1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-22 20:43:44 +02:00
llvm-mirror/lib/Transforms/IPO/FunctionAttrs.cpp

1229 lines
42 KiB
C++
Raw Normal View History

//===- FunctionAttrs.cpp - Pass which marks functions attributes ----------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
///
/// \file
/// This file implements interprocedural passes which walk the
/// call-graph deducing and/or propagating function attributes.
///
//===----------------------------------------------------------------------===//
#include "llvm/Transforms/IPO/FunctionAttrs.h"
#include "llvm/Transforms/IPO.h"
#include "llvm/ADT/SCCIterator.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/StringSwitch.h"
#include "llvm/Analysis/AliasAnalysis.h"
[PM/AA] Rebuild LLVM's alias analysis infrastructure in a way compatible with the new pass manager, and no longer relying on analysis groups. This builds essentially a ground-up new AA infrastructure stack for LLVM. The core ideas are the same that are used throughout the new pass manager: type erased polymorphism and direct composition. The design is as follows: - FunctionAAResults is a type-erasing alias analysis results aggregation interface to walk a single query across a range of results from different alias analyses. Currently this is function-specific as we always assume that aliasing queries are *within* a function. - AAResultBase is a CRTP utility providing stub implementations of various parts of the alias analysis result concept, notably in several cases in terms of other more general parts of the interface. This can be used to implement only a narrow part of the interface rather than the entire interface. This isn't really ideal, this logic should be hoisted into FunctionAAResults as currently it will cause a significant amount of redundant work, but it faithfully models the behavior of the prior infrastructure. - All the alias analysis passes are ported to be wrapper passes for the legacy PM and new-style analysis passes for the new PM with a shared result object. In some cases (most notably CFL), this is an extremely naive approach that we should revisit when we can specialize for the new pass manager. - BasicAA has been restructured to reflect that it is much more fundamentally a function analysis because it uses dominator trees and loop info that need to be constructed for each function. All of the references to getting alias analysis results have been updated to use the new aggregation interface. All the preservation and other pass management code has been updated accordingly. The way the FunctionAAResultsWrapperPass works is to detect the available alias analyses when run, and add them to the results object. This means that we should be able to continue to respect when various passes are added to the pipeline, for example adding CFL or adding TBAA passes should just cause their results to be available and to get folded into this. The exception to this rule is BasicAA which really needs to be a function pass due to using dominator trees and loop info. As a consequence, the FunctionAAResultsWrapperPass directly depends on BasicAA and always includes it in the aggregation. This has significant implications for preserving analyses. Generally, most passes shouldn't bother preserving FunctionAAResultsWrapperPass because rebuilding the results just updates the set of known AA passes. The exception to this rule are LoopPass instances which need to preserve all the function analyses that the loop pass manager will end up needing. This means preserving both BasicAAWrapperPass and the aggregating FunctionAAResultsWrapperPass. Now, when preserving an alias analysis, you do so by directly preserving that analysis. This is only necessary for non-immutable-pass-provided alias analyses though, and there are only three of interest: BasicAA, GlobalsAA (formerly GlobalsModRef), and SCEVAA. Usually BasicAA is preserved when needed because it (like DominatorTree and LoopInfo) is marked as a CFG-only pass. I've expanded GlobalsAA into the preserved set everywhere we previously were preserving all of AliasAnalysis, and I've added SCEVAA in the intersection of that with where we preserve SCEV itself. One significant challenge to all of this is that the CGSCC passes were actually using the alias analysis implementations by taking advantage of a pretty amazing set of loop holes in the old pass manager's analysis management code which allowed analysis groups to slide through in many cases. Moving away from analysis groups makes this problem much more obvious. To fix it, I've leveraged the flexibility the design of the new PM components provides to just directly construct the relevant alias analyses for the relevant functions in the IPO passes that need them. This is a bit hacky, but should go away with the new pass manager, and is already in many ways cleaner than the prior state. Another significant challenge is that various facilities of the old alias analysis infrastructure just don't fit any more. The most significant of these is the alias analysis 'counter' pass. That pass relied on the ability to snoop on AA queries at different points in the analysis group chain. Instead, I'm planning to build printing functionality directly into the aggregation layer. I've not included that in this patch merely to keep it smaller. Note that all of this needs a nearly complete rewrite of the AA documentation. I'm planning to do that, but I'd like to make sure the new design settles, and to flesh out a bit more of what it looks like in the new pass manager first. Differential Revision: http://reviews.llvm.org/D12080 llvm-svn: 247167
2015-09-09 19:55:00 +02:00
#include "llvm/Analysis/AssumptionCache.h"
#include "llvm/Analysis/BasicAliasAnalysis.h"
#include "llvm/Analysis/CallGraph.h"
#include "llvm/Analysis/CallGraphSCCPass.h"
#include "llvm/Analysis/CaptureTracking.h"
[PM/AA] Rebuild LLVM's alias analysis infrastructure in a way compatible with the new pass manager, and no longer relying on analysis groups. This builds essentially a ground-up new AA infrastructure stack for LLVM. The core ideas are the same that are used throughout the new pass manager: type erased polymorphism and direct composition. The design is as follows: - FunctionAAResults is a type-erasing alias analysis results aggregation interface to walk a single query across a range of results from different alias analyses. Currently this is function-specific as we always assume that aliasing queries are *within* a function. - AAResultBase is a CRTP utility providing stub implementations of various parts of the alias analysis result concept, notably in several cases in terms of other more general parts of the interface. This can be used to implement only a narrow part of the interface rather than the entire interface. This isn't really ideal, this logic should be hoisted into FunctionAAResults as currently it will cause a significant amount of redundant work, but it faithfully models the behavior of the prior infrastructure. - All the alias analysis passes are ported to be wrapper passes for the legacy PM and new-style analysis passes for the new PM with a shared result object. In some cases (most notably CFL), this is an extremely naive approach that we should revisit when we can specialize for the new pass manager. - BasicAA has been restructured to reflect that it is much more fundamentally a function analysis because it uses dominator trees and loop info that need to be constructed for each function. All of the references to getting alias analysis results have been updated to use the new aggregation interface. All the preservation and other pass management code has been updated accordingly. The way the FunctionAAResultsWrapperPass works is to detect the available alias analyses when run, and add them to the results object. This means that we should be able to continue to respect when various passes are added to the pipeline, for example adding CFL or adding TBAA passes should just cause their results to be available and to get folded into this. The exception to this rule is BasicAA which really needs to be a function pass due to using dominator trees and loop info. As a consequence, the FunctionAAResultsWrapperPass directly depends on BasicAA and always includes it in the aggregation. This has significant implications for preserving analyses. Generally, most passes shouldn't bother preserving FunctionAAResultsWrapperPass because rebuilding the results just updates the set of known AA passes. The exception to this rule are LoopPass instances which need to preserve all the function analyses that the loop pass manager will end up needing. This means preserving both BasicAAWrapperPass and the aggregating FunctionAAResultsWrapperPass. Now, when preserving an alias analysis, you do so by directly preserving that analysis. This is only necessary for non-immutable-pass-provided alias analyses though, and there are only three of interest: BasicAA, GlobalsAA (formerly GlobalsModRef), and SCEVAA. Usually BasicAA is preserved when needed because it (like DominatorTree and LoopInfo) is marked as a CFG-only pass. I've expanded GlobalsAA into the preserved set everywhere we previously were preserving all of AliasAnalysis, and I've added SCEVAA in the intersection of that with where we preserve SCEV itself. One significant challenge to all of this is that the CGSCC passes were actually using the alias analysis implementations by taking advantage of a pretty amazing set of loop holes in the old pass manager's analysis management code which allowed analysis groups to slide through in many cases. Moving away from analysis groups makes this problem much more obvious. To fix it, I've leveraged the flexibility the design of the new PM components provides to just directly construct the relevant alias analyses for the relevant functions in the IPO passes that need them. This is a bit hacky, but should go away with the new pass manager, and is already in many ways cleaner than the prior state. Another significant challenge is that various facilities of the old alias analysis infrastructure just don't fit any more. The most significant of these is the alias analysis 'counter' pass. That pass relied on the ability to snoop on AA queries at different points in the analysis group chain. Instead, I'm planning to build printing functionality directly into the aggregation layer. I've not included that in this patch merely to keep it smaller. Note that all of this needs a nearly complete rewrite of the AA documentation. I'm planning to do that, but I'd like to make sure the new design settles, and to flesh out a bit more of what it looks like in the new pass manager first. Differential Revision: http://reviews.llvm.org/D12080 llvm-svn: 247167
2015-09-09 19:55:00 +02:00
#include "llvm/Analysis/TargetLibraryInfo.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/IR/GlobalVariable.h"
#include "llvm/IR/InstIterator.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/LLVMContext.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Analysis/TargetLibraryInfo.h"
using namespace llvm;
#define DEBUG_TYPE "functionattrs"
STATISTIC(NumReadNone, "Number of functions marked readnone");
STATISTIC(NumReadOnly, "Number of functions marked readonly");
STATISTIC(NumNoCapture, "Number of arguments marked nocapture");
STATISTIC(NumReadNoneArg, "Number of arguments marked readnone");
STATISTIC(NumReadOnlyArg, "Number of arguments marked readonly");
STATISTIC(NumNoAlias, "Number of function returns marked noalias");
STATISTIC(NumNonNullReturn, "Number of function returns marked nonnull");
STATISTIC(NumNoRecurse, "Number of functions marked as norecurse");
namespace {
typedef SmallSetVector<Function *, 8> SCCNodeSet;
}
namespace {
/// The three kinds of memory access relevant to 'readonly' and
/// 'readnone' attributes.
