mirror of
https://github.com/RPCS3/llvm.git
synced 2024-12-27 14:45:50 +00:00
Remove the long, long defunct IR block placement pass.
This pass was based on the previous (essentially unused) profiling infrastructure and the assumption that by ordering the basic blocks at the IR level in a particular way, the correct layout would happen in the end. This sometimes worked, and mostly didn't. It also was a really naive implementation of the classical paper that dates from when branch predictors were primarily directional and when loop structure wasn't commonly available. It also didn't factor into the equation non-fallthrough branches and other machine level details. Anyways, for all of these reasons and more, I wrote MachineBlockPlacement, which completely supercedes this pass. It both uses modern profile information infrastructure, and actually works. =] git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@190748 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
parent
dcc425c630
commit
3748de6e2d
@ -76,7 +76,6 @@ void initializeBasicCallGraphPass(PassRegistry&);
|
||||
void initializeBasicTTIPass(PassRegistry&);
|
||||
void initializeBlockExtractorPassPass(PassRegistry&);
|
||||
void initializeBlockFrequencyInfoPass(PassRegistry&);
|
||||
void initializeBlockPlacementPass(PassRegistry&);
|
||||
void initializeBoundsCheckingPass(PassRegistry&);
|
||||
void initializeBranchFolderPassPass(PassRegistry&);
|
||||
void initializeBranchProbabilityInfoPass(PassRegistry&);
|
||||
|
@ -56,7 +56,6 @@ namespace {
|
||||
(void) llvm::createLibCallAliasAnalysisPass(0);
|
||||
(void) llvm::createScalarEvolutionAliasAnalysisPass();
|
||||
(void) llvm::createTypeBasedAliasAnalysisPass();
|
||||
(void) llvm::createBlockPlacementPass();
|
||||
(void) llvm::createBoundsCheckingPass();
|
||||
(void) llvm::createBreakCriticalEdgesPass();
|
||||
(void) llvm::createCallGraphPrinterPass();
|
||||
|
@ -265,13 +265,6 @@ FunctionPass *createLowerInvokePass(const TargetMachine *TM = 0,
|
||||
bool useExpensiveEHSupport = false);
|
||||
extern char &LowerInvokePassID;
|
||||
|
||||
//===----------------------------------------------------------------------===//
|
||||
//
|
||||
// BlockPlacement - This pass reorders basic blocks in order to increase the
|
||||
// number of fall-through conditional branches.
|
||||
//
|
||||
FunctionPass *createBlockPlacementPass();
|
||||
|
||||
//===----------------------------------------------------------------------===//
|
||||
//
|
||||
// LCSSA - This pass inserts phi nodes at loop boundaries to simplify other loop
|
||||
|
@ -1,152 +0,0 @@
|
||||
//===-- BasicBlockPlacement.cpp - Basic Block Code Layout optimization ----===//
|
||||
//
|
||||
// The LLVM Compiler Infrastructure
|
||||
//
|
||||
// This file is distributed under the University of Illinois Open Source
|
||||
// License. See LICENSE.TXT for details.
|
||||
//
|
||||
//===----------------------------------------------------------------------===//
|
||||
//
|
||||
// This file implements a very simple profile guided basic block placement
|
||||
// algorithm. The idea is to put frequently executed blocks together at the
|
||||
// start of the function, and hopefully increase the number of fall-through
|
||||
// conditional branches. If there is no profile information for a particular
|
||||
// function, this pass basically orders blocks in depth-first order
|
||||
//
|
||||
// The algorithm implemented here is basically "Algo1" from "Profile Guided Code
|
||||
// Positioning" by Pettis and Hansen, except that it uses basic block counts
|
||||
// instead of edge counts. This should be improved in many ways, but is very
|
||||
// simple for now.
|
||||
//
|
||||
// Basically we "place" the entry block, then loop over all successors in a DFO,
|
||||
// placing the most frequently executed successor until we run out of blocks. I
|
||||
// told you this was _extremely_ simplistic. :) This is also much slower than it
|
||||
// could be. When it becomes important, this pass will be rewritten to use a
|
||||
// better algorithm, and then we can worry about efficiency.
|
||||
//
|
||||
//===----------------------------------------------------------------------===//
|
||||
|
||||
#define DEBUG_TYPE "block-placement"
|
||||
#include "llvm/Transforms/Scalar.h"
|
||||
#include "llvm/ADT/Statistic.h"
|
||||
#include "llvm/Analysis/ProfileInfo.h"
|
||||
#include "llvm/IR/Function.h"
|
||||
#include "llvm/Pass.h"
|
||||
#include "llvm/Support/CFG.h"
|
||||
#include <set>
|
||||
using namespace llvm;
|
||||
|
||||
STATISTIC(NumMoved, "Number of basic blocks moved");
|
||||
|
||||
namespace {
|
||||
struct BlockPlacement : public FunctionPass {
|
||||
static char ID; // Pass identification, replacement for typeid
|
||||
BlockPlacement() : FunctionPass(ID) {
|
||||
initializeBlockPlacementPass(*PassRegistry::getPassRegistry());
|
||||
}
|
||||
|
||||
virtual bool runOnFunction(Function &F);
|
||||
|
||||
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
|
||||
AU.setPreservesCFG();
|
||||
AU.addRequired<ProfileInfo>();
|
||||
//AU.addPreserved<ProfileInfo>(); // Does this work?
|
||||
}
|
||||
private:
|
||||
/// PI - The profile information that is guiding us.
|
||||
///
|
||||
ProfileInfo *PI;
|
||||
|
||||
/// NumMovedBlocks - Every time we move a block, increment this counter.
