xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/SeparateConstOffsetFromGEP.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 //===- SeparateConstOffsetFromGEP.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 // Loop unrolling may create many similar GEPs for array accesses.
10 // e.g., a 2-level loop
11 //
12 // float a[32][32]; // global variable
13 //
14 // for (int i = 0; i < 2; ++i) {
15 //   for (int j = 0; j < 2; ++j) {
16 //     ...
17 //     ... = a[x + i][y + j];
18 //     ...
19 //   }
20 // }
21 //
22 // will probably be unrolled to:
23 //
24 // gep %a, 0, %x, %y; load
25 // gep %a, 0, %x, %y + 1; load
26 // gep %a, 0, %x + 1, %y; load
27 // gep %a, 0, %x + 1, %y + 1; load
28 //
29 // LLVM's GVN does not use partial redundancy elimination yet, and is thus
30 // unable to reuse (gep %a, 0, %x, %y). As a result, this misoptimization incurs
31 // significant slowdown in targets with limited addressing modes. For instance,
32 // because the PTX target does not support the reg+reg addressing mode, the
33 // NVPTX backend emits PTX code that literally computes the pointer address of
34 // each GEP, wasting tons of registers. It emits the following PTX for the
35 // first load and similar PTX for other loads.
36 //
37 // mov.u32         %r1, %x;
38 // mov.u32         %r2, %y;
39 // mul.wide.u32    %rl2, %r1, 128;
40 // mov.u64         %rl3, a;
41 // add.s64         %rl4, %rl3, %rl2;
42 // mul.wide.u32    %rl5, %r2, 4;
43 // add.s64         %rl6, %rl4, %rl5;
44 // ld.global.f32   %f1, [%rl6];
45 //
46 // To reduce the register pressure, the optimization implemented in this file
47 // merges the common part of a group of GEPs, so we can compute each pointer
48 // address by adding a simple offset to the common part, saving many registers.
49 //
50 // It works by splitting each GEP into a variadic base and a constant offset.
51 // The variadic base can be computed once and reused by multiple GEPs, and the
52 // constant offsets can be nicely folded into the reg+immediate addressing mode
53 // (supported by most targets) without using any extra register.
54 //
55 // For instance, we transform the four GEPs and four loads in the above example
56 // into:
57 //
58 // base = gep a, 0, x, y
59 // load base
60 // load base + 1  * sizeof(float)
61 // load base + 32 * sizeof(float)
62 // load base + 33 * sizeof(float)
63 //
64 // Given the transformed IR, a backend that supports the reg+immediate
65 // addressing mode can easily fold the pointer arithmetics into the loads. For
66 // example, the NVPTX backend can easily fold the pointer arithmetics into the
67 // ld.global.f32 instructions, and the resultant PTX uses much fewer registers.
68 //
69 // mov.u32         %r1, %tid.x;
70 // mov.u32         %r2, %tid.y;
71 // mul.wide.u32    %rl2, %r1, 128;
72 // mov.u64         %rl3, a;
73 // add.s64         %rl4, %rl3, %rl2;
74 // mul.wide.u32    %rl5, %r2, 4;
75 // add.s64         %rl6, %rl4, %rl5;
76 // ld.global.f32   %f1, [%rl6]; // so far the same as unoptimized PTX
77 // ld.global.f32   %f2, [%rl6+4]; // much better
78 // ld.global.f32   %f3, [%rl6+128]; // much better
79 // ld.global.f32   %f4, [%rl6+132]; // much better
80 //
81 // Another improvement enabled by the LowerGEP flag is to lower a GEP with
82 // multiple indices to either multiple GEPs with a single index or arithmetic
83 // operations (depending on whether the target uses alias analysis in codegen).
84 // Such transformation can have following benefits:
85 // (1) It can always extract constants in the indices of structure type.
86 // (2) After such Lowering, there are more optimization opportunities such as
87 //     CSE, LICM and CGP.
88 //
89 // E.g. The following GEPs have multiple indices:
90 //  BB1:
91 //    %p = getelementptr [10 x %struct]* %ptr, i64 %i, i64 %j1, i32 3
92 //    load %p
93 //    ...
94 //  BB2:
95 //    %p2 = getelementptr [10 x %struct]* %ptr, i64 %i, i64 %j1, i32 2
96 //    load %p2
97 //    ...
98 //
99 // We can not do CSE to the common part related to index "i64 %i". Lowering
100 // GEPs can achieve such goals.
101 // If the target does not use alias analysis in codegen, this pass will
102 // lower a GEP with multiple indices into arithmetic operations:
103 //  BB1:
104 //    %1 = ptrtoint [10 x %struct]* %ptr to i64    ; CSE opportunity
105 //    %2 = mul i64 %i, length_of_10xstruct         ; CSE opportunity
106 //    %3 = add i64 %1, %2                          ; CSE opportunity
107 //    %4 = mul i64 %j1, length_of_struct
108 //    %5 = add i64 %3, %4
109 //    %6 = add i64 %3, struct_field_3              ; Constant offset
110 //    %p = inttoptr i64 %6 to i32*
111 //    load %p
112 //    ...
113 //  BB2:
114 //    %7 = ptrtoint [10 x %struct]* %ptr to i64    ; CSE opportunity
115 //    %8 = mul i64 %i, length_of_10xstruct         ; CSE opportunity
116 //    %9 = add i64 %7, %8                          ; CSE opportunity
117 //    %10 = mul i64 %j2, length_of_struct
118 //    %11 = add i64 %9, %10
119 //    %12 = add i64 %11, struct_field_2            ; Constant offset
120 //    %p = inttoptr i64 %12 to i32*
121 //    load %p2
122 //    ...
123 //
124 // If the target uses alias analysis in codegen, this pass will lower a GEP
125 // with multiple indices into multiple GEPs with a single index:
126 //  BB1:
127 //    %1 = bitcast [10 x %struct]* %ptr to i8*     ; CSE opportunity
128 //    %2 = mul i64 %i, length_of_10xstruct         ; CSE opportunity
129 //    %3 = getelementptr i8* %1, i64 %2            ; CSE opportunity
130 //    %4 = mul i64 %j1, length_of_struct
131 //    %5 = getelementptr i8* %3, i64 %4
132 //    %6 = getelementptr i8* %5, struct_field_3    ; Constant offset
133 //    %p = bitcast i8* %6 to i32*
134 //    load %p
135 //    ...
136 //  BB2:
137 //    %7 = bitcast [10 x %struct]* %ptr to i8*     ; CSE opportunity
138 //    %8 = mul i64 %i, length_of_10xstruct         ; CSE opportunity
139 //    %9 = getelementptr i8* %7, i64 %8            ; CSE opportunity
140 //    %10 = mul i64 %j2, length_of_struct
141 //    %11 = getelementptr i8* %9, i64 %10
142 //    %12 = getelementptr i8* %11, struct_field_2  ; Constant offset
143 //    %p2 = bitcast i8* %12 to i32*
144 //    load %p2
145 //    ...
146 //
147 // Lowering GEPs can also benefit other passes such as LICM and CGP.
148 // LICM (Loop Invariant Code Motion) can not hoist/sink a GEP of multiple
149 // indices if one of the index is variant. If we lower such GEP into invariant
150 // parts and variant parts, LICM can hoist/sink those invariant parts.
151 // CGP (CodeGen Prepare) tries to sink address calculations that match the
152 // target's addressing modes. A GEP with multiple indices may not match and will
153 // not be sunk. If we lower such GEP into smaller parts, CGP may sink some of
154 // them. So we end up with a better addressing mode.
