[llvm] r325630 - [MemoryBuiltins] Check nobuiltin status when identifying calls to free.
Sam Clegg via llvm-commits
llvm-commits at lists.llvm.org
Wed Feb 21 14:28:56 PST 2018
On Wed, Feb 21, 2018 at 2:20 PM, Philip Reames
<listmail at philipreames.com> wrote:
>
>
> On 02/21/2018 12:44 PM, Sam Clegg via llvm-commits wrote:
>>
>> Interestingly it doesn't elide the call if I return a non-const value.
>> If I write `return &n+s;` instead it works as expected. So it seems
>> as if it is assuming that because new returns a const value it can
>> elide it, despite the fact that new contains other side effects (in
>> this case the printf statement).
>
> Probably not the actual reasoning, just FYI. If we can show we returning a
> value derived from an alloca, that's UB. I'd phrase this as simply a missed
> optimization. (i.e. an offset from the alloca is still, by assumption,
> within the alloca and thus UB)
This is not an alloca but a static. it behaves the same if I move
"int n" out into file scope.
>
>>
>> On Wed, Feb 21, 2018 at 12:29 PM, Sam Clegg <sbc at google.com> wrote:
>>>
>>> On Wed, Feb 21, 2018 at 2:46 AM, Benjamin Kramer <benny.kra at gmail.com>
>>> wrote:
>>>>
>>>> I'm relatively certain that the test is bad wrt the C++ standard. The
>>>> compiler is allowed to elide new even if it is user-defined. The
>>>> standard's
>>>> escape hatch is calling "::operator new(size)" instead of using a new
>>>> expression. Building with -fno-builtin should also work. Or feed the
>>>> pointer
>>>> into an __asm__ barrier to prevent the compiler from removing it.
>>>>
>>> Interesting. Why is it allowed to elide the call to new? I would have
>>> thought that since the result of the new is used (its printed) the
>>> call can't be elided. But I don't know the spec so if you are sure
>>> this is allowed I can simply disable the tests that use this pattern.
>>>
>>> Would you have expected the change in question to have caused this to
>>> start failing?
>>>
>>>
>>>> On Wed, Feb 21, 2018 at 2:49 AM Sam Clegg <sbc at google.com> wrote:
>>>>>
>>>>> We have some test code on the WebAssembly waterfall the broke as a
>>>>> result of this. Perhaps the test is bad but I've distilled to the
>>>>> following:
>>>>>
>>>>> ```
>>>>> #include <stdio.h>
>>>>> #include <stdlib.h>
>>>>>
>>>>> void* operator new(size_t s) {
>>>>> static int n = 1;
>>>>> printf("new %zu\n", s);
>>>>> return &n;
>>>>> }
>>>>>
>>>>> int main() {
>>>>> int* foo = new int;
>>>>> printf("%p\n", foo);
>>>>> return 0;
>>>>> }
>>>>> ```
>>>>>
>>>>> Before this change the above program will always print "new ...".
>>>>> After this change it only prints this O0. In O1 and O2 the call to
>>>>> new is elided. Is this intended?
>>>>>
>>>>> On Tue, Feb 20, 2018 at 2:00 PM, Benjamin Kramer via llvm-commits
>>>>> <llvm-commits at lists.llvm.org> wrote:
>>>>>>
>>>>>> Author: d0k
>>>>>> Date: Tue Feb 20 14:00:33 2018
>>>>>> New Revision: 325630
>>>>>>
>>>>>> URL: http://llvm.org/viewvc/llvm-project?rev=325630&view=rev
>>>>>> Log:
>>>>>> [MemoryBuiltins] Check nobuiltin status when identifying calls to
>>>>>> free.
>>>>>>
>>>>>> This is usually not a problem because this code's main purpose is
>>>>>> eliminating unused new/delete pairs. We got deletes of nullptr or
>>>>>> nobuiltin deletes of builtin new wrong though.
