[llvm-commits] CVS: llvm-www/pubs/2004-09-22-LCPCLLVMTutorial-Handout.pdf 2004-09-22-LCPCLLVMTutorial-Handout.txt 2004-09-22-LCPCLLVMTutorial.html 2004-09-22-LCPCLLVMTutorial.pdf 2004-09-22-LCPCLLVMTutorial.ppt
Chris Lattner
lattner at cs.uiuc.edu
Tue Sep 21 22:30:37 PDT 2004
Changes in directory llvm-www/pubs:
2004-09-22-LCPCLLVMTutorial-Handout.pdf added (r1.1)
2004-09-22-LCPCLLVMTutorial-Handout.txt added (r1.1)
2004-09-22-LCPCLLVMTutorial.html added (r1.1)
2004-09-22-LCPCLLVMTutorial.pdf added (r1.1)
2004-09-22-LCPCLLVMTutorial.ppt added (r1.1)
---
Log message:
Initial checkin of the LLVM tutorial I'm presenting tommorow.
This is intentionally not linked into the LLVM web pages until tommorow.
---
Diffs of the changes: (+476 -0)
Index: llvm-www/pubs/2004-09-22-LCPCLLVMTutorial-Handout.pdf
Index: llvm-www/pubs/2004-09-22-LCPCLLVMTutorial-Handout.txt
diff -c /dev/null llvm-www/pubs/2004-09-22-LCPCLLVMTutorial-Handout.txt:1.1
*** /dev/null Wed Sep 22 00:30:37 2004
--- llvm-www/pubs/2004-09-22-LCPCLLVMTutorial-Handout.txt Wed Sep 22 00:30:25 2004
***************
*** 0 ****
--- 1,425 ----
+ //===-- SimpleArgumentPromotion.cpp - Promote by-reference arguments ------===//
+ //
+ // The LLVM Compiler Infrastructure
+ //
+ // This file was developed by the LLVM research group and is distributed under
+ // the University of Illinois Open Source License. See LICENSE.TXT for details.
+ //
+ //===----------------------------------------------------------------------===//
+ //
+ // This pass promotes "by reference" arguments to be "by value" arguments. In
+ // practice, this means looking for internal functions that have pointer
+ // arguments. If we can prove, through the use of alias analysis, that an
+ // argument is *only* loaded, then we can pass the value into the function
+ // instead of the address of the value. This can cause recursive simplification
+ // of code and lead to the elimination of allocas (especially in C++ template
+ // code like the STL).
+ //
+ // This pass is a simplified version of the LLVM argpromotion pass (it
+ // invalidates alias analysis instead of updating it, and can not promote
+ // pointers to aggregates).
+ //
+ //===----------------------------------------------------------------------===//
+
+ #include "llvm/CallGraphSCCPass.h"
+ #include "llvm/DerivedTypes.h"
+ #include "llvm/Instructions.h"
+ #include "llvm/Module.h"
+ #include "llvm/Analysis/AliasAnalysis.h"
+ #include "llvm/Analysis/CallGraph.h"
+ #include "llvm/Target/TargetData.h"
+ #include "llvm/Support/CallSite.h"
+ #include "llvm/Support/CFG.h"
+ #include "llvm/Support/Debug.h"
+ #include "llvm/ADT/DepthFirstIterator.h"
+ #include "llvm/ADT/Statistic.h"
+ #include <set>
+ using namespace llvm;
+
+ namespace {
+ Statistic<> NumArgumentsPromoted("simpleargpromotion",
+ "Number of pointer arguments promoted");
+ Statistic<> NumArgumentsDead("simpleargpromotion",
+ "Number of dead pointer args eliminated");
+
+ /// SimpleArgPromotion - Convert 'by reference' arguments to 'by value'.
+ ///
+ struct SimpleArgPromotion : public CallGraphSCCPass {
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.addRequired<AliasAnalysis>();
+ AU.addRequired<TargetData>();
+ CallGraphSCCPass::getAnalysisUsage(AU);
+ }
+
+ virtual bool runOnSCC(const std::vector<CallGraphNode*> &SCC);
+ private:
+ bool PromoteArguments(CallGraphNode *CGN);
+ bool isSafeToPromoteArgument(Argument *Arg) const;
+ Function *DoPromotion(Function *F, std::vector<Argument*> &ArgsToPromote);
+ };
+
+ RegisterOpt<SimpleArgPromotion> X("simpleargpromotion",
+ "Promote 'by reference' arguments to 'by value'");
+ }
+
+ bool SimpleArgPromotion::runOnSCC(const std::vector<CallGraphNode*> &SCC) {
+ bool Changed = false, LocalChange;
+
+ do { // Iterate until we stop promoting from this SCC.
