| 1 | //===- StackColoring.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 | // This pass implements the stack-coloring optimization that looks for |
| 10 | // lifetime markers machine instructions (LIFETIME_START and LIFETIME_END), |
| 11 | // which represent the possible lifetime of stack slots. It attempts to |
| 12 | // merge disjoint stack slots and reduce the used stack space. |
| 13 | // NOTE: This pass is not StackSlotColoring, which optimizes spill slots. |
| 14 | // |
| 15 | // TODO: In the future we plan to improve stack coloring in the following ways: |
| 16 | // 1. Allow merging multiple small slots into a single larger slot at different |
| 17 | // offsets. |
| 18 | // 2. Merge this pass with StackSlotColoring and allow merging of allocas with |
| 19 | // spill slots. |
| 20 | // |
| 21 | //===----------------------------------------------------------------------===// |
| 22 | |
| 23 | #include "llvm/ADT/BitVector.h" |
| 24 | #include "llvm/ADT/DenseMap.h" |
| 25 | #include "llvm/ADT/DepthFirstIterator.h" |
| 26 | #include "llvm/ADT/SmallPtrSet.h" |
| 27 | #include "llvm/ADT/SmallVector.h" |
| 28 | #include "llvm/ADT/Statistic.h" |
| 29 | #include "llvm/Analysis/ValueTracking.h" |
| 30 | #include "llvm/CodeGen/LiveInterval.h" |
| 31 | #include "llvm/CodeGen/MachineBasicBlock.h" |
| 32 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 33 | #include "llvm/CodeGen/MachineFunction.h" |
| 34 | #include "llvm/CodeGen/MachineFunctionPass.h" |
| 35 | #include "llvm/CodeGen/MachineInstr.h" |
| 36 | #include "llvm/CodeGen/MachineMemOperand.h" |
| 37 | #include "llvm/CodeGen/MachineOperand.h" |
| 38 | #include "llvm/CodeGen/Passes.h" |
| 39 | #include "llvm/CodeGen/PseudoSourceValueManager.h" |
| 40 | #include "llvm/CodeGen/SlotIndexes.h" |
| 41 | #include "llvm/CodeGen/TargetOpcodes.h" |
| 42 | #include "llvm/CodeGen/WinEHFuncInfo.h" |
| 43 | #include "llvm/Config/llvm-config.h" |
| 44 | #include "llvm/IR/Constants.h" |
| 45 | #include "llvm/IR/DebugInfoMetadata.h" |
| 46 | #include "llvm/IR/Instructions.h" |
| 47 | #include "llvm/IR/Metadata.h" |
| 48 | #include "llvm/IR/Use.h" |
| 49 | #include "llvm/IR/Value.h" |
| 50 | #include "llvm/InitializePasses.h" |
| 51 | #include "llvm/Pass.h" |
| 52 | #include "llvm/Support/Casting.h" |
| 53 | #include "llvm/Support/CommandLine.h" |
| 54 | #include "llvm/Support/Compiler.h" |
| 55 | #include "llvm/Support/Debug.h" |
| 56 | #include "llvm/Support/raw_ostream.h" |
| 57 | #include <algorithm> |
| 58 | #include <cassert> |
| 59 | #include <limits> |
| 60 | #include <memory> |
| 61 | #include <utility> |
| 62 | |
| 63 | using namespace llvm; |
| 64 | |
| 65 | #define DEBUG_TYPE "stack-coloring" |
| 66 | |
| 67 | static cl::opt<bool> |
| 68 | DisableColoring("no-stack-coloring", |
| 69 | cl::init(Val: false), cl::Hidden, |
| 70 | cl::desc("Disable stack coloring")); |
| 71 | |
| 72 | /// The user may write code that uses allocas outside of the declared lifetime |
| 73 | /// zone. This can happen when the user returns a reference to a local |
| 74 | /// data-structure. We can detect these cases and decide not to optimize the |
| 75 | /// code. If this flag is enabled, we try to save the user. This option |
| 76 | /// is treated as overriding LifetimeStartOnFirstUse below. |
| 77 | static cl::opt<bool> |
| 78 | ProtectFromEscapedAllocas("protect-from-escaped-allocas", |
| 79 | cl::init(Val: false), cl::Hidden, |
| 80 | cl::desc("Do not optimize lifetime zones that " |
| 81 | "are broken")); |
| 82 | |
| 83 | /// Enable enhanced dataflow scheme for lifetime analysis (treat first |
| 84 | /// use of stack slot as start of slot lifetime, as opposed to looking |
| 85 | /// for LIFETIME_START marker). See "Implementation notes" below for |
| 86 | /// more info. |
| 87 | static cl::opt<bool> |
| 88 | LifetimeStartOnFirstUse("stackcoloring-lifetime-start-on-first-use", |
| 89 | cl::init(Val: true), cl::Hidden, |
| 90 | cl::desc("Treat stack lifetimes as starting on first use, not on START marker.")); |
| 91 | |
| 92 | |
| 93 | STATISTIC(NumMarkerSeen, "Number of lifetime markers found."); |
| 94 | STATISTIC(StackSpaceSaved, "Number of bytes saved due to merging slots."); |
| 95 | STATISTIC(StackSlotMerged, "Number of stack slot merged."); |
| 96 | STATISTIC(EscapedAllocas, "Number of allocas that escaped the lifetime region"); |
| 97 | |
| 98 | //===----------------------------------------------------------------------===// |
| 99 | // StackColoring Pass |
| 100 | //===----------------------------------------------------------------------===// |
| 101 | // |
| 102 | // Stack Coloring reduces stack usage by merging stack slots when they |
| 103 | // can't be used together. For example, consider the following C program: |
| 104 | // |
| 105 | // void bar(char *, int); |
| 106 | // void foo(bool var) { |
| 107 | // A: { |
| 108 | // char z[4096]; |
| 109 | // bar(z, 0); |
| 110 | // } |
| 111 | // |
| 112 | // char *p; |
| 113 | // char x[4096]; |
| 114 | // char y[4096]; |
| 115 | // if (var) { |
| 116 | // p = x; |
| 117 | // } else { |
| 118 | // bar(y, 1); |
| 119 | // p = y + 1024; |
| 120 | // } |
| 121 | // B: |
| 122 | // bar(p, 2); |
| 123 | // } |
| 124 | // |
| 125 | // Naively-compiled, this program would use 12k of stack space. However, the |
| 126 | // stack slot corresponding to `z` is always destroyed before either of the |
| 127 | // stack slots for `x` or `y` are used, and then `x` is only used if `var` |
| 128 | // is true, while `y` is only used if `var` is false. So in no time are 2 |
| 129 | // of the stack slots used together, and therefore we can merge them, |
| 130 | // compiling the function using only a single 4k alloca: |
| 131 | // |
| 132 | // void foo(bool var) { // equivalent |
| 133 | // char x[4096]; |
| 134 | // char *p; |
| 135 | // bar(x, 0); |
| 136 | // if (var) { |
| 137 | // p = x; |
| 138 | // } else { |
| 139 | // bar(x, 1); |
| 140 | // p = x + 1024; |
| 141 | // } |
| 142 | // bar(p, 2); |
| 143 | // } |
| 144 | // |
| 145 | // This is an important optimization if we want stack space to be under |
| 146 | // control in large functions, both open-coded ones and ones created by |
| 147 | // inlining. |
| 148 | // |
| 149 | // Implementation Notes: |
| 150 | // --------------------- |
| 151 | // |
| 152 | // An important part of the above reasoning is that `z` can't be accessed |
| 153 | // while the latter 2 calls to `bar` are running. This is justified because |
| 154 | // `z`'s lifetime is over after we exit from block `A:`, so any further |
| 155 | // accesses to it would be UB. The way we represent this information |
| 156 | // in LLVM is by having frontends delimit blocks with `lifetime.start` |
| 157 | // and `lifetime.end` intrinsics. |
| 158 | // |
| 159 | // The effect of these intrinsics seems to be as follows (maybe I should |
| 160 | // specify this in the reference?): |
| 161 | // |
| 162 | // L1) at start, each stack-slot is marked as *out-of-scope*, unless no |
| 163 | // lifetime intrinsic refers to that stack slot, in which case |
| 164 | // it is marked as *in-scope*. |
| 165 | // L2) on a `lifetime.start`, a stack slot is marked as *in-scope* and |
| 166 | // the stack slot is overwritten with `undef`. |
| 167 | // L3) on a `lifetime.end`, a stack slot is marked as *out-of-scope*. |
| 168 | // L4) on function exit, all stack slots are marked as *out-of-scope*. |