155 //
156 //===----------------------------------------------------------------------===//
157 
158 #include "llvm/Transforms/Scalar/SeparateConstOffsetFromGEP.h"
159 #include "llvm/ADT/APInt.h"
160 #include "llvm/ADT/DenseMap.h"
161 #include "llvm/ADT/DepthFirstIterator.h"
162 #include "llvm/ADT/SmallVector.h"
163 #include "llvm/Analysis/LoopInfo.h"
164 #include "llvm/Analysis/MemoryBuiltins.h"
165 #include "llvm/Analysis/TargetLibraryInfo.h"
166 #include "llvm/Analysis/TargetTransformInfo.h"
167 #include "llvm/Analysis/ValueTracking.h"
168 #include "llvm/IR/BasicBlock.h"
169 #include "llvm/IR/Constant.h"
170 #include "llvm/IR/Constants.h"
171 #include "llvm/IR/DataLayout.h"
172 #include "llvm/IR/DerivedTypes.h"
173 #include "llvm/IR/Dominators.h"
174 #include "llvm/IR/Function.h"
175 #include "llvm/IR/GetElementPtrTypeIterator.h"
176 #include "llvm/IR/IRBuilder.h"
177 #include "llvm/IR/InstrTypes.h"
178 #include "llvm/IR/Instruction.h"
179 #include "llvm/IR/Instructions.h"
180 #include "llvm/IR/Module.h"
181 #include "llvm/IR/PassManager.h"
182 #include "llvm/IR/PatternMatch.h"
183 #include "llvm/IR/Type.h"
184 #include "llvm/IR/User.h"
185 #include "llvm/IR/Value.h"
186 #include "llvm/InitializePasses.h"
187 #include "llvm/Pass.h"
188 #include "llvm/Support/Casting.h"
189 #include "llvm/Support/CommandLine.h"
190 #include "llvm/Support/ErrorHandling.h"
191 #include "llvm/Support/raw_ostream.h"
192 #include "llvm/Transforms/Scalar.h"
193 #include "llvm/Transforms/Utils/Local.h"
194 #include <cassert>
195 #include <cstdint>
196 #include <string>
197 
198 using namespace llvm;
199 using namespace llvm::PatternMatch;
200 
201 static cl::opt<bool> DisableSeparateConstOffsetFromGEP(
202     "disable-separate-const-offset-from-gep", cl::init(false),
203     cl::desc("Do not separate the constant offset from a GEP instruction"),
204     cl::Hidden);
205 
206 // Setting this flag may emit false positives when the input module already
207 // contains dead instructions. Therefore, we set it only in unit tests that are
208 // free of dead code.
209 static cl::opt<bool>
210     VerifyNoDeadCode("reassociate-geps-verify-no-dead-code", cl::init(false),
211                      cl::desc("Verify this pass produces no dead code"),
212                      cl::Hidden);
213 
214 namespace {
215 
216 /// A helper class for separating a constant offset from a GEP index.
217 ///
218 /// In real programs, a GEP index may be more complicated than a simple addition
219 /// of something and a constant integer which can be trivially splitted. For
220 /// example, to split ((a << 3) | 5) + b, we need to search deeper for the
221 /// constant offset, so that we can separate the index to (a << 3) + b and 5.
222 ///
223 /// Therefore, this class looks into the expression that computes a given GEP
224 /// index, and tries to find a constant integer that can be hoisted to the
225 /// outermost level of the expression as an addition. Not every constant in an
226 /// expression can jump out. e.g., we cannot transform (b * (a + 5)) to (b * a +
227 /// 5); nor can we transform (3 * (a + 5)) to (3 * a + 5), however in this case,
228 /// -instcombine probably already optimized (3 * (a + 5)) to (3 * a + 15).
229 class ConstantOffsetExtractor {
230 public:
231   /// Extracts a constant offset from the given GEP index. It returns the
232   /// new index representing the remainder (equal to the original index minus
233   /// the constant offset), or nullptr if we cannot extract a constant offset.
234   /// \p Idx The given GEP index
235   /// \p GEP The given GEP
236   /// \p UserChainTail Outputs the tail of UserChain so that we can
237   ///                  garbage-collect unused instructions in UserChain.
238   static Value *Extract(Value *Idx, GetElementPtrInst *GEP,
239                         User *&UserChainTail);
240 
241   /// Looks for a constant offset from the given GEP index without extracting
242   /// it. It returns the numeric value of the extracted constant offset (0 if
243   /// failed). The meaning of the arguments are the same as Extract.
244   static int64_t Find(Value *Idx, GetElementPtrInst *GEP);
245 
246 private:
ConstantOffsetExtractor(BasicBlock::iterator InsertionPt)247   ConstantOffsetExtractor(BasicBlock::iterator InsertionPt)
248       : IP(InsertionPt), DL(InsertionPt->getDataLayout()) {}
249 
250   /// Searches the expression that computes V for a non-zero constant C s.t.
251   /// V can be reassociated into the form V' + C. If the searching is
252   /// successful, returns C and update UserChain as a def-use chain from C to V;
253   /// otherwise, UserChain is empty.
254   ///
255   /// \p V            The given expression
256   /// \p SignExtended Whether V will be sign-extended in the computation of the
257   ///                 GEP index
258   /// \p ZeroExtended Whether V will be zero-extended in the computation of the
259   ///                 GEP index
260   /// \p NonNegative  Whether V is guaranteed to be non-negative. For example,
261   ///                 an index of an inbounds GEP is guaranteed to be
262   ///                 non-negative. Levaraging this, we can better split
263   ///                 inbounds GEPs.
264   APInt find(Value *V, bool SignExtended, bool ZeroExtended, bool NonNegative);
265 
266   /// A helper function to look into both operands of a binary operator.
267   APInt findInEitherOperand(BinaryOperator *BO, bool SignExtended,
268                             bool ZeroExtended);
269 
270   /// After finding the constant offset C from the GEP index I, we build a new
271   /// index I' s.t. I' + C = I. This function builds and returns the new
272   /// index I' according to UserChain produced by function "find".
273   ///
274   /// The building conceptually takes two steps:
275   /// 1) iteratively distribute s/zext towards the leaves of the expression tree
276   /// that computes I
277   /// 2) reassociate the expression tree to the form I' + C.
278   ///
279   /// For example, to extract the 5 from sext(a + (b + 5)), we first distribute
280   /// sext to a, b and 5 so that we have
281   ///   sext(a) + (sext(b) + 5).
282   /// Then, we reassociate it to
283   ///   (sext(a) + sext(b)) + 5.
284   /// Given this form, we know I' is sext(a) + sext(b).
285   Value *rebuildWithoutConstOffset();
286 
287   /// After the first step of rebuilding the GEP index without the constant
288   /// offset, distribute s/zext to the operands of all operators in UserChain.
289   /// e.g., zext(sext(a + (b + 5)) (assuming no overflow) =>
290   /// zext(sext(a)) + (zext(sext(b)) + zext(sext(5))).
291   ///
292   /// The function also updates UserChain to point to new subexpressions after
293   /// distributing s/zext. e.g., the old UserChain of the above example is
294   /// 5 -> b + 5 -> a + (b + 5) -> sext(...) -> zext(sext(...)),
295   /// and the new UserChain is
296   /// zext(sext(5)) -> zext(sext(b)) + zext(sext(5)) ->
297   ///   zext(sext(a)) + (zext(sext(b)) + zext(sext(5))
298   ///
299   /// \p ChainIndex The index to UserChain. ChainIndex is initially
300   ///               UserChain.size() - 1, and is decremented during
301   ///               the recursion.
302   Value *distributeExtsAndCloneChain(unsigned ChainIndex);
303 
304   /// Reassociates the GEP index to the form I' + C and returns I'.
305   Value *removeConstOffset(unsigned ChainIndex);
306 
307   /// A helper function to apply ExtInsts, a list of s/zext, to value V.
308   /// e.g., if ExtInsts = [sext i32 to i64, zext i16 to i32], this function
309   /// returns "sext i32 (zext i16 V to i32) to i64".
310   Value *applyExts(Value *V);
311 
312   /// A helper function that returns whether we can trace into the operands
313   /// of binary operator BO for a constant offset.
314   ///
315   /// \p SignExtended Whether BO is surrounded by sext
316   /// \p ZeroExtended Whether BO is surrounded by zext
317   /// \p NonNegative Whether BO is known to be non-negative, e.g., an in-bound
318   ///                array index.
319   bool CanTraceInto(bool SignExtended, bool ZeroExtended, BinaryOperator *BO,
320                     bool NonNegative);
321 
322   /// The path from the constant offset to the old GEP index. e.g., if the GEP
323   /// index is "a * b + (c + 5)". After running function find, UserChain[0] will
324   /// be the constant 5, UserChain[1] will be the subexpression "c + 5", and
325   /// UserChain[2] will be the entire expression "a * b + (c + 5)".
326   ///
327   /// This path helps to rebuild the new GEP index.
328   SmallVector<User *, 8> UserChain;
329 
330   /// A data structure used in rebuildWithoutConstOffset. Contains all
331   /// sext/zext instructions along UserChain.
332   SmallVector<CastInst *, 16> ExtInsts;
333 
334   /// Insertion position of cloned instructions.
335   BasicBlock::iterator IP;
336 
337   const DataLayout &DL;
338 };
339 
340 /// A pass that tries to split every GEP in the function into a variadic
341 /// base and a constant offset. It is a FunctionPass because searching for the
342 /// constant offset may inspect other basic blocks.