+ LocalChange = false;
+ // Attempt to promote arguments from all functions in this SCC.
+ for (unsigned i = 0, e = SCC.size(); i != e; ++i)
+ LocalChange |= PromoteArguments(SCC[i]);
+ Changed |= LocalChange; // Remember that we changed something.
+ } while (LocalChange);
+
+ return Changed;
+ }
+
+ /// PromoteArguments - This method checks the specified function to see if there
+ /// are any promotable arguments and if it is safe to promote the function (for
+ /// example, all callers are direct). If safe to promote some arguments, it
+ /// calls the DoPromotion method.
+ ///
+ bool SimpleArgPromotion::PromoteArguments(CallGraphNode *CGN) {
+ Function *F = CGN->getFunction();
+
+ // Make sure that it is local to this module.
+ if (!F || !F->hasInternalLinkage()) return false;
+
+ // First check: see if there are any pointer arguments! If not, quick exit.
+ std::vector<Argument*> PointerArgs;
+ for (Function::aiterator I = F->abegin(), E = F->aend(); I != E; ++I)
+ if (isa<PointerType>(I->getType()))
+ PointerArgs.push_back(I);
+ if (PointerArgs.empty()) return false;
+
+ // Second check: make sure that all callers are direct callers. We can't
+ // transform functions that have indirect callers.
+ for (Value::use_iterator UI = F->use_begin(), E = F->use_end();
+ UI != E; ++UI) {
+ CallSite CS = CallSite::get(*UI);
+ if (!CS.getInstruction()) // "Taking the address" of the function
+ return false;
+
+ // Ensure that this call site is CALLING the function, not passing it as
+ // an argument.
+ for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end();
+ AI != E; ++AI)
+ if (*AI == F) return false; // Passing the function address in!
+ }
+
+ // Check to see which arguments are promotable. If an argument is not
+ // promotable, remove it from the PointerArgs vector.
+ for (unsigned i = 0; i != PointerArgs.size(); ++i)
+ if (!isSafeToPromoteArgument(PointerArgs[i])) {
+ std::swap(PointerArgs[i--], PointerArgs.back());
+ PointerArgs.pop_back();
+ }
+
+ // No promotable pointer arguments.
+ if (PointerArgs.empty()) return false;
+
+ // Okay, promote all of the arguments are rewrite the callees!
+ Function *NewF = DoPromotion(F, PointerArgs);
+
+ // Update the call graph to know that the old function is gone.
+ getAnalysis<CallGraph>().changeFunction(F, NewF);
+ return true;
+ }
+
+
+ /// isSafeToPromoteArgument - As you might guess from the name of this method,
+ /// it checks to see if it is both safe and useful to promote the argument.
+ bool SimpleArgPromotion::isSafeToPromoteArgument(Argument *Arg) const {
+ // We can only promote this argument if all of the uses are loads.
+ std::vector<LoadInst*> Loads;
+
+ for (Value::use_iterator UI = Arg->use_begin(), E = Arg->use_end();
+ UI != E; ++UI)
+ if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
+ if (LI->isVolatile()) return false; // Don't modify volatile loads.
+ Loads.push_back(LI);
+ } else {
+ return false; // Not a load.
+ }
+
+ // Okay, now we know that the argument is only used by load instructions. Use
+ // alias analysis to check to see if the pointer is guaranteed to not be
+ // modified from entry of the function to each of the load instructions.
+ Function &F = *Arg->getParent();
+
+ // Because there could be several/many load instructions, remember which
+ // blocks we know to be transparent to the load.
+ std::set<BasicBlock*> TranspBlocks;
+
+ AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
+ TargetData &TD = getAnalysis<TargetData>();
+
+ for (unsigned i = 0, e = Loads.size(); i != e; ++i) {
+ // Check to see if the load is invalidated from the start of the block to
+ // the load itself.
+ LoadInst *Load = Loads[i];
+ BasicBlock *BB = Load->getParent();
+
+ const PointerType *LoadTy =
+ cast<PointerType>(Load->getOperand(0)->getType());
+ unsigned LoadSize = TD.getTypeSize(LoadTy->getElementType());
+
+ if (AA.canInstructionRangeModify(BB->front(), *Load, Arg, LoadSize))
+ return false; // Pointer is invalidated!
+
+ // Now check every path from the entry block to the load for transparency.
