TailRecursionElimination.cpp revision b5d84d13bc338efc4eeed87d19c49dfaed38e036
12240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner//===- TailRecursionElimination.cpp - Eliminate Tail Calls ----------------===// 2fd93908ae8b9684fe71c239e3c6cfe13ff6a2663Misha Brukman// 3b576c94c15af9a440f69d9d03c2afead7971118cJohn Criswell// The LLVM Compiler Infrastructure 4b576c94c15af9a440f69d9d03c2afead7971118cJohn Criswell// 54ee451de366474b9c228b4e5fa573795a715216dChris Lattner// This file is distributed under the University of Illinois Open Source 64ee451de366474b9c228b4e5fa573795a715216dChris Lattner// License. See LICENSE.TXT for details. 7fd93908ae8b9684fe71c239e3c6cfe13ff6a2663Misha Brukman// 8b576c94c15af9a440f69d9d03c2afead7971118cJohn Criswell//===----------------------------------------------------------------------===// 92240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner// 107152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// This file transforms calls of the current function (self recursion) followed 117152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// by a return instruction with a branch to the entry of the function, creating 127152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// a loop. This pass also implements the following extensions to the basic 137152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// algorithm: 142240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner// 157152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// 1. Trivial instructions between the call and return do not prevent the 167152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// transformation from taking place, though currently the analysis cannot 177152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// support moving any really useful instructions (only dead ones). 18543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner// 2. This pass transforms functions that are prevented from being tail 1924080a98786cf393740cd035e27dd16c12827892Duncan Sands// recursive by an associative and commutative expression to use an 2024080a98786cf393740cd035e27dd16c12827892Duncan Sands// accumulator variable, thus compiling the typical naive factorial or 2124080a98786cf393740cd035e27dd16c12827892Duncan Sands// 'fib' implementation into efficient code. 22d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner// 3. TRE is performed if the function returns void, if the return 23d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner// returns the result returned by the call, or if the function returns a 24d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner// run-time constant on all exits from the function. It is possible, though 25d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner// unlikely, that the return returns something else (like constant 0), and 26d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner// can still be TRE'd. It can be TRE'd if ALL OTHER return instructions in 27d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner// the function return the exact same value. 280cade261322242132905914f613d25d16e1a9ff6Nick Lewycky// 4. If it can prove that callees do not access their caller stack frame, 297f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner// they are marked as eligible for tail call elimination (by the code 307f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner// generator). 312240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner// 327152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// There are several improvements that could be made: 337152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// 347152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// 1. If the function has any alloca instructions, these instructions will be 357152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// moved out of the entry block of the function, causing them to be 367152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// evaluated each time through the tail recursion. Safely keeping allocas 377152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// in the entry block requires analysis to proves that the tail-called 387152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// function does not read or write the stack object. 392240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner// 2. Tail recursion is only performed if the call immediately preceeds the 407152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// return instruction. It's possible that there could be a jump between 417152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// the call and the return. 42d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner// 3. There can be intervening operations between the call and the return that 437152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// prevent the TRE from occurring. For example, there could be GEP's and 447152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// stores to memory that will not be read or written by the call. This 457152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// requires some substantial analysis (such as with DSA) to prove safe to 467152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// move ahead of the call, but doing so could allow many more TREs to be 477152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner// performed, for example in TreeAdd/TreeAlloc from the treeadd benchmark. 487f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner// 4. The algorithm we use to detect if callees access their caller stack 497f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner// frames is very primitive. 