343 class SeparateConstOffsetFromGEPLegacyPass : public FunctionPass {
344 public:
345   static char ID;
346 
SeparateConstOffsetFromGEPLegacyPass(bool LowerGEP=false)347   SeparateConstOffsetFromGEPLegacyPass(bool LowerGEP = false)
348       : FunctionPass(ID), LowerGEP(LowerGEP) {
349     initializeSeparateConstOffsetFromGEPLegacyPassPass(
350         *PassRegistry::getPassRegistry());
351   }
352 
getAnalysisUsage(AnalysisUsage & AU) const353   void getAnalysisUsage(AnalysisUsage &AU) const override {
354     AU.addRequired<DominatorTreeWrapperPass>();
355     AU.addRequired<TargetTransformInfoWrapperPass>();
356     AU.addRequired<LoopInfoWrapperPass>();
357     AU.setPreservesCFG();
358     AU.addRequired<TargetLibraryInfoWrapperPass>();
359   }
360 
361   bool runOnFunction(Function &F) override;
362 
363 private:
364   bool LowerGEP;
365 };
366 
367 /// A pass that tries to split every GEP in the function into a variadic
368 /// base and a constant offset. It is a FunctionPass because searching for the
369 /// constant offset may inspect other basic blocks.
370 class SeparateConstOffsetFromGEP {
371 public:
SeparateConstOffsetFromGEP(DominatorTree * DT,LoopInfo * LI,TargetLibraryInfo * TLI,function_ref<TargetTransformInfo & (Function &)> GetTTI,bool LowerGEP)372   SeparateConstOffsetFromGEP(
373       DominatorTree *DT, LoopInfo *LI, TargetLibraryInfo *TLI,
374       function_ref<TargetTransformInfo &(Function &)> GetTTI, bool LowerGEP)
375       : DT(DT), LI(LI), TLI(TLI), GetTTI(GetTTI), LowerGEP(LowerGEP) {}
376 
377   bool run(Function &F);
378 
379 private:
380   /// Track the operands of an add or sub.
381   using ExprKey = std::pair<Value *, Value *>;
382 
383   /// Create a pair for use as a map key for a commutable operation.
createNormalizedCommutablePair(Value * A,Value * B)384   static ExprKey createNormalizedCommutablePair(Value *A, Value *B) {
385     if (A < B)
386       return {A, B};
387     return {B, A};
388   }
389 
390   /// Tries to split the given GEP into a variadic base and a constant offset,
391   /// and returns true if the splitting succeeds.
392   bool splitGEP(GetElementPtrInst *GEP);
393 
394   /// Tries to reorder the given GEP with the GEP that produces the base if
395   /// doing so results in producing a constant offset as the outermost
396   /// index.
397   bool reorderGEP(GetElementPtrInst *GEP, TargetTransformInfo &TTI);
398 
399   /// Lower a GEP with multiple indices into multiple GEPs with a single index.
400   /// Function splitGEP already split the original GEP into a variadic part and
401   /// a constant offset (i.e., AccumulativeByteOffset). This function lowers the
402   /// variadic part into a set of GEPs with a single index and applies
403   /// AccumulativeByteOffset to it.
404   /// \p Variadic                  The variadic part of the original GEP.
405   /// \p AccumulativeByteOffset    The constant offset.
406   void lowerToSingleIndexGEPs(GetElementPtrInst *Variadic,
407                               int64_t AccumulativeByteOffset);
408 
409   /// Lower a GEP with multiple indices into ptrtoint+arithmetics+inttoptr form.
410   /// Function splitGEP already split the original GEP into a variadic part and
411   /// a constant offset (i.e., AccumulativeByteOffset). This function lowers the
412   /// variadic part into a set of arithmetic operations and applies
413   /// AccumulativeByteOffset to it.
414   /// \p Variadic                  The variadic part of the original GEP.
415   /// \p AccumulativeByteOffset    The constant offset.
416   void lowerToArithmetics(GetElementPtrInst *Variadic,
417                           int64_t AccumulativeByteOffset);
418 
419   /// Finds the constant offset within each index and accumulates them. If
420   /// LowerGEP is true, it finds in indices of both sequential and structure
421   /// types, otherwise it only finds in sequential indices. The output
422   /// NeedsExtraction indicates whether we successfully find a non-zero constant
423   /// offset.
424   int64_t accumulateByteOffset(GetElementPtrInst *GEP, bool &NeedsExtraction);
425 
426   /// Canonicalize array indices to pointer-size integers. This helps to
427   /// simplify the logic of splitting a GEP. For example, if a + b is a
428   /// pointer-size integer, we have
429   ///   gep base, a + b = gep (gep base, a), b
430   /// However, this equality may not hold if the size of a + b is smaller than
431   /// the pointer size, because LLVM conceptually sign-extends GEP indices to
432   /// pointer size before computing the address
433   /// (http://llvm.org/docs/LangRef.html#id181).
434   ///
435   /// This canonicalization is very likely already done in clang and
436   /// instcombine. Therefore, the program will probably remain the same.
437   ///
438   /// Returns true if the module changes.
439   ///
440   /// Verified in @i32_add in split-gep.ll
441   bool canonicalizeArrayIndicesToIndexSize(GetElementPtrInst *GEP);
442 
443   /// Optimize sext(a)+sext(b) to sext(a+b) when a+b can't sign overflow.
444   /// SeparateConstOffsetFromGEP distributes a sext to leaves before extracting
445   /// the constant offset. After extraction, it becomes desirable to reunion the
446   /// distributed sexts. For example,
447   ///
448   ///                              &a[sext(i +nsw (j +nsw 5)]
449   ///   => distribute              &a[sext(i) +nsw (sext(j) +nsw 5)]
450   ///   => constant extraction     &a[sext(i) + sext(j)] + 5
451   ///   => reunion                 &a[sext(i +nsw j)] + 5
452   bool reuniteExts(Function &F);
453 
454   /// A helper that reunites sexts in an instruction.
455   bool reuniteExts(Instruction *I);
456 
457   /// Find the closest dominator of <Dominatee> that is equivalent to <Key>.
458   Instruction *findClosestMatchingDominator(
459       ExprKey Key, Instruction *Dominatee,
460       DenseMap<ExprKey, SmallVector<Instruction *, 2>> &DominatingExprs);
461 
462   /// Verify F is free of dead code.
463   void verifyNoDeadCode(Function &F);
464 
465   bool hasMoreThanOneUseInLoop(Value *v, Loop *L);
466 
467   // Swap the index operand of two GEP.
468   void swapGEPOperand(GetElementPtrInst *First, GetElementPtrInst *Second);
469 
470   // Check if it is safe to swap operand of two GEP.
471   bool isLegalToSwapOperand(GetElementPtrInst *First, GetElementPtrInst *Second,
472                             Loop *CurLoop);
473 
474   const DataLayout *DL = nullptr;
475   DominatorTree *DT = nullptr;
476   LoopInfo *LI;
477   TargetLibraryInfo *TLI;
478   // Retrieved lazily since not always used.
479   function_ref<TargetTransformInfo &(Function &)> GetTTI;
480 
481   /// Whether to lower a GEP with multiple indices into arithmetic operations or
482   /// multiple GEPs with a single index.
483   bool LowerGEP;
484 
485   DenseMap<ExprKey, SmallVector<Instruction *, 2>> DominatingAdds;
486   DenseMap<ExprKey, SmallVector<Instruction *, 2>> DominatingSubs;
487 };
488 
489 } // end anonymous namespace
490 
491 char SeparateConstOffsetFromGEPLegacyPass::ID = 0;
492 
493 INITIALIZE_PASS_BEGIN(
494     SeparateConstOffsetFromGEPLegacyPass, "separate-const-offset-from-gep",
495     "Split GEPs to a variadic base and a constant offset for better CSE", false,
496     false)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)497 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
498 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
499 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
500 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
501 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
502 INITIALIZE_PASS_END(
503     SeparateConstOffsetFromGEPLegacyPass, "separate-const-offset-from-gep",
504     "Split GEPs to a variadic base and a constant offset for better CSE", false,
505     false)
506 
507 FunctionPass *llvm::createSeparateConstOffsetFromGEPPass(bool LowerGEP) {
508   return new SeparateConstOffsetFromGEPLegacyPass(LowerGEP);
509 }
510 
CanTraceInto(bool SignExtended,bool ZeroExtended,BinaryOperator * BO,bool NonNegative)511 bool ConstantOffsetExtractor::CanTraceInto(bool SignExtended,
512                                             bool ZeroExtended,
513                                             BinaryOperator *BO,
514                                             bool NonNegative) {
515   // We only consider ADD, SUB and OR, because a non-zero constant found in
516   // expressions composed of these operations can be easily hoisted as a
517   // constant offset by reassociation.
518   if (BO->getOpcode() != Instruction::Add &&
519       BO->getOpcode() != Instruction::Sub &&
520       BO->getOpcode() != Instruction::Or) {
521     return false;
522   }
523 
524   Value *LHS = BO->getOperand(0), *RHS = BO->getOperand(1);
525   // Do not trace into "or" unless it is equivalent to "add".