| 169 | // L5) `lifetime.end` is a no-op when called on a slot that is already |
| 170 | // *out-of-scope*. |
| 171 | // L6) memory accesses to *out-of-scope* stack slots are UB. |
| 172 | // L7) when a stack-slot is marked as *out-of-scope*, all pointers to it |
| 173 | // are invalidated, unless the slot is "degenerate". This is used to |
| 174 | // justify not marking slots as in-use until the pointer to them is |
| 175 | // used, but feels a bit hacky in the presence of things like LICM. See |
| 176 | // the "Degenerate Slots" section for more details. |
| 177 | // |
| 178 | // Now, let's ground stack coloring on these rules. We'll define a slot |
| 179 | // as *in-use* at a (dynamic) point in execution if it either can be |
| 180 | // written to at that point, or if it has a live and non-undef content |
| 181 | // at that point. |
| 182 | // |
| 183 | // Obviously, slots that are never *in-use* together can be merged, and |
| 184 | // in our example `foo`, the slots for `x`, `y` and `z` are never |
| 185 | // in-use together (of course, sometimes slots that *are* in-use together |
| 186 | // might still be mergable, but we don't care about that here). |
| 187 | // |
| 188 | // In this implementation, we successively merge pairs of slots that are |
| 189 | // not *in-use* together. We could be smarter - for example, we could merge |
| 190 | // a single large slot with 2 small slots, or we could construct the |
| 191 | // interference graph and run a "smart" graph coloring algorithm, but with |
| 192 | // that aside, how do we find out whether a pair of slots might be *in-use* |
| 193 | // together? |
| 194 | // |
| 195 | // From our rules, we see that *out-of-scope* slots are never *in-use*, |
| 196 | // and from (L7) we see that "non-degenerate" slots remain non-*in-use* |
| 197 | // until their address is taken. Therefore, we can approximate slot activity |
| 198 | // using dataflow. |
| 199 | // |
| 200 | // A subtle point: naively, we might try to figure out which pairs of |
| 201 | // stack-slots interfere by propagating `S in-use` through the CFG for every |
| 202 | // stack-slot `S`, and having `S` and `T` interfere if there is a CFG point in |
| 203 | // which they are both *in-use*. |
| 204 | // |
| 205 | // That is sound, but overly conservative in some cases: in our (artificial) |
| 206 | // example `foo`, either `x` or `y` might be in use at the label `B:`, but |
| 207 | // as `x` is only in use if we came in from the `var` edge and `y` only |
| 208 | // if we came from the `!var` edge, they still can't be in use together. |
| 209 | // See PR32488 for an important real-life case. |
| 210 | // |
| 211 | // If we wanted to find all points of interference precisely, we could |
| 212 | // propagate `S in-use` and `S&T in-use` predicates through the CFG. That |
| 213 | // would be precise, but requires propagating `O(n^2)` dataflow facts. |
| 214 | // |
| 215 | // However, we aren't interested in the *set* of points of interference |
| 216 | // between 2 stack slots, only *whether* there *is* such a point. So we |
| 217 | // can rely on a little trick: for `S` and `T` to be in-use together, |
| 218 | // one of them needs to become in-use while the other is in-use (or |
| 219 | // they might both become in use simultaneously). We can check this |
| 220 | // by also keeping track of the points at which a stack slot might *start* |
| 221 | // being in-use. |
| 222 | // |
| 223 | // Exact first use: |
| 224 | // ---------------- |
| 225 | // |
| 226 | // Consider the following motivating example: |
| 227 | // |
| 228 | // int foo() { |
| 229 | // char b1[1024], b2[1024]; |
| 230 | // if (...) { |
| 231 | // char b3[1024]; |
| 232 | // <uses of b1, b3>; |
| 233 | // return x; |
| 234 | // } else { |
| 235 | // char b4[1024], b5[1024]; |
| 236 | // <uses of b2, b4, b5>; |
| 237 | // return y; |
| 238 | // } |
| 239 | // } |
| 240 | // |
| 241 | // In the code above, "b3" and "b4" are declared in distinct lexical |
| 242 | // scopes, meaning that it is easy to prove that they can share the |
| 243 | // same stack slot. Variables "b1" and "b2" are declared in the same |
| 244 | // scope, meaning that from a lexical point of view, their lifetimes |
| 245 | // overlap. From a control flow pointer of view, however, the two |
| 246 | // variables are accessed in disjoint regions of the CFG, thus it |
| 247 | // should be possible for them to share the same stack slot. An ideal |
| 248 | // stack allocation for the function above would look like: |
| 249 | // |
| 250 | // slot 0: b1, b2 |
| 251 | // slot 1: b3, b4 |
| 252 | // slot 2: b5 |
| 253 | // |
| 254 | // Achieving this allocation is tricky, however, due to the way |
| 255 | // lifetime markers are inserted. Here is a simplified view of the |
| 256 | // control flow graph for the code above: |
| 257 | // |
| 258 | // +------ block 0 -------+ |
| 259 | // 0| LIFETIME_START b1, b2 | |
| 260 | // 1| <test 'if' condition> | |
| 261 | // +-----------------------+ |
| 262 | // ./ \. |
| 263 | // +------ block 1 -------+ +------ block 2 -------+ |
| 264 | // 2| LIFETIME_START b3 | 5| LIFETIME_START b4, b5 | |
| 265 | // 3| <uses of b1, b3> | 6| <uses of b2, b4, b5> | |
| 266 | // 4| LIFETIME_END b3 | 7| LIFETIME_END b4, b5 | |
| 267 | // +-----------------------+ +-----------------------+ |
| 268 | // \. /. |
| 269 | // +------ block 3 -------+ |
| 270 | // 8| <cleanupcode> | |
| 271 | // 9| LIFETIME_END b1, b2 | |
| 272 | // 10| return | |
| 273 | // +-----------------------+ |
| 274 | // |
| 275 | // If we create live intervals for the variables above strictly based |
| 276 | // on the lifetime markers, we'll get the set of intervals on the |
| 277 | // left. If we ignore the lifetime start markers and instead treat a |
| 278 | // variable's lifetime as beginning with the first reference to the |
| 279 | // var, then we get the intervals on the right. |
| 280 | // |
| 281 | // LIFETIME_START First Use |
| 282 | // b1: [0,9] [3,4] [8,9] |
| 283 | // b2: [0,9] [6,9] |
| 284 | // b3: [2,4] [3,4] |
| 285 | // b4: [5,7] [6,7] |
| 286 | // b5: [5,7] [6,7] |
| 287 | // |
| 288 | // For the intervals on the left, the best we can do is overlap two |
| 289 | // variables (b3 and b4, for example); this gives us a stack size of |
| 290 | // 4*1024 bytes, not ideal. When treating first-use as the start of a |
| 291 | // lifetime, we can additionally overlap b1 and b5, giving us a 3*1024 |
| 292 | // byte stack (better). |
| 293 | // |
| 294 | // Degenerate Slots: |
| 295 | // ----------------- |
| 296 | // |
| 297 | // Relying entirely on first-use of stack slots is problematic, |
| 298 | // however, due to the fact that optimizations can sometimes migrate |
| 299 | // uses of a variable outside of its lifetime start/end region. Here |
| 300 | // is an example: |
| 301 | // |
| 302 | // int bar() { |
| 303 | // char b1[1024], b2[1024]; |
| 304 | // if (...) { |
| 305 | // <uses of b2> |
| 306 | // return y; |
| 307 | // } else { |
| 308 | // <uses of b1> |
| 309 | // while (...) { |
| 310 | // char b3[1024]; |
| 311 | // <uses of b3> |
| 312 | // } |
| 313 | // } |
| 314 | // } |
| 315 | // |
| 316 | // Before optimization, the control flow graph for the code above |
| 317 | // might look like the following: |
| 318 | // |
| 319 | // +------ block 0 -------+ |
