1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-11-26 12:43:36 +01:00
llvm-mirror/lib/Analysis/CodeMetrics.cpp

210 lines
7.7 KiB
C++
Raw Normal View History

//===- CodeMetrics.cpp - Code cost measurements ---------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file implements code cost measurement utilities.
//
//===----------------------------------------------------------------------===//
#include "llvm/Analysis/CodeMetrics.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/Analysis/AssumptionCache.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/IR/Function.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/InstructionCost.h"
#define DEBUG_TYPE "code-metrics"
using namespace llvm;
static void
appendSpeculatableOperands(const Value *V,
SmallPtrSetImpl<const Value *> &Visited,
SmallVectorImpl<const Value *> &Worklist) {
const User *U = dyn_cast<User>(V);
if (!U)
return;
for (const Value *Operand : U->operands())
if (Visited.insert(Operand).second)
if (isSafeToSpeculativelyExecute(Operand))
Worklist.push_back(Operand);
}
static void completeEphemeralValues(SmallPtrSetImpl<const Value *> &Visited,
SmallVectorImpl<const Value *> &Worklist,
SmallPtrSetImpl<const Value *> &EphValues) {
// Note: We don't speculate PHIs here, so we'll miss instruction chains kept
// alive only by ephemeral values.
// Walk the worklist using an index but without caching the size so we can
// append more entries as we process the worklist. This forms a queue without
// quadratic behavior by just leaving processed nodes at the head of the
// worklist forever.
for (int i = 0; i < (int)Worklist.size(); ++i) {
const Value *V = Worklist[i];
assert(Visited.count(V) &&
"Failed to add a worklist entry to our visited set!");
// If all uses of this value are ephemeral, then so is this value.
if (!all_of(V->users(), [&](const User *U) { return EphValues.count(U); }))
continue;
EphValues.insert(V);
LLVM_DEBUG(dbgs() << "Ephemeral Value: " << *V << "\n");
// Append any more operands to consider.
appendSpeculatableOperands(V, Visited, Worklist);
}
}
// Find all ephemeral values.
void CodeMetrics::collectEphemeralValues(
const Loop *L, AssumptionCache *AC,
SmallPtrSetImpl<const Value *> &EphValues) {
SmallPtrSet<const Value *, 32> Visited;
SmallVector<const Value *, 16> Worklist;
for (auto &AssumeVH : AC->assumptions()) {
if (!AssumeVH)
continue;
Instruction *I = cast<Instruction>(AssumeVH);
// Filter out call sites outside of the loop so we don't do a function's
// worth of work for each of its loops (and, in the common case, ephemeral
// values in the loop are likely due to @llvm.assume calls in the loop).
if (!L->contains(I->getParent()))
continue;
if (EphValues.insert(I).second)
appendSpeculatableOperands(I, Visited, Worklist);
}
completeEphemeralValues(Visited, Worklist, EphValues);
}
void CodeMetrics::collectEphemeralValues(
const Function *F, AssumptionCache *AC,
SmallPtrSetImpl<const Value *> &EphValues) {
SmallPtrSet<const Value *, 32> Visited;
SmallVector<const Value *, 16> Worklist;
for (auto &AssumeVH : AC->assumptions()) {
if (!AssumeVH)
continue;
Instruction *I = cast<Instruction>(AssumeVH);
assert(I->getParent()->getParent() == F &&
"Found assumption for the wrong function!");
if (EphValues.insert(I).second)
appendSpeculatableOperands(I, Visited, Worklist);
}
completeEphemeralValues(Visited, Worklist, EphValues);
}
/// Fill in the current structure with information gleaned from the specified
/// block.