526   // This is the case if the or's disjoint flag is set.
527   if (BO->getOpcode() == Instruction::Or &&
528       !cast<PossiblyDisjointInst>(BO)->isDisjoint())
529     return false;
530 
531   // FIXME: We don't currently support constants from the RHS of subs,
532   // when we are zero-extended, because we need a way to zero-extended
533   // them before they are negated.
534   if (ZeroExtended && !SignExtended && BO->getOpcode() == Instruction::Sub)
535     return false;
536 
537   // In addition, tracing into BO requires that its surrounding s/zext (if
538   // any) is distributable to both operands.
539   //
540   // Suppose BO = A op B.
541   //  SignExtended | ZeroExtended | Distributable?
542   // --------------+--------------+----------------------------------
543   //       0       |      0       | true because no s/zext exists
544   //       0       |      1       | zext(BO) == zext(A) op zext(B)
545   //       1       |      0       | sext(BO) == sext(A) op sext(B)
546   //       1       |      1       | zext(sext(BO)) ==
547   //               |              |     zext(sext(A)) op zext(sext(B))
548   if (BO->getOpcode() == Instruction::Add && !ZeroExtended && NonNegative) {
549     // If a + b >= 0 and (a >= 0 or b >= 0), then
550     //   sext(a + b) = sext(a) + sext(b)
551     // even if the addition is not marked nsw.
552     //
553     // Leveraging this invariant, we can trace into an sext'ed inbound GEP
554     // index if the constant offset is non-negative.
555     //
556     // Verified in @sext_add in split-gep.ll.
557     if (ConstantInt *ConstLHS = dyn_cast<ConstantInt>(LHS)) {
558       if (!ConstLHS->isNegative())
559         return true;
560     }
561     if (ConstantInt *ConstRHS = dyn_cast<ConstantInt>(RHS)) {
562       if (!ConstRHS->isNegative())
563         return true;
564     }
565   }
566 
567   // sext (add/sub nsw A, B) == add/sub nsw (sext A), (sext B)
568   // zext (add/sub nuw A, B) == add/sub nuw (zext A), (zext B)
569   if (BO->getOpcode() == Instruction::Add ||
570       BO->getOpcode() == Instruction::Sub) {
571     if (SignExtended && !BO->hasNoSignedWrap())
572       return false;
573     if (ZeroExtended && !BO->hasNoUnsignedWrap())
574       return false;
575   }
576 
577   return true;
578 }
579 
findInEitherOperand(BinaryOperator * BO,bool SignExtended,bool ZeroExtended)580 APInt ConstantOffsetExtractor::findInEitherOperand(BinaryOperator *BO,
581                                                    bool SignExtended,
582                                                    bool ZeroExtended) {
583   // Save off the current height of the chain, in case we need to restore it.
584   size_t ChainLength = UserChain.size();
585 
586   // BO being non-negative does not shed light on whether its operands are
587   // non-negative. Clear the NonNegative flag here.
588   APInt ConstantOffset = find(BO->getOperand(0), SignExtended, ZeroExtended,
589                               /* NonNegative */ false);
590   // If we found a constant offset in the left operand, stop and return that.
591   // This shortcut might cause us to miss opportunities of combining the
592   // constant offsets in both operands, e.g., (a + 4) + (b + 5) => (a + b) + 9.
593   // However, such cases are probably already handled by -instcombine,
594   // given this pass runs after the standard optimizations.
595   if (ConstantOffset != 0) return ConstantOffset;
596 
597   // Reset the chain back to where it was when we started exploring this node,
598   // since visiting the LHS didn't pan out.
599   UserChain.resize(ChainLength);
600 
601   ConstantOffset = find(BO->getOperand(1), SignExtended, ZeroExtended,
602                         /* NonNegative */ false);
603   // If U is a sub operator, negate the constant offset found in the right
604   // operand.
605   if (BO->getOpcode() == Instruction::Sub)
606     ConstantOffset = -ConstantOffset;
607 
608   // If RHS wasn't a suitable candidate either, reset the chain again.
609   if (ConstantOffset == 0)
610     UserChain.resize(ChainLength);
611 
612   return ConstantOffset;
613 }
614 
find(Value * V,bool SignExtended,bool ZeroExtended,bool NonNegative)615 APInt ConstantOffsetExtractor::find(Value *V, bool SignExtended,
616                                     bool ZeroExtended, bool NonNegative) {
617   // TODO(jingyue): We could trace into integer/pointer casts, such as
618   // inttoptr, ptrtoint, bitcast, and addrspacecast. We choose to handle only
619   // integers because it gives good enough results for our benchmarks.
620   unsigned BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
621 
622   // We cannot do much with Values that are not a User, such as an Argument.
623   User *U = dyn_cast<User>(V);
624   if (U == nullptr) return APInt(BitWidth, 0);
625 
626   APInt ConstantOffset(BitWidth, 0);
627   if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
628     // Hooray, we found it!
629     ConstantOffset = CI->getValue();
630   } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(V)) {
631     // Trace into subexpressions for more hoisting opportunities.
632     if (CanTraceInto(SignExtended, ZeroExtended, BO, NonNegative))
633       ConstantOffset = findInEitherOperand(BO, SignExtended, ZeroExtended);
634   } else if (isa<TruncInst>(V)) {
635     ConstantOffset =
636         find(U->getOperand(0), SignExtended, ZeroExtended, NonNegative)
637             .trunc(BitWidth);
638   } else if (isa<SExtInst>(V)) {
639     ConstantOffset = find(U->getOperand(0), /* SignExtended */ true,
640                           ZeroExtended, NonNegative).sext(BitWidth);
641   } else if (isa<ZExtInst>(V)) {
642     // As an optimization, we can clear the SignExtended flag because
643     // sext(zext(a)) = zext(a). Verified in @sext_zext in split-gep.ll.
644     //
645     // Clear the NonNegative flag, because zext(a) >= 0 does not imply a >= 0.
646     ConstantOffset =
647         find(U->getOperand(0), /* SignExtended */ false,
648              /* ZeroExtended */ true, /* NonNegative */ false).zext(BitWidth);
649   }
650 
651   // If we found a non-zero constant offset, add it to the path for
652   // rebuildWithoutConstOffset. Zero is a valid constant offset, but doesn't
653   // help this optimization.
654   if (ConstantOffset != 0)
655     UserChain.push_back(U);
656   return ConstantOffset;
657 }
658 
applyExts(Value * V)659 Value *ConstantOffsetExtractor::applyExts(Value *V) {
660   Value *Current = V;
661   // ExtInsts is built in the use-def order. Therefore, we apply them to V
662   // in the reversed order.
663   for (CastInst *I : llvm::reverse(ExtInsts)) {
664     if (Constant *C = dyn_cast<Constant>(Current)) {
665       // Try to constant fold the cast.
666       Current = ConstantFoldCastOperand(I->getOpcode(), C, I->getType(), DL);
667       if (Current)
668         continue;
669     }
670 
671     Instruction *Ext = I->clone();
672     Ext->setOperand(0, Current);
673     Ext->insertBefore(*IP->getParent(), IP);
674     Current = Ext;
675   }
676   return Current;
677 }
678 
rebuildWithoutConstOffset()679 Value *ConstantOffsetExtractor::rebuildWithoutConstOffset() {
680   distributeExtsAndCloneChain(UserChain.size() - 1);
681   // Remove all nullptrs (used to be s/zext) from UserChain.
682   unsigned NewSize = 0;
683   for (User *I : UserChain) {
684     if (I != nullptr) {
685       UserChain[NewSize] = I;
686       NewSize++;
687     }
688   }
689   UserChain.resize(NewSize);
690   return removeConstOffset(UserChain.size() - 1);
691 }
692 
693 Value *
distributeExtsAndCloneChain(unsigned ChainIndex)694 ConstantOffsetExtractor::distributeExtsAndCloneChain(unsigned ChainIndex) {
695   User *U = UserChain[ChainIndex];
696   if (ChainIndex == 0) {
697     assert(isa<ConstantInt>(U));
698     // If U is a ConstantInt, applyExts will return a ConstantInt as well.
699     return UserChain[ChainIndex] = cast<ConstantInt>(applyExts(U));
700   }
701 
702   if (CastInst *Cast = dyn_cast<CastInst>(U)) {
703     assert(
704         (isa<SExtInst>(Cast) || isa<ZExtInst>(Cast) || isa<TruncInst>(Cast)) &&
705         "Only following instructions can be traced: sext, zext & trunc");
706     ExtInsts.push_back(Cast);
707     UserChain[ChainIndex] = nullptr;
708     return distributeExtsAndCloneChain(ChainIndex - 1);
709   }
710 
711   // Function find only trace into BinaryOperator and CastInst.