502240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner// 512240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner//===----------------------------------------------------------------------===// 522240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner 530e5f499638c8d277b9dc4a4385712177c53b5681Chris Lattner#define DEBUG_TYPE "tailcallelim" 543fc6ef1bb96d9a3194cef667a2d3cbc94e3fb189Chris Lattner#include "llvm/Transforms/Scalar.h" 556a35b40250735a50efe66c88414cdd3b79185019Chris Lattner#include "llvm/Transforms/Utils/Local.h" 56ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner#include "llvm/Constants.h" 572240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner#include "llvm/DerivedTypes.h" 582240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner#include "llvm/Function.h" 592240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner#include "llvm/Instructions.h" 602240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner#include "llvm/Pass.h" 610cade261322242132905914f613d25d16e1a9ff6Nick Lewycky#include "llvm/Analysis/CaptureTracking.h" 62ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman#include "llvm/Analysis/InlineCost.h" 63dd9344f3face8f1978a7f9f393c31b628144d1f6Dan Gohman#include "llvm/Analysis/Loads.h" 64ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman#include "llvm/Support/CallSite.h" 65543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner#include "llvm/Support/CFG.h" 66551ccae044b0ff658fe629dd67edd5ffe75d10e8Reid Spencer#include "llvm/ADT/Statistic.h" 67f8485c643412dbff46fe87ea2867445169a5c28eChris Lattnerusing namespace llvm; 68d0fde30ce850b78371fd1386338350591f9ff494Brian Gaeke 690e5f499638c8d277b9dc4a4385712177c53b5681Chris LattnerSTATISTIC(NumEliminated, "Number of tail calls removed"); 700e5f499638c8d277b9dc4a4385712177c53b5681Chris LattnerSTATISTIC(NumAccumAdded, "Number of accumulators introduced"); 712240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner 720e5f499638c8d277b9dc4a4385712177c53b5681Chris Lattnernamespace { 733e8b6631e67e01e4960a7ba4668a50c596607473Chris Lattner struct TailCallElim : public FunctionPass { 74ecd94c804a563f2a86572dcf1d2e81f397e19daaNick Lewycky static char ID; // Pass identification, replacement for typeid 7590c579de5a383cee278acc3f7e7b9d0a656e6a35Owen Anderson TailCallElim() : FunctionPass(ID) {} 76794fd75c67a2cdc128d67342c6d88a504d186896Devang Patel 772240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner virtual bool runOnFunction(Function &F); 787152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner 797152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner private: 807152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner bool ProcessReturningBlock(ReturnInst *RI, BasicBlock *&OldEntry, 81ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner bool &TailCallsAreMarkedTail, 820cade261322242132905914f613d25d16e1a9ff6Nick Lewycky SmallVector<PHINode*, 8> &ArgumentPHIs, 83ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner bool CannotTailCallElimCallsMarkedTail); 847152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner bool CanMoveAboveCall(Instruction *I, CallInst *CI); 85543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner Value *CanTransformAccumulatorRecursion(Instruction *I, CallInst *CI); 862240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner }; 872240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner} 882240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner 89844731a7f1909f55935e3514c9e713a62d67662eDan Gohmanchar TailCallElim::ID = 0; 90d13db2c59cc94162d6cf0a04187d408bfef6d4a7Owen AndersonINITIALIZE_PASS(TailCallElim, "tailcallelim", 91d13db2c59cc94162d6cf0a04187d408bfef6d4a7Owen Anderson "Tail Call Elimination", false, false); 92844731a7f1909f55935e3514c9e713a62d67662eDan Gohman 93d0fde30ce850b78371fd1386338350591f9ff494Brian Gaeke// Public interface to the TailCallElimination pass 94f8485c643412dbff46fe87ea2867445169a5c28eChris LattnerFunctionPass *llvm::createTailCallEliminationPass() { 95f8485c643412dbff46fe87ea2867445169a5c28eChris Lattner return new TailCallElim(); 96f8485c643412dbff46fe87ea2867445169a5c28eChris Lattner} 973fc6ef1bb96d9a3194cef667a2d3cbc94e3fb189Chris Lattner 98cb19438f63019eb557c1fbaeae8a6de78b03c584Nick Lewycky/// AllocaMightEscapeToCalls - Return true if this alloca may be accessed by 99cb19438f63019eb557c1fbaeae8a6de78b03c584Nick Lewycky/// callees of this function. We only do very simple analysis right now, this 100cb19438f63019eb557c1fbaeae8a6de78b03c584Nick Lewycky/// could be expanded in the future to use mod/ref information for particular 101cb19438f63019eb557c1fbaeae8a6de78b03c584Nick Lewycky/// call sites if desired. 102cb19438f63019eb557c1fbaeae8a6de78b03c584Nick Lewyckystatic bool AllocaMightEscapeToCalls(AllocaInst *AI) { 103cb19438f63019eb557c1fbaeae8a6de78b03c584Nick Lewycky // FIXME: do simple 'address taken' analysis. 104cb19438f63019eb557c1fbaeae8a6de78b03c584Nick Lewycky return true; 105cb19438f63019eb557c1fbaeae8a6de78b03c584Nick Lewycky} 106cb19438f63019eb557c1fbaeae8a6de78b03c584Nick Lewycky 1070cade261322242132905914f613d25d16e1a9ff6Nick Lewycky/// CheckForEscapingAllocas - Scan the specified basic block for alloca 1087f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner/// instructions. If it contains any that might be accessed by calls, return 1097f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner/// true. 110ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattnerstatic bool CheckForEscapingAllocas(BasicBlock *BB, 111ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner bool &CannotTCETailMarkedCall) { 112ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner bool RetVal = false; 1137f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 114ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) { 115cb19438f63019eb557c1fbaeae8a6de78b03c584Nick Lewycky RetVal |= AllocaMightEscapeToCalls(AI); 116ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner 117ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // If this alloca is in the body of the function, or if it is a variable 118ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // sized allocation, we cannot tail call eliminate calls marked 'tail' 119ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // with this mechanism. 120ecb7a77885b174cf4d001a9b48533b3979e7810dDan Gohman if (BB != &BB->getParent()->getEntryBlock() || 121ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner !isa<ConstantInt>(AI->getArraySize())) 122ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner CannotTCETailMarkedCall = true; 123ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner } 124ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner return RetVal; 1257f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner} 1267f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner 1272240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattnerbool TailCallElim::runOnFunction(Function &F) { 1282240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner // If this function is a varargs function, we won't be able to PHI the args 1292240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner // right, so don't even try to convert it... 1302240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner if (F.getFunctionType()->isVarArg()) return false; 1312240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner 1322240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner BasicBlock *OldEntry = 0; 133ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner bool TailCallsAreMarkedTail = false; 1340cade261322242132905914f613d25d16e1a9ff6Nick Lewycky SmallVector<PHINode*, 8> ArgumentPHIs; 1352240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner bool MadeChange = false; 1362240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner 1377f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner bool FunctionContainsEscapingAllocas = false; 1387f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner 139ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // CannotTCETailMarkedCall - If true, we cannot perform TCE on tail calls 140ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // marked with the 'tail' attribute, because doing so would cause the stack 141ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // size to increase (real TCE would deallocate variable sized allocas, TCE 142ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // doesn't). 143ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner bool CannotTCETailMarkedCall = false; 144ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner 1457f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner // Loop over the function, looking for any returning blocks, and keeping track 1467f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner // of whether this function has any non-trivially used allocas. 1477f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { 148ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner if (FunctionContainsEscapingAllocas && CannotTCETailMarkedCall) 149ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner break; 15000b16889ab461b7ecef1c91ade101186b7f1fce2Jeff Cohen 1514c0e4cdc40bbf94766242294fe386f4eb2d32d2eChris Lattner FunctionContainsEscapingAllocas |= 152ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner CheckForEscapingAllocas(BB, CannotTCETailMarkedCall); 1537f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner } 15432b1e87f111d79f3f5b0e86c8f138c02bef8aa18Chris Lattner 15532b1e87f111d79f3f5b0e86c8f138c02bef8aa18Chris Lattner /// FIXME: The code generator produces really bad code when an 'escaping 15632b1e87f111d79f3f5b0e86c8f138c02bef8aa18Chris Lattner /// alloca' is changed from being a static alloca to being a dynamic alloca. 15732b1e87f111d79f3f5b0e86c8f138c02bef8aa18Chris Lattner /// Until this is resolved, disable this transformation if that would ever 15832b1e87f111d79f3f5b0e86c8f138c02bef8aa18Chris Lattner /// happen. This bug is PR962. 15932b1e87f111d79f3f5b0e86c8f138c02bef8aa18Chris Lattner if (FunctionContainsEscapingAllocas) 16032b1e87f111d79f3f5b0e86c8f138c02bef8aa18Chris Lattner return false; 161fd93908ae8b9684fe71c239e3c6cfe13ff6a2663Misha Brukman 162ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // Second pass, change any tail calls to loops. 163ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 164ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB->getTerminator())) 165ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner MadeChange |= ProcessReturningBlock(Ret, OldEntry, TailCallsAreMarkedTail, 166ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner ArgumentPHIs,CannotTCETailMarkedCall); 167ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner 168cf2f89251d45bc9783917874adb9029f71d50068Chris Lattner // If we eliminated any tail recursions, it's possible that we inserted some 169cf2f89251d45bc9783917874adb9029f71d50068Chris Lattner // silly PHI nodes which just merge an initial value (the incoming operand) 170cf2f89251d45bc9783917874adb9029f71d50068Chris Lattner // with themselves. Check to see if we did and clean up our mess if so. This 171cf2f89251d45bc9783917874adb9029f71d50068Chris Lattner // occurs when a function passes an argument straight through to its tail 172cf2f89251d45bc9783917874adb9029f71d50068Chris Lattner // call. 173cf2f89251d45bc9783917874adb9029f71d50068Chris Lattner if (!ArgumentPHIs.empty()) { 174cf2f89251d45bc9783917874adb9029f71d50068Chris Lattner for (unsigned i = 0, e = ArgumentPHIs.size(); i != e; ++i) { 175cf2f89251d45bc9783917874adb9029f71d50068Chris Lattner PHINode *PN = ArgumentPHIs[i]; 176cf2f89251d45bc9783917874adb9029f71d50068Chris Lattner 177cf2f89251d45bc9783917874adb9029f71d50068Chris Lattner // If the PHI Node is a dynamic constant, replace it with the value it is. 178ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner if (Value *PNV = PN->hasConstantValue()) { 179ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner PN->replaceAllUsesWith(PNV); 180ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner PN->eraseFromParent(); 181cf2f89251d45bc9783917874adb9029f71d50068Chris Lattner } 182cf2f89251d45bc9783917874adb9029f71d50068Chris Lattner } 183cf2f89251d45bc9783917874adb9029f71d50068Chris Lattner } 184cf2f89251d45bc9783917874adb9029f71d50068Chris Lattner 1857f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner // Finally, if this function contains no non-escaping allocas, mark all calls 1867f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner // in the function as eligible for tail calls (there is no stack memory for 1877f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner // them to access). 1887f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner if (!FunctionContainsEscapingAllocas) 1897f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 1907f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 191febc81680c80a53f15f85b1812cba07fc179b9fdEvan Cheng if (CallInst *CI = dyn_cast<CallInst>(I)) { 192febc81680c80a53f15f85b1812cba07fc179b9fdEvan Cheng CI->setTailCall(); 193febc81680c80a53f15f85b1812cba07fc179b9fdEvan Cheng MadeChange = true; 194febc81680c80a53f15f85b1812cba07fc179b9fdEvan Cheng } 1957f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner 1962240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner return MadeChange; 1972240d2b3f7b9f5835868c83ce78d125d1b65212bChris Lattner} 1987152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner 1997152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner 200543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner/// CanMoveAboveCall - Return true if it is safe to move the specified 201543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner/// instruction from after the call to before the call, assuming that all 202543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner/// instructions between the call and this instruction are movable. 203543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner/// 2047152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattnerbool TailCallElim::CanMoveAboveCall(Instruction *I, CallInst *CI) { 2057152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // FIXME: We can move load/store/call/free instructions above the call if the 2067152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // call does not mod/ref the memory location being processed. 