |
971 | DL.getIndexSizeInBits(AS: PtrTy->getPointerAddressSpace()); |
972 | |
973 | /// Skip pointer casts. Return Zero polynomial otherwise |
974 | if (isa<CastInst>(Val: &Ptr)) { |
975 | CastInst &CI = *cast<CastInst>(Val: &Ptr); |
976 | switch (CI.getOpcode()) { |
977 | case Instruction::BitCast: |
978 | computePolynomialFromPointer(Ptr&: *CI.getOperand(i_nocapture: 0), Result, BasePtr, DL); |
979 | break; |
980 | default: |
981 | BasePtr = &Ptr; |
982 | Polynomial(PointerBits, 0); |
983 | break; |
984 | } |
985 | } |
986 | /// Resolve GetElementPtrInst. |
987 | else if (isa<GetElementPtrInst>(Val: &Ptr)) { |
988 | GetElementPtrInst &GEP = *cast<GetElementPtrInst>(Val: &Ptr); |
989 | |
990 | APInt BaseOffset(PointerBits, 0); |
991 | |
992 | // Check if we can compute the Offset with accumulateConstantOffset |
993 | if (GEP.accumulateConstantOffset(DL, Offset&: BaseOffset)) { |
994 | Result = Polynomial(BaseOffset); |
995 | BasePtr = GEP.getPointerOperand(); |
996 | return; |
997 | } else { |
998 | // Otherwise we allow that the last index operand of the GEP is |
999 | // non-constant. |
1000 | unsigned idxOperand, e; |
1001 | SmallVector<Value *, 4> Indices; |
1002 | for (idxOperand = 1, e = GEP.getNumOperands(); idxOperand < e; |
1003 | idxOperand++) { |
1004 | ConstantInt *IDX = dyn_cast<ConstantInt>(Val: GEP.getOperand(i_nocapture: idxOperand)); |
1005 | if (!IDX) |
1006 | break; |
1007 | Indices.push_back(Elt: IDX); |
1008 | } |
1009 | |
1010 | // It must also be the last operand. |
1011 | if (idxOperand + 1 != e) { |
1012 | Result = Polynomial(); |
1013 | BasePtr = nullptr; |
1014 | return; |
1015 | } |
1016 | |
1017 | // Compute the polynomial of the index operand. |
1018 | computePolynomial(V&: *GEP.getOperand(i_nocapture: idxOperand), Result); |
1019 | |
1020 | // Compute base offset from zero based index, excluding the last |
1021 | // variable operand. |
1022 | BaseOffset = |
1023 | DL.getIndexedOffsetInType(ElemTy: GEP.getSourceElementType(), Indices); |
1024 | |
1025 | // Apply the operations of GEP to the polynomial. |
1026 | unsigned ResultSize = DL.getTypeAllocSize(Ty: GEP.getResultElementType()); |
1027 | Result.sextOrTrunc(n: PointerBits); |
1028 | Result.mul(C: APInt(PointerBits, ResultSize)); |
1029 | Result.add(C: BaseOffset); |
1030 | BasePtr = GEP.getPointerOperand(); |
1031 | } |
1032 | } |
1033 | // All other instructions are handled by using the value as base pointer and |
1034 | // a zero polynomial. |
1035 | else { |
1036 | BasePtr = &Ptr; |
1037 | Polynomial(DL.getIndexSizeInBits(AS: PtrTy->getPointerAddressSpace()), 0); |
1038 | } |
1039 | } |
1040 | |
1041 | #ifndef NDEBUG |
1042 | void print(raw_ostream &OS) const { |
1043 | if (PV) |
1044 | OS << *PV; |
1045 | else |
1046 | OS << "(none)" ; |
1047 | OS << " + " ; |
1048 | for (unsigned i = 0; i < getDimension(); i++) |
1049 | OS << ((i == 0) ? "[" : ", " ) << EI[i].Ofs; |
1050 | OS << "]" ; |
1051 | } |
1052 | #endif |
1053 | }; |
1054 | |
1055 | } // anonymous namespace |
1056 | |
1057 | bool InterleavedLoadCombineImpl::findPattern( |
1058 | std::list<VectorInfo> &Candidates, std::list<VectorInfo> &InterleavedLoad, |