712   BinaryOperator *BO = cast<BinaryOperator>(U);
713   // OpNo = which operand of BO is UserChain[ChainIndex - 1]
714   unsigned OpNo = (BO->getOperand(0) == UserChain[ChainIndex - 1] ? 0 : 1);
715   Value *TheOther = applyExts(BO->getOperand(1 - OpNo));
716   Value *NextInChain = distributeExtsAndCloneChain(ChainIndex - 1);
717 
718   BinaryOperator *NewBO = nullptr;
719   if (OpNo == 0) {
720     NewBO = BinaryOperator::Create(BO->getOpcode(), NextInChain, TheOther,
721                                    BO->getName(), IP);
722   } else {
723     NewBO = BinaryOperator::Create(BO->getOpcode(), TheOther, NextInChain,
724                                    BO->getName(), IP);
725   }
726   return UserChain[ChainIndex] = NewBO;
727 }
728 
removeConstOffset(unsigned ChainIndex)729 Value *ConstantOffsetExtractor::removeConstOffset(unsigned ChainIndex) {
730   if (ChainIndex == 0) {
731     assert(isa<ConstantInt>(UserChain[ChainIndex]));
732     return ConstantInt::getNullValue(UserChain[ChainIndex]->getType());
733   }
734 
735   BinaryOperator *BO = cast<BinaryOperator>(UserChain[ChainIndex]);
736   assert((BO->use_empty() || BO->hasOneUse()) &&
737          "distributeExtsAndCloneChain clones each BinaryOperator in "
738          "UserChain, so no one should be used more than "
739          "once");
740 
741   unsigned OpNo = (BO->getOperand(0) == UserChain[ChainIndex - 1] ? 0 : 1);
742   assert(BO->getOperand(OpNo) == UserChain[ChainIndex - 1]);
743   Value *NextInChain = removeConstOffset(ChainIndex - 1);
744   Value *TheOther = BO->getOperand(1 - OpNo);
745 
746   // If NextInChain is 0 and not the LHS of a sub, we can simplify the
747   // sub-expression to be just TheOther.
748   if (ConstantInt *CI = dyn_cast<ConstantInt>(NextInChain)) {
749     if (CI->isZero() && !(BO->getOpcode() == Instruction::Sub && OpNo == 0))
750       return TheOther;
751   }
752 
753   BinaryOperator::BinaryOps NewOp = BO->getOpcode();
754   if (BO->getOpcode() == Instruction::Or) {
755     // Rebuild "or" as "add", because "or" may be invalid for the new
756     // expression.
757     //
758     // For instance, given
759     //   a | (b + 5) where a and b + 5 have no common bits,
760     // we can extract 5 as the constant offset.
761     //
762     // However, reusing the "or" in the new index would give us
763     //   (a | b) + 5
764     // which does not equal a | (b + 5).
765     //
766     // Replacing the "or" with "add" is fine, because
767     //   a | (b + 5) = a + (b + 5) = (a + b) + 5
768     NewOp = Instruction::Add;
769   }
770 
771   BinaryOperator *NewBO;
772   if (OpNo == 0) {
773     NewBO = BinaryOperator::Create(NewOp, NextInChain, TheOther, "", IP);
774   } else {
775     NewBO = BinaryOperator::Create(NewOp, TheOther, NextInChain, "", IP);
776   }
777   NewBO->takeName(BO);
778   return NewBO;
779 }
780 
Extract(Value * Idx,GetElementPtrInst * GEP,User * & UserChainTail)781 Value *ConstantOffsetExtractor::Extract(Value *Idx, GetElementPtrInst *GEP,
782                                         User *&UserChainTail) {
783   ConstantOffsetExtractor Extractor(GEP->getIterator());
784   // Find a non-zero constant offset first.
785   APInt ConstantOffset =
786       Extractor.find(Idx, /* SignExtended */ false, /* ZeroExtended */ false,
787                      GEP->isInBounds());
788   if (ConstantOffset == 0) {
789     UserChainTail = nullptr;
790     return nullptr;
791   }
792   // Separates the constant offset from the GEP index.
793   Value *IdxWithoutConstOffset = Extractor.rebuildWithoutConstOffset();
794   UserChainTail = Extractor.UserChain.back();
795   return IdxWithoutConstOffset;
796 }
797 
Find(Value * Idx,GetElementPtrInst * GEP)798 int64_t ConstantOffsetExtractor::Find(Value *Idx, GetElementPtrInst *GEP) {
799   // If Idx is an index of an inbound GEP, Idx is guaranteed to be non-negative.
800   return ConstantOffsetExtractor(GEP->getIterator())
801       .find(Idx, /* SignExtended */ false, /* ZeroExtended */ false,
802             GEP->isInBounds())
803       .getSExtValue();
804 }
805 
canonicalizeArrayIndicesToIndexSize(GetElementPtrInst * GEP)806 bool SeparateConstOffsetFromGEP::canonicalizeArrayIndicesToIndexSize(
807     GetElementPtrInst *GEP) {
808   bool Changed = false;
809   Type *PtrIdxTy = DL->getIndexType(GEP->getType());
810   gep_type_iterator GTI = gep_type_begin(*GEP);
811   for (User::op_iterator I = GEP->op_begin() + 1, E = GEP->op_end();
812        I != E; ++I, ++GTI) {
813     // Skip struct member indices which must be i32.
814     if (GTI.isSequential()) {
815       if ((*I)->getType() != PtrIdxTy) {
816         *I = CastInst::CreateIntegerCast(*I, PtrIdxTy, true, "idxprom",
817                                          GEP->getIterator());
818         Changed = true;
819       }
820     }
821   }
822   return Changed;
823 }
824 
825 int64_t
accumulateByteOffset(GetElementPtrInst * GEP,bool & NeedsExtraction)826 SeparateConstOffsetFromGEP::accumulateByteOffset(GetElementPtrInst *GEP,
827                                                  bool &NeedsExtraction) {
828   NeedsExtraction = false;
829   int64_t AccumulativeByteOffset = 0;
830   gep_type_iterator GTI = gep_type_begin(*GEP);
831   for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I, ++GTI) {
832     if (GTI.isSequential()) {
833       // Constant offsets of scalable types are not really constant.
834       if (GTI.getIndexedType()->isScalableTy())
835         continue;
836 
837       // Tries to extract a constant offset from this GEP index.
838       int64_t ConstantOffset =
839           ConstantOffsetExtractor::Find(GEP->getOperand(I), GEP);
840       if (ConstantOffset != 0) {
841         NeedsExtraction = true;
842         // A GEP may have multiple indices.  We accumulate the extracted
843         // constant offset to a byte offset, and later offset the remainder of
844         // the original GEP with this byte offset.
845         AccumulativeByteOffset +=
846             ConstantOffset * GTI.getSequentialElementStride(*DL);
847       }
848     } else if (LowerGEP) {
849       StructType *StTy = GTI.getStructType();
850       uint64_t Field = cast<ConstantInt>(GEP->getOperand(I))->getZExtValue();
851       // Skip field 0 as the offset is always 0.
852       if (Field != 0) {
853         NeedsExtraction = true;
854         AccumulativeByteOffset +=
855             DL->getStructLayout(StTy)->getElementOffset(Field);
856       }
857     }
858   }
859   return AccumulativeByteOffset;
860 }
861 
lowerToSingleIndexGEPs(GetElementPtrInst * Variadic,int64_t AccumulativeByteOffset)862 void SeparateConstOffsetFromGEP::lowerToSingleIndexGEPs(
863     GetElementPtrInst *Variadic, int64_t AccumulativeByteOffset) {
864   IRBuilder<> Builder(Variadic);
865   Type *PtrIndexTy = DL->getIndexType(Variadic->getType());
866 
867   Value *ResultPtr = Variadic->getOperand(0);
868   Loop *L = LI->getLoopFor(Variadic->getParent());
869   // Check if the base is not loop invariant or used more than once.
870   bool isSwapCandidate =
871       L && L->isLoopInvariant(ResultPtr) &&
872       !hasMoreThanOneUseInLoop(ResultPtr, L);
873   Value *FirstResult = nullptr;
874 
875   gep_type_iterator GTI = gep_type_begin(*Variadic);
876   // Create an ugly GEP for each sequential index. We don't create GEPs for
877   // structure indices, as they are accumulated in the constant offset index.
878   for (unsigned I = 1, E = Variadic->getNumOperands(); I != E; ++I, ++GTI) {
879     if (GTI.isSequential()) {
880       Value *Idx = Variadic->getOperand(I);
881       // Skip zero indices.