enum MemoryAccessKind {
MAK_ReadNone = 0,
MAK_ReadOnly = 1,
MAK_MayWrite = 2
};
}
static MemoryAccessKind checkFunctionMemoryAccess(Function &F, AAResults &AAR,
const SCCNodeSet &SCCNodes) {
FunctionModRefBehavior MRB = AAR.getModRefBehavior(&F);
if (MRB == FMRB_DoesNotAccessMemory)
// Already perfect!
return MAK_ReadNone;
Don't IPO over functions that can be de-refined Summary: Fixes PR26774. If you're aware of the issue, feel free to skip the "Motivation" section and jump directly to "This patch". Motivation: I define "refinement" as discarding behaviors from a program that the optimizer has license to discard. So transforming: ``` void f(unsigned x) { unsigned t = 5 / x; (void)t; } ``` to ``` void f(unsigned x) { } ``` is refinement, since the behavior went from "if x == 0 then undefined else nothing" to "nothing" (the optimizer has license to discard undefined behavior). Refinement is a fundamental aspect of many mid-level optimizations done by LLVM. For instance, transforming `x == (x + 1)` to `false` also involves refinement since the expression's value went from "if x is `undef` then { `true` or `false` } else { `false` }" to "`false`" (by definition, the optimizer has license to fold `undef` to any non-`undef` value). Unfortunately, refinement implies that the optimizer cannot assume that the implementation of a function it can see has all of the behavior an unoptimized or a differently optimized version of the same function can have. This is a problem for functions with comdat linkage, where a function can be replaced by an unoptimized or a differently optimized version of the same source level function. For instance, FunctionAttrs cannot assume a comdat function is actually `readnone` even if it does not have any loads or stores in it; since there may have been loads and stores in the "original function" that were refined out in the currently visible variant, and at the link step the linker may in fact choose an implementation with a load or a store. As an example, consider a function that does two atomic loads from the same memory location, and writes to memory only if the two values are not equal. The optimizer is allowed to refine this function by first CSE'ing the two loads, and the folding the comparision to always report that the two values are equal. Such a refined variant will look like it is `readonly`. However, the unoptimized version of the function can still write to memory (since the two loads //can// result in different values), and selecting the unoptimized version at link time will retroactively invalidate transforms we may have done under the assumption that the function does not write to memory. Note: this is not just a problem with atomics or with linking differently optimized object files. See PR26774 for more realistic examples that involved neither. This patch: This change introduces a new set of linkage types, predicated as `GlobalValue::mayBeDerefined` that returns true if the linkage type allows a function to be replaced by a differently optimized variant at link time. It then changes a set of IPO passes to bail out if they see such a function. Reviewers: chandlerc, hfinkel, dexonsmith, joker.eph, rnk Subscribers: mcrosier, llvm-commits Differential Revision: http://reviews.llvm.org/D18634 llvm-svn: 265762
2016-04-08 02:48:30 +02:00
// Non-exact function definitions may not be selected at link time, and an
// alternative version that writes to memory may be selected. See the comment
// on GlobalValue::isDefinitionExact for more details.
if (!F.hasExactDefinition()) {
if (AliasAnalysis::onlyReadsMemory(MRB))
return MAK_ReadOnly;
// Conservatively assume it writes to memory.
return MAK_MayWrite;
}
// Scan the function body for instructions that may read or write memory.
bool ReadsMemory = false;
for (inst_iterator II = inst_begin(F), E = inst_end(F); II != E; ++II) {
Instruction *I = &*II;
// Some instructions can be ignored even if they read or write memory.
// Detect these now, skipping to the next instruction if one is found.
CallSite CS(cast<Value>(I));
if (CS) {
// Ignore calls to functions in the same SCC, as long as the call sites
// don't have operand bundles. Calls with operand bundles are allowed to
// have memory effects not described by the memory effects of the call
// target.
if (!CS.hasOperandBundles() && CS.getCalledFunction() &&
SCCNodes.count(CS.getCalledFunction()))
continue;
FunctionModRefBehavior MRB = AAR.getModRefBehavior(CS);
// If the call doesn't access memory, we're done.
if (!(MRB & MRI_ModRef))
continue;
if (!AliasAnalysis::onlyAccessesArgPointees(MRB)) {
// The call could access any memory. If that includes writes, give up.
if (MRB & MRI_Mod)
return MAK_MayWrite;
// If it reads, note it.
if (MRB & MRI_Ref)
ReadsMemory = true;
continue;
}
// Check whether all pointer arguments point to local memory, and
// ignore calls that only access local memory.
for (CallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
CI != CE; ++CI) {
Value *Arg = *CI;
if (!Arg->getType()->isPtrOrPtrVectorTy())
continue;
AAMDNodes AAInfo;
I->getAAMetadata(AAInfo);
MemoryLocation Loc(Arg, MemoryLocation::UnknownSize, AAInfo);
// Skip accesses to local or constant memory as they don't impact the
// externally visible mod/ref behavior.
if (AAR.pointsToConstantMemory(Loc, /*OrLocal=*/true))
continue;
if (MRB & MRI_Mod)
// Writes non-local memory. Give up.
return MAK_MayWrite;
if (MRB & MRI_Ref)
// Ok, it reads non-local memory.
ReadsMemory = true;
}
continue;
} else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
// Ignore non-volatile loads from local memory. (Atomic is okay here.)
if (!LI->isVolatile()) {
MemoryLocation Loc = MemoryLocation::get(LI);
if (AAR.pointsToConstantMemory(Loc, /*OrLocal=*/true))
continue;
}
} else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
// Ignore non-volatile stores to local memory. (Atomic is okay here.)
if (!SI->isVolatile()) {
MemoryLocation Loc = MemoryLocation::get(SI);
if (AAR.pointsToConstantMemory(Loc, /*OrLocal=*/true))
continue;
}
} else if (VAArgInst *VI = dyn_cast<VAArgInst>(I)) {
// Ignore vaargs on local memory.
MemoryLocation Loc = MemoryLocation::get(VI);
if (AAR.pointsToConstantMemory(Loc, /*OrLocal=*/true))
continue;
}
// Any remaining instructions need to be taken seriously! Check if they
// read or write memory.
if (I->mayWriteToMemory())
// Writes memory. Just give up.
return MAK_MayWrite;
// If this instruction may read memory, remember that.
ReadsMemory |= I->mayReadFromMemory();
}
return ReadsMemory ? MAK_ReadOnly : MAK_ReadNone;
}
/// Deduce readonly/readnone attributes for the SCC.
template <typename AARGetterT>
static bool addReadAttrs(const SCCNodeSet &SCCNodes, AARGetterT AARGetter) {
2008-09-29 15:35:31 +02:00
// Check if any of the functions in the SCC read or write memory. If they
// write memory then they can't be marked readnone or readonly.
bool ReadsMemory = false;
for (Function *F : SCCNodes) {
// Call the callable parameter to look up AA results for this function.