[LoopRotate] Add PrepareForLTO stage, avoid rotating with inline cands. D84108 exposed a bad interaction between inlining and loop-rotation during regular LTO, which is causing notable regressions in at least CINT2006/473.astar. The problem boils down to: we now rotate a loop just before the vectorizer which requires duplicating a function call in the preheader when compiling the individual files ('prepare for LTO'). But this then prevents further inlining of the function during LTO. This patch tries to resolve this issue by making LoopRotate more conservative with respect to rotating loops that have inline-able calls during the 'prepare for LTO' stage. I think this change intuitively improves the current situation in general. Loop-rotate tries hard to avoid creating headers that are 'too big'. At the moment, it assumes all inlining already happened and the cost of duplicating a call is equal to just doing the call. But with LTO, inlining also happens during full LTO and it is possible that a previously duplicated call is actually a huge function which gets inlined during LTO. From the perspective of LV, not much should change overall. Most loops calling user-provided functions won't get vectorized to start with (unless we can infer that the function does not touch memory, has no other side effects). If we do not inline the 'inline-able' call during the LTO stage, we merely delayed loop-rotation & vectorization. If we inline during LTO, chances should be very high that the inlined code is itself vectorizable or the user call was not vectorizable to start with. There could of course be scenarios where we inline a sufficiently large function with code not profitable to vectorize, which would have be vectorized earlier (by scalarzing the call). But even in that case, there probably is no big performance impact, because it should be mostly down to the cost-model to reject vectorization in that case. And then the version with scalarized calls should also not be beneficial. In a way, LV should have strictly more information after inlining and make more accurate decisions (barring cost-model issues). There is of course plenty of room for things to go wrong unexpectedly, so we need to keep a close look at actual performance and address any follow-up issues. I took a look at the impact on statistics for MultiSource/SPEC2000/SPEC2006. There are a few benchmarks with fewer loops rotated, but no change to the number of loops vectorized. Reviewed By: sanwou01 Differential Revision: https://reviews.llvm.org/D94232
2021-01-19 10:22:40 +01:00
void CodeMetrics::analyzeBasicBlock(
const BasicBlock *BB, const TargetTransformInfo &TTI,
const SmallPtrSetImpl<const Value *> &EphValues, bool PrepareForLTO) {
++NumBlocks;
// Use a proxy variable for NumInsts of type InstructionCost, so that it can
// use InstructionCost's arithmetic properties such as saturation when this
// feature is added to InstructionCost.
// When storing the value back to NumInsts, we can assume all costs are Valid
// because the IR should not contain any nodes that cannot be costed. If that
// happens the cost-model is broken.
InstructionCost NumInstsProxy = NumInsts;
InstructionCost NumInstsBeforeThisBB = NumInsts;
for (const Instruction &I : *BB) {
// Skip ephemeral values.
if (EphValues.count(&I))
continue;
// Special handling for calls.
if (const auto *Call = dyn_cast<CallBase>(&I)) {
if (const Function *F = Call->getCalledFunction()) {
bool IsLoweredToCall = TTI.isLoweredToCall(F);
// If a function is both internal and has a single use, then it is
// extremely likely to get inlined in the future (it was probably
// exposed by an interleaved devirtualization pass).
[LoopRotate] Add PrepareForLTO stage, avoid rotating with inline cands. D84108 exposed a bad interaction between inlining and loop-rotation during regular LTO, which is causing notable regressions in at least CINT2006/473.astar. The problem boils down to: we now rotate a loop just before the vectorizer which requires duplicating a function call in the preheader when compiling the individual files ('prepare for LTO'). But this then prevents further inlining of the function during LTO. This patch tries to resolve this issue by making LoopRotate more conservative with respect to rotating loops that have inline-able calls during the 'prepare for LTO' stage. I think this change intuitively improves the current situation in general. Loop-rotate tries hard to avoid creating headers that are 'too big'. At the moment, it assumes all inlining already happened and the cost of duplicating a call is equal to just doing the call. But with LTO, inlining also happens during full LTO and it is possible that a previously duplicated call is actually a huge function which gets inlined during LTO. From the perspective of LV, not much should change overall. Most loops calling user-provided functions won't get vectorized to start with (unless we can infer that the function does not touch memory, has no other side effects). If we do not inline the 'inline-able' call during the LTO stage, we merely delayed loop-rotation & vectorization. If we inline during LTO, chances should be very high that the inlined code is itself vectorizable or the user call was not vectorizable to start with. There could of course be scenarios where we inline a sufficiently large function with code not profitable to vectorize, which would have be vectorized earlier (by scalarzing the call). But even in that case, there probably is no big performance impact, because it should be mostly down to the cost-model to reject vectorization in that case. And then the version with scalarized calls should also not be beneficial. In a way, LV should have strictly more information after inlining and make more accurate decisions (barring cost-model issues). There is of course plenty of room for things to go wrong unexpectedly, so we need to keep a close look at actual performance and address any follow-up issues. I took a look at the impact on statistics for MultiSource/SPEC2000/SPEC2006. There are a few benchmarks with fewer loops rotated, but no change to the number of loops vectorized. Reviewed By: sanwou01 Differential Revision: https://reviews.llvm.org/D94232
2021-01-19 10:22:40 +01:00
// When preparing for LTO, liberally consider calls as inline
// candidates.