882       if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx))
883         if (CI->isZero())
884           continue;
885 
886       APInt ElementSize = APInt(PtrIndexTy->getIntegerBitWidth(),
887                                 GTI.getSequentialElementStride(*DL));
888       // Scale the index by element size.
889       if (ElementSize != 1) {
890         if (ElementSize.isPowerOf2()) {
891           Idx = Builder.CreateShl(
892               Idx, ConstantInt::get(PtrIndexTy, ElementSize.logBase2()));
893         } else {
894           Idx =
895               Builder.CreateMul(Idx, ConstantInt::get(PtrIndexTy, ElementSize));
896         }
897       }
898       // Create an ugly GEP with a single index for each index.
899       ResultPtr = Builder.CreatePtrAdd(ResultPtr, Idx, "uglygep");
900       if (FirstResult == nullptr)
901         FirstResult = ResultPtr;
902     }
903   }
904 
905   // Create a GEP with the constant offset index.
906   if (AccumulativeByteOffset != 0) {
907     Value *Offset = ConstantInt::get(PtrIndexTy, AccumulativeByteOffset);
908     ResultPtr = Builder.CreatePtrAdd(ResultPtr, Offset, "uglygep");
909   } else
910     isSwapCandidate = false;
911 
912   // If we created a GEP with constant index, and the base is loop invariant,
913   // then we swap the first one with it, so LICM can move constant GEP out
914   // later.
915   auto *FirstGEP = dyn_cast_or_null<GetElementPtrInst>(FirstResult);
916   auto *SecondGEP = dyn_cast<GetElementPtrInst>(ResultPtr);
917   if (isSwapCandidate && isLegalToSwapOperand(FirstGEP, SecondGEP, L))
918     swapGEPOperand(FirstGEP, SecondGEP);
919 
920   Variadic->replaceAllUsesWith(ResultPtr);
921   Variadic->eraseFromParent();
922 }
923 
924 void
lowerToArithmetics(GetElementPtrInst * Variadic,int64_t AccumulativeByteOffset)925 SeparateConstOffsetFromGEP::lowerToArithmetics(GetElementPtrInst *Variadic,
926                                                int64_t AccumulativeByteOffset) {
927   IRBuilder<> Builder(Variadic);
928   Type *IntPtrTy = DL->getIntPtrType(Variadic->getType());
929   assert(IntPtrTy == DL->getIndexType(Variadic->getType()) &&
930          "Pointer type must match index type for arithmetic-based lowering of "
931          "split GEPs");
932 
933   Value *ResultPtr = Builder.CreatePtrToInt(Variadic->getOperand(0), IntPtrTy);
934   gep_type_iterator GTI = gep_type_begin(*Variadic);
935   // Create ADD/SHL/MUL arithmetic operations for each sequential indices. We
936   // don't create arithmetics for structure indices, as they are accumulated
937   // in the constant offset index.
938   for (unsigned I = 1, E = Variadic->getNumOperands(); I != E; ++I, ++GTI) {
939     if (GTI.isSequential()) {
940       Value *Idx = Variadic->getOperand(I);
941       // Skip zero indices.
942       if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx))
943         if (CI->isZero())
944           continue;
945 
946       APInt ElementSize = APInt(IntPtrTy->getIntegerBitWidth(),
947                                 GTI.getSequentialElementStride(*DL));
948       // Scale the index by element size.
949       if (ElementSize != 1) {
950         if (ElementSize.isPowerOf2()) {
951           Idx = Builder.CreateShl(
952               Idx, ConstantInt::get(IntPtrTy, ElementSize.logBase2()));
953         } else {
954           Idx = Builder.CreateMul(Idx, ConstantInt::get(IntPtrTy, ElementSize));
955         }
956       }
957       // Create an ADD for each index.
958       ResultPtr = Builder.CreateAdd(ResultPtr, Idx);
959     }
960   }
961 
962   // Create an ADD for the constant offset index.
963   if (AccumulativeByteOffset != 0) {
964     ResultPtr = Builder.CreateAdd(
965         ResultPtr, ConstantInt::get(IntPtrTy, AccumulativeByteOffset));
966   }
967 
968   ResultPtr = Builder.CreateIntToPtr(ResultPtr, Variadic->getType());
969   Variadic->replaceAllUsesWith(ResultPtr);
970   Variadic->eraseFromParent();
971 }
972 
reorderGEP(GetElementPtrInst * GEP,TargetTransformInfo & TTI)973 bool SeparateConstOffsetFromGEP::reorderGEP(GetElementPtrInst *GEP,
974                                             TargetTransformInfo &TTI) {
975   auto PtrGEP = dyn_cast<GetElementPtrInst>(GEP->getPointerOperand());
976   if (!PtrGEP)
977     return false;
978 
979   bool NestedNeedsExtraction;
980   int64_t NestedByteOffset =
981       accumulateByteOffset(PtrGEP, NestedNeedsExtraction);
982   if (!NestedNeedsExtraction)
983     return false;
984 
985   unsigned AddrSpace = PtrGEP->getPointerAddressSpace();
986   if (!TTI.isLegalAddressingMode(GEP->getResultElementType(),
987                                  /*BaseGV=*/nullptr, NestedByteOffset,
988                                  /*HasBaseReg=*/true, /*Scale=*/0, AddrSpace))
989     return false;
990 
991   bool GEPInBounds = GEP->isInBounds();
992   bool PtrGEPInBounds = PtrGEP->isInBounds();
993   bool IsChainInBounds = GEPInBounds && PtrGEPInBounds;
994   if (IsChainInBounds) {
995     auto IsKnownNonNegative = [this](Value *V) {
996       return isKnownNonNegative(V, *DL);
997     };
998     IsChainInBounds &= all_of(GEP->indices(), IsKnownNonNegative);
999     if (IsChainInBounds)
1000       IsChainInBounds &= all_of(PtrGEP->indices(), IsKnownNonNegative);
1001   }
1002 
1003   IRBuilder<> Builder(GEP);
1004   // For trivial GEP chains, we can swap the indices.
1005   Value *NewSrc = Builder.CreateGEP(
1006       GEP->getSourceElementType(), PtrGEP->getPointerOperand(),
1007       SmallVector<Value *, 4>(GEP->indices()), "", IsChainInBounds);
1008   Value *NewGEP = Builder.CreateGEP(PtrGEP->getSourceElementType(), NewSrc,
1009                                     SmallVector<Value *, 4>(PtrGEP->indices()),
1010                                     "", IsChainInBounds);
1011   GEP->replaceAllUsesWith(NewGEP);
1012   RecursivelyDeleteTriviallyDeadInstructions(GEP);
1013   return true;
1014 }
1015 
splitGEP(GetElementPtrInst * GEP)1016 bool SeparateConstOffsetFromGEP::splitGEP(GetElementPtrInst *GEP) {
1017   // Skip vector GEPs.
1018   if (GEP->getType()->isVectorTy())
1019     return false;
1020 
1021   // The backend can already nicely handle the case where all indices are
1022   // constant.
1023   if (GEP->hasAllConstantIndices())
1024     return false;
1025 
1026   bool Changed = canonicalizeArrayIndicesToIndexSize(GEP);
1027 
1028   bool NeedsExtraction;
1029   int64_t AccumulativeByteOffset = accumulateByteOffset(GEP, NeedsExtraction);
1030 
1031   TargetTransformInfo &TTI = GetTTI(*GEP->getFunction());
1032 
1033   if (!NeedsExtraction) {
1034     Changed |= reorderGEP(GEP, TTI);
1035     return Changed;
1036   }
1037 
1038   // If LowerGEP is disabled, before really splitting the GEP, check whether the
1039   // backend supports the addressing mode we are about to produce. If no, this
1040   // splitting probably won't be beneficial.
1041   // If LowerGEP is enabled, even the extracted constant offset can not match
1042   // the addressing mode, we can still do optimizations to other lowered parts
1043   // of variable indices. Therefore, we don't check for addressing modes in that
1044   // case.
1045   if (!LowerGEP) {
1046     unsigned AddrSpace = GEP->getPointerAddressSpace();
1047     if (!TTI.isLegalAddressingMode(GEP->getResultElementType(),
1048                                    /*BaseGV=*/nullptr, AccumulativeByteOffset,
1049                                    /*HasBaseReg=*/true, /*Scale=*/0,
1050                                    AddrSpace)) {
1051       return Changed;
1052     }
1053   }
1054 
1055   // Remove the constant offset in each sequential index. The resultant GEP
1056   // computes the variadic base.
1057   // Notice that we don't remove struct field indices here. If LowerGEP is
1058   // disabled, a structure index is not accumulated and we still use the old
1059   // one. If LowerGEP is enabled, a structure index is accumulated in the
1060   // constant offset. LowerToSingleIndexGEPs or lowerToArithmetics will later
1061   // handle the constant offset and won't need a new structure index.