AAResults &AAR = AARGetter(*F);
[PM/AA] Rebuild LLVM's alias analysis infrastructure in a way compatible with the new pass manager, and no longer relying on analysis groups. This builds essentially a ground-up new AA infrastructure stack for LLVM. The core ideas are the same that are used throughout the new pass manager: type erased polymorphism and direct composition. The design is as follows: - FunctionAAResults is a type-erasing alias analysis results aggregation interface to walk a single query across a range of results from different alias analyses. Currently this is function-specific as we always assume that aliasing queries are *within* a function. - AAResultBase is a CRTP utility providing stub implementations of various parts of the alias analysis result concept, notably in several cases in terms of other more general parts of the interface. This can be used to implement only a narrow part of the interface rather than the entire interface. This isn't really ideal, this logic should be hoisted into FunctionAAResults as currently it will cause a significant amount of redundant work, but it faithfully models the behavior of the prior infrastructure. - All the alias analysis passes are ported to be wrapper passes for the legacy PM and new-style analysis passes for the new PM with a shared result object. In some cases (most notably CFL), this is an extremely naive approach that we should revisit when we can specialize for the new pass manager. - BasicAA has been restructured to reflect that it is much more fundamentally a function analysis because it uses dominator trees and loop info that need to be constructed for each function. All of the references to getting alias analysis results have been updated to use the new aggregation interface. All the preservation and other pass management code has been updated accordingly. The way the FunctionAAResultsWrapperPass works is to detect the available alias analyses when run, and add them to the results object. This means that we should be able to continue to respect when various passes are added to the pipeline, for example adding CFL or adding TBAA passes should just cause their results to be available and to get folded into this. The exception to this rule is BasicAA which really needs to be a function pass due to using dominator trees and loop info. As a consequence, the FunctionAAResultsWrapperPass directly depends on BasicAA and always includes it in the aggregation. This has significant implications for preserving analyses. Generally, most passes shouldn't bother preserving FunctionAAResultsWrapperPass because rebuilding the results just updates the set of known AA passes. The exception to this rule are LoopPass instances which need to preserve all the function analyses that the loop pass manager will end up needing. This means preserving both BasicAAWrapperPass and the aggregating FunctionAAResultsWrapperPass. Now, when preserving an alias analysis, you do so by directly preserving that analysis. This is only necessary for non-immutable-pass-provided alias analyses though, and there are only three of interest: BasicAA, GlobalsAA (formerly GlobalsModRef), and SCEVAA. Usually BasicAA is preserved when needed because it (like DominatorTree and LoopInfo) is marked as a CFG-only pass. I've expanded GlobalsAA into the preserved set everywhere we previously were preserving all of AliasAnalysis, and I've added SCEVAA in the intersection of that with where we preserve SCEV itself. One significant challenge to all of this is that the CGSCC passes were actually using the alias analysis implementations by taking advantage of a pretty amazing set of loop holes in the old pass manager's analysis management code which allowed analysis groups to slide through in many cases. Moving away from analysis groups makes this problem much more obvious. To fix it, I've leveraged the flexibility the design of the new PM components provides to just directly construct the relevant alias analyses for the relevant functions in the IPO passes that need them. This is a bit hacky, but should go away with the new pass manager, and is already in many ways cleaner than the prior state. Another significant challenge is that various facilities of the old alias analysis infrastructure just don't fit any more. The most significant of these is the alias analysis 'counter' pass. That pass relied on the ability to snoop on AA queries at different points in the analysis group chain. Instead, I'm planning to build printing functionality directly into the aggregation layer. I've not included that in this patch merely to keep it smaller. Note that all of this needs a nearly complete rewrite of the AA documentation. I'm planning to do that, but I'd like to make sure the new design settles, and to flesh out a bit more of what it looks like in the new pass manager first. Differential Revision: http://reviews.llvm.org/D12080 llvm-svn: 247167
2015-09-09 19:55:00 +02:00
switch (checkFunctionMemoryAccess(*F, AAR, SCCNodes)) {
case MAK_MayWrite:
return false;
case MAK_ReadOnly:
ReadsMemory = true;
break;
case MAK_ReadNone:
// Nothing to do!
break;
}
}
// Success! Functions in this SCC do not access memory, or only read memory.
// Give them the appropriate attribute.
bool MadeChange = false;
for (Function *F : SCCNodes) {
if (F->doesNotAccessMemory())
// Already perfect!
continue;
if (F->onlyReadsMemory() && ReadsMemory)
// No change.
continue;
MadeChange = true;
// Clear out any existing attributes.
AttrBuilder B;
B.addAttribute(Attribute::ReadOnly).addAttribute(Attribute::ReadNone);
F->removeAttributes(
AttributeSet::FunctionIndex,
AttributeSet::get(F->getContext(), AttributeSet::FunctionIndex, B));
// Add in the new attribute.
F->addAttribute(AttributeSet::FunctionIndex,
ReadsMemory ? Attribute::ReadOnly : Attribute::ReadNone);
if (ReadsMemory)
++NumReadOnly;
else
++NumReadNone;
}
return MadeChange;
}
namespace {
/// For a given pointer Argument, this retains a list of Arguments of functions
/// in the same SCC that the pointer data flows into. We use this to build an
/// SCC of the arguments.
struct ArgumentGraphNode {
Argument *Definition;
SmallVector<ArgumentGraphNode *, 4> Uses;
};
class ArgumentGraph {
// We store pointers to ArgumentGraphNode objects, so it's important that
// that they not move around upon insert.
typedef std::map<Argument *, ArgumentGraphNode> ArgumentMapTy;
ArgumentMapTy ArgumentMap;
// There is no root node for the argument graph, in fact:
// void f(int *x, int *y) { if (...) f(x, y); }
// is an example where the graph is disconnected. The SCCIterator requires a
// single entry point, so we maintain a fake ("synthetic") root node that
// uses every node. Because the graph is directed and nothing points into
// the root, it will not participate in any SCCs (except for its own).
ArgumentGraphNode SyntheticRoot;
public:
ArgumentGraph() { SyntheticRoot.Definition = nullptr; }
typedef SmallVectorImpl<ArgumentGraphNode *>::iterator iterator;
iterator begin() { return SyntheticRoot.Uses.begin(); }
iterator end() { return SyntheticRoot.Uses.end(); }
ArgumentGraphNode *getEntryNode() { return &SyntheticRoot; }
ArgumentGraphNode *operator[](Argument *A) {
ArgumentGraphNode &Node = ArgumentMap[A];
Node.Definition = A;
SyntheticRoot.Uses.push_back(&Node);
return &Node;
}
};
/// This tracker checks whether callees are in the SCC, and if so it does not
/// consider that a capture, instead adding it to the "Uses" list and
/// continuing with the analysis.
struct ArgumentUsesTracker : public CaptureTracker {
ArgumentUsesTracker(const SCCNodeSet &SCCNodes)
: Captured(false), SCCNodes(SCCNodes) {}
void tooManyUses() override { Captured = true; }
bool captured(const Use *U) override {
CallSite CS(U->getUser());
if (!CS.getInstruction()) {
Captured = true;
return true;
}
Function *F = CS.getCalledFunction();
Don't IPO over functions that can be de-refined Summary: Fixes PR26774. If you're aware of the issue, feel free to skip the "Motivation" section and jump directly to "This patch". Motivation: I define "refinement" as discarding behaviors from a program that the optimizer has license to discard. So transforming: ``` void f(unsigned x) { unsigned t = 5 / x; (void)t; } ``` to ``` void f(unsigned x) { } ``` is refinement, since the behavior went from "if x == 0 then undefined else nothing" to "nothing" (the optimizer has license to discard undefined behavior). Refinement is a fundamental aspect of many mid-level optimizations done by LLVM. For instance, transforming `x == (x + 1)` to `false` also involves refinement since the expression's value went from "if x is `undef` then { `true` or `false` } else { `false` }" to "`false`" (by definition, the optimizer has license to fold `undef` to any non-`undef` value). Unfortunately, refinement implies that the optimizer cannot assume that the implementation of a function it can see has all of the behavior an unoptimized or a differently optimized version of the same function can have. This is a problem for functions with comdat linkage, where a function can be replaced by an unoptimized or a differently optimized version of the same source level function. For instance, FunctionAttrs cannot assume a comdat function is actually `readnone` even if it does not have any loads or stores in it; since there may have been loads and stores in the "original function" that were refined out in the currently visible variant, and at the link step the linker may in fact choose an implementation with a load or a store. As an example, consider a function that does two atomic loads from the same memory location, and writes to memory only if the two values are not equal. The optimizer is allowed to refine this function by first CSE'ing the two loads, and the folding the comparision to always report that the two values are equal. Such a refined variant will look like it is `readonly`. However, the unoptimized version of the function can still write to memory (since the two loads //can// result in different values), and selecting the unoptimized version at link time will retroactively invalidate transforms we may have done under the assumption that the function does not write to memory. Note: this is not just a problem with atomics or with linking differently optimized object files. See PR26774 for more realistic examples that involved neither. This patch: This change introduces a new set of linkage types, predicated as `GlobalValue::mayBeDerefined` that returns true if the linkage type allows a function to be replaced by a differently optimized variant at link time. It then changes a set of IPO passes to bail out if they see such a function. Reviewers: chandlerc, hfinkel, dexonsmith, joker.eph, rnk Subscribers: mcrosier, llvm-commits Differential Revision: http://reviews.llvm.org/D18634 llvm-svn: 265762
2016-04-08 02:48:30 +02:00
if (!F || !F->hasExactDefinition() || !SCCNodes.count(F)) {
Captured = true;
return true;
}
// Note: the callee and the two successor blocks *follow* the argument
// operands. This means there is no need to adjust UseIndex to account for
// these.
unsigned UseIndex =
std::distance(const_cast<const Use *>(CS.arg_begin()), U);
assert(UseIndex < CS.data_operands_size() &&
"Indirect function calls should have been filtered above!");
if (UseIndex >= CS.getNumArgOperands()) {
// Data operand, but not a argument operand -- must be a bundle operand
assert(CS.hasOperandBundles() && "Must be!");
// CaptureTracking told us that we're being captured by an operand bundle
// use. In this case it does not matter if the callee is within our SCC
// or not -- we've been captured in some unknown way, and we have to be
// conservative.