+ // To do this, we perform a depth first search on the inverse CFG from the
+ // loading block.
+ for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
+ for (idf_ext_iterator<BasicBlock*> I = idf_ext_begin(*PI, TranspBlocks),
+ E = idf_ext_end(*PI, TranspBlocks); I != E; ++I)
+ if (AA.canBasicBlockModify(**I, Arg, LoadSize))
+ return false;
+ }
+
+ // If the path from the entry of the function to each load is free of
+ // instructions that potentially invalidate the load, we can make the
+ // transformation!
+ return true;
+ }
+
+ /// DoPromotion - This method actually performs the promotion of the specified
+ /// arguments, and returns the new function. At this point, we know that it's
+ /// safe to do so.
+ Function *SimpleArgPromotion::DoPromotion(Function *F,
+ std::vector<Argument*> &Args2Prom) {
+ std::set<Argument*> ArgsToPromote(Args2Prom.begin(), Args2Prom.end());
+
+ // Start by computing a new prototype for the function, which is the same as
+ // the old function, but has modified arguments.
+ const FunctionType *FTy = F->getFunctionType();
+ std::vector<const Type*> Params;
+
+ for (Function::aiterator I = F->abegin(), E = F->aend(); I != E; ++I)
+ if (!ArgsToPromote.count(I)) {
+ Params.push_back(I->getType());
+ } else if (I->use_empty()) {
+ ++NumArgumentsDead;
+ } else {
+ // Add a parameter to the function for each element passed in.
+ Params.push_back(cast<PointerType>(I->getType())->getElementType());
+ ++NumArgumentsPromoted;
+ }
+
+ // Create the new function body and insert it into the module.
+ FunctionType *NFTy = FunctionType::get(FTy->getReturnType(), Params,
+ FTy->isVarArg());
+ Function *NF = new Function(NFTy, F->getLinkage(), F->getName());
+ F->getParent()->getFunctionList().insert(F, NF);
+
+ // Loop over all of the callers of the function, transforming the call sites
+ // to pass in the loaded pointers.
+ //
+ std::vector<Value*> Args;
+ while (!F->use_empty()) {
+ CallSite CS = CallSite::get(F->use_back());
+ Instruction *Call = CS.getInstruction();
+
+ // Loop over the operands, inserting the loads in the caller as needed.
+ CallSite::arg_iterator AI = CS.arg_begin();
+ for (Function::aiterator I = F->abegin(), E = F->aend(); I != E; ++I, ++AI)
+ if (!ArgsToPromote.count(I)) // Unmodified argument.
+ Args.push_back(*AI);
+ else if (!I->use_empty()) // Non-dead argument: insert the load.
+ Args.push_back(new LoadInst(*AI, (*AI)->getName()+".val", Call));
+
+ // Push any varargs arguments on the list
+ for (; AI != CS.arg_end(); ++AI)
+ Args.push_back(*AI);
+
+ Instruction *New; // Create the new call or invoke instruction.
+ if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
+ New = new InvokeInst(NF, II->getNormalDest(), II->getUnwindDest(),
+ Args, "", Call);
+ } else {
+ New = new CallInst(NF, Args, "", Call);
+ }
+ Args.clear();
+
+ if (!Call->use_empty()) {
+ Call->replaceAllUsesWith(New);
+ New->setName(Call->getName());
+ }
+
+ // Finally, remove the old call from the program, reducing the use-count of
+ // F.
+ Call->getParent()->getInstList().erase(Call);
+ }
+
+ // Since we have now created the new function, splice the body of the old
+ // function right into the new function, leaving the old rotting hulk of the
+ // function empty.
+ NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
+
+ // Loop over the argument list, transfering uses of the old arguments over to
+ // the new arguments, also transfering over the names as well.
+ //
+ for (Function::aiterator I = F->abegin(), E = F->aend(), I2 = NF->abegin();
+ I != E; ++I, ++I2)
+ if (!ArgsToPromote.count(I)) {
+ // If this is an unmodified argument, move the name and users over to the
+ // new version.
+ I->replaceAllUsesWith(I2);
+ I2->setName(I->getName());
+ } else if (!I->use_empty()) {
+ // Otherwise, if we promoted this argument, then all users are load
+ // instructions, and all loads should be using the new argument that we
+ // added.