if (!Call->isNoInline() && IsLoweredToCall &&
[LoopRotate] Add PrepareForLTO stage, avoid rotating with inline cands. D84108 exposed a bad interaction between inlining and loop-rotation during regular LTO, which is causing notable regressions in at least CINT2006/473.astar. The problem boils down to: we now rotate a loop just before the vectorizer which requires duplicating a function call in the preheader when compiling the individual files ('prepare for LTO'). But this then prevents further inlining of the function during LTO. This patch tries to resolve this issue by making LoopRotate more conservative with respect to rotating loops that have inline-able calls during the 'prepare for LTO' stage. I think this change intuitively improves the current situation in general. Loop-rotate tries hard to avoid creating headers that are 'too big'. At the moment, it assumes all inlining already happened and the cost of duplicating a call is equal to just doing the call. But with LTO, inlining also happens during full LTO and it is possible that a previously duplicated call is actually a huge function which gets inlined during LTO. From the perspective of LV, not much should change overall. Most loops calling user-provided functions won't get vectorized to start with (unless we can infer that the function does not touch memory, has no other side effects). If we do not inline the 'inline-able' call during the LTO stage, we merely delayed loop-rotation & vectorization. If we inline during LTO, chances should be very high that the inlined code is itself vectorizable or the user call was not vectorizable to start with. There could of course be scenarios where we inline a sufficiently large function with code not profitable to vectorize, which would have be vectorized earlier (by scalarzing the call). But even in that case, there probably is no big performance impact, because it should be mostly down to the cost-model to reject vectorization in that case. And then the version with scalarized calls should also not be beneficial. In a way, LV should have strictly more information after inlining and make more accurate decisions (barring cost-model issues). There is of course plenty of room for things to go wrong unexpectedly, so we need to keep a close look at actual performance and address any follow-up issues. I took a look at the impact on statistics for MultiSource/SPEC2000/SPEC2006. There are a few benchmarks with fewer loops rotated, but no change to the number of loops vectorized. Reviewed By: sanwou01 Differential Revision: https://reviews.llvm.org/D94232
2021-01-19 10:22:40 +01:00
((F->hasInternalLinkage() && F->hasOneUse()) || PrepareForLTO)) {
++NumInlineCandidates;
[LoopRotate] Add PrepareForLTO stage, avoid rotating with inline cands. D84108 exposed a bad interaction between inlining and loop-rotation during regular LTO, which is causing notable regressions in at least CINT2006/473.astar. The problem boils down to: we now rotate a loop just before the vectorizer which requires duplicating a function call in the preheader when compiling the individual files ('prepare for LTO'). But this then prevents further inlining of the function during LTO. This patch tries to resolve this issue by making LoopRotate more conservative with respect to rotating loops that have inline-able calls during the 'prepare for LTO' stage. I think this change intuitively improves the current situation in general. Loop-rotate tries hard to avoid creating headers that are 'too big'. At the moment, it assumes all inlining already happened and the cost of duplicating a call is equal to just doing the call. But with LTO, inlining also happens during full LTO and it is possible that a previously duplicated call is actually a huge function which gets inlined during LTO. From the perspective of LV, not much should change overall. Most loops calling user-provided functions won't get vectorized to start with (unless we can infer that the function does not touch memory, has no other side effects). If we do not inline the 'inline-able' call during the LTO stage, we merely delayed loop-rotation & vectorization. If we inline during LTO, chances should be very high that the inlined code is itself vectorizable or the user call was not vectorizable to start with. There could of course be scenarios where we inline a sufficiently large function with code not profitable to vectorize, which would have be vectorized earlier (by scalarzing the call). But even in that case, there probably is no big performance impact, because it should be mostly down to the cost-model to reject vectorization in that case. And then the version with scalarized calls should also not be beneficial. In a way, LV should have strictly more information after inlining and make more accurate decisions (barring cost-model issues). There is of course plenty of room for things to go wrong unexpectedly, so we need to keep a close look at actual performance and address any follow-up issues. I took a look at the impact on statistics for MultiSource/SPEC2000/SPEC2006. There are a few benchmarks with fewer loops rotated, but no change to the number of loops vectorized. Reviewed By: sanwou01 Differential Revision: https://reviews.llvm.org/D94232
2021-01-19 10:22:40 +01:00
}
// If this call is to function itself, then the function is recursive.