Captured = true;
return true;
}
if (UseIndex >= F->arg_size()) {
assert(F->isVarArg() && "More params than args in non-varargs call");
Captured = true;
return true;
}
Uses.push_back(&*std::next(F->arg_begin(), UseIndex));
return false;
}
bool Captured; // True only if certainly captured (used outside our SCC).
SmallVector<Argument *, 4> Uses; // Uses within our SCC.
const SCCNodeSet &SCCNodes;
};
}
namespace llvm {
template <> struct GraphTraits<ArgumentGraphNode *> {
typedef ArgumentGraphNode NodeType;
typedef SmallVectorImpl<ArgumentGraphNode *>::iterator ChildIteratorType;
static inline NodeType *getEntryNode(NodeType *A) { return A; }
static inline ChildIteratorType child_begin(NodeType *N) {
return N->Uses.begin();
}
static inline ChildIteratorType child_end(NodeType *N) {
return N->Uses.end();
}
};
template <>
struct GraphTraits<ArgumentGraph *> : public GraphTraits<ArgumentGraphNode *> {
static NodeType *getEntryNode(ArgumentGraph *AG) {
return AG->getEntryNode();
}
static ChildIteratorType nodes_begin(ArgumentGraph *AG) {
return AG->begin();
}
static ChildIteratorType nodes_end(ArgumentGraph *AG) { return AG->end(); }
};
}
/// Returns Attribute::None, Attribute::ReadOnly or Attribute::ReadNone.
static Attribute::AttrKind
determinePointerReadAttrs(Argument *A,
const SmallPtrSet<Argument *, 8> &SCCNodes) {
SmallVector<Use *, 32> Worklist;
SmallSet<Use *, 32> Visited;
// inalloca arguments are always clobbered by the call.
if (A->hasInAllocaAttr())
return Attribute::None;
bool IsRead = false;
// We don't need to track IsWritten. If A is written to, return immediately.
[C++11] Add range based accessors for the Use-Def chain of a Value. This requires a number of steps. 1) Move value_use_iterator into the Value class as an implementation detail 2) Change it to actually be a *Use* iterator rather than a *User* iterator. 3) Add an adaptor which is a User iterator that always looks through the Use to the User. 4) Wrap these in Value::use_iterator and Value::user_iterator typedefs. 5) Add the range adaptors as Value::uses() and Value::users(). 6) Update *all* of the callers to correctly distinguish between whether they wanted a use_iterator (and to explicitly dig out the User when needed), or a user_iterator which makes the Use itself totally opaque. Because #6 requires churning essentially everything that walked the Use-Def chains, I went ahead and added all of the range adaptors and switched them to range-based loops where appropriate. Also because the renaming requires at least churning every line of code, it didn't make any sense to split these up into multiple commits -- all of which would touch all of the same lies of code. The result is still not quite optimal. The Value::use_iterator is a nice regular iterator, but Value::user_iterator is an iterator over User*s rather than over the User objects themselves. As a consequence, it fits a bit awkwardly into the range-based world and it has the weird extra-dereferencing 'operator->' that so many of our iterators have. I think this could be fixed by providing something which transforms a range of T&s into a range of T*s, but that *can* be separated into another patch, and it isn't yet 100% clear whether this is the right move. However, this change gets us most of the benefit and cleans up a substantial amount of code around Use and User. =] llvm-svn: 203364
2014-03-09 04:16:01 +01:00
for (Use &U : A->uses()) {
Visited.insert(&U);
Worklist.push_back(&U);
}
while (!Worklist.empty()) {
Use *U = Worklist.pop_back_val();
Instruction *I = cast<Instruction>(U->getUser());
switch (I->getOpcode()) {
case Instruction::BitCast:
case Instruction::GetElementPtr:
case Instruction::PHI:
case Instruction::Select:
case Instruction::AddrSpaceCast:
// The original value is not read/written via this if the new value isn't.
[C++11] Add range based accessors for the Use-Def chain of a Value. This requires a number of steps. 1) Move value_use_iterator into the Value class as an implementation detail 2) Change it to actually be a *Use* iterator rather than a *User* iterator. 3) Add an adaptor which is a User iterator that always looks through the Use to the User. 4) Wrap these in Value::use_iterator and Value::user_iterator typedefs. 5) Add the range adaptors as Value::uses() and Value::users(). 6) Update *all* of the callers to correctly distinguish between whether they wanted a use_iterator (and to explicitly dig out the User when needed), or a user_iterator which makes the Use itself totally opaque. Because #6 requires churning essentially everything that walked the Use-Def chains, I went ahead and added all of the range adaptors and switched them to range-based loops where appropriate. Also because the renaming requires at least churning every line of code, it didn't make any sense to split these up into multiple commits -- all of which would touch all of the same lies of code. The result is still not quite optimal. The Value::use_iterator is a nice regular iterator, but Value::user_iterator is an iterator over User*s rather than over the User objects themselves. As a consequence, it fits a bit awkwardly into the range-based world and it has the weird extra-dereferencing 'operator->' that so many of our iterators have. I think this could be fixed by providing something which transforms a range of T&s into a range of T*s, but that *can* be separated into another patch, and it isn't yet 100% clear whether this is the right move. However, this change gets us most of the benefit and cleans up a substantial amount of code around Use and User. =] llvm-svn: 203364
2014-03-09 04:16:01 +01:00
for (Use &UU : I->uses())
if (Visited.insert(&UU).second)
[C++11] Add range based accessors for the Use-Def chain of a Value. This requires a number of steps. 1) Move value_use_iterator into the Value class as an implementation detail 2) Change it to actually be a *Use* iterator rather than a *User* iterator. 3) Add an adaptor which is a User iterator that always looks through the Use to the User. 4) Wrap these in Value::use_iterator and Value::user_iterator typedefs. 5) Add the range adaptors as Value::uses() and Value::users(). 6) Update *all* of the callers to correctly distinguish between whether they wanted a use_iterator (and to explicitly dig out the User when needed), or a user_iterator which makes the Use itself totally opaque. Because #6 requires churning essentially everything that walked the Use-Def chains, I went ahead and added all of the range adaptors and switched them to range-based loops where appropriate. Also because the renaming requires at least churning every line of code, it didn't make any sense to split these up into multiple commits -- all of which would touch all of the same lies of code. The result is still not quite optimal. The Value::use_iterator is a nice regular iterator, but Value::user_iterator is an iterator over User*s rather than over the User objects themselves. As a consequence, it fits a bit awkwardly into the range-based world and it has the weird extra-dereferencing 'operator->' that so many of our iterators have. I think this could be fixed by providing something which transforms a range of T&s into a range of T*s, but that *can* be separated into another patch, and it isn't yet 100% clear whether this is the right move. However, this change gets us most of the benefit and cleans up a substantial amount of code around Use and User. =] llvm-svn: 203364
2014-03-09 04:16:01 +01:00
Worklist.push_back(&UU);
break;
case Instruction::Call:
case Instruction::Invoke: {
bool Captures = true;
if (I->getType()->isVoidTy())
Captures = false;
auto AddUsersToWorklistIfCapturing = [&] {
if (Captures)
for (Use &UU : I->uses())
if (Visited.insert(&UU).second)
Worklist.push_back(&UU);
};
CallSite CS(I);
if (CS.doesNotAccessMemory()) {
AddUsersToWorklistIfCapturing();
continue;
}
Function *F = CS.getCalledFunction();
if (!F) {
if (CS.onlyReadsMemory()) {
IsRead = true;
AddUsersToWorklistIfCapturing();
continue;
}
return Attribute::None;
}
// Note: the callee and the two successor blocks *follow* the argument
// operands. This means there is no need to adjust UseIndex to account
// for these.
unsigned UseIndex = std::distance(CS.arg_begin(), U);
// U cannot be the callee operand use: since we're exploring the
// transitive uses of an Argument, having such a use be a callee would
// imply the CallSite is an indirect call or invoke; and we'd take the
// early exit above.
assert(UseIndex < CS.data_operands_size() &&
"Data operand use expected!");
bool IsOperandBundleUse = UseIndex >= CS.getNumArgOperands();
if (UseIndex >= F->arg_size() && !IsOperandBundleUse) {
assert(F->isVarArg() && "More params than args in non-varargs call");
return Attribute::None;
}
Captures &= !CS.doesNotCapture(UseIndex);
// Since the optimizer (by design) cannot see the data flow corresponding
// to a operand bundle use, these cannot participate in the optimistic SCC
// analysis. Instead, we model the operand bundle uses as arguments in
// call to a function external to the SCC.
if (!SCCNodes.count(&*std::next(F->arg_begin(), UseIndex)) ||
IsOperandBundleUse) {
// The accessors used on CallSite here do the right thing for calls and
// invokes with operand bundles.
if (!CS.onlyReadsMemory() && !CS.onlyReadsMemory(UseIndex))
return Attribute::None;
if (!CS.doesNotAccessMemory(UseIndex))
IsRead = true;
}
AddUsersToWorklistIfCapturing();
break;
}
case Instruction::Load:
IsRead = true;
break;
case Instruction::ICmp:
case Instruction::Ret:
break;
default:
return Attribute::None;
}
}
return IsRead ? Attribute::ReadOnly : Attribute::ReadNone;
}
/// Deduce nocapture attributes for the SCC.