| 320 | // 0| LIFETIME_START b1, b2 | |
| 321 | // 1| <test 'if' condition> | |
| 322 | // +-----------------------+ |
| 323 | // ./ \. |
| 324 | // +------ block 1 -------+ +------- block 2 -------+ |
| 325 | // 2| <uses of b2> | 3| <uses of b1> | |
| 326 | // +-----------------------+ +-----------------------+ |
| 327 | // | | |
| 328 | // | +------- block 3 -------+ <-\. |
| 329 | // | 4| <while condition> | | |
| 330 | // | +-----------------------+ | |
| 331 | // | / | | |
| 332 | // | / +------- block 4 -------+ |
| 333 | // \ / 5| LIFETIME_START b3 | | |
| 334 | // \ / 6| <uses of b3> | | |
| 335 | // \ / 7| LIFETIME_END b3 | | |
| 336 | // \ | +------------------------+ | |
| 337 | // \ | \ / |
| 338 | // +------ block 5 -----+ \--------------- |
| 339 | // 8| <cleanupcode> | |
| 340 | // 9| LIFETIME_END b1, b2 | |
| 341 | // 10| return | |
| 342 | // +---------------------+ |
| 343 | // |
| 344 | // During optimization, however, it can happen that an instruction |
| 345 | // computing an address in "b3" (for example, a loop-invariant GEP) is |
| 346 | // hoisted up out of the loop from block 4 to block 2. [Note that |
| 347 | // this is not an actual load from the stack, only an instruction that |
| 348 | // computes the address to be loaded]. If this happens, there is now a |
| 349 | // path leading from the first use of b3 to the return instruction |
| 350 | // that does not encounter the b3 LIFETIME_END, hence b3's lifetime is |
| 351 | // now larger than if we were computing live intervals strictly based |
| 352 | // on lifetime markers. In the example above, this lengthened lifetime |
| 353 | // would mean that it would appear illegal to overlap b3 with b2. |
| 354 | // |
| 355 | // To deal with this such cases, the code in ::collectMarkers() below |
| 356 | // tries to identify "degenerate" slots -- those slots where on a single |
| 357 | // forward pass through the CFG we encounter a first reference to slot |
| 358 | // K before we hit the slot K lifetime start marker. For such slots, |
| 359 | // we fall back on using the lifetime start marker as the beginning of |
| 360 | // the variable's lifetime. NB: with this implementation, slots can |
| 361 | // appear degenerate in cases where there is unstructured control flow: |
| 362 | // |
| 363 | // if (q) goto mid; |
| 364 | // if (x > 9) { |
| 365 | // int b[100]; |
| 366 | // memcpy(&b[0], ...); |
| 367 | // mid: b[k] = ...; |
| 368 | // abc(&b); |
| 369 | // } |
| 370 | // |
| 371 | // If in RPO ordering chosen to walk the CFG we happen to visit the b[k] |
| 372 | // before visiting the memcpy block (which will contain the lifetime start |
| 373 | // for "b" then it will appear that 'b' has a degenerate lifetime. |
| 374 | |
| 375 | namespace { |
| 376 | |
| 377 | /// StackColoring - A machine pass for merging disjoint stack allocations, |
| 378 | /// marked by the LIFETIME_START and LIFETIME_END pseudo instructions. |
| 379 | class StackColoring : public MachineFunctionPass { |
| 380 | MachineFrameInfo *MFI = nullptr; |
| 381 | MachineFunction *MF = nullptr; |
| 382 | |
| 383 | /// A class representing liveness information for a single basic block. |
| 384 | /// Each bit in the BitVector represents the liveness property |
| 385 | /// for a different stack slot. |
| 386 | struct BlockLifetimeInfo { |
| 387 | /// Which slots BEGINs in each basic block. |
| 388 | BitVector Begin; |
| 389 | |
| 390 | /// Which slots ENDs in each basic block. |
| 391 | BitVector End; |
| 392 | |
| 393 | /// Which slots are marked as LIVE_IN, coming into each basic block. |
| 394 | BitVector LiveIn; |
| 395 | |
| 396 | /// Which slots are marked as LIVE_OUT, coming out of each basic block. |
| 397 | BitVector LiveOut; |
| 398 | }; |
| 399 | |
| 400 | /// Maps active slots (per bit) for each basic block. |
| 401 | using LivenessMap = DenseMap<const MachineBasicBlock *, BlockLifetimeInfo>; |
| 402 | LivenessMap BlockLiveness; |
| 403 | |
| 404 | /// Maps serial numbers to basic blocks. |
| 405 | DenseMap<const MachineBasicBlock *, int> BasicBlocks; |
| 406 | |
| 407 | /// Maps basic blocks to a serial number. |
| 408 | SmallVector<const MachineBasicBlock *, 8> BasicBlockNumbering; |
| 409 | |
| 410 | /// Maps slots to their use interval. Outside of this interval, slots |
| 411 | /// values are either dead or `undef` and they will not be written to. |
| 412 | SmallVector<std::unique_ptr<LiveInterval>, 16> Intervals; |
| 413 | |
| 414 | /// Maps slots to the points where they can become in-use. |
| 415 | SmallVector<SmallVector<SlotIndex, 4>, 16> LiveStarts; |
| 416 | |
| 417 | /// VNInfo is used for the construction of LiveIntervals. |
| 418 | VNInfo::Allocator VNInfoAllocator; |
| 419 | |
| 420 | /// SlotIndex analysis object. |
| 421 | SlotIndexes *Indexes = nullptr; |
| 422 | |
| 423 | /// The list of lifetime markers found. These markers are to be removed |
| 424 | /// once the coloring is done. |
| 425 | SmallVector<MachineInstr*, 8> Markers; |
| 426 | |
| 427 | /// Record the FI slots for which we have seen some sort of |
| 428 | /// lifetime marker (either start or end). |
| 429 | BitVector InterestingSlots; |
| 430 | |
| 431 | /// FI slots that need to be handled conservatively (for these |
| 432 | /// slots lifetime-start-on-first-use is disabled). |
| 433 | BitVector ConservativeSlots; |
| 434 | |
| 435 | /// Number of iterations taken during data flow analysis. |
| 436 | unsigned NumIterations; |
| 437 | |
| 438 | public: |
| 439 | static char ID; |
| 440 | |
| 441 | StackColoring() : MachineFunctionPass(ID) { |
| 442 | initializeStackColoringPass(*PassRegistry::getPassRegistry()); |
| 443 | } |
| 444 | |
| 445 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
| 446 | bool runOnMachineFunction(MachineFunction &Func) override; |
| 447 | |
| 448 | private: |
| 449 | /// Used in collectMarkers |
| 450 | using BlockBitVecMap = DenseMap<const MachineBasicBlock *, BitVector>; |
| 451 | |
| 452 | /// Debug. |
| 453 | void dump() const; |
| 454 | void dumpIntervals() const; |
| 455 | void dumpBB(MachineBasicBlock *MBB) const; |
| 456 | void dumpBV(const char *tag, const BitVector &BV) const; |
| 457 | |
| 458 | /// Removes all of the lifetime marker instructions from the function. |
| 459 | /// \returns true if any markers were removed. |
| 460 | bool removeAllMarkers(); |
| 461 | |
| 462 | /// Scan the machine function and find all of the lifetime markers. |
| 463 | /// Record the findings in the BEGIN and END vectors. |
| 464 | /// \returns the number of markers found. |
| 465 | unsigned collectMarkers(unsigned NumSlot); |
| 466 | |
| 467 | /// Perform the dataflow calculation and calculate the lifetime for each of |
| 468 | /// the slots, based on the BEGIN/END vectors. Set the LifetimeLIVE_IN and |
| 469 | /// LifetimeLIVE_OUT maps that represent which stack slots are live coming |
| 470 | /// in and out blocks. |
| 471 | void calculateLocalLiveness(); |
| 472 | |
| 473 | /// Returns TRUE if we're using the first-use-begins-lifetime method for |
| 474 | /// this slot (if FALSE, then the start marker is treated as start of lifetime). |
| 475 | bool applyFirstUse(int Slot) { |
| 476 | if (!LifetimeStartOnFirstUse || ProtectFromEscapedAllocas) |
| 477 | return false; |
| 478 | if (ConservativeSlots.test(Idx: Slot)) |
| 479 | return false; |
| 480 | return true; |
| 481 | } |
