Strict Aliasing Problem

[BryceMehring]

Using the -O2 flag for g++ 4.8.1 with glm, I get aliasing warnings everywhere:

warning: dereferencing type-punned pointer will break strict-aliasing rules

Edit: I'm also using the -Wall flag

[Groovounet]

Can you point me to the code that generate that warning?

[BryceMehring]

Including either glm/gtx/transform.hpp or glm/geometric.hpp will cause the warnings. There might be more.

[ghost]

I also get a lot of those in Android NDK 9c (only in release build).

An example from 0.9.5.1 release:

In file included from   glm/detail/func_exponential.hpp:127:0,
                 from   glm/exponential.hpp:32,
                 from   glm/glm.hpp:109,
  glm/detail/func_exponential.inl: In instantiation of 'static vecType<float, (glm::precision)2u> glm::detail::compute_inversesqrt<vecType, float, (glm::precision)2u>::call(const vecType<float, (glm::precision)2u>&) [with vecType = glm::detail::tvec1]':
  glm/gtx/fast_square_root.inl:29:67:   required from here
  glm/detail/func_exponential.inl:176:9: warning: dereferencing type-punned pointer will break strict-aliasing rules [-Wstrict-aliasing]
     tmp = *reinterpret_cast<vecType<float, lowp>*>(&i);
         ^

However, it occurs in tons of places.

[Groovounet]

So I am not 100% sure I fixed the issue but at least a apply a tentative fix to resolve it.

This fix is available in GLM 0.9.5 branch of the repository.

Thanks for contributing,
Christophe

[bagage]

The issue still occurs with the Android NDK, using glm branch 0.9.5 at func_integer.inl:234/175:

lsb = reinterpret_cast<int32&>(Value64);

Full warning:

In file included from libs/glm/glm/detail/func_integer.hpp:200:0,
                 from libs/glm/glm/integer.hpp:32,
                 from libs/glm/glm/glm.hpp:115,
libs/glm/glm/detail/func_integer.inl: In function 'void glm::umulExtended(const genUType&, const genUType&, genUType&, genUType&) [with genUType = unsigned int]':
libs/glm/glm/detail/func_integer.inl:175:42: warning: dereferencing type-punned pointer will break strict-aliasing rules [-Wstrict-aliasing]
libs/glm/glm/detail/func_integer.inl: In function 'void glm::imulExtended(const genIType&, const genIType&, genIType&, genIType&) [with genIType = int]':
libs/glm/glm/detail/func_integer.inl:234:41: warning: dereferencing type-punned pointer will break strict-aliasing rules [-Wstrict-aliasing]

[Groovounet]

If that was the only two issues left I think this time the problem should be fixed.

This changes will be part of GLM 0.9.5.3.

Thanks for contributing,
Christophe

[bagage]

I guess you fixed the wrong line. The faulty one was the line below it:
lsb = reinterpret_cast<int32&>(Value64);

[klemmster]

@Groovounet
We've encountered a similar issue in libavg.
I've added a fix(libavg/libavg@a894a9f) to our local glm copy for now, motivated by these two sources:

• http://blog.regehr.org/archives/959
• http://dbp-consulting.com/tutorials/StrictAliasing.html

and a bit of research it seems to be the appropriate way.

If you would like to integrate this solution, just let me know which branch I should open a PR against.

[Groovounet]

I'll read this before coming to a conclusion.

I am not convinced that the aliasing here in GLM is really an issue because it's aliasing from a temporary variable created in the function generating that aliasing. Aliasing of the arguments is an issue that should be fixed in GLM 0.9.5.

[Groovounet]

This issue should be fixed in GLM 0.9.5 branch for GLM 0.9.5.4 release.

Thanks,
Christophe