static bool addArgumentAttrs(const SCCNodeSet &SCCNodes) {
bool Changed = false;
ArgumentGraph AG;
AttrBuilder B;
B.addAttribute(Attribute::NoCapture);
// Check each function in turn, determining which pointer arguments are not
// captured.
for (Function *F : SCCNodes) {
Don't IPO over functions that can be de-refined Summary: Fixes PR26774. If you're aware of the issue, feel free to skip the "Motivation" section and jump directly to "This patch". Motivation: I define "refinement" as discarding behaviors from a program that the optimizer has license to discard. So transforming: ``` void f(unsigned x) { unsigned t = 5 / x; (void)t; } ``` to ``` void f(unsigned x) { } ``` is refinement, since the behavior went from "if x == 0 then undefined else nothing" to "nothing" (the optimizer has license to discard undefined behavior). Refinement is a fundamental aspect of many mid-level optimizations done by LLVM. For instance, transforming `x == (x + 1)` to `false` also involves refinement since the expression's value went from "if x is `undef` then { `true` or `false` } else { `false` }" to "`false`" (by definition, the optimizer has license to fold `undef` to any non-`undef` value). Unfortunately, refinement implies that the optimizer cannot assume that the implementation of a function it can see has all of the behavior an unoptimized or a differently optimized version of the same function can have. This is a problem for functions with comdat linkage, where a function can be replaced by an unoptimized or a differently optimized version of the same source level function. For instance, FunctionAttrs cannot assume a comdat function is actually `readnone` even if it does not have any loads or stores in it; since there may have been loads and stores in the "original function" that were refined out in the currently visible variant, and at the link step the linker may in fact choose an implementation with a load or a store. As an example, consider a function that does two atomic loads from the same memory location, and writes to memory only if the two values are not equal. The optimizer is allowed to refine this function by first CSE'ing the two loads, and the folding the comparision to always report that the two values are equal. Such a refined variant will look like it is `readonly`. However, the unoptimized version of the function can still write to memory (since the two loads //can// result in different values), and selecting the unoptimized version at link time will retroactively invalidate transforms we may have done under the assumption that the function does not write to memory. Note: this is not just a problem with atomics or with linking differently optimized object files. See PR26774 for more realistic examples that involved neither. This patch: This change introduces a new set of linkage types, predicated as `GlobalValue::mayBeDerefined` that returns true if the linkage type allows a function to be replaced by a differently optimized variant at link time. It then changes a set of IPO passes to bail out if they see such a function. Reviewers: chandlerc, hfinkel, dexonsmith, joker.eph, rnk Subscribers: mcrosier, llvm-commits Differential Revision: http://reviews.llvm.org/D18634 llvm-svn: 265762
2016-04-08 02:48:30 +02:00
// We can infer and propagate function attributes only when we know that the
// definition we'll get at link time is *exactly* the definition we see now.
// For more details, see GlobalValue::mayBeDerefined.
if (!F->hasExactDefinition())
continue;
// Functions that are readonly (or readnone) and nounwind and don't return
// a value can't capture arguments. Don't analyze them.
if (F->onlyReadsMemory() && F->doesNotThrow() &&
F->getReturnType()->isVoidTy()) {
for (Function::arg_iterator A = F->arg_begin(), E = F->arg_end(); A != E;
++A) {
if (A->getType()->isPointerTy() && !A->hasNoCaptureAttr()) {
A->addAttr(AttributeSet::get(F->getContext(), A->getArgNo() + 1, B));
++NumNoCapture;
Changed = true;
}
}
continue;
}
for (Function::arg_iterator A = F->arg_begin(), E = F->arg_end(); A != E;
++A) {
if (!A->getType()->isPointerTy())
continue;
bool HasNonLocalUses = false;
if (!A->hasNoCaptureAttr()) {
ArgumentUsesTracker Tracker(SCCNodes);
PointerMayBeCaptured(&*A, &Tracker);
if (!Tracker.Captured) {
if (Tracker.Uses.empty()) {
// If it's trivially not captured, mark it nocapture now.
A->addAttr(
AttributeSet::get(F->getContext(), A->getArgNo() + 1, B));
++NumNoCapture;
Changed = true;
} else {
// If it's not trivially captured and not trivially not captured,
// then it must be calling into another function in our SCC. Save
// its particulars for Argument-SCC analysis later.
ArgumentGraphNode *Node = AG[&*A];
for (SmallVectorImpl<Argument *>::iterator
UI = Tracker.Uses.begin(),
UE = Tracker.Uses.end();
UI != UE; ++UI) {
Node->Uses.push_back(AG[*UI]);
if (*UI != &*A)
HasNonLocalUses = true;
}
}
}
// Otherwise, it's captured. Don't bother doing SCC analysis on it.
}
if (!HasNonLocalUses && !A->onlyReadsMemory()) {
// Can we determine that it's readonly/readnone without doing an SCC?
// Note that we don't allow any calls at all here, or else our result
// will be dependent on the iteration order through the functions in the
// SCC.
SmallPtrSet<Argument *, 8> Self;
Self.insert(&*A);
Attribute::AttrKind R = determinePointerReadAttrs(&*A, Self);
if (R != Attribute::None) {
AttrBuilder B;
B.addAttribute(R);
A->addAttr(AttributeSet::get(A->getContext(), A->getArgNo() + 1, B));
Changed = true;
R == Attribute::ReadOnly ? ++NumReadOnlyArg : ++NumReadNoneArg;
}
}
}
}
// The graph we've collected is partial because we stopped scanning for
// argument uses once we solved the argument trivially. These partial nodes
// show up as ArgumentGraphNode objects with an empty Uses list, and for
// these nodes the final decision about whether they capture has already been
// made. If the definition doesn't have a 'nocapture' attribute by now, it
// captures.
for (scc_iterator<ArgumentGraph *> I = scc_begin(&AG); !I.isAtEnd(); ++I) {
const std::vector<ArgumentGraphNode *> &ArgumentSCC = *I;
if (ArgumentSCC.size() == 1) {
if (!ArgumentSCC[0]->Definition)
continue; // synthetic root node
// eg. "void f(int* x) { if (...) f(x); }"
if (ArgumentSCC[0]->Uses.size() == 1 &&
ArgumentSCC[0]->Uses[0] == ArgumentSCC[0]) {
Argument *A = ArgumentSCC[0]->Definition;
A->addAttr(AttributeSet::get(A->getContext(), A->getArgNo() + 1, B));
++NumNoCapture;
Changed = true;
}
continue;
}
bool SCCCaptured = false;
for (auto I = ArgumentSCC.begin(), E = ArgumentSCC.end();
I != E && !SCCCaptured; ++I) {
ArgumentGraphNode *Node = *I;
if (Node->Uses.empty()) {
if (!Node->Definition->hasNoCaptureAttr())
SCCCaptured = true;
}
}
if (SCCCaptured)
continue;
SmallPtrSet<Argument *, 8> ArgumentSCCNodes;
// Fill ArgumentSCCNodes with the elements of the ArgumentSCC. Used for
// quickly looking up whether a given Argument is in this ArgumentSCC.
for (auto I = ArgumentSCC.begin(), E = ArgumentSCC.end(); I != E; ++I) {
ArgumentSCCNodes.insert((*I)->Definition);
}
for (auto I = ArgumentSCC.begin(), E = ArgumentSCC.end();
I != E && !SCCCaptured; ++I) {
ArgumentGraphNode *N = *I;
for (SmallVectorImpl<ArgumentGraphNode *>::iterator UI = N->Uses.begin(),
UE = N->Uses.end();
UI != UE; ++UI) {
Argument *A = (*UI)->Definition;
if (A->hasNoCaptureAttr() || ArgumentSCCNodes.count(A))
continue;
SCCCaptured = true;
break;
}
}
if (SCCCaptured)
continue;
for (unsigned i = 0, e = ArgumentSCC.size(); i != e; ++i) {
Argument *A = ArgumentSCC[i]->Definition;
A->addAttr(AttributeSet::get(A->getContext(), A->getArgNo() + 1, B));
++NumNoCapture;
Changed = true;
}
// We also want to compute readonly/readnone. With a small number of false
// negatives, we can assume that any pointer which is captured isn't going
// to be provably readonly or readnone, since by definition we can't
// analyze all uses of a captured pointer.
//
// The false negatives happen when the pointer is captured by a function
// that promises readonly/readnone behaviour on the pointer, then the
// pointer's lifetime ends before anything that writes to arbitrary memory.
// Also, a readonly/readnone pointer may be returned, but returning a
// pointer is capturing it.