| 482 | |
| 483 | /// Examines the specified instruction and returns TRUE if the instruction |
| 484 | /// represents the start or end of an interesting lifetime. The slot or slots |
| 485 | /// starting or ending are added to the vector "slots" and "isStart" is set |
| 486 | /// accordingly. |
| 487 | /// \returns True if inst contains a lifetime start or end |
| 488 | bool isLifetimeStartOrEnd(const MachineInstr &MI, |
| 489 | SmallVector<int, 4> &slots, |
| 490 | bool &isStart); |
| 491 | |
| 492 | /// Construct the LiveIntervals for the slots. |
| 493 | void calculateLiveIntervals(unsigned NumSlots); |
| 494 | |
| 495 | /// Go over the machine function and change instructions which use stack |
| 496 | /// slots to use the joint slots. |
| 497 | void remapInstructions(DenseMap<int, int> &SlotRemap); |
| 498 | |
| 499 | /// The input program may contain instructions which are not inside lifetime |
| 500 | /// markers. This can happen due to a bug in the compiler or due to a bug in |
| 501 | /// user code (for example, returning a reference to a local variable). |
| 502 | /// This procedure checks all of the instructions in the function and |
| 503 | /// invalidates lifetime ranges which do not contain all of the instructions |
| 504 | /// which access that frame slot. |
| 505 | void removeInvalidSlotRanges(); |
| 506 | |
| 507 | /// Map entries which point to other entries to their destination. |
| 508 | /// A->B->C becomes A->C. |
| 509 | void expungeSlotMap(DenseMap<int, int> &SlotRemap, unsigned NumSlots); |
| 510 | }; |
| 511 | |
| 512 | } // end anonymous namespace |
| 513 | |
| 514 | char StackColoring::ID = 0; |
| 515 | |
| 516 | char &llvm::StackColoringID = StackColoring::ID; |
| 517 | |
| 518 | INITIALIZE_PASS_BEGIN(StackColoring, DEBUG_TYPE, |
| 519 | "Merge disjoint stack slots", false, false) |
| 520 | INITIALIZE_PASS_DEPENDENCY(SlotIndexesWrapperPass) |
| 521 | INITIALIZE_PASS_END(StackColoring, DEBUG_TYPE, |
| 522 | "Merge disjoint stack slots", false, false) |
| 523 | |
| 524 | void StackColoring::getAnalysisUsage(AnalysisUsage &AU) const { |
| 525 | AU.addRequired<SlotIndexesWrapperPass>(); |
| 526 | MachineFunctionPass::getAnalysisUsage(AU); |
| 527 | } |
| 528 | |
| 529 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
| 530 | LLVM_DUMP_METHOD void StackColoring::dumpBV(const char *tag, |
| 531 | const BitVector &BV) const { |
| 532 | dbgs() << tag << " : { "; |
| 533 | for (unsigned I = 0, E = BV.size(); I != E; ++I) |
| 534 | dbgs() << BV.test(I) << " "; |
| 535 | dbgs() << "}\n"; |
| 536 | } |
| 537 | |
| 538 | LLVM_DUMP_METHOD void StackColoring::dumpBB(MachineBasicBlock *MBB) const { |
| 539 | LivenessMap::const_iterator BI = BlockLiveness.find(MBB); |
| 540 | assert(BI != BlockLiveness.end() && "Block not found"); |
| 541 | const BlockLifetimeInfo &BlockInfo = BI->second; |
| 542 | |
| 543 | dumpBV("BEGIN", BlockInfo.Begin); |
| 544 | dumpBV("END", BlockInfo.End); |
| 545 | dumpBV("LIVE_IN", BlockInfo.LiveIn); |
| 546 | dumpBV("LIVE_OUT", BlockInfo.LiveOut); |
| 547 | } |
| 548 | |
| 549 | LLVM_DUMP_METHOD void StackColoring::dump() const { |
| 550 | for (MachineBasicBlock *MBB : depth_first(MF)) { |
| 551 | dbgs() << "Inspecting block #"<< MBB->getNumber() << " [" |
| 552 | << MBB->getName() << "]\n"; |
| 553 | dumpBB(MBB); |
| 554 | } |
| 555 | } |
| 556 | |
| 557 | LLVM_DUMP_METHOD void StackColoring::dumpIntervals() const { |
| 558 | for (unsigned I = 0, E = Intervals.size(); I != E; ++I) { |
| 559 | dbgs() << "Interval["<< I << "]:\n"; |
| 560 | Intervals[I]->dump(); |
| 561 | } |
| 562 | } |
| 563 | #endif |
| 564 | |
| 565 | static inline int getStartOrEndSlot(const MachineInstr &MI) |
| 566 | { |
| 567 | assert((MI.getOpcode() == TargetOpcode::LIFETIME_START || |
| 568 | MI.getOpcode() == TargetOpcode::LIFETIME_END) && |
| 569 | "Expected LIFETIME_START or LIFETIME_END op"); |
| 570 | const MachineOperand &MO = MI.getOperand(i: 0); |
| 571 | int Slot = MO.getIndex(); |
| 572 | if (Slot >= 0) |
| 573 | return Slot; |
| 574 | return -1; |
| 575 | } |
| 576 | |
| 577 | // At the moment the only way to end a variable lifetime is with |
| 578 | // a VARIABLE_LIFETIME op (which can't contain a start). If things |
| 579 | // change and the IR allows for a single inst that both begins |
| 580 | // and ends lifetime(s), this interface will need to be reworked. |
| 581 | bool StackColoring::isLifetimeStartOrEnd(const MachineInstr &MI, |
| 582 | SmallVector<int, 4> &slots, |
| 583 | bool &isStart) { |
| 584 | if (MI.getOpcode() == TargetOpcode::LIFETIME_START || |
| 585 | MI.getOpcode() == TargetOpcode::LIFETIME_END) { |
| 586 | int Slot = getStartOrEndSlot(MI); |
| 587 | if (Slot < 0) |
| 588 | return false; |
| 589 | if (!InterestingSlots.test(Idx: Slot)) |
| 590 | return false; |
| 591 | slots.push_back(Elt: Slot); |
| 592 | if (MI.getOpcode() == TargetOpcode::LIFETIME_END) { |
| 593 | isStart = false; |
| 594 | return true; |
| 595 | } |
| 596 | if (!applyFirstUse(Slot)) { |
| 597 | isStart = true; |
| 598 | return true; |
| 599 | } |
| 600 | } else if (LifetimeStartOnFirstUse && !ProtectFromEscapedAllocas) { |
| 601 | if (!MI.isDebugInstr()) { |
| 602 | bool found = false; |
| 603 | for (const MachineOperand &MO : MI.operands()) { |
| 604 | if (!MO.isFI()) |
| 605 | continue; |
| 606 | int Slot = MO.getIndex(); |
| 607 | if (Slot<0) |
| 608 | continue; |
| 609 | if (InterestingSlots.test(Idx: Slot) && applyFirstUse(Slot)) { |
| 610 | slots.push_back(Elt: Slot); |
| 611 | found = true; |
| 612 | } |
| 613 | } |
| 614 | if (found) { |
| 615 | isStart = true; |
| 616 | return true; |
| 617 | } |
| 618 | } |
| 619 | } |
| 620 | return false; |
| 621 | } |
| 622 | |
| 623 | unsigned StackColoring::collectMarkers(unsigned NumSlot) { |
| 624 | unsigned MarkersFound = 0; |
| 625 | BlockBitVecMap SeenStartMap; |
| 626 | InterestingSlots.clear(); |
| 627 | InterestingSlots.resize(N: NumSlot); |
| 628 | ConservativeSlots.clear(); |
| 629 | ConservativeSlots.resize(N: NumSlot); |
| 630 | |
| 631 | // number of start and end lifetime ops for each slot |
| 632 | SmallVector<int, 8> NumStartLifetimes(NumSlot, 0); |
| 633 | SmallVector<int, 8> NumEndLifetimes(NumSlot, 0); |
| 634 | |
| 635 | // Step 1: collect markers and populate the "InterestingSlots" |
| 636 | // and "ConservativeSlots" sets. |
| 637 | for (MachineBasicBlock *MBB : depth_first(G: MF)) { |
| 638 | // Compute the set of slots for which we've seen a START marker but have |
| 639 | // not yet seen an END marker at this point in the walk (e.g. on entry |
| 640 | // to this bb). |
| 641 | BitVector BetweenStartEnd; |
| 642 | BetweenStartEnd.resize(N: NumSlot); |
| 643 | for (const MachineBasicBlock *Pred : MBB->predecessors()) { |
| 644 | BlockBitVecMap::const_iterator I = SeenStartMap.find(Val: Pred); |
| 645 | if (I != SeenStartMap.end()) { |
| 646 | BetweenStartEnd |= I->second; |
| 647 | } |
| 648 | } |
| 649 | |
| 650 | // Walk the instructions in the block to look for start/end ops. |
| 651 | for (MachineInstr &MI : *MBB) { |
| 652 | if (MI.isDebugInstr()) |
| 653 | continue; |
| 654 | if (MI.getOpcode() == TargetOpcode::LIFETIME_START || |
| 655 | MI.getOpcode() == TargetOpcode::LIFETIME_END) { |
| 656 | int Slot = getStartOrEndSlot(MI); |
| 657 | if (Slot < 0) |
| 658 | continue; |
| 659 | InterestingSlots.set(Slot); |