|
||||
///
|
||||
unsigned NumMovedBlocks;
|
||||
|
||||
/// PlacedBlocks - Every time we place a block, remember it so we don't get
|
||||
/// into infinite loops.
|
||||
std::set<BasicBlock*> PlacedBlocks;
|
||||
|
||||
/// InsertPos - This an iterator to the next place we want to insert a
|
||||
/// block.
|
||||
Function::iterator InsertPos;
|
||||
|
||||
/// PlaceBlocks - Recursively place the specified blocks and any unplaced
|
||||
/// successors.
|
||||
void PlaceBlocks(BasicBlock *BB);
|
||||
};
|
||||
}
|
||||
|
||||
char BlockPlacement::ID = 0;
|
||||
INITIALIZE_PASS_BEGIN(BlockPlacement, "block-placement",
|
||||
"Profile Guided Basic Block Placement", false, false)
|
||||
INITIALIZE_AG_DEPENDENCY(ProfileInfo)
|
||||
INITIALIZE_PASS_END(BlockPlacement, "block-placement",
|
||||
"Profile Guided Basic Block Placement", false, false)
|
||||
|
||||
FunctionPass *llvm::createBlockPlacementPass() { return new BlockPlacement(); }
|
||||
|
||||
bool BlockPlacement::runOnFunction(Function &F) {
|
||||
PI = &getAnalysis<ProfileInfo>();
|
||||
|
||||
NumMovedBlocks = 0;
|
||||
InsertPos = F.begin();
|
||||
|
||||
// Recursively place all blocks.
|
||||
PlaceBlocks(F.begin());
|
||||
|
||||
PlacedBlocks.clear();
|
||||
NumMoved += NumMovedBlocks;
|
||||
return NumMovedBlocks != 0;
|
||||
}
|
||||
|
||||
|
||||
/// PlaceBlocks - Recursively place the specified blocks and any unplaced
|
||||
/// successors.
|
||||
void BlockPlacement::PlaceBlocks(BasicBlock *BB) {
|
||||
assert(!PlacedBlocks.count(BB) && "Already placed this block!");
|
||||
PlacedBlocks.insert(BB);
|
||||
|
||||
// Place the specified block.
|
||||
if (&*InsertPos != BB) {
|
||||
// Use splice to move the block into the right place. This avoids having to
|
||||
// remove the block from the function then readd it, which causes a bunch of
|
||||
// symbol table traffic that is entirely pointless.
|
||||
Function::BasicBlockListType &Blocks = BB->getParent()->getBasicBlockList();
|
||||
Blocks.splice(InsertPos, Blocks, BB);
|
||||
|
||||
++NumMovedBlocks;
|
||||
} else {
|
||||
// This block is already in the right place, we don't have to do anything.
|
||||
++InsertPos;
|
||||
}
|
||||
|
||||
// Keep placing successors until we run out of ones to place. Note that this
|
||||
// loop is very inefficient (N^2) for blocks with many successors, like switch
|
||||
// statements. FIXME!
|
||||
while (1) {
|
||||
// Okay, now place any unplaced successors.
|
||||
succ_iterator SI = succ_begin(BB), E = succ_end(BB);
|
||||
|
||||
// Scan for the first unplaced successor.
|
||||
for (; SI != E && PlacedBlocks.count(*SI); ++SI)
|
||||
/*empty*/;
|
||||
if (SI == E) return; // No more successors to place.
|
||||
|
||||
double MaxExecutionCount = PI->getExecutionCount(*SI);
|
||||
BasicBlock *MaxSuccessor = *SI;
|
||||
|
||||
// Scan for more frequently executed successors
|
||||
for (; SI != E; ++SI)
|
||||
if (!PlacedBlocks.count(*SI)) {
|
||||
double Count = PI->getExecutionCount(*SI);
|
||||
if (Count > MaxExecutionCount ||
|
||||
// Prefer to not disturb the code.
|
||||
(Count == MaxExecutionCount && *SI == &*InsertPos)) {
|
||||
MaxExecutionCount = Count;
|
||||
MaxSuccessor = *SI;
|
||||
}
|
||||
}
|
||||
|
||||
// Now that we picked the maximally executed successor, place it.
|
||||
PlaceBlocks(MaxSuccessor);
|
||||
}
|
||||
}
|
@ -1,6 +1,5 @@
|
||||
add_llvm_library(LLVMScalarOpts
|
||||
ADCE.cpp
|
||||
BasicBlockPlacement.cpp
|
||||
CodeGenPrepare.cpp
|
||||
ConstantProp.cpp
|
||||
CorrelatedValuePropagation.cpp
|
||||
|
@ -28,7 +28,6 @@ using namespace llvm;
|
||||
/// ScalarOpts library.
|
||||
void llvm::initializeScalarOpts(PassRegistry &Registry) {
|
||||
initializeADCEPass(Registry);
|
||||
initializeBlockPlacementPass(Registry);
|
||||
initializeCodeGenPreparePass(Registry);
|
||||
initializeConstantPropagationPass(Registry);
|
||||
initializeCorrelatedValuePropagationPass(Registry);
|
||||
|
@ -1,15 +0,0 @@
|
||||
; RUN: opt < %s -block-placement -disable-output -print-function 2> /dev/null
|
||||
|
||||
define i32 @test() {
|
||||
br i1 true, label %X, label %Y
|
||||
|
||||
A: ; preds = %Y, %X
|
||||
ret i32 0
|
||||
|
||||
X: ; preds = %0
|
||||
br label %A
|
||||
|
||||
Y: ; preds = %0
|
||||
br label %A
|
||||
}
|
||||
|
Loading…
Reference in New Issue
Block a user