2076a35b40250735a50efe66c88414cdd3b79185019Chris Lattner if (I->mayHaveSideEffects()) // This also handles volatile loads. 2087152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner return false; 2096a35b40250735a50efe66c88414cdd3b79185019Chris Lattner 2100cade261322242132905914f613d25d16e1a9ff6Nick Lewycky if (LoadInst *L = dyn_cast<LoadInst>(I)) { 2116a35b40250735a50efe66c88414cdd3b79185019Chris Lattner // Loads may always be moved above calls without side effects. 2126a35b40250735a50efe66c88414cdd3b79185019Chris Lattner if (CI->mayHaveSideEffects()) { 2136a35b40250735a50efe66c88414cdd3b79185019Chris Lattner // Non-volatile loads may be moved above a call with side effects if it 2146a35b40250735a50efe66c88414cdd3b79185019Chris Lattner // does not write to memory and the load provably won't trap. 2156a35b40250735a50efe66c88414cdd3b79185019Chris Lattner // FIXME: Writes to memory only matter if they may alias the pointer 2166a35b40250735a50efe66c88414cdd3b79185019Chris Lattner // being loaded from. 2176a35b40250735a50efe66c88414cdd3b79185019Chris Lattner if (CI->mayWriteToMemory() || 21849db68fba01722ca032dc5170f8248a9d25f0199Bob Wilson !isSafeToLoadUnconditionally(L->getPointerOperand(), L, 21949db68fba01722ca032dc5170f8248a9d25f0199Bob Wilson L->getAlignment())) 2206a35b40250735a50efe66c88414cdd3b79185019Chris Lattner return false; 2216a35b40250735a50efe66c88414cdd3b79185019Chris Lattner } 2226a35b40250735a50efe66c88414cdd3b79185019Chris Lattner } 2237152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner 2247152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // Otherwise, if this is a side-effect free instruction, check to make sure 2257152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // that it does not use the return value of the call. If it doesn't use the 2267152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // return value of the call, it must only use things that are defined before 2277152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // the call, or movable instructions between the call and the instruction 2287152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // itself. 2297152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 2307152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner if (I->getOperand(i) == CI) 2317152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner return false; 2327152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner return true; 2337152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner} 2347152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner 235d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner// isDynamicConstant - Return true if the specified value is the same when the 236d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner// return would exit as it was when the initial iteration of the recursive 237d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner// function was executed. 238d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner// 239d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner// We currently handle static constants and arguments that are not modified as 240d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner// part of the recursion. 241d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner// 242f80fcd00b3590626a6110fe5fbe9b5fd2a8be1aeNick Lewyckystatic bool isDynamicConstant(Value *V, CallInst *CI, ReturnInst *RI) { 243d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner if (isa<Constant>(V)) return true; // Static constants are always dyn consts 244d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner 245d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner // Check to see if this is an immutable argument, if so, the value 246d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner // will be available to initialize the accumulator. 247d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner if (Argument *Arg = dyn_cast<Argument>(V)) { 248d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner // Figure out which argument number this is... 249d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner unsigned ArgNo = 0; 250d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner Function *F = CI->getParent()->getParent(); 251e4d5c441e04bdc00ccf1804744af670655123b07Chris Lattner for (Function::arg_iterator AI = F->arg_begin(); &*AI != Arg; ++AI) 252d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner ++ArgNo; 253fd93908ae8b9684fe71c239e3c6cfe13ff6a2663Misha Brukman 254d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner // If we are passing this argument into call as the corresponding 255d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner // argument operand, then the argument is dynamically constant. 256d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner // Otherwise, we cannot transform this function safely. 257de9f5452d3ae894bb7fdd455cec5af50e2560aa5Gabor Greif if (CI->getArgOperand(ArgNo) == Arg) 258d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner return true; 259d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner } 260f80fcd00b3590626a6110fe5fbe9b5fd2a8be1aeNick Lewycky 261f80fcd00b3590626a6110fe5fbe9b5fd2a8be1aeNick Lewycky // Switch cases are always constant integers. If the value is being switched 262f80fcd00b3590626a6110fe5fbe9b5fd2a8be1aeNick Lewycky // on and the return is only reachable from one of its cases, it's 263f80fcd00b3590626a6110fe5fbe9b5fd2a8be1aeNick Lewycky // effectively constant. 264f80fcd00b3590626a6110fe5fbe9b5fd2a8be1aeNick Lewycky if (BasicBlock *UniquePred = RI->getParent()->getUniquePredecessor()) 265f80fcd00b3590626a6110fe5fbe9b5fd2a8be1aeNick Lewycky if (SwitchInst *SI = dyn_cast<SwitchInst>(UniquePred->getTerminator())) 266f80fcd00b3590626a6110fe5fbe9b5fd2a8be1aeNick Lewycky if (SI->getCondition() == V) 267f80fcd00b3590626a6110fe5fbe9b5fd2a8be1aeNick Lewycky return SI->getDefaultDest() != RI->getParent(); 268f80fcd00b3590626a6110fe5fbe9b5fd2a8be1aeNick Lewycky 269d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner // Not a constant or immutable argument, we can't safely transform. 270d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner return false; 271d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner} 272d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner 273d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner// getCommonReturnValue - Check to see if the function containing the specified 274d50e9e2566479301e687fc81e16dab7e3c2b50eaDuncan Sands// tail call consistently returns the same runtime-constant value at all exit 275d50e9e2566479301e687fc81e16dab7e3c2b50eaDuncan Sands// points except for IgnoreRI. If so, return the returned value. 