1062   gep_type_iterator GTI = gep_type_begin(*GEP);
1063   for (unsigned I = 1, E = GEP->getNumOperands(); I != E; ++I, ++GTI) {
1064     if (GTI.isSequential()) {
1065       // Constant offsets of scalable types are not really constant.
1066       if (GTI.getIndexedType()->isScalableTy())
1067         continue;
1068 
1069       // Splits this GEP index into a variadic part and a constant offset, and
1070       // uses the variadic part as the new index.
1071       Value *OldIdx = GEP->getOperand(I);
1072       User *UserChainTail;
1073       Value *NewIdx =
1074           ConstantOffsetExtractor::Extract(OldIdx, GEP, UserChainTail);
1075       if (NewIdx != nullptr) {
1076         // Switches to the index with the constant offset removed.
1077         GEP->setOperand(I, NewIdx);
1078         // After switching to the new index, we can garbage-collect UserChain
1079         // and the old index if they are not used.
1080         RecursivelyDeleteTriviallyDeadInstructions(UserChainTail);
1081         RecursivelyDeleteTriviallyDeadInstructions(OldIdx);
1082       }
1083     }
1084   }
1085 
1086   // Clear the inbounds attribute because the new index may be off-bound.
1087   // e.g.,
1088   //
1089   //   b     = add i64 a, 5
1090   //   addr  = gep inbounds float, float* p, i64 b
1091   //
1092   // is transformed to:
1093   //
1094   //   addr2 = gep float, float* p, i64 a ; inbounds removed
1095   //   addr  = gep inbounds float, float* addr2, i64 5
1096   //
1097   // If a is -4, although the old index b is in bounds, the new index a is
1098   // off-bound. http://llvm.org/docs/LangRef.html#id181 says "if the
1099   // inbounds keyword is not present, the offsets are added to the base
1100   // address with silently-wrapping two's complement arithmetic".
1101   // Therefore, the final code will be a semantically equivalent.
1102   //
1103   // TODO(jingyue): do some range analysis to keep as many inbounds as
1104   // possible. GEPs with inbounds are more friendly to alias analysis.
1105   // TODO(gep_nowrap): Preserve nuw at least.
1106   bool GEPWasInBounds = GEP->isInBounds();
1107   GEP->setNoWrapFlags(GEPNoWrapFlags::none());
1108 
1109   // Lowers a GEP to either GEPs with a single index or arithmetic operations.
1110   if (LowerGEP) {
1111     // As currently BasicAA does not analyze ptrtoint/inttoptr, do not lower to
1112     // arithmetic operations if the target uses alias analysis in codegen.
1113     // Additionally, pointers that aren't integral (and so can't be safely
1114     // converted to integers) or those whose offset size is different from their
1115     // pointer size (which means that doing integer arithmetic on them could
1116     // affect that data) can't be lowered in this way.
1117     unsigned AddrSpace = GEP->getPointerAddressSpace();
1118     bool PointerHasExtraData = DL->getPointerSizeInBits(AddrSpace) !=
1119                                DL->getIndexSizeInBits(AddrSpace);
1120     if (TTI.useAA() || DL->isNonIntegralAddressSpace(AddrSpace) ||
1121         PointerHasExtraData)
1122       lowerToSingleIndexGEPs(GEP, AccumulativeByteOffset);
1123     else
1124       lowerToArithmetics(GEP, AccumulativeByteOffset);
1125     return true;
1126   }
1127 
1128   // No need to create another GEP if the accumulative byte offset is 0.
1129   if (AccumulativeByteOffset == 0)
1130     return true;
1131 
1132   // Offsets the base with the accumulative byte offset.
1133   //
1134   //   %gep                        ; the base
1135   //   ... %gep ...
1136   //
1137   // => add the offset
1138   //
1139   //   %gep2                       ; clone of %gep
1140   //   %new.gep = gep i8, %gep2, %offset
1141   //   %gep                        ; will be removed
1142   //   ... %gep ...
1143   //
1144   // => replace all uses of %gep with %new.gep and remove %gep
1145   //
1146   //   %gep2                       ; clone of %gep
1147   //   %new.gep = gep i8, %gep2, %offset
1148   //   ... %new.gep ...
1149   Instruction *NewGEP = GEP->clone();
1150   NewGEP->insertBefore(GEP);
1151 
1152   Type *PtrIdxTy = DL->getIndexType(GEP->getType());
1153   IRBuilder<> Builder(GEP);
1154   NewGEP = cast<Instruction>(Builder.CreatePtrAdd(
1155       NewGEP, ConstantInt::get(PtrIdxTy, AccumulativeByteOffset, true),
1156       GEP->getName(), GEPWasInBounds));
1157   NewGEP->copyMetadata(*GEP);
1158 
1159   GEP->replaceAllUsesWith(NewGEP);
1160   GEP->eraseFromParent();
1161 
1162   return true;
1163 }
1164 
runOnFunction(Function & F)1165 bool SeparateConstOffsetFromGEPLegacyPass::runOnFunction(Function &F) {
1166   if (skipFunction(F))
1167     return false;
1168   auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1169   auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1170   auto *TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
1171   auto GetTTI = [this](Function &F) -> TargetTransformInfo & {
1172     return this->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1173   };
1174   SeparateConstOffsetFromGEP Impl(DT, LI, TLI, GetTTI, LowerGEP);
1175   return Impl.run(F);
1176 }
1177 
run(Function & F)1178 bool SeparateConstOffsetFromGEP::run(Function &F) {
1179   if (DisableSeparateConstOffsetFromGEP)
1180     return false;
1181 
1182   DL = &F.getDataLayout();
1183   bool Changed = false;
1184   for (BasicBlock &B : F) {
1185     if (!DT->isReachableFromEntry(&B))
1186       continue;
1187 
1188     for (Instruction &I : llvm::make_early_inc_range(B))
1189       if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(&I))
1190         Changed |= splitGEP(GEP);
1191     // No need to split GEP ConstantExprs because all its indices are constant
1192     // already.
1193   }
1194 
1195   Changed |= reuniteExts(F);
1196 
1197   if (VerifyNoDeadCode)
1198     verifyNoDeadCode(F);
1199 
1200   return Changed;
1201 }
1202 
findClosestMatchingDominator(ExprKey Key,Instruction * Dominatee,DenseMap<ExprKey,SmallVector<Instruction *,2>> & DominatingExprs)1203 Instruction *SeparateConstOffsetFromGEP::findClosestMatchingDominator(
1204     ExprKey Key, Instruction *Dominatee,
1205     DenseMap<ExprKey, SmallVector<Instruction *, 2>> &DominatingExprs) {
1206   auto Pos = DominatingExprs.find(Key);
1207   if (Pos == DominatingExprs.end())
1208     return nullptr;
1209 
1210   auto &Candidates = Pos->second;
1211   // Because we process the basic blocks in pre-order of the dominator tree, a
1212   // candidate that doesn't dominate the current instruction won't dominate any
1213   // future instruction either. Therefore, we pop it out of the stack. This
1214   // optimization makes the algorithm O(n).
1215   while (!Candidates.empty()) {
1216     Instruction *Candidate = Candidates.back();
1217     if (DT->dominates(Candidate, Dominatee))
1218       return Candidate;
1219     Candidates.pop_back();
1220   }
1221   return nullptr;
1222 }
1223 
reuniteExts(Instruction * I)1224 bool SeparateConstOffsetFromGEP::reuniteExts(Instruction *I) {
1225   if (!I->getType()->isIntOrIntVectorTy())
1226     return false;
1227 
1228   //   Dom: LHS+RHS
1229   //   I: sext(LHS)+sext(RHS)
1230   // If Dom can't sign overflow and Dom dominates I, optimize I to sext(Dom).