1059 | unsigned Factor, const DataLayout &DL) { |
1060 | for (auto C0 = Candidates.begin(), E0 = Candidates.end(); C0 != E0; ++C0) { |
1061 | unsigned i; |
1062 | // Try to find an interleaved load using the front of Worklist as first line |
1063 | unsigned Size = DL.getTypeAllocSize(Ty: C0->VTy->getElementType()); |
1064 | |
1065 | // List containing iterators pointing to the VectorInfos of the candidates |
1066 | std::vector<std::list<VectorInfo>::iterator> Res(Factor, Candidates.end()); |
1067 | |
1068 | for (auto C = Candidates.begin(), E = Candidates.end(); C != E; C++) { |
1069 | if (C->VTy != C0->VTy) |
1070 | continue; |
1071 | if (C->BB != C0->BB) |
1072 | continue; |
1073 | if (C->PV != C0->PV) |
1074 | continue; |
1075 | |
1076 | // Check the current value matches any of factor - 1 remaining lines |
1077 | for (i = 1; i < Factor; i++) { |
1078 | if (C->EI[0].Ofs.isProvenEqualTo(o: C0->EI[0].Ofs + i * Size)) { |
1079 | Res[i] = C; |
1080 | } |
1081 | } |
1082 | |
1083 | for (i = 1; i < Factor; i++) { |
1084 | if (Res[i] == Candidates.end()) |
1085 | break; |
1086 | } |
1087 | if (i == Factor) { |
1088 | Res[0] = C0; |
1089 | break; |
1090 | } |
1091 | } |
1092 | |
1093 | if (Res[0] != Candidates.end()) { |
1094 | // Move the result into the output |
1095 | for (unsigned i = 0; i < Factor; i++) { |
1096 | InterleavedLoad.splice(position: InterleavedLoad.end(), x&: Candidates, i: Res[i]); |
1097 | } |
1098 | |
1099 | return true; |
1100 | } |
1101 | } |
1102 | return false; |
1103 | } |
1104 | |
1105 | LoadInst * |
1106 | InterleavedLoadCombineImpl::findFirstLoad(const std::set<LoadInst *> &LIs) { |
1107 | assert(!LIs.empty() && "No load instructions given." ); |
1108 | |
1109 | // All LIs are within the same BB. Select the first for a reference. |
1110 | BasicBlock *BB = (*LIs.begin())->getParent(); |
1111 | BasicBlock::iterator FLI = llvm::find_if( |
1112 | Range&: *BB, P: [&LIs](Instruction &I) -> bool { return is_contained(Range: LIs, Element: &I); }); |
1113 | assert(FLI != BB->end()); |
1114 | |
1115 | return cast<LoadInst>(Val&: FLI); |
1116 | } |
1117 | |
1118 | bool InterleavedLoadCombineImpl::(std::list<VectorInfo> &InterleavedLoad, |
1119 | OptimizationRemarkEmitter &ORE) { |
1120 | LLVM_DEBUG(dbgs() << "Checking interleaved load\n" ); |
1121 | |
1122 | // The insertion point is the LoadInst which loads the first values. The |
1123 | // following tests are used to proof that the combined load can be inserted |
1124 | // just before InsertionPoint. |
1125 | LoadInst *InsertionPoint = InterleavedLoad.front().EI[0].LI; |
1126 | |
1127 | // Test if the offset is computed |
1128 | if (!InsertionPoint) |
1129 | return false; |
1130 | |
1131 | std::set<LoadInst *> LIs; |
1132 | std::set<Instruction *> Is; |
1133 | std::set<Instruction *> SVIs; |
1134 | |
1135 | InstructionCost InterleavedCost; |
1136 | InstructionCost InstructionCost = 0; |
1137 | const TTI::TargetCostKind CostKind = TTI::TCK_SizeAndLatency; |
1138 | |
1139 | // Get the interleave factor |
1140 | unsigned Factor = InterleavedLoad.size(); |
1141 | |
1142 | // Merge all input sets used in analysis |