+ while (!I->use_empty()) {
+ LoadInst *LI = cast<LoadInst>(I->use_back());
+ I2->setName(I->getName()+".val");
+ LI->replaceAllUsesWith(I2);
+ LI->getParent()->getInstList().erase(LI);
+ DEBUG(std::cerr << "*** Promoted load of argument '" << I->getName()
+ << "' in function '" << F->getName() << "'\n");
+ }
+ }
+
+ // Now that the old function is dead, delete it.
+ F->getParent()->getFunctionList().erase(F);
+ return NF;
+ }
+
+
+
+
+ /************************ Loading the pass into 'opt' **************************
+
+ $ opt -load ~/llvm/lib/Debug/libsimpleargpromote.so -help
+ OVERVIEW: llvm .bc -> .bc modular optimizer
+
+ USAGE: opt [options] <input bytecode>
+
+ OPTIONS:
+ Optimizations available:
+ ...
+ -sccp - Sparse Conditional Constant Propagation
+ -simpleargpromotion - Promote 'by reference' arguments to 'by value'
+ -simplifycfg - Simplify the CFG
+ ...
+ -load=<pluginfilename> - Load the specified plugin
+ ...
+ -stats - Enable statistics output from program
+ ...
+
+
+ ************************ Simple LLVM Example ***********************************
+
+ --------- basictest.ll ---------
+ internal int %test(int *%X, int* %Y) {
+ %A = load int* %X
+ %B = load int* %Y
+ %C = add int %A, %B
+ ret int %C
+ }
+
+ internal int %caller(int* %B) {
+ %A = alloca int
+ store int 1, int* %A
+ %C = call int %test(int* %A, int* %B)
+ ret int %C
+ }
+
+ int %callercaller() {
+ %B = alloca int
+ store int 2, int* %B
+ %X = call int %caller(int* %B)
+ ret int %X
+ }
+ --------- basictest.ll ---------
+
+ *********************** Run with simpleargpromotion ****************************
+
+ $ llvm-as < basictest.ll | opt -load ~/llvm/lib/Debug/libsimpleargpromote.so \
+ -simpleargpromotion -stats | llvm-dis
+
+ ===-------------------------------------------------------------------------===
+ ... Statistics Collected ...
+ ===-------------------------------------------------------------------------===
+
+ 248 bytecodewriter - Number of bytecode bytes written
+ 3 simpleargpromotion - Number of pointer arguments promoted
+
+
+ internal int %test(int %X.val, int %Y.val) {
+ %C = add int %X.val, %Y.val
+ ret int %C
+ }
+
+ internal int %caller(int %B.val) {
+ %A = alloca int
+ store int 1, int* %A
+ %A.val = load int* %A
+ %C1 = call int %test( int %A.val, int %B.val )
+ ret int %C1
+ }
+
+ int %callercaller() {
+ %B = alloca int
+ store int 2, int* %B
+ %B.val = load int* %B
+ %X1 = call int %caller( int %B.val )
+ ret int %X1
+ }
+
+ *********************** Run with simpleargpromotion & mem2reg ******************
+
+ $ llvm-as < basictest.ll | opt -load ~/llvm/lib/Debug/libsimpleargpromote.so \
+ -simpleargpromotion -mem2reg -stats | llvm-dis
+
+ ===-------------------------------------------------------------------------===
+ ... Statistics Collected ...
+ ===-------------------------------------------------------------------------===
+
+ 194 bytecodewriter - Number of bytecode bytes written
+ 2 mem2reg - Number of alloca's promoted
+ 3 simpleargpromotion - Number of pointer arguments promoted
+
+ internal int %test(int %X.val, int %Y.val) {
+ %C = add int %X.val, %Y.val
+ ret int %C
+ }
+
+ internal int %caller(int %B.val) {
+ %C1 = call int %test( int 1, int %B.val )
+ ret int %C1
+ }
+
+ int %callercaller() {
+ %X1 = call int %caller( int 2 )
+ ret int %X1
+ }
+
+ ****************************** Simple C++ Example ******************************
+
+ void test(std::vector<int> &V) {
+ V.push_back(7);
+ }
+
+ ... compiles to this LLVM code:
+
+ void %_Z4testRSt6vectorIiSaIiEE("std::vector<int>"* %V) {
+ %mem_tmp = alloca int
+ store int 7, int* %mem_tmp
+ call void %_ZNSt6vectorIiSaIiEE9push_backERKi("std::vector<int>"* %V,
+ int* %mem_tmp)
+ ret void
+ }
+
+ ... arg promotion and mem2reg result in this, eliminating the stack allocation
+ and simplifying the code.