| 660 | if (MI.getOpcode() == TargetOpcode::LIFETIME_START) { |
| 661 | BetweenStartEnd.set(Slot); |
| 662 | NumStartLifetimes[Slot] += 1; |
| 663 | } else { |
| 664 | BetweenStartEnd.reset(Idx: Slot); |
| 665 | NumEndLifetimes[Slot] += 1; |
| 666 | } |
| 667 | const AllocaInst *Allocation = MFI->getObjectAllocation(ObjectIdx: Slot); |
| 668 | if (Allocation) { |
| 669 | LLVM_DEBUG(dbgs() << "Found a lifetime "); |
| 670 | LLVM_DEBUG(dbgs() << (MI.getOpcode() == TargetOpcode::LIFETIME_START |
| 671 | ? "start" |
| 672 | : "end")); |
| 673 | LLVM_DEBUG(dbgs() << " marker for slot #"<< Slot); |
| 674 | LLVM_DEBUG(dbgs() |
| 675 | << " with allocation: "<< Allocation->getName() << "\n"); |
| 676 | } |
| 677 | Markers.push_back(Elt: &MI); |
| 678 | MarkersFound += 1; |
| 679 | } else { |
| 680 | for (const MachineOperand &MO : MI.operands()) { |
| 681 | if (!MO.isFI()) |
| 682 | continue; |
| 683 | int Slot = MO.getIndex(); |
| 684 | if (Slot < 0) |
| 685 | continue; |
| 686 | if (! BetweenStartEnd.test(Idx: Slot)) { |
| 687 | ConservativeSlots.set(Slot); |
| 688 | } |
| 689 | } |
| 690 | } |
| 691 | } |
| 692 | BitVector &SeenStart = SeenStartMap[MBB]; |
| 693 | SeenStart |= BetweenStartEnd; |
| 694 | } |
| 695 | if (!MarkersFound) { |
| 696 | return 0; |
| 697 | } |
| 698 | |
| 699 | // PR27903: slots with multiple start or end lifetime ops are not |
| 700 | // safe to enable for "lifetime-start-on-first-use". |
| 701 | for (unsigned slot = 0; slot < NumSlot; ++slot) { |
| 702 | if (NumStartLifetimes[slot] > 1 || NumEndLifetimes[slot] > 1) |
| 703 | ConservativeSlots.set(slot); |
| 704 | } |
| 705 | |
| 706 | // The write to the catch object by the personality function is not propely |
| 707 | // modeled in IR: It happens before any cleanuppads are executed, even if the |
| 708 | // first mention of the catch object is in a catchpad. As such, mark catch |
| 709 | // object slots as conservative, so they are excluded from first-use analysis. |
| 710 | if (WinEHFuncInfo *EHInfo = MF->getWinEHFuncInfo()) |
| 711 | for (WinEHTryBlockMapEntry &TBME : EHInfo->TryBlockMap) |
| 712 | for (WinEHHandlerType &H : TBME.HandlerArray) |
| 713 | if (H.CatchObj.FrameIndex != std::numeric_limits<int>::max() && |
| 714 | H.CatchObj.FrameIndex >= 0) |
| 715 | ConservativeSlots.set(H.CatchObj.FrameIndex); |
| 716 | |
| 717 | LLVM_DEBUG(dumpBV("Conservative slots", ConservativeSlots)); |
| 718 | |
| 719 | // Step 2: compute begin/end sets for each block |
| 720 | |
| 721 | // NOTE: We use a depth-first iteration to ensure that we obtain a |
| 722 | // deterministic numbering. |
| 723 | for (MachineBasicBlock *MBB : depth_first(G: MF)) { |
| 724 | // Assign a serial number to this basic block. |
| 725 | BasicBlocks[MBB] = BasicBlockNumbering.size(); |
| 726 | BasicBlockNumbering.push_back(Elt: MBB); |
| 727 | |
| 728 | // Keep a reference to avoid repeated lookups. |
| 729 | BlockLifetimeInfo &BlockInfo = BlockLiveness[MBB]; |
| 730 | |
| 731 | BlockInfo.Begin.resize(N: NumSlot); |
| 732 | BlockInfo.End.resize(N: NumSlot); |
| 733 | |
| 734 | SmallVector<int, 4> slots; |
| 735 | for (MachineInstr &MI : *MBB) { |
| 736 | bool isStart = false; |
| 737 | slots.clear(); |
| 738 | if (isLifetimeStartOrEnd(MI, slots, isStart)) { |
| 739 | if (!isStart) { |
| 740 | assert(slots.size() == 1 && "unexpected: MI ends multiple slots"); |
| 741 | int Slot = slots[0]; |
| 742 | if (BlockInfo.Begin.test(Idx: Slot)) { |
| 743 | BlockInfo.Begin.reset(Idx: Slot); |
| 744 | } |
| 745 | BlockInfo.End.set(Slot); |
| 746 | } else { |
| 747 | for (auto Slot : slots) { |
| 748 | LLVM_DEBUG(dbgs() << "Found a use of slot #"<< Slot); |
| 749 | LLVM_DEBUG(dbgs() |
| 750 | << " at "<< printMBBReference(*MBB) << " index "); |
| 751 | LLVM_DEBUG(Indexes->getInstructionIndex(MI).print(dbgs())); |
| 752 | const AllocaInst *Allocation = MFI->getObjectAllocation(ObjectIdx: Slot); |
| 753 | if (Allocation) { |
| 754 | LLVM_DEBUG(dbgs() |
| 755 | << " with allocation: "<< Allocation->getName()); |
| 756 | } |
| 757 | LLVM_DEBUG(dbgs() << "\n"); |
| 758 | if (BlockInfo.End.test(Idx: Slot)) { |
| 759 | BlockInfo.End.reset(Idx: Slot); |
| 760 | } |
| 761 | BlockInfo.Begin.set(Slot); |
| 762 | } |
| 763 | } |
| 764 | } |
| 765 | } |
| 766 | } |
| 767 | |
| 768 | // Update statistics. |
| 769 | NumMarkerSeen += MarkersFound; |
| 770 | return MarkersFound; |
| 771 | } |
| 772 | |
| 773 | void StackColoring::calculateLocalLiveness() { |
| 774 | unsigned NumIters = 0; |
| 775 | bool changed = true; |
| 776 | // Create BitVector outside the loop and reuse them to avoid repeated heap |
| 777 | // allocations. |
| 778 | BitVector LocalLiveIn; |
| 779 | BitVector LocalLiveOut; |
| 780 | while (changed) { |
| 781 | changed = false; |
| 782 | ++NumIters; |
| 783 | |
| 784 | for (const MachineBasicBlock *BB : BasicBlockNumbering) { |
| 785 | // Use an iterator to avoid repeated lookups. |
| 786 | LivenessMap::iterator BI = BlockLiveness.find(Val: BB); |
| 787 | assert(BI != BlockLiveness.end() && "Block not found"); |
| 788 | BlockLifetimeInfo &BlockInfo = BI->second; |
| 789 | |
| 790 | // Compute LiveIn by unioning together the LiveOut sets of all preds. |
| 791 | LocalLiveIn.clear(); |
| 792 | for (MachineBasicBlock *Pred : BB->predecessors()) { |
| 793 | LivenessMap::const_iterator I = BlockLiveness.find(Val: Pred); |
| 794 | // PR37130: transformations prior to stack coloring can |
| 795 | // sometimes leave behind statically unreachable blocks; these |
| 796 | // can be safely skipped here. |
| 797 | if (I != BlockLiveness.end()) |
| 798 | LocalLiveIn |= I->second.LiveOut; |
| 799 | } |
| 800 | |
| 801 | // Compute LiveOut by subtracting out lifetimes that end in this |
| 802 | // block, then adding in lifetimes that begin in this block. If |
| 803 | // we have both BEGIN and END markers in the same basic block |
| 804 | // then we know that the BEGIN marker comes after the END, |
| 805 | // because we already handle the case where the BEGIN comes |
| 806 | // before the END when collecting the markers (and building the |
| 807 | // BEGIN/END vectors). |
| 808 | LocalLiveOut = LocalLiveIn; |
| 809 | LocalLiveOut.reset(RHS: BlockInfo.End); |
| 810 | LocalLiveOut |= BlockInfo.Begin; |
| 811 | |
| 812 | // Update block LiveIn set, noting whether it has changed. |
| 813 | if (LocalLiveIn.test(RHS: BlockInfo.LiveIn)) { |
| 814 | changed = true; |
| 815 | BlockInfo.LiveIn |= LocalLiveIn; |
| 816 | } |
| 817 | |
| 818 | // Update block LiveOut set, noting whether it has changed. |
| 819 | if (LocalLiveOut.test(RHS: BlockInfo.LiveOut)) { |
| 820 | changed = true; |
| 821 | BlockInfo.LiveOut |= LocalLiveOut; |
| 822 | } |
| 823 | } |
| 824 | } // while changed. |
| 825 | |
| 826 | NumIterations = NumIters; |
| 827 | } |
| 828 | |
| 829 | void StackColoring::calculateLiveIntervals(unsigned NumSlots) { |
| 830 | SmallVector<SlotIndex, 16> Starts; |
| 831 | SmallVector<bool, 16> DefinitelyInUse; |
| 832 | |
| 833 | // For each block, find which slots are active within this block |
| 834 | // and update the live intervals. |
| 835 | for (const MachineBasicBlock &MBB : *MF) { |
| 836 | Starts.clear(); |
| 837 | Starts.resize(N: NumSlots); |
| 838 | DefinitelyInUse.clear(); |
| 839 | DefinitelyInUse.resize(N: NumSlots); |
| 840 | |
| 841 | // Start the interval of the slots that we previously found to be 'in-use'. |