Attribute::AttrKind ReadAttr = Attribute::ReadNone;
for (unsigned i = 0, e = ArgumentSCC.size(); i != e; ++i) {
Argument *A = ArgumentSCC[i]->Definition;
Attribute::AttrKind K = determinePointerReadAttrs(A, ArgumentSCCNodes);
if (K == Attribute::ReadNone)
continue;
if (K == Attribute::ReadOnly) {
ReadAttr = Attribute::ReadOnly;
continue;
}
ReadAttr = K;
break;
}
if (ReadAttr != Attribute::None) {
AttrBuilder B, R;
B.addAttribute(ReadAttr);
R.addAttribute(Attribute::ReadOnly).addAttribute(Attribute::ReadNone);
for (unsigned i = 0, e = ArgumentSCC.size(); i != e; ++i) {
Argument *A = ArgumentSCC[i]->Definition;
// Clear out existing readonly/readnone attributes
A->removeAttr(AttributeSet::get(A->getContext(), A->getArgNo() + 1, R));
A->addAttr(AttributeSet::get(A->getContext(), A->getArgNo() + 1, B));
ReadAttr == Attribute::ReadOnly ? ++NumReadOnlyArg : ++NumReadNoneArg;
Changed = true;
}
}
}
return Changed;
}
/// Tests whether a function is "malloc-like".
///
/// A function is "malloc-like" if it returns either null or a pointer that
/// doesn't alias any other pointer visible to the caller.
static bool isFunctionMallocLike(Function *F, const SCCNodeSet &SCCNodes) {
SmallSetVector<Value *, 8> FlowsToReturn;
for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
if (ReturnInst *Ret = dyn_cast<ReturnInst>(I->getTerminator()))
FlowsToReturn.insert(Ret->getReturnValue());
for (unsigned i = 0; i != FlowsToReturn.size(); ++i) {
Value *RetVal = FlowsToReturn[i];
if (Constant *C = dyn_cast<Constant>(RetVal)) {
if (!C->isNullValue() && !isa<UndefValue>(C))
return false;
continue;
}
if (isa<Argument>(RetVal))
return false;
if (Instruction *RVI = dyn_cast<Instruction>(RetVal))
switch (RVI->getOpcode()) {
// Extend the analysis by looking upwards.
case Instruction::BitCast:
case Instruction::GetElementPtr:
case Instruction::AddrSpaceCast:
FlowsToReturn.insert(RVI->getOperand(0));
continue;
case Instruction::Select: {
SelectInst *SI = cast<SelectInst>(RVI);
FlowsToReturn.insert(SI->getTrueValue());
FlowsToReturn.insert(SI->getFalseValue());
continue;
}
case Instruction::PHI: {
PHINode *PN = cast<PHINode>(RVI);
for (Value *IncValue : PN->incoming_values())
FlowsToReturn.insert(IncValue);
continue;
}
// Check whether the pointer came from an allocation.
case Instruction::Alloca:
break;
case Instruction::Call:
case Instruction::Invoke: {
CallSite CS(RVI);
if (CS.paramHasAttr(0, Attribute::NoAlias))
break;
if (CS.getCalledFunction() && SCCNodes.count(CS.getCalledFunction()))
break;
} // fall-through
default:
return false; // Did not come from an allocation.
}
if (PointerMayBeCaptured(RetVal, false, /*StoreCaptures=*/false))
return false;
}
return true;
}
/// Deduce noalias attributes for the SCC.
static bool addNoAliasAttrs(const SCCNodeSet &SCCNodes) {
// Check each function in turn, determining which functions return noalias
// pointers.
for (Function *F : SCCNodes) {
// Already noalias.
if (F->doesNotAlias(0))
continue;
Don't IPO over functions that can be de-refined Summary: Fixes PR26774. If you're aware of the issue, feel free to skip the "Motivation" section and jump directly to "This patch". Motivation: I define "refinement" as discarding behaviors from a program that the optimizer has license to discard. So transforming: ``` void f(unsigned x) { unsigned t = 5 / x; (void)t; } ``` to ``` void f(unsigned x) { } ``` is refinement, since the behavior went from "if x == 0 then undefined else nothing" to "nothing" (the optimizer has license to discard undefined behavior). Refinement is a fundamental aspect of many mid-level optimizations done by LLVM. For instance, transforming `x == (x + 1)` to `false` also involves refinement since the expression's value went from "if x is `undef` then { `true` or `false` } else { `false` }" to "`false`" (by definition, the optimizer has license to fold `undef` to any non-`undef` value). Unfortunately, refinement implies that the optimizer cannot assume that the implementation of a function it can see has all of the behavior an unoptimized or a differently optimized version of the same function can have. This is a problem for functions with comdat linkage, where a function can be replaced by an unoptimized or a differently optimized version of the same source level function. For instance, FunctionAttrs cannot assume a comdat function is actually `readnone` even if it does not have any loads or stores in it; since there may have been loads and stores in the "original function" that were refined out in the currently visible variant, and at the link step the linker may in fact choose an implementation with a load or a store. As an example, consider a function that does two atomic loads from the same memory location, and writes to memory only if the two values are not equal. The optimizer is allowed to refine this function by first CSE'ing the two loads, and the folding the comparision to always report that the two values are equal. Such a refined variant will look like it is `readonly`. However, the unoptimized version of the function can still write to memory (since the two loads //can// result in different values), and selecting the unoptimized version at link time will retroactively invalidate transforms we may have done under the assumption that the function does not write to memory. Note: this is not just a problem with atomics or with linking differently optimized object files. See PR26774 for more realistic examples that involved neither. This patch: This change introduces a new set of linkage types, predicated as `GlobalValue::mayBeDerefined` that returns true if the linkage type allows a function to be replaced by a differently optimized variant at link time. It then changes a set of IPO passes to bail out if they see such a function. Reviewers: chandlerc, hfinkel, dexonsmith, joker.eph, rnk Subscribers: mcrosier, llvm-commits Differential Revision: http://reviews.llvm.org/D18634 llvm-svn: 265762
2016-04-08 02:48:30 +02:00
// We can infer and propagate function attributes only when we know that the
// definition we'll get at link time is *exactly* the definition we see now.
// For more details, see GlobalValue::mayBeDerefined.
if (!F->hasExactDefinition())
return false;
// We annotate noalias return values, which are only applicable to
// pointer types.
if (!F->getReturnType()->isPointerTy())
continue;
if (!isFunctionMallocLike(F, SCCNodes))
return false;
}
bool MadeChange = false;
for (Function *F : SCCNodes) {
if (F->doesNotAlias(0) || !F->getReturnType()->isPointerTy())
continue;
F->setDoesNotAlias(0);
++NumNoAlias;
MadeChange = true;
}
return MadeChange;
}
/// Tests whether this function is known to not return null.
///
/// Requires that the function returns a pointer.
///
/// Returns true if it believes the function will not return a null, and sets
/// \p Speculative based on whether the returned conclusion is a speculative
/// conclusion due to SCC calls.
static bool isReturnNonNull(Function *F, const SCCNodeSet &SCCNodes,
const TargetLibraryInfo &TLI, bool &Speculative) {
assert(F->getReturnType()->isPointerTy() &&
"nonnull only meaningful on pointer types");
Speculative = false;
SmallSetVector<Value *, 8> FlowsToReturn;
for (BasicBlock &BB : *F)
if (auto *Ret = dyn_cast<ReturnInst>(BB.getTerminator()))
FlowsToReturn.insert(Ret->getReturnValue());
for (unsigned i = 0; i != FlowsToReturn.size(); ++i) {
Value *RetVal = FlowsToReturn[i];
// If this value is locally known to be non-null, we're good
if (isKnownNonNull(RetVal, &TLI))
continue;
// Otherwise, we need to look upwards since we can't make any local
// conclusions.
Instruction *RVI = dyn_cast<Instruction>(RetVal);
if (!RVI)
return false;
switch (RVI->getOpcode()) {
// Extend the analysis by looking upwards.
case Instruction::BitCast:
case Instruction::GetElementPtr:
case Instruction::AddrSpaceCast:
FlowsToReturn.insert(RVI->getOperand(0));
continue;
case Instruction::Select: {
SelectInst *SI = cast<SelectInst>(RVI);
FlowsToReturn.insert(SI->getTrueValue());
FlowsToReturn.insert(SI->getFalseValue());
continue;
}
case Instruction::PHI: {
PHINode *PN = cast<PHINode>(RVI);
for (int i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
FlowsToReturn.insert(PN->getIncomingValue(i));
continue;
}
case Instruction::Call:
case Instruction::Invoke: {
CallSite CS(RVI);
Function *Callee = CS.getCalledFunction();
// A call to a node within the SCC is assumed to return null until
// proven otherwise
if (Callee && SCCNodes.count(Callee)) {
Speculative = true;
continue;
}
return false;
}
default:
return false; // Unknown source, may be null
};
llvm_unreachable("should have either continued or returned");
}
return true;
}
/// Deduce nonnull attributes for the SCC.
static bool addNonNullAttrs(const SCCNodeSet &SCCNodes,
const TargetLibraryInfo &TLI) {
// Speculative that all functions in the SCC return only nonnull
// pointers. We may refute this as we analyze functions.
