<div dir="ltr">Thanks for the review!<div><br><div>I think your analysis is in general correct. This was definitely an unintended effect of my change. I thought I had checked all four of the operations, but I had missed a return in the routine that adds readonly and readnone.</div></div><div><br></div><div>However, I can't see any way this changes behavior because of what the readonly and readnone analysis itself does: it scans the function bodies, and should make a conservative choice about any calls to external functions.</div><div><br></div><div>I thought this might break in an SCC containing an optnone node because we would never scan the body of the optnone function... But that also means we would never infer any attributes on it, and so when it was called by some other function in the SCC, when we scanned *that* function, we would see a call to a may-write function and not make any assumptions.</div><div><br></div><div>So while this wasn't intended (and I plan to change it to match the original behavior) as far as I can tell this is benign. Does that make sense to you? I have tried to form a test case to hit this as well, and I can't produce anything that is *actually* more interesting than the existing test cases, and it all seems to remain correct even after this change.</div><div><br></div><div><br></div><div>It is also worth noting that I think this code behaves a bit differently than I had originally assumed, and my comments around ExternalNode are, I think, misleading at best. The old call graph does not actually take the form I had expected in a particularly subtle but important corner case and so I'm having to re-evaluate how I understand its behavior. I'll try to update this code to have better comments once I've fully re-internalized the model.</div></div><br><div class="gmail_quote"><div dir="ltr">On Tue, Nov 3, 2015 at 6:15 PM Philip Reames via llvm-commits <<a href="mailto:llvm-commits@lists.llvm.org">llvm-commits@lists.llvm.org</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">Chandler,<br>
<br>
I noticed the handling in this change for ExternalNode. I *think* that<br>
the code in question is only unnecessarily conservative for non-null and<br>
no alias inference, but is incorrect for readonly, readnone, and<br>
argument attributes.<br>
<br>
In particular, none of these analyzes should be able to make any<br>
speculative conclusions about an SCC which contains an external call.<br>
By filtering the external node out of the SCC, I think you've allowed<br>
them to do so. Am I wrong here?<br>
<br>
Philip<br>
<br>
On 10/29/2015 11:29 AM, Chandler Carruth via llvm-commits wrote:<br>
> Author: chandlerc<br>
> Date: Thu Oct 29 13:29:15 2015<br>
> New Revision: 251640<br>
><br>
> URL: <a href="http://llvm.org/viewvc/llvm-project?rev=251640&view=rev" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-project?rev=251640&view=rev</a><br>
> Log:<br>
> [FunctionAttrs] Provide a single SCC node set to all of the<br>
> transformations in FunctionAttrs rather than building a new one each<br>
> time.<br>
><br>
> This isn't trivial because there are different heuristics from different<br>
> passes for exactly what set they want. The primary difference is whether<br>
> an *overridable* function completely disables the synthesis of<br>
> attributes. I've modeled this by directly testing for overridable, and<br>
> using the common set that excludes external and opt-none functions.<br>
><br>
> This does cause some changes by disabling more optimizations in the face<br>
> of opt-none. Specifically, we were still optimizing *calls* to opt-none<br>