276d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner// 277d50e9e2566479301e687fc81e16dab7e3c2b50eaDuncan Sandsstatic Value *getCommonReturnValue(ReturnInst *IgnoreRI, CallInst *CI) { 278d50e9e2566479301e687fc81e16dab7e3c2b50eaDuncan Sands Function *F = CI->getParent()->getParent(); 279d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner Value *ReturnedValue = 0; 280d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner 281b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) { 282b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator()); 283b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner if (RI == 0 || RI == IgnoreRI) continue; 284b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner 285b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner // We can only perform this transformation if the value returned is 286b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner // evaluatable at the start of the initial invocation of the function, 287b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner // instead of at the end of the evaluation. 288b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner // 289b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner Value *RetOp = RI->getOperand(0); 290b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner if (!isDynamicConstant(RetOp, CI, RI)) 291b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner return 0; 292b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner 293b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner if (ReturnedValue && RetOp != ReturnedValue) 294b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner return 0; // Cannot transform if differing values are returned. 295b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner ReturnedValue = RetOp; 296b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner } 297d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner return ReturnedValue; 298d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner} 2997152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner 300543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner/// CanTransformAccumulatorRecursion - If the specified instruction can be 301543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner/// transformed using accumulator recursion elimination, return the constant 302543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner/// which is the start of the accumulator value. Otherwise return null. 303543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner/// 304543d622ef7505910c1cdc09ada0ab797c3437590Chris LattnerValue *TailCallElim::CanTransformAccumulatorRecursion(Instruction *I, 305543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner CallInst *CI) { 30624080a98786cf393740cd035e27dd16c12827892Duncan Sands if (!I->isAssociative() || !I->isCommutative()) return 0; 307543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner assert(I->getNumOperands() == 2 && 30824080a98786cf393740cd035e27dd16c12827892Duncan Sands "Associative/commutative operations should have 2 args!"); 309543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner 310b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner // Exactly one operand should be the result of the call instruction. 31107e6e56f57e8781a8d7bc601cc9034a3741d84c2Anton Korobeynikov if ((I->getOperand(0) == CI && I->getOperand(1) == CI) || 31207e6e56f57e8781a8d7bc601cc9034a3741d84c2Anton Korobeynikov (I->getOperand(0) != CI && I->getOperand(1) != CI)) 313543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner return 0; 314543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner 315543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // The only user of this instruction we allow is a single return instruction. 316543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner if (!I->hasOneUse() || !isa<ReturnInst>(I->use_back())) 317543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner return 0; 318543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner 319543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // Ok, now we have to check all of the other return instructions in this 320543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // function. If they return non-constants or differing values, then we cannot 321543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // transform the function safely. 322d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner return getCommonReturnValue(cast<ReturnInst>(I->use_back()), CI); 323543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner} 324543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner 3257152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattnerbool TailCallElim::ProcessReturningBlock(ReturnInst *Ret, BasicBlock *&OldEntry, 326ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner bool &TailCallsAreMarkedTail, 3270cade261322242132905914f613d25d16e1a9ff6Nick Lewycky SmallVector<PHINode*, 8> &ArgumentPHIs, 328ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner bool CannotTailCallElimCallsMarkedTail) { 3297152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner BasicBlock *BB = Ret->getParent(); 3307152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner Function *F = BB->getParent(); 3317152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner 3327152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner if (&BB->front() == Ret) // Make sure there is something before the ret... 3337152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner return false; 3348d9455d4e4f63c368f17b74a36b472fc61685310Chris Lattner 3357152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // Scan backwards from the return, checking to see if there is a tail call in 3367152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // this block. If so, set CI to it. 3377152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner CallInst *CI; 3387152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner BasicBlock::iterator BBI = Ret; 3397152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner while (1) { 3407152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner CI = dyn_cast<CallInst>(BBI); 3417152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner if (CI && CI->getCalledFunction() == F) 3427152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner break; 3437152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner 3447152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner if (BBI == BB->begin()) 3457152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner return false; // Didn't find a potential tail call. 3467152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner --BBI; 3477152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner } 3487152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner 349ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // If this call is marked as a tail call, and if there are dynamic allocas in 350ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // the function, we cannot perform this optimization. 351ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner if (CI->isTailCall() && CannotTailCallElimCallsMarkedTail) 352ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner return false; 353ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner 354ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman // As a special case, detect code like this: 355ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman // double fabs(double f) { return __builtin_fabs(f); } // a 'fabs' call 356ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman // and disable this xform in this case, because the code generator will 357ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman // lower the call to fabs into inline code. 358ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman if (BB == &F->getEntryBlock() && 359ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman &BB->front() == CI && &*++BB->begin() == Ret && 360ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman callIsSmall(F)) { 361ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman // A single-block function with just a call and a return. Check that 362ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman // the arguments match. 363ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman CallSite::arg_iterator I = CallSite(CI).arg_begin(), 364ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman E = CallSite(CI).arg_end(); 365ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman Function::arg_iterator FI = F->arg_begin(), 366ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman FE = F->arg_end(); 367ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman for (; I != E && FI != FE; ++I, ++FI) 368ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman if (*I != &*FI) break; 369ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman if (I == E && FI == FE) 370ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman return false; 371ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman } 372ea25b48af33be42e19236d8eac26bd42b45bcc1bDan Gohman 37324080a98786cf393740cd035e27dd16c12827892Duncan Sands // If we are introducing accumulator recursion to eliminate operations after 37424080a98786cf393740cd035e27dd16c12827892Duncan Sands // the call instruction that are both associative and commutative, the initial 37524080a98786cf393740cd035e27dd16c12827892Duncan Sands // value for the accumulator is placed in this variable. If this value is set 37624080a98786cf393740cd035e27dd16c12827892Duncan Sands // then we actually perform accumulator recursion elimination instead of 377d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands // simple tail recursion elimination. If the operation is an LLVM instruction 378d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands // (eg: "add") then it is recorded in AccumulatorRecursionInstr. If not, then 379d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands // we are handling the case when the return instruction returns a constant C 380d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands // which is different to the constant returned by other return instructions 381d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands // (which is recorded in AccumulatorRecursionEliminationInitVal). This is a 382d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands // special case of accumulator recursion, the operation being "return C". 383543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner Value *AccumulatorRecursionEliminationInitVal = 0; 384543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner Instruction *AccumulatorRecursionInstr = 0; 385543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner 3867152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // Ok, we found a potential tail call. We can currently only transform the 3877152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // tail call if all of the instructions between the call and the return are 3887152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // movable to above the call itself, leaving the call next to the return. 3897152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // Check that this is the case now. 390b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner for (BBI = CI, ++BBI; &*BBI != Ret; ++BBI) { 391b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner if (CanMoveAboveCall(BBI, CI)) continue; 392b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner 393b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner // If we can't move the instruction above the call, it might be because it 394b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner // is an associative and commutative operation that could be tranformed 395b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner // using accumulator recursion elimination. Check to see if this is the 396b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner // case, and if so, remember the initial accumulator value for later. 397b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner if ((AccumulatorRecursionEliminationInitVal = 398b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner CanTransformAccumulatorRecursion(BBI, CI))) { 399b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner // Yes, this is accumulator recursion. Remember which instruction 400b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner // accumulates. 401b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner AccumulatorRecursionInstr = BBI; 402b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner } else { 403b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner return false; // Otherwise, we cannot eliminate the tail recursion! 