1231   // TODO: handle zext
1232   Value *LHS = nullptr, *RHS = nullptr;
1233   if (match(I, m_Add(m_SExt(m_Value(LHS)), m_SExt(m_Value(RHS))))) {
1234     if (LHS->getType() == RHS->getType()) {
1235       ExprKey Key = createNormalizedCommutablePair(LHS, RHS);
1236       if (auto *Dom = findClosestMatchingDominator(Key, I, DominatingAdds)) {
1237         Instruction *NewSExt =
1238             new SExtInst(Dom, I->getType(), "", I->getIterator());
1239         NewSExt->takeName(I);
1240         I->replaceAllUsesWith(NewSExt);
1241         NewSExt->setDebugLoc(I->getDebugLoc());
1242         RecursivelyDeleteTriviallyDeadInstructions(I);
1243         return true;
1244       }
1245     }
1246   } else if (match(I, m_Sub(m_SExt(m_Value(LHS)), m_SExt(m_Value(RHS))))) {
1247     if (LHS->getType() == RHS->getType()) {
1248       if (auto *Dom =
1249               findClosestMatchingDominator({LHS, RHS}, I, DominatingSubs)) {
1250         Instruction *NewSExt =
1251             new SExtInst(Dom, I->getType(), "", I->getIterator());
1252         NewSExt->takeName(I);
1253         I->replaceAllUsesWith(NewSExt);
1254         NewSExt->setDebugLoc(I->getDebugLoc());
1255         RecursivelyDeleteTriviallyDeadInstructions(I);
1256         return true;
1257       }
1258     }
1259   }
1260 
1261   // Add I to DominatingExprs if it's an add/sub that can't sign overflow.
1262   if (match(I, m_NSWAdd(m_Value(LHS), m_Value(RHS)))) {
1263     if (programUndefinedIfPoison(I)) {
1264       ExprKey Key = createNormalizedCommutablePair(LHS, RHS);
1265       DominatingAdds[Key].push_back(I);
1266     }
1267   } else if (match(I, m_NSWSub(m_Value(LHS), m_Value(RHS)))) {
1268     if (programUndefinedIfPoison(I))
1269       DominatingSubs[{LHS, RHS}].push_back(I);
1270   }
1271   return false;
1272 }
1273 
reuniteExts(Function & F)1274 bool SeparateConstOffsetFromGEP::reuniteExts(Function &F) {
1275   bool Changed = false;
1276   DominatingAdds.clear();
1277   DominatingSubs.clear();
1278   for (const auto Node : depth_first(DT)) {
1279     BasicBlock *BB = Node->getBlock();
1280     for (Instruction &I : llvm::make_early_inc_range(*BB))
1281       Changed |= reuniteExts(&I);
1282   }
1283   return Changed;
1284 }
1285 
verifyNoDeadCode(Function & F)1286 void SeparateConstOffsetFromGEP::verifyNoDeadCode(Function &F) {
1287   for (BasicBlock &B : F) {
1288     for (Instruction &I : B) {
1289       if (isInstructionTriviallyDead(&I)) {
1290         std::string ErrMessage;
1291         raw_string_ostream RSO(ErrMessage);
1292         RSO << "Dead instruction detected!\n" << I << "\n";
1293         llvm_unreachable(RSO.str().c_str());
1294       }
1295     }
1296   }
1297 }
1298 
isLegalToSwapOperand(GetElementPtrInst * FirstGEP,GetElementPtrInst * SecondGEP,Loop * CurLoop)1299 bool SeparateConstOffsetFromGEP::isLegalToSwapOperand(
1300     GetElementPtrInst *FirstGEP, GetElementPtrInst *SecondGEP, Loop *CurLoop) {
1301   if (!FirstGEP || !FirstGEP->hasOneUse())
1302     return false;
1303 
1304   if (!SecondGEP || FirstGEP->getParent() != SecondGEP->getParent())
1305     return false;
1306 
1307   if (FirstGEP == SecondGEP)
1308     return false;
1309 
1310   unsigned FirstNum = FirstGEP->getNumOperands();
1311   unsigned SecondNum = SecondGEP->getNumOperands();
1312   // Give up if the number of operands are not 2.
1313   if (FirstNum != SecondNum || FirstNum != 2)
1314     return false;
1315 
1316   Value *FirstBase = FirstGEP->getOperand(0);
1317   Value *SecondBase = SecondGEP->getOperand(0);
1318   Value *FirstOffset = FirstGEP->getOperand(1);
1319   // Give up if the index of the first GEP is loop invariant.
1320   if (CurLoop->isLoopInvariant(FirstOffset))
1321     return false;
1322 
1323   // Give up if base doesn't have same type.
1324   if (FirstBase->getType() != SecondBase->getType())
1325     return false;
1326 
1327   Instruction *FirstOffsetDef = dyn_cast<Instruction>(FirstOffset);
1328 
1329   // Check if the second operand of first GEP has constant coefficient.
1330   // For an example, for the following code,  we won't gain anything by
1331   // hoisting the second GEP out because the second GEP can be folded away.
1332   //   %scevgep.sum.ur159 = add i64 %idxprom48.ur, 256
1333   //   %67 = shl i64 %scevgep.sum.ur159, 2
1334   //   %uglygep160 = getelementptr i8* %65, i64 %67
1335   //   %uglygep161 = getelementptr i8* %uglygep160, i64 -1024
1336 
1337   // Skip constant shift instruction which may be generated by Splitting GEPs.
1338   if (FirstOffsetDef && FirstOffsetDef->isShift() &&
1339       isa<ConstantInt>(FirstOffsetDef->getOperand(1)))
1340     FirstOffsetDef = dyn_cast<Instruction>(FirstOffsetDef->getOperand(0));
1341 
1342   // Give up if FirstOffsetDef is an Add or Sub with constant.
1343   // Because it may not profitable at all due to constant folding.
1344   if (FirstOffsetDef)
1345     if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FirstOffsetDef)) {
1346       unsigned opc = BO->getOpcode();
1347       if ((opc == Instruction::Add || opc == Instruction::Sub) &&
1348           (isa<ConstantInt>(BO->getOperand(0)) ||
1349            isa<ConstantInt>(BO->getOperand(1))))
1350         return false;
1351     }
1352   return true;
1353 }
1354 
hasMoreThanOneUseInLoop(Value * V,Loop * L)1355 bool SeparateConstOffsetFromGEP::hasMoreThanOneUseInLoop(Value *V, Loop *L) {
1356   int UsesInLoop = 0;
1357   for (User *U : V->users()) {
1358     if (Instruction *User = dyn_cast<Instruction>(U))
1359       if (L->contains(User))
1360         if (++UsesInLoop > 1)
1361           return true;
1362   }
1363   return false;
1364 }
1365 
swapGEPOperand(GetElementPtrInst * First,GetElementPtrInst * Second)1366 void SeparateConstOffsetFromGEP::swapGEPOperand(GetElementPtrInst *First,
1367                                                 GetElementPtrInst *Second) {
1368   Value *Offset1 = First->getOperand(1);
1369   Value *Offset2 = Second->getOperand(1);
1370   First->setOperand(1, Offset2);
1371   Second->setOperand(1, Offset1);
1372 
1373   // We changed p+o+c to p+c+o, p+c may not be inbound anymore.
1374   const DataLayout &DAL = First->getDataLayout();
1375   APInt Offset(DAL.getIndexSizeInBits(
1376                    cast<PointerType>(First->getType())->getAddressSpace()),
1377                0);
1378   Value *NewBase =
1379       First->stripAndAccumulateInBoundsConstantOffsets(DAL, Offset);
1380   uint64_t ObjectSize;
1381   if (!getObjectSize(NewBase, ObjectSize, DAL, TLI) ||
1382      Offset.ugt(ObjectSize)) {
1383     // TODO(gep_nowrap): Make flag preservation more precise.
1384     First->setNoWrapFlags(GEPNoWrapFlags::none());
1385     Second->setNoWrapFlags(GEPNoWrapFlags::none());
1386   } else
1387     First->setIsInBounds(true);
1388 }
1389 
printPipeline(raw_ostream & OS,function_ref<StringRef (StringRef)> MapClassName2PassName)1390 void SeparateConstOffsetFromGEPPass::printPipeline(
1391     raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
1392   static_cast<PassInfoMixin<SeparateConstOffsetFromGEPPass> *>(this)
1393       ->printPipeline(OS, MapClassName2PassName);
1394   OS << '<';
1395   if (LowerGEP)
1396     OS << "lower-gep";
1397   OS << '>';
1398 }
1399 
1400 PreservedAnalyses
run(Function & F,FunctionAnalysisManager & AM)1401 SeparateConstOffsetFromGEPPass::run(Function &F, FunctionAnalysisManager &AM) {
1402   auto *DT = &AM.getResult<DominatorTreeAnalysis>(F);
1403   auto *LI = &AM.getResult<LoopAnalysis>(F);
1404   auto *TLI = &AM.getResult<TargetLibraryAnalysis>(F);
1405   auto GetTTI = [&AM](Function &F) -> TargetTransformInfo & {
1406     return AM.getResult<TargetIRAnalysis>(F);
1407   };
1408   SeparateConstOffsetFromGEP Impl(DT, LI, TLI, GetTTI, LowerGEP);
1409   if (!Impl.run(F))
1410     return PreservedAnalyses::all();
1411   PreservedAnalyses PA;
1412   PA.preserveSet<CFGAnalyses>();
1413   return PA;
1414 }
1415