> functions based on their attributes, just not the bodies. It seems<br>
> better to be conservative on both fronts given the intended semanticas<br>
> here (best effort to not assume or disturb anything). I've not tried to<br>
> test this change as it seems complex, brittle, and not important to the<br>
> implicit contract of opt-none. Instead, it seems more like a choice that<br>
> should be dictated by the simplified implementation and the change to be<br>
> acceptable differences within the space of opt-none.<br>
><br>
> A big benefit here is that these transformations no longer rely on the<br>
> legacy pass manager's SCC types, they just work on generic sets of<br>
> function pointers. This will make it easy to re-use their logic in the<br>
> new pass manager.<br>
><br>
> I've also made the transforms static functions instead of members where<br>
> trivial while I was touching the signatures.<br>
><br>
> Modified:<br>
> llvm/trunk/lib/Transforms/IPO/FunctionAttrs.cpp<br>
><br>
> Modified: llvm/trunk/lib/Transforms/IPO/FunctionAttrs.cpp<br>
> URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/IPO/FunctionAttrs.cpp?rev=251640&r1=251639&r2=251640&view=diff" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/IPO/FunctionAttrs.cpp?rev=251640&r1=251639&r2=251640&view=diff</a><br>
> ==============================================================================<br>
> --- llvm/trunk/lib/Transforms/IPO/FunctionAttrs.cpp (original)<br>
> +++ llvm/trunk/lib/Transforms/IPO/FunctionAttrs.cpp Thu Oct 29 13:29:15 2015<br>
> @@ -52,6 +52,10 @@ STATISTIC(NumNonNullReturn, "Number of f<br>
> STATISTIC(NumAnnotated, "Number of attributes added to library functions");<br>
><br>
> namespace {<br>
> +typedef SmallSetVector<Function *, 8> SCCNodeSet;<br>
> +}<br>
> +<br>
> +namespace {<br>
> struct FunctionAttrs : public CallGraphSCCPass {<br>
> static char ID; // Pass identification, replacement for typeid<br>
> FunctionAttrs() : CallGraphSCCPass(ID) {<br>
> @@ -70,10 +74,7 @@ struct FunctionAttrs : public CallGraphS<br>
> private:<br>
> TargetLibraryInfo *TLI;<br>
><br>
> - bool AddReadAttrs(const CallGraphSCC &SCC);<br>
> - bool AddArgumentAttrs(const CallGraphSCC &SCC);<br>
> - bool AddNoAliasAttrs(const CallGraphSCC &SCC);<br>
> - bool AddNonNullAttrs(const CallGraphSCC &SCC);<br>
> + bool AddReadAttrs(const SCCNodeSet &SCCNodes);<br>
> bool annotateLibraryCalls(const CallGraphSCC &SCC);<br>
> };<br>
> }<br>
> @@ -99,9 +100,8 @@ enum MemoryAccessKind {<br>
> };<br>
> }<br>
><br>
> -static MemoryAccessKind<br>
> -checkFunctionMemoryAccess(Function &F, AAResults &AAR,<br>
> - const SmallPtrSetImpl<Function *> &SCCNodes) {<br>
> +static MemoryAccessKind checkFunctionMemoryAccess(Function &F, AAResults &AAR,<br>
> + const SCCNodeSet &SCCNodes) {<br>
> FunctionModRefBehavior MRB = AAR.getModRefBehavior(&F);<br>
> if (MRB == FMRB_DoesNotAccessMemory)<br>
> // Already perfect!<br>
> @@ -205,25 +205,11 @@ checkFunctionMemoryAccess(Function &F, A<br>
> }<br>
><br>
> /// Deduce readonly/readnone attributes for the SCC.<br>
> -bool FunctionAttrs::AddReadAttrs(const CallGraphSCC &SCC) {<br>
> - SmallPtrSet<Function *, 8> SCCNodes;<br>
> -<br>
> - // Fill SCCNodes with the elements of the SCC. Used for quickly<br>
> - // looking up whether a given CallGraphNode is in this SCC.<br>
> - for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I)<br>
> - SCCNodes.insert((*I)->getFunction());<br>
> -<br>
> +bool FunctionAttrs::AddReadAttrs(const SCCNodeSet &SCCNodes) {<br>
> // Check if any of the functions in the SCC read or write memory. If they<br>