404543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner } 405b5d84d13bc338efc4eeed87d19c49dfaed38e036Chris Lattner } 4067152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner 4077152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // We can only transform call/return pairs that either ignore the return value 408d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner // of the call and return void, ignore the value of the call and return a 409d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner // constant, return the value returned by the tail call, or that are being 410d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner // accumulator recursion variable eliminated. 411826c49132a9cba1e08a5fce78d05561d9a3746ceDevang Patel if (Ret->getNumOperands() == 1 && Ret->getReturnValue() != CI && 4123b5f45042b17ad52815e3dd6c0c1df99f196dd04Chris Lattner !isa<UndefValue>(Ret->getReturnValue()) && 413d64152a70842b2f4186aa912938e69ca09c1434cChris Lattner AccumulatorRecursionEliminationInitVal == 0 && 414d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands !getCommonReturnValue(0, CI)) { 415d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands // One case remains that we are able to handle: the current return 416d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands // instruction returns a constant, and all other return instructions 417d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands // return a different constant. 418d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands if (!isDynamicConstant(Ret->getReturnValue(), CI, Ret)) 419d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands return false; // Current return instruction does not return a constant. 420d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands // Check that all other return instructions return a common constant. If 421d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands // so, record it in AccumulatorRecursionEliminationInitVal. 422d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands AccumulatorRecursionEliminationInitVal = getCommonReturnValue(Ret, CI); 423d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands if (!AccumulatorRecursionEliminationInitVal) 424d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands return false; 425d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands } 4267152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner 4277152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // OK! We can transform this tail call. If this is the first one found, 4287152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // create the new entry block, allowing us to branch back to the old entry. 4297152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner if (OldEntry == 0) { 4307152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner OldEntry = &F->getEntryBlock(); 4311d0be15f89cb5056e20e2d24faa8d6afb1573bcaOwen Anderson BasicBlock *NewEntry = BasicBlock::Create(F->getContext(), "", F, OldEntry); 4326934a04a8c15e9971cd1ea4d5c8df2d7afdd5be5Chris Lattner NewEntry->takeName(OldEntry); 4336934a04a8c15e9971cd1ea4d5c8df2d7afdd5be5Chris Lattner OldEntry->setName("tailrecurse"); 434051a950000e21935165db56695e35bade668193bGabor Greif BranchInst::Create(OldEntry, NewEntry); 435fd93908ae8b9684fe71c239e3c6cfe13ff6a2663Misha Brukman 436ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // If this tail call is marked 'tail' and if there are any allocas in the 437ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // entry block, move them up to the new entry block. 438ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner TailCallsAreMarkedTail = CI->isTailCall(); 439ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner if (TailCallsAreMarkedTail) 440ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // Move all fixed sized allocas from OldEntry to NewEntry. 441ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner for (BasicBlock::iterator OEBI = OldEntry->begin(), E = OldEntry->end(), 442ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner NEBI = NewEntry->begin(); OEBI != E; ) 443ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner if (AllocaInst *AI = dyn_cast<AllocaInst>(OEBI++)) 444ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner if (isa<ConstantInt>(AI->getArraySize())) 4454bc5f8071a28b6fc4f4c2207dd03a5f747d0d84bChris Lattner AI->moveBefore(NEBI); 446ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner 4477152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // Now that we have created a new block, which jumps to the entry 4487152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // block, insert a PHI node for each argument of the function. 4497152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // For now, we initialize each PHI to only have the real arguments 4507152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // which are passed in. 4517152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner Instruction *InsertPos = OldEntry->begin(); 4527f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); 4537f78f218fdb3515a61b995ac64f909bfd8ee84a7Chris Lattner I != E; ++I) { 454b1dbcd886a4b5597a839f299054b78b33fb2d6dfGabor Greif PHINode *PN = PHINode::Create(I->getType(), 455b1dbcd886a4b5597a839f299054b78b33fb2d6dfGabor Greif I->getName() + ".tr", InsertPos); 4567152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner I->replaceAllUsesWith(PN); // Everyone use the PHI node now! 4577152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner PN->addIncoming(I, NewEntry); 4587152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner ArgumentPHIs.push_back(PN); 4597152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner } 4607152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner } 461fd93908ae8b9684fe71c239e3c6cfe13ff6a2663Misha Brukman 462ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // If this function has self recursive calls in the tail position where some 463ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // are marked tail and some are not, only transform one flavor or another. We 464ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // have to choose whether we move allocas in the entry block to the new entry 465ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // block or not, so we can't make a good choice for both. NOTE: We could do 466ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // slightly better here in the case that the function has no entry block 467ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner // allocas. 468ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner if (TailCallsAreMarkedTail && !CI->isTailCall()) 469ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner return false; 470ce869ee05bfcb9d4750a3d0919a8260a727841c3Chris Lattner 4717152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // Ok, now that we know we have a pseudo-entry block WITH all of the 4727152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // required PHI nodes, add entries into the PHI node for the actual 4737152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // parameters passed into the tail-recursive call. 474407014f9a5f05f5a5867e5992a036358acc4a441Gabor Greif for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) 475de9f5452d3ae894bb7fdd455cec5af50e2560aa5Gabor Greif ArgumentPHIs[i]->addIncoming(CI->getArgOperand(i), BB); 476fd93908ae8b9684fe71c239e3c6cfe13ff6a2663Misha Brukman 477543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // If we are introducing an accumulator variable to eliminate the recursion, 478543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // do so now. Note that we _know_ that no subsequent tail recursion 479543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // eliminations will happen on this function because of the way the 480543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // accumulator recursion predicate is set up. 481543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // 482543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner if (AccumulatorRecursionEliminationInitVal) { 483543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner Instruction *AccRecInstr = AccumulatorRecursionInstr; 484543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // Start by inserting a new PHI node for the accumulator. 485d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands PHINode *AccPN = 486d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands PHINode::Create(AccumulatorRecursionEliminationInitVal->getType(), 487d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands "accumulator.tr", OldEntry->begin()); 488543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner 489543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // Loop over all of the predecessors of the tail recursion block. For the 490543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // real entry into the function we seed the PHI with the initial value, 491543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // computed earlier. For any other existing branches to this block (due to 492543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // other tail recursions eliminated) the accumulator is not modified. 493543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // Because we haven't added the branch in the current block to OldEntry yet, 494543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // it will not show up as a predecessor. 495543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner for (pred_iterator PI = pred_begin(OldEntry), PE = pred_end(OldEntry); 496543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner PI != PE; ++PI) { 497a8b9df7bd96e0b0bc6dec448d30b7d72180b6595Gabor Greif BasicBlock *P = *PI; 498a8b9df7bd96e0b0bc6dec448d30b7d72180b6595Gabor Greif if (P == &F->getEntryBlock()) 499a8b9df7bd96e0b0bc6dec448d30b7d72180b6595Gabor Greif AccPN->addIncoming(AccumulatorRecursionEliminationInitVal, P); 500543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner else 501a8b9df7bd96e0b0bc6dec448d30b7d72180b6595Gabor Greif AccPN->addIncoming(AccPN, P); 502543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner } 503543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner 504d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands if (AccRecInstr) { 505d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands // Add an incoming argument for the current block, which is computed by 506d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands // our associative and commutative accumulator instruction. 507d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands AccPN->addIncoming(AccRecInstr, BB); 508d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands 509d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands // Next, rewrite the accumulator recursion instruction so that it does not 510d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands // use the result of the call anymore, instead, use the PHI node we just 511d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands // inserted. 512d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands AccRecInstr->setOperand(AccRecInstr->getOperand(0) != CI, AccPN); 513d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands } else { 514d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands // Add an incoming argument for the current block, which is just the 515d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands // constant returned by the current return instruction. 516d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands AccPN->addIncoming(Ret->getReturnValue(), BB); 517d0d3ccc827de69de3d9f666aa922f5f191219a06Duncan Sands } 518543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner 519543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // Finally, rewrite any return instructions in the program to return the PHI 520543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // node instead of the "initval" that they do currently. This loop will 521543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner // actually rewrite the return value we are destroying, but that's ok. 522543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner for (Function::iterator BBI = F->begin(), E = F->end(); BBI != E; ++BBI) 523543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner if (ReturnInst *RI = dyn_cast<ReturnInst>(BBI->getTerminator())) 524543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner RI->setOperand(0, AccPN); 525543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner ++NumAccumAdded; 526543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner } 527543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner 5287152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // Now that all of the PHI nodes are in place, remove the call and 5297152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner // ret instructions, replacing them with an unconditional branch. 530051a950000e21935165db56695e35bade668193bGabor Greif BranchInst::Create(OldEntry, Ret); 5317152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner BB->getInstList().erase(Ret); // Remove return. 5327152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner BB->getInstList().erase(CI); // Remove call. 533543d622ef7505910c1cdc09ada0ab797c3437590Chris Lattner ++NumEliminated; 5347152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner return true; 5357152da38ef2c1f9f8ee407dc7740dfd43cbc41a6Chris Lattner} 536