bool SCCReturnsNonNull = true;
bool MadeChange = false;
// Check each function in turn, determining which functions return nonnull
// pointers.
for (Function *F : SCCNodes) {
// Already nonnull.
if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
Attribute::NonNull))
continue;
Don't IPO over functions that can be de-refined Summary: Fixes PR26774. If you're aware of the issue, feel free to skip the "Motivation" section and jump directly to "This patch". Motivation: I define "refinement" as discarding behaviors from a program that the optimizer has license to discard. So transforming: ``` void f(unsigned x) { unsigned t = 5 / x; (void)t; } ``` to ``` void f(unsigned x) { } ``` is refinement, since the behavior went from "if x == 0 then undefined else nothing" to "nothing" (the optimizer has license to discard undefined behavior). Refinement is a fundamental aspect of many mid-level optimizations done by LLVM. For instance, transforming `x == (x + 1)` to `false` also involves refinement since the expression's value went from "if x is `undef` then { `true` or `false` } else { `false` }" to "`false`" (by definition, the optimizer has license to fold `undef` to any non-`undef` value). Unfortunately, refinement implies that the optimizer cannot assume that the implementation of a function it can see has all of the behavior an unoptimized or a differently optimized version of the same function can have. This is a problem for functions with comdat linkage, where a function can be replaced by an unoptimized or a differently optimized version of the same source level function. For instance, FunctionAttrs cannot assume a comdat function is actually `readnone` even if it does not have any loads or stores in it; since there may have been loads and stores in the "original function" that were refined out in the currently visible variant, and at the link step the linker may in fact choose an implementation with a load or a store. As an example, consider a function that does two atomic loads from the same memory location, and writes to memory only if the two values are not equal. The optimizer is allowed to refine this function by first CSE'ing the two loads, and the folding the comparision to always report that the two values are equal. Such a refined variant will look like it is `readonly`. However, the unoptimized version of the function can still write to memory (since the two loads //can// result in different values), and selecting the unoptimized version at link time will retroactively invalidate transforms we may have done under the assumption that the function does not write to memory. Note: this is not just a problem with atomics or with linking differently optimized object files. See PR26774 for more realistic examples that involved neither. This patch: This change introduces a new set of linkage types, predicated as `GlobalValue::mayBeDerefined` that returns true if the linkage type allows a function to be replaced by a differently optimized variant at link time. It then changes a set of IPO passes to bail out if they see such a function. Reviewers: chandlerc, hfinkel, dexonsmith, joker.eph, rnk Subscribers: mcrosier, llvm-commits Differential Revision: http://reviews.llvm.org/D18634 llvm-svn: 265762
2016-04-08 02:48:30 +02:00
// We can infer and propagate function attributes only when we know that the
// definition we'll get at link time is *exactly* the definition we see now.
// For more details, see GlobalValue::mayBeDerefined.
if (!F->hasExactDefinition())
return false;
// We annotate nonnull return values, which are only applicable to
// pointer types.
if (!F->getReturnType()->isPointerTy())
continue;
bool Speculative = false;
if (isReturnNonNull(F, SCCNodes, TLI, Speculative)) {
if (!Speculative) {
// Mark the function eagerly since we may discover a function
// which prevents us from speculating about the entire SCC
DEBUG(dbgs() << "Eagerly marking " << F->getName() << " as nonnull\n");
F->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull);
++NumNonNullReturn;
MadeChange = true;
}
continue;
}
// At least one function returns something which could be null, can't
// speculate any more.
SCCReturnsNonNull = false;
}
if (SCCReturnsNonNull) {
for (Function *F : SCCNodes) {
if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
Attribute::NonNull) ||
!F->getReturnType()->isPointerTy())
continue;
DEBUG(dbgs() << "SCC marking " << F->getName() << " as nonnull\n");
F->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull);
++NumNonNullReturn;
MadeChange = true;
}
}
return MadeChange;
}
/// Remove the convergent attribute from all functions in the SCC if every
/// callsite within the SCC is not convergent (except for calls to functions
/// within the SCC). Returns true if changes were made.
static bool removeConvergentAttrs(const SCCNodeSet &SCCNodes) {
// For every function in SCC, ensure that either
// * it is not convergent, or
// * we can remove its convergent attribute.
bool HasConvergentFn = false;
for (Function *F : SCCNodes) {
if (!F->isConvergent()) continue;
HasConvergentFn = true;
// Can't remove convergent from function declarations.
if (F->isDeclaration()) return false;
// Can't remove convergent if any of our functions has a convergent call to a
// function not in the SCC.
for (Instruction &I : instructions(*F)) {
CallSite CS(&I);
// Bail if CS is a convergent call to a function not in the SCC.
if (CS && CS.isConvergent() &&
SCCNodes.count(CS.getCalledFunction()) == 0)
return false;
}
}
// If the SCC doesn't have any convergent functions, we have nothing to do.
if (!HasConvergentFn) return false;
// If we got here, all of the calls the SCC makes to functions not in the SCC
// are non-convergent. Therefore all of the SCC's functions can also be made
// non-convergent. We'll remove the attr from the callsites in
// InstCombineCalls.
for (Function *F : SCCNodes) {
if (!F->isConvergent()) continue;
DEBUG(dbgs() << "Removing convergent attr from fn " << F->getName()
<< "\n");
F->setNotConvergent();
}
return true;
}
static bool setDoesNotRecurse(Function &F) {
if (F.doesNotRecurse())
return false;
F.setDoesNotRecurse();
++NumNoRecurse;
return true;
}
static bool addNoRecurseAttrs(const SCCNodeSet &SCCNodes) {
// Try and identify functions that do not recurse.
// If the SCC contains multiple nodes we know for sure there is recursion.
if (SCCNodes.size() != 1)
return false;
Function *F = *SCCNodes.begin();
if (!F || F->isDeclaration() || F->doesNotRecurse())
return false;
// If all of the calls in F are identifiable and are to norecurse functions, F
// is norecurse. This check also detects self-recursion as F is not currently
// marked norecurse, so any called from F to F will not be marked norecurse.
for (Instruction &I : instructions(*F))
if (auto CS = CallSite(&I)) {
Function *Callee = CS.getCalledFunction();
if (!Callee || Callee == F || !Callee->doesNotRecurse())
// Function calls a potentially recursive function.
return false;
}
// Every call was to a non-recursive function other than this function, and
// we have no indirect recursion as the SCC size is one. This function cannot
// recurse.
return setDoesNotRecurse(*F);
}
PreservedAnalyses PostOrderFunctionAttrsPass::run(LazyCallGraph::SCC &C,
CGSCCAnalysisManager &AM) {
Module &M = *C.begin()->getFunction().getParent();
const ModuleAnalysisManager &MAM =
AM.getResult<ModuleAnalysisManagerCGSCCProxy>(C).getManager();
FunctionAnalysisManager &FAM =
AM.getResult<FunctionAnalysisManagerCGSCCProxy>(C).getManager();
// FIXME: Need some way to make it more reasonable to assume that this is
// always cached.
TargetLibraryInfo &TLI = *MAM.getCachedResult<TargetLibraryAnalysis>(M);
// We pass a lambda into functions to wire them up to the analysis manager
// for getting function analyses.
auto AARGetter = [&](Function &F) -> AAResults & {
return FAM.getResult<AAManager>(F);
};
// Fill SCCNodes with the elements of the SCC. Also track whether there are
// any external or opt-none nodes that will prevent us from optimizing any
// part of the SCC.
SCCNodeSet SCCNodes;
bool HasUnknownCall = false;
for (LazyCallGraph::Node &N : C) {
Function &F = N.getFunction();
if (F.hasFnAttribute(Attribute::OptimizeNone)) {
// Treat any function we're trying not to optimize as if it were an
// indirect call and omit it from the node set used below.
HasUnknownCall = true;
continue;
}
// Track whether any functions in this SCC have an unknown call edge.