| 842 | BlockLifetimeInfo &MBBLiveness = BlockLiveness[&MBB]; |
| 843 | for (int pos = MBBLiveness.LiveIn.find_first(); pos != -1; |
| 844 | pos = MBBLiveness.LiveIn.find_next(Prev: pos)) { |
| 845 | Starts[pos] = Indexes->getMBBStartIdx(mbb: &MBB); |
| 846 | } |
| 847 | |
| 848 | // Create the interval for the basic blocks containing lifetime begin/end. |
| 849 | for (const MachineInstr &MI : MBB) { |
| 850 | SmallVector<int, 4> slots; |
| 851 | bool IsStart = false; |
| 852 | if (!isLifetimeStartOrEnd(MI, slots, isStart&: IsStart)) |
| 853 | continue; |
| 854 | SlotIndex ThisIndex = Indexes->getInstructionIndex(MI); |
| 855 | for (auto Slot : slots) { |
| 856 | if (IsStart) { |
| 857 | // If a slot is already definitely in use, we don't have to emit |
| 858 | // a new start marker because there is already a pre-existing |
| 859 | // one. |
| 860 | if (!DefinitelyInUse[Slot]) { |
| 861 | LiveStarts[Slot].push_back(Elt: ThisIndex); |
| 862 | DefinitelyInUse[Slot] = true; |
| 863 | } |
| 864 | if (!Starts[Slot].isValid()) |
| 865 | Starts[Slot] = ThisIndex; |
| 866 | } else { |
| 867 | if (Starts[Slot].isValid()) { |
| 868 | VNInfo *VNI = Intervals[Slot]->getValNumInfo(ValNo: 0); |
| 869 | Intervals[Slot]->addSegment( |
| 870 | S: LiveInterval::Segment(Starts[Slot], ThisIndex, VNI)); |
| 871 | Starts[Slot] = SlotIndex(); // Invalidate the start index |
| 872 | DefinitelyInUse[Slot] = false; |
| 873 | } |
| 874 | } |
| 875 | } |
| 876 | } |
| 877 | |
| 878 | // Finish up started segments |
| 879 | for (unsigned i = 0; i < NumSlots; ++i) { |
| 880 | if (!Starts[i].isValid()) |
| 881 | continue; |
| 882 | |
| 883 | SlotIndex EndIdx = Indexes->getMBBEndIdx(mbb: &MBB); |
| 884 | VNInfo *VNI = Intervals[i]->getValNumInfo(ValNo: 0); |
| 885 | Intervals[i]->addSegment(S: LiveInterval::Segment(Starts[i], EndIdx, VNI)); |
| 886 | } |
| 887 | } |
| 888 | } |
| 889 | |
| 890 | bool StackColoring::removeAllMarkers() { |
| 891 | unsigned Count = 0; |
| 892 | for (MachineInstr *MI : Markers) { |
| 893 | MI->eraseFromParent(); |
| 894 | Count++; |
| 895 | } |
| 896 | Markers.clear(); |
| 897 | |
| 898 | LLVM_DEBUG(dbgs() << "Removed "<< Count << " markers.\n"); |
| 899 | return Count; |
| 900 | } |
| 901 | |
| 902 | void StackColoring::remapInstructions(DenseMap<int, int> &SlotRemap) { |
| 903 | unsigned FixedInstr = 0; |
| 904 | unsigned FixedMemOp = 0; |
| 905 | unsigned FixedDbg = 0; |
| 906 | |
| 907 | // Remap debug information that refers to stack slots. |
| 908 | for (auto &VI : MF->getVariableDbgInfo()) { |
| 909 | if (!VI.Var || !VI.inStackSlot()) |
| 910 | continue; |
| 911 | int Slot = VI.getStackSlot(); |
| 912 | if (SlotRemap.count(Val: Slot)) { |
| 913 | LLVM_DEBUG(dbgs() << "Remapping debug info for [" |
| 914 | << cast<DILocalVariable>(VI.Var)->getName() << "].\n"); |
| 915 | VI.updateStackSlot(NewSlot: SlotRemap[Slot]); |
| 916 | FixedDbg++; |
| 917 | } |
| 918 | } |
| 919 | |
| 920 | // Keep a list of *allocas* which need to be remapped. |
| 921 | DenseMap<const AllocaInst*, const AllocaInst*> Allocas; |
| 922 | |
| 923 | // Keep a list of allocas which has been affected by the remap. |
| 924 | SmallPtrSet<const AllocaInst*, 32> MergedAllocas; |
| 925 | |
| 926 | for (const std::pair<int, int> &SI : SlotRemap) { |
| 927 | const AllocaInst *From = MFI->getObjectAllocation(ObjectIdx: SI.first); |
| 928 | const AllocaInst *To = MFI->getObjectAllocation(ObjectIdx: SI.second); |
| 929 | assert(To && From && "Invalid allocation object"); |
| 930 | Allocas[From] = To; |
| 931 | |
| 932 | // If From is before wo, its possible that there is a use of From between |
| 933 | // them. |
| 934 | if (From->comesBefore(Other: To)) |
| 935 | const_cast<AllocaInst*>(To)->moveBefore(MovePos: const_cast<AllocaInst*>(From)); |
| 936 | |
| 937 | // AA might be used later for instruction scheduling, and we need it to be |
| 938 | // able to deduce the correct aliasing releationships between pointers |
| 939 | // derived from the alloca being remapped and the target of that remapping. |
| 940 | // The only safe way, without directly informing AA about the remapping |
| 941 | // somehow, is to directly update the IR to reflect the change being made |
| 942 | // here. |
| 943 | Instruction *Inst = const_cast<AllocaInst *>(To); |
| 944 | if (From->getType() != To->getType()) { |
| 945 | BitCastInst *Cast = new BitCastInst(Inst, From->getType()); |
| 946 | Cast->insertAfter(InsertPos: Inst); |
| 947 | Inst = Cast; |
| 948 | } |
| 949 | |
| 950 | // We keep both slots to maintain AliasAnalysis metadata later. |
| 951 | MergedAllocas.insert(Ptr: From); |
| 952 | MergedAllocas.insert(Ptr: To); |
| 953 | |
| 954 | // Transfer the stack protector layout tag, but make sure that SSPLK_AddrOf |
| 955 | // does not overwrite SSPLK_SmallArray or SSPLK_LargeArray, and make sure |
| 956 | // that SSPLK_SmallArray does not overwrite SSPLK_LargeArray. |
| 957 | MachineFrameInfo::SSPLayoutKind FromKind |
| 958 | = MFI->getObjectSSPLayout(ObjectIdx: SI.first); |
| 959 | MachineFrameInfo::SSPLayoutKind ToKind = MFI->getObjectSSPLayout(ObjectIdx: SI.second); |
| 960 | if (FromKind != MachineFrameInfo::SSPLK_None && |
| 961 | (ToKind == MachineFrameInfo::SSPLK_None || |
| 962 | (ToKind != MachineFrameInfo::SSPLK_LargeArray && |
| 963 | FromKind != MachineFrameInfo::SSPLK_AddrOf))) |
| 964 | MFI->setObjectSSPLayout(ObjectIdx: SI.second, Kind: FromKind); |
| 965 | |
| 966 | // The new alloca might not be valid in a llvm.dbg.declare for this |
| 967 | // variable, so poison out the use to make the verifier happy. |
| 968 | AllocaInst *FromAI = const_cast<AllocaInst *>(From); |
| 969 | if (FromAI->isUsedByMetadata()) |
| 970 | ValueAsMetadata::handleRAUW(From: FromAI, To: PoisonValue::get(T: FromAI->getType())); |
| 971 | for (auto &Use : FromAI->uses()) { |
| 972 | if (BitCastInst *BCI = dyn_cast<BitCastInst>(Val: Use.get())) |
| 973 | if (BCI->isUsedByMetadata()) |
| 974 | ValueAsMetadata::handleRAUW(From: BCI, To: PoisonValue::get(T: BCI->getType())); |
| 975 | } |
| 976 | |
| 977 | // Note that this will not replace uses in MMOs (which we'll update below), |
| 978 | // or anywhere else (which is why we won't delete the original |
| 979 | // instruction). |
| 980 | FromAI->replaceAllUsesWith(V: Inst); |
| 981 | } |
| 982 | |
| 983 | // Remap all instructions to the new stack slots. |
| 984 | std::vector<std::vector<MachineMemOperand *>> SSRefs( |
| 985 | MFI->getObjectIndexEnd()); |
| 986 | for (MachineBasicBlock &BB : *MF) |
| 987 | for (MachineInstr &I : BB) { |
| 988 | // Skip lifetime markers. We'll remove them soon. |
| 989 | if (I.getOpcode() == TargetOpcode::LIFETIME_START || |
| 990 | I.getOpcode() == TargetOpcode::LIFETIME_END) |
| 991 | continue; |
| 992 | |
| 993 | // Update the MachineMemOperand to use the new alloca. |
| 994 | for (MachineMemOperand *MMO : I.memoperands()) { |
| 995 | // We've replaced IR-level uses of the remapped allocas, so we only |
| 996 | // need to replace direct uses here. |
| 997 | const AllocaInst *AI = dyn_cast_or_null<AllocaInst>(Val: MMO->getValue()); |
| 998 | if (!AI) |
| 999 | continue; |
| 1000 | |
| 1001 | if (!Allocas.count(Val: AI)) |
| 1002 | continue; |
| 1003 | |
| 1004 | MMO->setValue(Allocas[AI]); |
| 1005 | FixedMemOp++; |
| 1006 | } |
| 1007 | |
| 1008 | // Update all of the machine instruction operands. |
| 1009 | for (MachineOperand &MO : I.operands()) { |