> // write memory then they can't be marked readnone or readonly.<br>
> bool ReadsMemory = false;<br>
> - for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {<br>
> - Function *F = (*I)->getFunction();<br>
> -<br>
> - if (!F || F->hasFnAttribute(Attribute::OptimizeNone))<br>
> - // External node or node we don't want to optimize - assume it may write<br>
> - // memory and give up.<br>
> - return false;<br>
> -<br>
> + for (Function *F : SCCNodes) {<br>
> // We need to manually construct BasicAA directly in order to disable its<br>
> // use of other function analyses.<br>
> BasicAAResult BAR(createLegacyPMBasicAAResult(*this, *F));<br>
> @@ -247,9 +233,7 @@ bool FunctionAttrs::AddReadAttrs(const C<br>
> // Success! Functions in this SCC do not access memory, or only read memory.<br>
> // Give them the appropriate attribute.<br>
> bool MadeChange = false;<br>
> - for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {<br>
> - Function *F = (*I)->getFunction();<br>
> -<br>
> + for (Function *F : SCCNodes) {<br>
> if (F->doesNotAccessMemory())<br>
> // Already perfect!<br>
> continue;<br>
> @@ -325,7 +309,7 @@ public:<br>
> /// consider that a capture, instead adding it to the "Uses" list and<br>
> /// continuing with the analysis.<br>
> struct ArgumentUsesTracker : public CaptureTracker {<br>
> - ArgumentUsesTracker(const SmallPtrSet<Function *, 8> &SCCNodes)<br>
> + ArgumentUsesTracker(const SCCNodeSet &SCCNodes)<br>
> : Captured(false), SCCNodes(SCCNodes) {}<br>
><br>
> void tooManyUses() override { Captured = true; }<br>
> @@ -338,7 +322,8 @@ struct ArgumentUsesTracker : public Capt<br>
> }<br>
><br>
> Function *F = CS.getCalledFunction();<br>
> - if (!F || !SCCNodes.count(F)) {<br>
> + if (!F || F->isDeclaration() || F->mayBeOverridden() ||<br>
> + !SCCNodes.count(F)) {<br>
> Captured = true;<br>
> return true;<br>
> }<br>
> @@ -366,7 +351,7 @@ struct ArgumentUsesTracker : public Capt<br>
> bool Captured; // True only if certainly captured (used outside our SCC).<br>
> SmallVector<Argument *, 4> Uses; // Uses within our SCC.<br>
><br>
> - const SmallPtrSet<Function *, 8> &SCCNodes;<br>
> + const SCCNodeSet &SCCNodes;<br>
> };<br>
> }<br>
><br>
> @@ -501,20 +486,9 @@ determinePointerReadAttrs(Argument *A,<br>
> }<br>
><br>
> /// Deduce nocapture attributes for the SCC.<br>
> -bool FunctionAttrs::AddArgumentAttrs(const CallGraphSCC &SCC) {<br>
> +static bool addArgumentAttrs(const SCCNodeSet &SCCNodes) {<br>
> bool Changed = false;<br>
><br>
> - SmallPtrSet<Function *, 8> SCCNodes;<br>
> -<br>
> - // Fill SCCNodes with the elements of the SCC. Used for quickly<br>
> - // looking up whether a given CallGraphNode is in this SCC.<br>
> - for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {<br>
> - Function *F = (*I)->getFunction();<br>
> - if (F && !F->isDeclaration() && !F->mayBeOverridden() &&<br>
> - !F->hasFnAttribute(Attribute::OptimizeNone))<br>
> - SCCNodes.insert(F);<br>
> - }<br>
> -<br>
> ArgumentGraph AG;<br>
><br>
> AttrBuilder B;<br>
> @@ -522,14 +496,7 @@ bool FunctionAttrs::AddArgumentAttrs(con<br>
><br>
> // Check each function in turn, determining which pointer arguments are not<br>
> // captured.<br>
> - for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {<br>
> - Function *F = (*I)->getFunction();<br>
> -<br>
> - if (!F || F->hasFnAttribute(Attribute::OptimizeNone))<br>
> - // External node or function we're trying not to optimize - only a problem<br>
> - // for arguments that we pass to it.<br>
> - continue;<br>
> -<br>
> + for (Function *F : SCCNodes) {<br>