// Note: if this is ever a performance hit, we can common it with
// subsequent routines which also do scans over the instructions of the
// function.
if (!HasUnknownCall)
for (Instruction &I : instructions(F))
if (auto CS = CallSite(&I))
if (!CS.getCalledFunction()) {
HasUnknownCall = true;
break;
}
SCCNodes.insert(&F);
}
bool Changed = false;
Changed |= addReadAttrs(SCCNodes, AARGetter);
Changed |= addArgumentAttrs(SCCNodes);
// If we have no external nodes participating in the SCC, we can deduce some
// more precise attributes as well.
if (!HasUnknownCall) {
Changed |= addNoAliasAttrs(SCCNodes);
Changed |= addNonNullAttrs(SCCNodes, TLI);
Changed |= removeConvergentAttrs(SCCNodes);
Changed |= addNoRecurseAttrs(SCCNodes);
}
return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
}
namespace {
struct PostOrderFunctionAttrsLegacyPass : public CallGraphSCCPass {
static char ID; // Pass identification, replacement for typeid
PostOrderFunctionAttrsLegacyPass() : CallGraphSCCPass(ID) {
initializePostOrderFunctionAttrsLegacyPassPass(*PassRegistry::getPassRegistry());
}
bool runOnSCC(CallGraphSCC &SCC) override;
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.setPreservesCFG();
AU.addRequired<AssumptionCacheTracker>();
AU.addRequired<TargetLibraryInfoWrapperPass>();
[AA] Hoist the logic to reformulate various AA queries in terms of other parts of the AA interface out of the base class of every single AA result object. Because this logic reformulates the query in terms of some other aspect of the API, it would easily cause O(n^2) query patterns in alias analysis. These could in turn be magnified further based on the number of call arguments, and then further based on the number of AA queries made for a particular call. This ended up causing problems for Rust that were actually noticable enough to get a bug (PR26564) and probably other places as well. When originally re-working the AA infrastructure, the desire was to regularize the pattern of refinement without losing any generality. While I think it was successful, that is clearly proving to be too costly. And the cost is needless: we gain no actual improvement for this generality of making a direct query to tbaa actually be able to re-use some other alias analysis's refinement logic for one of the other APIs, or some such. In short, this is entirely wasted work. To the extent possible, delegation to other API surfaces should be done at the aggregation layer so that we can avoid re-walking the aggregation. In fact, this significantly simplifies the logic as we no longer need to smuggle the aggregation layer into each alias analysis (or the TargetLibraryInfo into each alias analysis just so we can form argument memory locations!). However, we also have some delegation logic inside of BasicAA and some of it even makes sense. When the delegation logic is baking in specific knowledge of aliasing properties of the LLVM IR, as opposed to simply reformulating the query to utilize a different alias analysis interface entry point, it makes a lot of sense to restrict that logic to a different layer such as BasicAA. So one aspect of the delegation that was in every AA base class is that when we don't have operand bundles, we re-use function AA results as a fallback for callsite alias results. This relies on the IR properties of calls and functions w.r.t. aliasing, and so seems a better fit to BasicAA. I've lifted the logic up to that point where it seems to be a natural fit. This still does a bit of redundant work (we query function attributes twice, once via the callsite and once via the function AA query) but it is *exactly* twice here, no more. The end result is that all of the delegation logic is hoisted out of the base class and into either the aggregation layer when it is a pure retargeting to a different API surface, or into BasicAA when it relies on the IR's aliasing properties. This should fix the quadratic query pattern reported in PR26564, although I don't have a stand-alone test case to reproduce it. It also seems general goodness. Now the numerous AAs that don't need target library info don't carry it around and depend on it. I think I can even rip out the general access to the aggregation layer and only expose that in BasicAA as it is the only place where we re-query in that manner. However, this is a non-trivial change to the AA infrastructure so I want to get some additional eyes on this before it lands. Sadly, it can't wait long because we should really cherry pick this into 3.8 if we're going to go this route. Differential Revision: http://reviews.llvm.org/D17329 llvm-svn: 262490
2016-03-02 16:56:53 +01:00
getAAResultsAnalysisUsage(AU);
CallGraphSCCPass::getAnalysisUsage(AU);
}
private:
TargetLibraryInfo *TLI;
};
}
char PostOrderFunctionAttrsLegacyPass::ID = 0;
INITIALIZE_PASS_BEGIN(PostOrderFunctionAttrsLegacyPass, "functionattrs",
"Deduce function attributes", false, false)
INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)
INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
INITIALIZE_PASS_END(PostOrderFunctionAttrsLegacyPass, "functionattrs",
"Deduce function attributes", false, false)
Pass *llvm::createPostOrderFunctionAttrsLegacyPass() { return new PostOrderFunctionAttrsLegacyPass(); }
bool PostOrderFunctionAttrsLegacyPass::runOnSCC(CallGraphSCC &SCC) {
if (skipSCC(SCC))
return false;
TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
bool Changed = false;
// We compute dedicated AA results for each function in the SCC as needed. We
// use a lambda referencing external objects so that they live long enough to
// be queried, but we re-use them each time.
Optional<BasicAAResult> BAR;
Optional<AAResults> AAR;
auto AARGetter = [&](Function &F) -> AAResults & {
BAR.emplace(createLegacyPMBasicAAResult(*this, F));
AAR.emplace(createLegacyPMAAResults(*this, F, *BAR));
return *AAR;
};
// Fill SCCNodes with the elements of the SCC. Used for quickly looking up
// whether a given CallGraphNode is in this SCC. Also track whether there are
// any external or opt-none nodes that will prevent us from optimizing any
// part of the SCC.
SCCNodeSet SCCNodes;
bool ExternalNode = false;
for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
Function *F = (*I)->getFunction();
if (!F || F->hasFnAttribute(Attribute::OptimizeNone)) {
// External node or function we're trying not to optimize - we both avoid
// transform them and avoid leveraging information they provide.
ExternalNode = true;
continue;
}
SCCNodes.insert(F);
}
Changed |= addReadAttrs(SCCNodes, AARGetter);
Changed |= addArgumentAttrs(SCCNodes);
// If we have no external nodes participating in the SCC, we can deduce some
// more precise attributes as well.
if (!ExternalNode) {
Changed |= addNoAliasAttrs(SCCNodes);
Changed |= addNonNullAttrs(SCCNodes, *TLI);
Changed |= removeConvergentAttrs(SCCNodes);
Changed |= addNoRecurseAttrs(SCCNodes);
}
return Changed;
}
namespace {
/// A pass to do RPO deduction and propagation of function attributes.
///
/// This pass provides a general RPO or "top down" propagation of
/// function attributes. For a few (rare) cases, we can deduce significantly
/// more about function attributes by working in RPO, so this pass
/// provides the compliment to the post-order pass above where the majority of
/// deduction is performed.
// FIXME: Currently there is no RPO CGSCC pass structure to slide into and so
// this is a boring module pass, but eventually it should be an RPO CGSCC pass
// when such infrastructure is available.
struct ReversePostOrderFunctionAttrs : public ModulePass {
static char ID; // Pass identification, replacement for typeid
ReversePostOrderFunctionAttrs() : ModulePass(ID) {
initializeReversePostOrderFunctionAttrsPass(*PassRegistry::getPassRegistry());
}
bool runOnModule(Module &M) override;
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.setPreservesCFG();
AU.addRequired<CallGraphWrapperPass>();
}
};
}
char ReversePostOrderFunctionAttrs::ID = 0;
INITIALIZE_PASS_BEGIN(ReversePostOrderFunctionAttrs, "rpo-functionattrs",
"Deduce function attributes in RPO", false, false)
INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)
INITIALIZE_PASS_END(ReversePostOrderFunctionAttrs, "rpo-functionattrs",
"Deduce function attributes in RPO", false, false)
Pass *llvm::createReversePostOrderFunctionAttrsPass() {
return new ReversePostOrderFunctionAttrs();
}
static bool addNoRecurseAttrsTopDown(Function &F) {
// We check the preconditions for the function prior to calling this to avoid
// the cost of building up a reversible post-order list. We assert them here
// to make sure none of the invariants this relies on were violated.
assert(!F.isDeclaration() && "Cannot deduce norecurse without a definition!");
assert(!F.doesNotRecurse() &&
"This function has already been deduced as norecurs!");
assert(F.hasInternalLinkage() &&
"Can only do top-down deduction for internal linkage functions!");
// If F is internal and all of its uses are calls from a non-recursive
// functions, then none of its calls could in fact recurse without going
// through a function marked norecurse, and so we can mark this function too
// as norecurse. Note that the uses must actually be calls -- otherwise
// a pointer to this function could be returned from a norecurse function but
// this function could be recursively (indirectly) called. Note that this
// also detects if F is directly recursive as F is not yet marked as
// a norecurse function.
for (auto *U : F.users()) {
auto *I = dyn_cast<Instruction>(U);
if (!I)
return false;
CallSite CS(I);
if (!CS || !CS.getParent()->getParent()->doesNotRecurse())
return false;
}
return setDoesNotRecurse(F);
}
bool ReversePostOrderFunctionAttrs::runOnModule(Module &M) {
if (skipModule(M))
return false;
// We only have a post-order SCC traversal (because SCCs are inherently
// discovered in post-order), so we accumulate them in a vector and then walk
// it in reverse. This is simpler than using the RPO iterator infrastructure
// because we need to combine SCC detection and the PO walk of the call
// graph. We can also cheat egregiously because we're primarily interested in
// synthesizing norecurse and so we can only save the singular SCCs as SCCs
// with multiple functions in them will clearly be recursive.
auto &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
SmallVector<Function *, 16> Worklist;
for (scc_iterator<CallGraph *> I = scc_begin(&CG); !I.isAtEnd(); ++I) {
if (I->size() != 1)
continue;
Function *F = I->front()->getFunction();
if (F && !F->isDeclaration() && !F->doesNotRecurse() &&
F->hasInternalLinkage())
Worklist.push_back(F);
}
bool Changed = false;
for (auto *F : reverse(Worklist))
Changed |= addNoRecurseAttrsTopDown(*F);
return Changed;
}