// Inlining it into other functions is a bad idea, because this is
// basically just a form of loop peeling, and our metrics aren't useful
// for that case.
if (F == BB->getParent())
isRecursive = true;
if (IsLoweredToCall)
++NumCalls;
} else {
// We don't want inline asm to count as a call - that would prevent loop
// unrolling. The argument setup cost is still real, though.
if (!Call->isInlineAsm())
++NumCalls;
}
}
if (const AllocaInst *AI = dyn_cast<AllocaInst>(&I)) {
if (!AI->isStaticAlloca())
this->usesDynamicAlloca = true;
}
if (isa<ExtractElementInst>(I) || I.getType()->isVectorTy())
++NumVectorInsts;
if (I.getType()->isTokenTy() && I.isUsedOutsideOfBlock(BB))
[IR] Add token types This introduces the basic functionality to support "token types". The motivation stems from the need to perform operations on a Value whose provenance cannot be obscured. There are several applications for such a type but my immediate motivation stems from WinEH. Our personality routine enforces a single-entry - single-exit regime for cleanups. After several rounds of optimizations, we may be left with a terminator whose "cleanup-entry block" is not entirely clear because control flow has merged two cleanups together. We have experimented with using labels as operands inside of instructions which are not terminators to indicate where we came from but found that LLVM does not expect such exotic uses of BasicBlocks. Instead, we can use this new type to clearly associate the "entry point" and "exit point" of our cleanup. This is done by having the cleanuppad yield a Token and consuming it at the cleanupret. The token type makes it impossible to obscure or otherwise hide the Value, making it trivial to track the relationship between the two points. What is the burden to the optimizer? Well, it turns out we have already paid down this cost by accepting that there are certain calls that we are not permitted to duplicate, optimizations have to watch out for such instructions anyway. There are additional places in the optimizer that we will probably have to update but early examination has given me the impression that this will not be heroic. Differential Revision: http://reviews.llvm.org/D11861 llvm-svn: 245029
2015-08-14 07:09:07 +02:00
notDuplicatable = true;
if (const CallInst *CI = dyn_cast<CallInst>(&I)) {
if (CI->cannotDuplicate())
notDuplicatable = true;
if (CI->isConvergent())
convergent = true;
}
if (const InvokeInst *InvI = dyn_cast<InvokeInst>(&I))
if (InvI->cannotDuplicate())
notDuplicatable = true;
NumInstsProxy += TTI.getUserCost(&I, TargetTransformInfo::TCK_CodeSize);
NumInsts = *NumInstsProxy.getValue();
}
if (isa<ReturnInst>(BB->getTerminator()))
++NumRets;
// We never want to inline functions that contain an indirectbr. This is
// incorrect because all the blockaddress's (in static global initializers
// for example) would be referring to the original function, and this indirect
// jump would jump from the inlined copy of the function into the original
// function which is extremely undefined behavior.
// FIXME: This logic isn't really right; we can safely inline functions
// with indirectbr's as long as no other function or global references the
// blockaddress of a block within the current function. And as a QOI issue,
// if someone is using a blockaddress without an indirectbr, and that
// reference somehow ends up in another function or global, we probably
// don't want to inline this function.
notDuplicatable |= isa<IndirectBrInst>(BB->getTerminator());
// Remember NumInsts for this BB.
InstructionCost NumInstsThisBB = NumInstsProxy - NumInstsBeforeThisBB;
NumBBInsts[BB] = *NumInstsThisBB.getValue();
}