1143 | for (auto &VI : InterleavedLoad) { |
1144 | // Generate a set of all load instructions to be combined |
1145 | LIs.insert(first: VI.LIs.begin(), last: VI.LIs.end()); |
1146 | |
1147 | // Generate a set of all instructions taking part in load |
1148 | // interleaved. This list excludes the instructions necessary for the |
1149 | // polynomial construction. |
1150 | Is.insert(first: VI.Is.begin(), last: VI.Is.end()); |
1151 | |
1152 | // Generate the set of the final ShuffleVectorInst. |
1153 | SVIs.insert(x: VI.SVI); |
1154 | } |
1155 | |
1156 | // There is nothing to combine. |
1157 | if (LIs.size() < 2) |
1158 | return false; |
1159 | |
1160 | // Test if all participating instruction will be dead after the |
1161 | // transformation. If intermediate results are used, no performance gain can |
1162 | // be expected. Also sum the cost of the Instructions beeing left dead. |
1163 | for (const auto &I : Is) { |
1164 | // Compute the old cost |
1165 | InstructionCost += TTI.getInstructionCost(U: I, CostKind); |
1166 | |
1167 | // The final SVIs are allowed not to be dead, all uses will be replaced |
1168 | if (SVIs.find(x: I) != SVIs.end()) |
1169 | continue; |
1170 | |
1171 | // If there are users outside the set to be eliminated, we abort the |
1172 | // transformation. No gain can be expected. |
1173 | for (auto *U : I->users()) { |
1174 | if (Is.find(x: dyn_cast<Instruction>(Val: U)) == Is.end()) |
1175 | return false; |
1176 | } |
1177 | } |
1178 | |
1179 | // We need to have a valid cost in order to proceed. |
1180 | if (!InstructionCost.isValid()) |
1181 | return false; |
1182 | |
1183 | // We know that all LoadInst are within the same BB. This guarantees that |
1184 | // either everything or nothing is loaded. |
1185 | LoadInst *First = findFirstLoad(LIs); |
1186 | |
1187 | // To be safe that the loads can be combined, iterate over all loads and test |
1188 | // that the corresponding defining access dominates first LI. This guarantees |
1189 | // that there are no aliasing stores in between the loads. |
1190 | auto FMA = MSSA.getMemoryAccess(I: First); |
1191 | for (auto *LI : LIs) { |
1192 | auto MADef = MSSA.getMemoryAccess(I: LI)->getDefiningAccess(); |
1193 | if (!MSSA.dominates(A: MADef, B: FMA)) |
1194 | return false; |
1195 | } |
1196 | assert(!LIs.empty() && "There are no LoadInst to combine" ); |
1197 | |
1198 | // It is necessary that insertion point dominates all final ShuffleVectorInst. |
1199 | for (auto &VI : InterleavedLoad) { |
1200 | if (!DT.dominates(Def: InsertionPoint, User: VI.SVI)) |
1201 | return false; |
1202 | } |
1203 | |
1204 | // All checks are done. Add instructions detectable by InterleavedAccessPass |
1205 | // The old instruction will are left dead. |
1206 | IRBuilder<> Builder(InsertionPoint); |
1207 | Type *ETy = InterleavedLoad.front().SVI->getType()->getElementType(); |
1208 | unsigned ElementsPerSVI = |
1209 | cast<FixedVectorType>(Val: InterleavedLoad.front().SVI->getType()) |
1210 | ->getNumElements(); |
1211 | FixedVectorType *ILTy = FixedVectorType::get(ElementType: ETy, NumElts: Factor * ElementsPerSVI); |
1212 | |