> // Definitions with weak linkage may be overridden at linktime with<br>
> // something that captures pointers, so treat them like declarations.<br>
> if (F->isDeclaration() || F->mayBeOverridden())<br>
> @@ -714,8 +681,7 @@ bool FunctionAttrs::AddArgumentAttrs(con<br>
> ///<br>
> /// A function is "malloc-like" if it returns either null or a pointer that<br>
> /// doesn't alias any other pointer visible to the caller.<br>
> -static bool isFunctionMallocLike(Function *F,<br>
> - SmallPtrSet<Function *, 8> &SCCNodes) {<br>
> +static bool isFunctionMallocLike(Function *F, const SCCNodeSet &SCCNodes) {<br>
> SmallSetVector<Value *, 8> FlowsToReturn;<br>
> for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)<br>
> if (ReturnInst *Ret = dyn_cast<ReturnInst>(I->getTerminator()))<br>
> @@ -778,23 +744,10 @@ static bool isFunctionMallocLike(Functio<br>
> }<br>
><br>
> /// Deduce noalias attributes for the SCC.<br>
> -bool FunctionAttrs::AddNoAliasAttrs(const CallGraphSCC &SCC) {<br>
> - SmallPtrSet<Function *, 8> SCCNodes;<br>
> -<br>
> - // Fill SCCNodes with the elements of the SCC. Used for quickly<br>
> - // looking up whether a given CallGraphNode is in this SCC.<br>
> - for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I)<br>
> - SCCNodes.insert((*I)->getFunction());<br>
> -<br>
> +static bool addNoAliasAttrs(const SCCNodeSet &SCCNodes) {<br>
> // Check each function in turn, determining which functions return noalias<br>
> // pointers.<br>
> - for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {<br>
> - Function *F = (*I)->getFunction();<br>
> -<br>
> - if (!F || F->hasFnAttribute(Attribute::OptimizeNone))<br>
> - // External node or node we don't want to optimize - skip it;<br>
> - return false;<br>
> -<br>
> + for (Function *F : SCCNodes) {<br>
> // Already noalias.<br>
> if (F->doesNotAlias(0))<br>
> continue;<br>
> @@ -814,8 +767,7 @@ bool FunctionAttrs::AddNoAliasAttrs(cons<br>
> }<br>
><br>
> bool MadeChange = false;<br>
> - for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {<br>
> - Function *F = (*I)->getFunction();<br>
> + for (Function *F : SCCNodes) {<br>
> if (F->doesNotAlias(0) || !F->getReturnType()->isPointerTy())<br>
> continue;<br>
><br>
> @@ -834,7 +786,7 @@ bool FunctionAttrs::AddNoAliasAttrs(cons<br>
> /// Returns true if it believes the function will not return a null, and sets<br>
> /// \p Speculative based on whether the returned conclusion is a speculative<br>
> /// conclusion due to SCC calls.<br>
> -static bool isReturnNonNull(Function *F, SmallPtrSet<Function *, 8> &SCCNodes,<br>
> +static bool isReturnNonNull(Function *F, const SCCNodeSet &SCCNodes,<br>
> const TargetLibraryInfo &TLI, bool &Speculative) {<br>
> assert(F->getReturnType()->isPointerTy() &&<br>
> "nonnull only meaningful on pointer types");<br>
> @@ -898,14 +850,8 @@ static bool isReturnNonNull(Function *F,<br>
> }<br>
><br>
> /// Deduce nonnull attributes for the SCC.<br>
> -bool FunctionAttrs::AddNonNullAttrs(const CallGraphSCC &SCC) {<br>
> - SmallPtrSet<Function *, 8> SCCNodes;<br>
> -<br>
> - // Fill SCCNodes with the elements of the SCC. Used for quickly<br>
> - // looking up whether a given CallGraphNode is in this SCC.<br>
> - for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I)<br>
> - SCCNodes.insert((*I)->getFunction());<br>
> -<br>
> +static bool addNonNullAttrs(const SCCNodeSet &SCCNodes,<br>
> + const TargetLibraryInfo &TLI) {<br>
> // Speculative that all functions in the SCC return only nonnull<br>
> // pointers. We may refute this as we analyze functions.<br>
> bool SCCReturnsNonNull = true;<br>