| 1010 | if (!MO.isFI()) |
| 1011 | continue; |
| 1012 | int FromSlot = MO.getIndex(); |
| 1013 | |
| 1014 | // Don't touch arguments. |
| 1015 | if (FromSlot<0) |
| 1016 | continue; |
| 1017 | |
| 1018 | // Only look at mapped slots. |
| 1019 | if (!SlotRemap.count(Val: FromSlot)) |
| 1020 | continue; |
| 1021 | |
| 1022 | // In a debug build, check that the instruction that we are modifying is |
| 1023 | // inside the expected live range. If the instruction is not inside |
| 1024 | // the calculated range then it means that the alloca usage moved |
| 1025 | // outside of the lifetime markers, or that the user has a bug. |
| 1026 | // NOTE: Alloca address calculations which happen outside the lifetime |
| 1027 | // zone are okay, despite the fact that we don't have a good way |
| 1028 | // for validating all of the usages of the calculation. |
| 1029 | #ifndef NDEBUG |
| 1030 | bool TouchesMemory = I.mayLoadOrStore(); |
| 1031 | // If we *don't* protect the user from escaped allocas, don't bother |
| 1032 | // validating the instructions. |
| 1033 | if (!I.isDebugInstr() && TouchesMemory && ProtectFromEscapedAllocas) { |
| 1034 | SlotIndex Index = Indexes->getInstructionIndex(I); |
| 1035 | const LiveInterval *Interval = &*Intervals[FromSlot]; |
| 1036 | assert(Interval->find(Index) != Interval->end() && |
| 1037 | "Found instruction usage outside of live range."); |
| 1038 | } |
| 1039 | #endif |
| 1040 | |
| 1041 | // Fix the machine instructions. |
| 1042 | int ToSlot = SlotRemap[FromSlot]; |
| 1043 | MO.setIndex(ToSlot); |
| 1044 | FixedInstr++; |
| 1045 | } |
| 1046 | |
| 1047 | // We adjust AliasAnalysis information for merged stack slots. |
| 1048 | SmallVector<MachineMemOperand *, 2> NewMMOs; |
| 1049 | bool ReplaceMemOps = false; |
| 1050 | for (MachineMemOperand *MMO : I.memoperands()) { |
| 1051 | // Collect MachineMemOperands which reference |
| 1052 | // FixedStackPseudoSourceValues with old frame indices. |
| 1053 | if (const auto *FSV = dyn_cast_or_null<FixedStackPseudoSourceValue>( |
| 1054 | Val: MMO->getPseudoValue())) { |
| 1055 | int FI = FSV->getFrameIndex(); |
| 1056 | auto To = SlotRemap.find(Val: FI); |
| 1057 | if (To != SlotRemap.end()) |
| 1058 | SSRefs[FI].push_back(x: MMO); |
| 1059 | } |
| 1060 | |
| 1061 | // If this memory location can be a slot remapped here, |
| 1062 | // we remove AA information. |
| 1063 | bool MayHaveConflictingAAMD = false; |
| 1064 | if (MMO->getAAInfo()) { |
| 1065 | if (const Value *MMOV = MMO->getValue()) { |
| 1066 | SmallVector<Value *, 4> Objs; |
| 1067 | getUnderlyingObjectsForCodeGen(V: MMOV, Objects&: Objs); |
| 1068 | |
| 1069 | if (Objs.empty()) |
| 1070 | MayHaveConflictingAAMD = true; |
| 1071 | else |
| 1072 | for (Value *V : Objs) { |
| 1073 | // If this memory location comes from a known stack slot |
| 1074 | // that is not remapped, we continue checking. |
| 1075 | // Otherwise, we need to invalidate AA infomation. |
| 1076 | const AllocaInst *AI = dyn_cast_or_null<AllocaInst>(Val: V); |
| 1077 | if (AI && MergedAllocas.count(Ptr: AI)) { |
| 1078 | MayHaveConflictingAAMD = true; |
| 1079 | break; |
| 1080 | } |
| 1081 | } |
| 1082 | } |
| 1083 | } |
| 1084 | if (MayHaveConflictingAAMD) { |
| 1085 | NewMMOs.push_back(Elt: MF->getMachineMemOperand(MMO, AAInfo: AAMDNodes())); |
| 1086 | ReplaceMemOps = true; |
| 1087 | } else { |
| 1088 | NewMMOs.push_back(Elt: MMO); |
| 1089 | } |
| 1090 | } |
| 1091 | |
| 1092 | // If any memory operand is updated, set memory references of |
| 1093 | // this instruction. |
| 1094 | if (ReplaceMemOps) |
| 1095 | I.setMemRefs(MF&: *MF, MemRefs: NewMMOs); |
| 1096 | } |
| 1097 | |
| 1098 | // Rewrite MachineMemOperands that reference old frame indices. |
| 1099 | for (auto E : enumerate(First&: SSRefs)) |
| 1100 | if (!E.value().empty()) { |
| 1101 | const PseudoSourceValue *NewSV = |
| 1102 | MF->getPSVManager().getFixedStack(FI: SlotRemap.find(Val: E.index())->second); |
| 1103 | for (MachineMemOperand *Ref : E.value()) |
| 1104 | Ref->setValue(NewSV); |
| 1105 | } |
| 1106 | |
| 1107 | // Update the location of C++ catch objects for the MSVC personality routine. |
| 1108 | if (WinEHFuncInfo *EHInfo = MF->getWinEHFuncInfo()) |
| 1109 | for (WinEHTryBlockMapEntry &TBME : EHInfo->TryBlockMap) |
| 1110 | for (WinEHHandlerType &H : TBME.HandlerArray) |
| 1111 | if (H.CatchObj.FrameIndex != std::numeric_limits<int>::max() && |
| 1112 | SlotRemap.count(Val: H.CatchObj.FrameIndex)) |
| 1113 | H.CatchObj.FrameIndex = SlotRemap[H.CatchObj.FrameIndex]; |
| 1114 | |
| 1115 | LLVM_DEBUG(dbgs() << "Fixed "<< FixedMemOp << " machine memory operands.\n"); |
| 1116 | LLVM_DEBUG(dbgs() << "Fixed "<< FixedDbg << " debug locations.\n"); |
| 1117 | LLVM_DEBUG(dbgs() << "Fixed "<< FixedInstr << " machine instructions.\n"); |
| 1118 | (void) FixedMemOp; |
| 1119 | (void) FixedDbg; |
| 1120 | (void) FixedInstr; |
| 1121 | } |
| 1122 | |
| 1123 | void StackColoring::removeInvalidSlotRanges() { |
| 1124 | for (MachineBasicBlock &BB : *MF) |
| 1125 | for (MachineInstr &I : BB) { |
| 1126 | if (I.getOpcode() == TargetOpcode::LIFETIME_START || |
| 1127 | I.getOpcode() == TargetOpcode::LIFETIME_END || I.isDebugInstr()) |
| 1128 | continue; |
| 1129 | |
| 1130 | // Some intervals are suspicious! In some cases we find address |
| 1131 | // calculations outside of the lifetime zone, but not actual memory |
| 1132 | // read or write. Memory accesses outside of the lifetime zone are a clear |
| 1133 | // violation, but address calculations are okay. This can happen when |
| 1134 | // GEPs are hoisted outside of the lifetime zone. |
| 1135 | // So, in here we only check instructions which can read or write memory. |
| 1136 | if (!I.mayLoad() && !I.mayStore()) |
| 1137 | continue; |
| 1138 | |
| 1139 | // Check all of the machine operands. |
| 1140 | for (const MachineOperand &MO : I.operands()) { |
| 1141 | if (!MO.isFI()) |
| 1142 | continue; |
| 1143 | |
| 1144 | int Slot = MO.getIndex(); |
| 1145 | |
| 1146 | if (Slot<0) |
| 1147 | continue; |
| 1148 | |
| 1149 | if (Intervals[Slot]->empty()) |
| 1150 | continue; |
| 1151 | |
| 1152 | // Check that the used slot is inside the calculated lifetime range. |
| 1153 | // If it is not, warn about it and invalidate the range. |
| 1154 | LiveInterval *Interval = &*Intervals[Slot]; |
| 1155 | SlotIndex Index = Indexes->getInstructionIndex(MI: I); |
| 1156 | if (Interval->find(Pos: Index) == Interval->end()) { |
| 1157 | Interval->clear(); |
| 1158 | LLVM_DEBUG(dbgs() << "Invalidating range #"<< Slot << "\n"); |
| 1159 | EscapedAllocas++; |
| 1160 | } |
| 1161 | } |
| 1162 | } |
| 1163 | } |
| 1164 | |
| 1165 | void StackColoring::expungeSlotMap(DenseMap<int, int> &SlotRemap, |
| 1166 | unsigned NumSlots) { |
| 1167 | // Expunge slot remap map. |
| 1168 | for (unsigned i=0; i < NumSlots; ++i) { |
| 1169 | // If we are remapping i |
| 1170 | if (SlotRemap.count(Val: i)) { |
| 1171 | int Target = SlotRemap[i]; |
| 1172 | // As long as our target is mapped to something else, follow it. |
| 1173 | while (SlotRemap.count(Val: Target)) { |
| 1174 | Target = SlotRemap[Target]; |
| 1175 | SlotRemap[i] = Target; |
| 1176 | } |
| 1177 | } |
| 1178 | } |
| 1179 | } |
| 1180 | |
| 1181 | bool StackColoring::runOnMachineFunction(MachineFunction &Func) { |
| 1182 | LLVM_DEBUG(dbgs() << "********** Stack Coloring **********\n" |