1213 | auto Indices = llvm::to_vector<4>(Range: llvm::seq<unsigned>(Begin: 0, End: Factor)); |
1214 | InterleavedCost = TTI.getInterleavedMemoryOpCost( |
1215 | Opcode: Instruction::Load, VecTy: ILTy, Factor, Indices, Alignment: InsertionPoint->getAlign(), |
1216 | AddressSpace: InsertionPoint->getPointerAddressSpace(), CostKind); |
1217 | |
1218 | if (InterleavedCost >= InstructionCost) { |
1219 | return false; |
1220 | } |
1221 | |
1222 | // Create the wide load and update the MemorySSA. |
1223 | auto Ptr = InsertionPoint->getPointerOperand(); |
1224 | auto LI = Builder.CreateAlignedLoad(Ty: ILTy, Ptr, Align: InsertionPoint->getAlign(), |
1225 | Name: "interleaved.wide.load" ); |
1226 | auto MSSAU = MemorySSAUpdater(&MSSA); |
1227 | MemoryUse *MSSALoad = cast<MemoryUse>(Val: MSSAU.createMemoryAccessBefore( |
1228 | I: LI, Definition: nullptr, InsertPt: MSSA.getMemoryAccess(I: InsertionPoint))); |
1229 | MSSAU.insertUse(Use: MSSALoad, /*RenameUses=*/ true); |
1230 | |
1231 | // Create the final SVIs and replace all uses. |
1232 | int i = 0; |
1233 | for (auto &VI : InterleavedLoad) { |
1234 | SmallVector<int, 4> Mask; |
1235 | for (unsigned j = 0; j < ElementsPerSVI; j++) |
1236 | Mask.push_back(Elt: i + j * Factor); |
1237 | |
1238 | Builder.SetInsertPoint(VI.SVI); |
1239 | auto SVI = Builder.CreateShuffleVector(V: LI, Mask, Name: "interleaved.shuffle" ); |
1240 | VI.SVI->replaceAllUsesWith(V: SVI); |
1241 | i++; |
1242 | } |
1243 | |
1244 | NumInterleavedLoadCombine++; |
1245 | ORE.emit(RemarkBuilder: [&]() { |
1246 | return OptimizationRemark(DEBUG_TYPE, "Combined Interleaved Load" , LI) |
1247 | << "Load interleaved combined with factor " |
1248 | << ore::NV("Factor" , Factor); |
1249 | }); |
1250 | |
1251 | return true; |
1252 | } |
1253 | |
1254 | bool InterleavedLoadCombineImpl::run() { |
1255 | OptimizationRemarkEmitter ORE(&F); |
1256 | bool changed = false; |
1257 | unsigned MaxFactor = TLI.getMaxSupportedInterleaveFactor(); |
1258 | |
1259 | auto &DL = F.getDataLayout(); |
1260 | |
1261 | // Start with the highest factor to avoid combining and recombining. |
1262 | for (unsigned Factor = MaxFactor; Factor >= 2; Factor--) { |
1263 | std::list<VectorInfo> Candidates; |
1264 | |
1265 | for (BasicBlock &BB : F) { |
1266 | for (Instruction &I : BB) { |
1267 | if (auto SVI = dyn_cast<ShuffleVectorInst>(Val: &I)) { |
1268 | // We don't support scalable vectors in this pass. |
1269 | if (isa<ScalableVectorType>(Val: SVI->getType())) |
1270 | continue; |
1271 | |
1272 | Candidates.emplace_back(args: cast<FixedVectorType>(Val: SVI->getType())); |
1273 | |
1274 | if (!VectorInfo::computeFromSVI(SVI, Result&: Candidates.back(), DL)) { |
1275 | Candidates.pop_back(); |
1276 | continue; |
1277 | } |
1278 | |
1279 | if (!Candidates.back().isInterleaved(Factor, DL)) { |
1280 | Candidates.pop_back(); |
1281 | } |
1282 | } |
1283 | } |
1284 | } |
1285 | |
1286 | std::list<VectorInfo> InterleavedLoad; |
1287 | while (findPattern(Candidates, InterleavedLoad, Factor, DL)) { |
1288 | if (combine(InterleavedLoad, ORE)) { |
1289 | changed = true; |
1290 | } else { |
1291 | // Remove the first element of the Interleaved Load but put the others |
