[PATCH] D119601: [analyzer] Refactor makeNull to makeNullWithWidth (NFC)

Artem Dergachev via Phabricator via cfe-commits cfe-commits at lists.llvm.org
Fri Mar 18 22:18:33 PDT 2022


NoQ added inline comments.


================
Comment at: clang/include/clang/StaticAnalyzer/Core/PathSensitive/SValBuilder.h:378
+    type = type->isReferenceType()
+               ? Context.getPointerType(type->getPointeeType())
+               : type;
----------------
vabridgers wrote:
> NoQ wrote:
> > How do you discover the pointer width by looking only at the pointee type? Are objects of specific type always restricted to a specific address space? I.e., does every `int *` have the same width everywhere in the program? If not, how is this code supposed to work?
> The case this code addresses is represented in the test case clang/test/Analysis/cast-value-notes.cpp, this function ...
> 
> ```
>  24 void evalReferences(const Shape &S) {
>  25   const auto &C = dyn_cast<Circle>(S);
>  26   // expected-note at -1 {{Assuming 'S' is not a 'Circle'}}
>  27   // expected-note at -2 {{Dereference of null pointer}}
>  28   // expected-warning at -3 {{Dereference of null pointer}}
>  29 }
> 
> ```
> The crash occurs from line 25 above. 
> 
> Debug printing what I see at this point in the code for this case ...
> 
> ```
> QualType type : LValueReferenceType 0xffea820 'const class clang::Circle &'
> `-QualType 0xffea7c1 'const class clang::Circle' const
>   `-SubstTemplateTypeParmType 0xffea7c0 'class clang::Circle' sugar
>     |-TemplateTypeParmType 0xffe4f90 'X' dependent depth 0 index 0
>     | `-TemplateTypeParm 0xffe4f40 'X'
>     `-RecordType 0xffea090 'class clang::Circle'
>       `-CXXRecord 0xffea000 'Circle'
> 
> Context.getPointerType(type) : PointerType 0x10006ab0 'const class clang::Circle &*'
> `-LValueReferenceType 0xffea820 'const class clang::Circle &'
>   `-QualType 0xffea7c1 'const class clang::Circle' const
>     `-SubstTemplateTypeParmType 0xffea7c0 'class clang::Circle' sugar
>       |-TemplateTypeParmType 0xffe4f90 'X' dependent depth 0 index 0
>       | `-TemplateTypeParm 0xffe4f40 'X'
>       `-RecordType 0xffea090 'class clang::Circle'
>         `-CXXRecord 0xffea000 'Circle'
> 
> Context.getPointerType(type->getPointeeType()) : PointerType 0x10006ae0 'const class clang::Circle *'
> `-QualType 0xffea7c1 'const class clang::Circle' const
>   `-SubstTemplateTypeParmType 0xffea7c0 'class clang::Circle' sugar
>     |-TemplateTypeParmType 0xffe4f90 'X' dependent depth 0 index 0
>     | `-TemplateTypeParm 0xffe4f40 'X'
>     `-RecordType 0xffea090 'class clang::Circle'
>       `-CXXRecord 0xffea000 'Circle'
> 
> ```
> The LValueReferenceType causes a crash if I do not get a pointer type, looks like ... 
> 
> ```
> clang:  <root>/clang/include/clang/StaticAnalyzer/Core/PathSensitive/BasicValueFactory.h:219: const llvm::APSInt& clang::ento::BasicValueFactory::getZeroWithTypeSize(clang::QualType): Assertion `T->isScalarType()' failed.
> 
> ```
> I'm assuming the pointer type retains the address space attribute of the LValueReferenceType. 
> 
> Context.getPointerType(type->getPointeeType()) produces a QualType :  PointerType 'const class clang::Circle *' from an original QualType : LValueReferenceType 'const class clang::Circle &'
> 
> Is this not what's wanted - a pointer type instead of a reference, with the same address space qualifiers, or am I missing something in the query? 
> 
> Thanks
Do I understand correctly that `__attribute__((address_space))` is a type attribute that gets applied to values rather than to pointers, so `__attribute__((address_space(3))) int *x` defines `x` as "pointer to (int that lives in address_space(3))", so when you unwrap the pointer type you get an "int that lives in address_space(3)"?

Can you share some dumps to demonstrate that the attribute is correctly preserved by this procedure? Damn, a test case could be great if we could have them.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D119601/new/

https://reviews.llvm.org/D119601



More information about the cfe-commits mailing list