+
+ void %_Z4testRSt6vectorIiSaIiEE("std::vector<int>"* %V) {
+ call void %_ZNSt6vectorIiSaIiEE9push_backERKi("std::vector<int>"* %V,
+ int 7)
+ ret void
+ }
+
+ */
Index: llvm-www/pubs/2004-09-22-LCPCLLVMTutorial.html
diff -c /dev/null llvm-www/pubs/2004-09-22-LCPCLLVMTutorial.html:1.1
*** /dev/null Wed Sep 22 00:30:37 2004
--- llvm-www/pubs/2004-09-22-LCPCLLVMTutorial.html Wed Sep 22 00:30:25 2004
***************
*** 0 ****
--- 1,51 ----
+ <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
+ <html>
+ <head>
+ <meta http-equiv="Content-Type" content="text/html; charset=UTF-8" />
+ <link rel="stylesheet" href="../llvm.css" type="text/css" media="screen" />
+ <title>The LLVM Compiler Framework and Infrastructure Tutorial</title>
+ </head>
+ <body>
+
+ <div class="pub_title">
+ The LLVM Compiler Framework and Infrastructure Tutorial
+ </div>
+ <div class="pub_author">
+ <a href="http://www.nondot.org/sabre/">Chris Lattner</a> and
+ <a href="http://www.cs.uiuc.edu/~vadve">Vikram Adve</a>
+ </div>
+
+ <h2>Abstract:</h2>
+ <blockquote>
+
+ <p>The Low-Level Virtual Machine (LLVM) is a collection of libraries and tools that make it easy to build compilers, optimizers, Just-In-Time code generators, and many other compiler-related programs. LLVM uses a single, language-independent virtual instruction set both as an offline code representation (to communicate code between compiler phases and to run-time systems) and as the compiler internal representation (to analyze and transform programs). This persistent code representation allows a common set of sophisticated compiler techniques to be applied at compile-time, link-time, install-time, run-time, or "idle-time" (between program runs).<p>
+ <p>The strengths of the LLVM infrastructure are its extremely simple design (which makes it easy to understand and use), source-language independence, powerful mid-level optimizer, automated compiler debugging support, extensibility, and its stability and reliability. LLVM is currently being used to host a wide variety of academic research projects and commercial projects. LLVM includes C and C++ front-ends (based on GCC 3.4), a front-end for a Forth-like language (Stacker), a young scheme front-end, and Java support is in development. LLVM can generate code for X86, SparcV9, PowerPC, or it can emit C code.</p>
+ <p>This tutorial describes the LLVM virtual instruction set and the high-level design of the LLVM compiler system. To illustrate the ideas in the LLVM IR, we use a running example (by-reference to by-value argument promotion) to illustrate several important API's in the LLVM system. Next, we describe some of the key tools provided by LLVM, and mention several projects that are natural targets for the LLVM system.</p>
+ </blockquote>
+
+ <h2>Presented:</h2>
+ <blockquote>
+ "The LLVM Compiler Framework and Infrastructure Tutorial", Chris Lattner and Vikram Adve.<br>
+ <a href="http://www.ecn.purdue.edu/LCPC2004/">LCPC'04</a> <a
+ href="http://www.ecn.purdue.edu/LCPC2004/miniws.html">Mini Workshop</a> on Compiler Research
+ Infrastructures, West Lafayette, Indiana, Sep. 2004.
+ </blockquote>
+
+ <h2>Download:</h2>
+
+ Slides:
+ <ul>
+ <li><a href="2004-09-22-LCPCLLVMTutorial.ppt">The LLVM Compiler Framework and Infrastructure
+ Tutorial Presentation</a> (PPT)</li>
+ <li><a href="2004-09-22-LCPCLLVMTutorial.pdf">The LLVM Compiler Framework and Infrastructure
+ Tutorial Presentation</a> (PDF) [note, some slides are impossible to read due to lack of animations]</li>
+ </ul>
+
+ Handout:
+ <ul>
+ <li><a href="2004-09-22-LCPCLLVMTutorial-Handout.pdf">The LLVM Compiler Framework and Infrastructure Tutorial Handout</a> (PDF)</li>
+ <li><a href="2004-09-22-LCPCLLVMTutorial-Handout.txt">The LLVM Compiler Framework and Infrastructure Tutorial Handout</a> (TXT)</li>
+ </ul>
+
+ </body>
+ </html>
Index: llvm-www/pubs/2004-09-22-LCPCLLVMTutorial.pdf
Index: llvm-www/pubs/2004-09-22-LCPCLLVMTutorial.ppt
More information about the llvm-commits
mailing list