>>>>>>
>>>>>> Modified:
>>>>>> llvm/trunk/lib/Analysis/MemoryBuiltins.cpp
>>>>>> llvm/trunk/test/Transforms/InstCombine/malloc-free-delete.ll
>>>>>>
>>>>>> Modified: llvm/trunk/lib/Analysis/MemoryBuiltins.cpp
>>>>>> URL:
>>>>>>
>>>>>> http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Analysis/MemoryBuiltins.cpp?rev=325630&r1=325629&r2=325630&view=diff
>>>>>>
>>>>>>
>>>>>> ==============================================================================
>>>>>> --- llvm/trunk/lib/Analysis/MemoryBuiltins.cpp (original)
>>>>>> +++ llvm/trunk/lib/Analysis/MemoryBuiltins.cpp Tue Feb 20 14:00:33
>>>>>> 2018
>>>>>> @@ -112,10 +112,9 @@ static const Function *getCalledFunction
>>>>>>
>>>>>> IsNoBuiltin = CS.isNoBuiltin();
>>>>>>
>>>>>> - const Function *Callee = CS.getCalledFunction();
>>>>>> - if (!Callee || !Callee->isDeclaration())
>>>>>> - return nullptr;
>>>>>> - return Callee;
>>>>>> + if (const Function *Callee = CS.getCalledFunction())
>>>>>> + return Callee;
>>>>>> + return nullptr;
>>>>>> }
>>>>>>
>>>>>> /// Returns the allocation data for the given value if it's either a
>>>>>> call to a
>>>>>> @@ -350,11 +349,10 @@ const CallInst *llvm::extractCallocCall(
>>>>>>
>>>>>> /// isFreeCall - Returns non-null if the value is a call to the
>>>>>> builtin
>>>>>> free()
>>>>>> const CallInst *llvm::isFreeCall(const Value *I, const
>>>>>> TargetLibraryInfo *TLI) {
>>>>>> - const CallInst *CI = dyn_cast<CallInst>(I);
>>>>>> - if (!CI || isa<IntrinsicInst>(CI))
>>>>>> - return nullptr;
>>>>>> - Function *Callee = CI->getCalledFunction();
>>>>>> - if (Callee == nullptr)
>>>>>> + bool IsNoBuiltinCall;
>>>>>> + const Function *Callee =
>>>>>> + getCalledFunction(I, /*LookThroughBitCast=*/false,
>>>>>> IsNoBuiltinCall);
>>>>>> + if (Callee == nullptr || IsNoBuiltinCall)
>>>>>> return nullptr;
>>>>>>
>>>>>> StringRef FnName = Callee->getName();
>>>>>> @@ -400,7 +398,7 @@ const CallInst *llvm::isFreeCall(const V
>>>>>> if (FTy->getParamType(0) !=
>>>>>> Type::getInt8PtrTy(Callee->getContext()))
>>>>>> return nullptr;
>>>>>>
>>>>>> - return CI;
>>>>>> + return dyn_cast<CallInst>(I);
>>>>>> }
>>>>>>
>>>>>>
>>>>>>
>>>>>> //===----------------------------------------------------------------------===//
>>>>>>
>>>>>> Modified: llvm/trunk/test/Transforms/InstCombine/malloc-free-delete.ll
>>>>>> URL:
>>>>>>
>>>>>> http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstCombine/malloc-free-delete.ll?rev=325630&r1=325629&r2=325630&view=diff
>>>>>>
>>>>>>
>>>>>> ==============================================================================
>>>>>> --- llvm/trunk/test/Transforms/InstCombine/malloc-free-delete.ll
>>>>>> (original)
>>>>>> +++ llvm/trunk/test/Transforms/InstCombine/malloc-free-delete.ll Tue
>>>>>> Feb
>>>>>> 20 14:00:33 2018
>>>>>> @@ -146,7 +146,11 @@ lpad.i:
>>>>>> }
>>>>>>
>>>>>> declare i8* @_Znwm(i64) nobuiltin
>>>>>> -declare i8* @_Znwj(i32) nobuiltin
>>>>>> +define i8* @_Znwj(i32 %n) nobuiltin {
>>>>>> + %z = zext i32 %n to i64
>>>>>> + call i8* @_Znwm(i64 %z)
>>>>>> + ret i8* %m
>>>>>> +}
>>>>>> declare i8* @_Znam(i64) nobuiltin
>>>>>> declare i8* @_Znaj(i32) nobuiltin
>>>>>> declare void @_ZdlPv(i8*) nobuiltin
>>>>>> @@ -197,3 +201,19 @@ define void @test9() {
>>>>>> call void @"\01??3 at YAXPEAX@Z"(i8* %new_long_long) builtin
>>>>>> ret void
>>>>>> }
>>>>>> +
>>>>>> +define void @test10() {
>>>>>> +; CHECK-LABEL: @test10
>>>>>> +; CHECK: call void @_ZdlPv
>>>>>> + call void @_ZdlPv(i8* null)
>>>>>> + ret void
>>>>>> +}
>>>>>> +
>>>>>> +define void @test11() {
>>>>>> +; CHECK-LABEL: @test11
>>>>>> +; CHECK: call i8* @_Znwm
>>>>>> +; CHECK: call void @_ZdlPv
>>>>>> + %call = call i8* @_Znwm(i64 8) builtin
>>>>>> + call void @_ZdlPv(i8* %call)
>>>>>> + ret void
>>>>>> +}
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> llvm-commits mailing list
>>>>>> llvm-commits at lists.llvm.org
>>>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>
>
More information about the llvm-commits
mailing list