1292 | // back on the list and continue searching |
1293 | Candidates.splice(position: Candidates.begin(), x&: InterleavedLoad, |
1294 | first: std::next(x: InterleavedLoad.begin()), |
1295 | last: InterleavedLoad.end()); |
1296 | } |
1297 | InterleavedLoad.clear(); |
1298 | } |
1299 | } |
1300 | |
1301 | return changed; |
1302 | } |
1303 | |
1304 | namespace { |
1305 | /// This pass combines interleaved loads into a pattern detectable by |
1306 | /// InterleavedAccessPass. |
1307 | struct InterleavedLoadCombine : public FunctionPass { |
1308 | static char ID; |
1309 | |
1310 | InterleavedLoadCombine() : FunctionPass(ID) { |
1311 | initializeInterleavedLoadCombinePass(*PassRegistry::getPassRegistry()); |
1312 | } |
1313 | |
1314 | StringRef getPassName() const override { |
1315 | return "Interleaved Load Combine Pass" ; |
1316 | } |
1317 | |
1318 | bool runOnFunction(Function &F) override { |
1319 | if (DisableInterleavedLoadCombine) |
1320 | return false; |
1321 | |
1322 | auto *TPC = getAnalysisIfAvailable<TargetPassConfig>(); |
1323 | if (!TPC) |
1324 | return false; |
1325 | |
1326 | LLVM_DEBUG(dbgs() << "*** " << getPassName() << ": " << F.getName() |
1327 | << "\n" ); |
1328 | |
1329 | return InterleavedLoadCombineImpl( |
1330 | F, getAnalysis<DominatorTreeWrapperPass>().getDomTree(), |
1331 | getAnalysis<MemorySSAWrapperPass>().getMSSA(), |
1332 | getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F), |
1333 | TPC->getTM<TargetMachine>()) |
1334 | .run(); |
1335 | } |
1336 | |
1337 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
1338 | AU.addRequired<MemorySSAWrapperPass>(); |
1339 | AU.addRequired<DominatorTreeWrapperPass>(); |
1340 | AU.addRequired<TargetTransformInfoWrapperPass>(); |
1341 | FunctionPass::getAnalysisUsage(AU); |
1342 | } |
1343 | |
1344 | private: |
1345 | }; |
1346 | } // anonymous namespace |
1347 | |
1348 | PreservedAnalyses |
1349 | InterleavedLoadCombinePass::run(Function &F, FunctionAnalysisManager &FAM) { |
1350 | |
1351 | auto &DT = FAM.getResult<DominatorTreeAnalysis>(IR&: F); |
1352 | auto &MemSSA = FAM.getResult<MemorySSAAnalysis>(IR&: F).getMSSA(); |
1353 | auto &TTI = FAM.getResult<TargetIRAnalysis>(IR&: F); |
1354 | bool Changed = InterleavedLoadCombineImpl(F, DT, MemSSA, TTI, *TM).run(); |
1355 | return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all(); |
1356 | } |
1357 | |
1358 | char InterleavedLoadCombine::ID = 0; |
1359 | |
1360 | INITIALIZE_PASS_BEGIN( |
1361 | InterleavedLoadCombine, DEBUG_TYPE, |
1362 | "Combine interleaved loads into wide loads and shufflevector instructions" , |
1363 | false, false) |
1364 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) |
1365 | INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass) |
1366 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) |
1367 | INITIALIZE_PASS_END( |
1368 | InterleavedLoadCombine, DEBUG_TYPE, |
1369 | "Combine interleaved loads into wide loads and shufflevector instructions" , |
1370 | false, false) |
1371 | |
1372 | FunctionPass * |
1373 | llvm::createInterleavedLoadCombinePass() { |
1374 | auto P = new InterleavedLoadCombine(); |
1375 | return P; |
1376 | } |
1377 | |