Add documentation for -fno-strict-aliasing

[simonbyrne]

Summary

This is probably good to document, as the behaviour of -fstrict-aliasing differs from gcc in a quite important way (see https://bugs.llvm.org//show_bug.cgi?id=31928) .

[hfinkel]

docs/UsersManual.rst

1106

Capitalize Unlike.

[hans]

+1 for documenting this, but I have to leave it to the language lawyers for how to fomulate it.

Enables/disables the strict aliasing assumption, which assumes that objects of different types do not share the same location in memory.

I think it needs to say incompatible types at least (and char and unsigned char are also special). And isn't it really about pointers -- the compiler assumes that when dereferncing two pointers of incompatible types, those pointers do not refer to the same memory?

clang does not allow "type-punning" by writing and reading from different union members

I thought clang does allow type-punning through unions, as long as it's in a single function, but that it fails when things get more complicated.

+dannyb who enjoys this stuff ;)

[• dberlin]

In D30538#690699, @hans wrote:

+1 for documenting this, but I have to leave it to the language lawyers for how to fomulate it.

Enables/disables the strict aliasing assumption, which assumes that objects of different types do not share the same location in memory.

I think it needs to say incompatible types at least (and char and unsigned char are also special). And isn't it really about pointers -- the compiler assumes that when dereferncing two pointers of incompatible types, those pointers do not refer to the same memory?

clang does not allow "type-punning" by writing and reading from different union members

I thought clang does allow type-punning through unions, as long as it's in a single function, but that it fails when things get more complicated.

The limitation is really "union accesses must be visibly through a union, and if you try to trick the compiler, you will lose".
That is what we meant to allow, it just is still broken *anyway*

+dannyb who enjoys this stuff ;)

You are so mean :)

[dim]

LGTM, though it would be nice if we implemented gcc's extension for this at some point... ;)

[simonbyrne]

Fix capitalisation.

[carlosgalvezp]

Hi, is this commit still valid? Why hasn't it been pushed?

[simonbyrne]

i have no idea. It looks like it may have been resolved in https://reviews.llvm.org/rL303851? In that case, the NOTE should be removed.

docs/UsersManual.rst

    the behavior of sanitizers in the ``cfi`` group to allow checking
    of cross-DSO virtual and indirect calls.
 
+.. option:: -f[no-]strict-aliasing
+
+   Enables/disables the strict aliasing assumption, which assumes that
+   objects of different types do not share the same location in memory.
+
+   NOTE: Unlike gcc, clang does not allow "type-punning" by writing and
+   reading from different union members without `-fno-strict-aliasing`
+   enabled.
 
 .. option:: -fstrict-vtable-pointers