> @@ -914,13 +860,7 @@ bool FunctionAttrs::AddNonNullAttrs(cons<br>
><br>
> // Check each function in turn, determining which functions return nonnull<br>
> // pointers.<br>
> - for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {<br>
> - Function *F = (*I)->getFunction();<br>
> -<br>
> - if (!F || F->hasFnAttribute(Attribute::OptimizeNone))<br>
> - // External node or node we don't want to optimize - skip it;<br>
> - return false;<br>
> -<br>
> + for (Function *F : SCCNodes) {<br>
> // Already nonnull.<br>
> if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,<br>
> Attribute::NonNull))<br>
> @@ -937,7 +877,7 @@ bool FunctionAttrs::AddNonNullAttrs(cons<br>
> continue;<br>
><br>
> bool Speculative = false;<br>
> - if (isReturnNonNull(F, SCCNodes, *TLI, Speculative)) {<br>
> + if (isReturnNonNull(F, SCCNodes, TLI, Speculative)) {<br>
> if (!Speculative) {<br>
> // Mark the function eagerly since we may discover a function<br>
> // which prevents us from speculating about the entire SCC<br>
> @@ -954,8 +894,7 @@ bool FunctionAttrs::AddNonNullAttrs(cons<br>
> }<br>
><br>
> if (SCCReturnsNonNull) {<br>
> - for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {<br>
> - Function *F = (*I)->getFunction();<br>
> + for (Function *F : SCCNodes) {<br>
> if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,<br>
> Attribute::NonNull) ||<br>
> !F->getReturnType()->isPointerTy())<br>
> @@ -1835,10 +1774,36 @@ bool FunctionAttrs::annotateLibraryCalls<br>
> bool FunctionAttrs::runOnSCC(CallGraphSCC &SCC) {<br>
> TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();<br>
><br>
> + // Annotate declarations for which we have special knowledge.<br>
> bool Changed = annotateLibraryCalls(SCC);<br>
> - Changed |= AddReadAttrs(SCC);<br>
> - Changed |= AddArgumentAttrs(SCC);<br>
> - Changed |= AddNoAliasAttrs(SCC);<br>
> - Changed |= AddNonNullAttrs(SCC);<br>
> +<br>
> + // Fill SCCNodes with the elements of the SCC. Used for quickly looking up<br>
> + // whether a given CallGraphNode is in this SCC. Also track whether there are<br>
> + // any external or opt-none nodes that will prevent us from optimizing any<br>
> + // part of the SCC.<br>
> + SCCNodeSet SCCNodes;<br>
> + bool ExternalNode = false;<br>
> + for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {<br>
> + Function *F = (*I)->getFunction();<br>
> + if (!F || F->hasFnAttribute(Attribute::OptimizeNone)) {<br>
> + // External node or function we're trying not to optimize - we both avoid<br>
> + // transform them and avoid leveraging information they provide.<br>
> + ExternalNode = true;<br>
> + continue;<br>
> + }<br>
> +<br>
> + SCCNodes.insert(F);<br>
> + }<br>
> +<br>
> + Changed |= AddReadAttrs(SCCNodes);<br>
> + Changed |= addArgumentAttrs(SCCNodes);<br>
> +<br>
> + // If we have no external nodes participating in the SCC, we can infer some<br>
> + // more precise attributes as well.<br>
> + if (!ExternalNode) {<br>
> + Changed |= addNoAliasAttrs(SCCNodes);<br>
> + Changed |= addNonNullAttrs(SCCNodes, *TLI);<br>
> + }<br>
> +<br>
> return Changed;<br>
> }<br>
><br>
><br>
> _______________________________________________<br>
> llvm-commits mailing list<br>
> <a href="mailto:llvm-commits@lists.llvm.org" target="_blank">llvm-commits@lists.llvm.org</a><br>
> <a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits</a><br>
<br>
_______________________________________________<br>
llvm-commits mailing list<br>
<a href="mailto:llvm-commits@lists.llvm.org" target="_blank">llvm-commits@lists.llvm.org</a><br>
<a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits</a><br>
</blockquote></div>