| 1183 | << "********** Function: "<< Func.getName() << '\n'); |
| 1184 | MF = &Func; |
| 1185 | MFI = &MF->getFrameInfo(); |
| 1186 | Indexes = &getAnalysis<SlotIndexesWrapperPass>().getSI(); |
| 1187 | BlockLiveness.clear(); |
| 1188 | BasicBlocks.clear(); |
| 1189 | BasicBlockNumbering.clear(); |
| 1190 | Markers.clear(); |
| 1191 | Intervals.clear(); |
| 1192 | LiveStarts.clear(); |
| 1193 | VNInfoAllocator.Reset(); |
| 1194 | |
| 1195 | unsigned NumSlots = MFI->getObjectIndexEnd(); |
| 1196 | |
| 1197 | // If there are no stack slots then there are no markers to remove. |
| 1198 | if (!NumSlots) |
| 1199 | return false; |
| 1200 | |
| 1201 | SmallVector<int, 8> SortedSlots; |
| 1202 | SortedSlots.reserve(N: NumSlots); |
| 1203 | Intervals.reserve(N: NumSlots); |
| 1204 | LiveStarts.resize(N: NumSlots); |
| 1205 | |
| 1206 | unsigned NumMarkers = collectMarkers(NumSlot: NumSlots); |
| 1207 | |
| 1208 | unsigned TotalSize = 0; |
| 1209 | LLVM_DEBUG(dbgs() << "Found "<< NumMarkers << " markers and "<< NumSlots |
| 1210 | << " slots\n"); |
| 1211 | LLVM_DEBUG(dbgs() << "Slot structure:\n"); |
| 1212 | |
| 1213 | for (int i=0; i < MFI->getObjectIndexEnd(); ++i) { |
| 1214 | LLVM_DEBUG(dbgs() << "Slot #"<< i << " - "<< MFI->getObjectSize(i) |
| 1215 | << " bytes.\n"); |
| 1216 | TotalSize += MFI->getObjectSize(ObjectIdx: i); |
| 1217 | } |
| 1218 | |
| 1219 | LLVM_DEBUG(dbgs() << "Total Stack size: "<< TotalSize << " bytes\n\n"); |
| 1220 | |
| 1221 | // Don't continue because there are not enough lifetime markers, or the |
| 1222 | // stack is too small, or we are told not to optimize the slots. |
| 1223 | if (NumMarkers < 2 || TotalSize < 16 || DisableColoring || |
| 1224 | skipFunction(F: Func.getFunction())) { |
| 1225 | LLVM_DEBUG(dbgs() << "Will not try to merge slots.\n"); |
| 1226 | return removeAllMarkers(); |
| 1227 | } |
| 1228 | |
| 1229 | for (unsigned i=0; i < NumSlots; ++i) { |
| 1230 | std::unique_ptr<LiveInterval> LI(new LiveInterval(i, 0)); |
| 1231 | LI->getNextValue(Def: Indexes->getZeroIndex(), VNInfoAllocator); |
| 1232 | Intervals.push_back(Elt: std::move(LI)); |
| 1233 | SortedSlots.push_back(Elt: i); |
| 1234 | } |
| 1235 | |
| 1236 | // Calculate the liveness of each block. |
| 1237 | calculateLocalLiveness(); |
| 1238 | LLVM_DEBUG(dbgs() << "Dataflow iterations: "<< NumIterations << "\n"); |
| 1239 | LLVM_DEBUG(dump()); |
| 1240 | |
| 1241 | // Propagate the liveness information. |
| 1242 | calculateLiveIntervals(NumSlots); |
| 1243 | LLVM_DEBUG(dumpIntervals()); |
| 1244 | |
| 1245 | // Search for allocas which are used outside of the declared lifetime |
| 1246 | // markers. |
| 1247 | if (ProtectFromEscapedAllocas) |
| 1248 | removeInvalidSlotRanges(); |
| 1249 | |
| 1250 | // Maps old slots to new slots. |
| 1251 | DenseMap<int, int> SlotRemap; |
| 1252 | unsigned RemovedSlots = 0; |
| 1253 | unsigned ReducedSize = 0; |
| 1254 | |
| 1255 | // Do not bother looking at empty intervals. |
| 1256 | for (unsigned I = 0; I < NumSlots; ++I) { |
| 1257 | if (Intervals[SortedSlots[I]]->empty()) |
| 1258 | SortedSlots[I] = -1; |
| 1259 | } |
| 1260 | |
| 1261 | // This is a simple greedy algorithm for merging allocas. First, sort the |
| 1262 | // slots, placing the largest slots first. Next, perform an n^2 scan and look |
| 1263 | // for disjoint slots. When you find disjoint slots, merge the smaller one |
| 1264 | // into the bigger one and update the live interval. Remove the small alloca |
| 1265 | // and continue. |
| 1266 | |
| 1267 | // Sort the slots according to their size. Place unused slots at the end. |
| 1268 | // Use stable sort to guarantee deterministic code generation. |
| 1269 | llvm::stable_sort(Range&: SortedSlots, C: [this](int LHS, int RHS) { |
| 1270 | // We use -1 to denote a uninteresting slot. Place these slots at the end. |
| 1271 | if (LHS == -1) |
| 1272 | return false; |
| 1273 | if (RHS == -1) |
| 1274 | return true; |
| 1275 | // Sort according to size. |
| 1276 | return MFI->getObjectSize(ObjectIdx: LHS) > MFI->getObjectSize(ObjectIdx: RHS); |
| 1277 | }); |
| 1278 | |
| 1279 | for (auto &s : LiveStarts) |
| 1280 | llvm::sort(C&: s); |
| 1281 | |
| 1282 | bool Changed = true; |
| 1283 | while (Changed) { |
| 1284 | Changed = false; |
| 1285 | for (unsigned I = 0; I < NumSlots; ++I) { |
| 1286 | if (SortedSlots[I] == -1) |
| 1287 | continue; |
| 1288 | |
| 1289 | for (unsigned J=I+1; J < NumSlots; ++J) { |
| 1290 | if (SortedSlots[J] == -1) |
| 1291 | continue; |
| 1292 | |
| 1293 | int FirstSlot = SortedSlots[I]; |
| 1294 | int SecondSlot = SortedSlots[J]; |
| 1295 | |
| 1296 | // Objects with different stack IDs cannot be merged. |
| 1297 | if (MFI->getStackID(ObjectIdx: FirstSlot) != MFI->getStackID(ObjectIdx: SecondSlot)) |
| 1298 | continue; |
| 1299 | |
| 1300 | LiveInterval *First = &*Intervals[FirstSlot]; |
| 1301 | LiveInterval *Second = &*Intervals[SecondSlot]; |
| 1302 | auto &FirstS = LiveStarts[FirstSlot]; |
| 1303 | auto &SecondS = LiveStarts[SecondSlot]; |
| 1304 | assert(!First->empty() && !Second->empty() && "Found an empty range"); |
| 1305 | |
| 1306 | // Merge disjoint slots. This is a little bit tricky - see the |
| 1307 | // Implementation Notes section for an explanation. |
| 1308 | if (!First->isLiveAtIndexes(Slots: SecondS) && |
| 1309 | !Second->isLiveAtIndexes(Slots: FirstS)) { |
| 1310 | Changed = true; |
| 1311 | First->MergeSegmentsInAsValue(RHS: *Second, LHSValNo: First->getValNumInfo(ValNo: 0)); |
| 1312 | |
| 1313 | int OldSize = FirstS.size(); |
| 1314 | FirstS.append(in_start: SecondS.begin(), in_end: SecondS.end()); |
| 1315 | auto Mid = FirstS.begin() + OldSize; |
| 1316 | std::inplace_merge(first: FirstS.begin(), middle: Mid, last: FirstS.end()); |
| 1317 | |
| 1318 | SlotRemap[SecondSlot] = FirstSlot; |
| 1319 | SortedSlots[J] = -1; |
| 1320 | LLVM_DEBUG(dbgs() << "Merging #"<< FirstSlot << " and slots #" |
| 1321 | << SecondSlot << " together.\n"); |
| 1322 | Align MaxAlignment = std::max(a: MFI->getObjectAlign(ObjectIdx: FirstSlot), |
| 1323 | b: MFI->getObjectAlign(ObjectIdx: SecondSlot)); |
| 1324 | |
| 1325 | assert(MFI->getObjectSize(FirstSlot) >= |
| 1326 | MFI->getObjectSize(SecondSlot) && |
| 1327 | "Merging a small object into a larger one"); |
| 1328 | |
| 1329 | RemovedSlots+=1; |
| 1330 | ReducedSize += MFI->getObjectSize(ObjectIdx: SecondSlot); |
| 1331 | MFI->setObjectAlignment(ObjectIdx: FirstSlot, Alignment: MaxAlignment); |
| 1332 | MFI->RemoveStackObject(ObjectIdx: SecondSlot); |
| 1333 | } |
| 1334 | } |
| 1335 | } |
| 1336 | }// While changed. |
| 1337 | |
| 1338 | // Record statistics. |
| 1339 | StackSpaceSaved += ReducedSize; |
| 1340 | StackSlotMerged += RemovedSlots; |
| 1341 | LLVM_DEBUG(dbgs() << "Merge "<< RemovedSlots << " slots. Saved " |
| 1342 | << ReducedSize << " bytes\n"); |
| 1343 | |
| 1344 | // Scan the entire function and update all machine operands that use frame |
| 1345 | // indices to use the remapped frame index. |
| 1346 | if (!SlotRemap.empty()) { |
| 1347 | expungeSlotMap(SlotRemap, NumSlots); |
| 1348 | remapInstructions(SlotRemap); |
| 1349 | } |
| 1350 | |
| 1351 | return removeAllMarkers(); |
| 1352 | } |
| 1353 |
Generated on 2024-Oct-10 from project llvm_projects revision release-19.x-64075837b
Powered by 
Code Browser 2.1
Generator usage only permitted with license.