[llvm-commits] [llvm-gcc-4.2] r62076 - /llvm-gcc-4.2/trunk/gcc/llvm-convert.cpp

[Duncan Sands]

Hi Evan,

> Load / store alignments should be the alignments of the memory object used to construct the LValue, not the alignments of the values being loaded / stored.

I've reverted this because it was causing several failures in the
testsuite, and also broke the Ada build.  Also, I don't see how this
approach can work without major changes.  Consider the following part
of your patch:

> @@ -5956,8 +5975,10 @@
...
>    } else {
>      ArrayAddr = Emit(Array, 0);
> +    ArrayAlign = expr_align(ArrayTreeType) / 8;
>    }

This is the case of when we are doing "a[i]", where "a" is a pointer.
There's no reason to think that "a" points to aligned memory, so
using the type alignment is wrong.  So where to get the alignment
from?  We get the pointer "a" from Emit, which doesn't provide alignment
info for the pointee.  As far as I can see either Emit needs to return
more info (complicated to implement) or we need to dig down into the gcc
tree to get the underlying object and extract its alignment.  In other
words, we need to do what we expected expr_align to do (only it wasn't).
I think we should just make expr_align more intelligent.  Since we are the
only users of it, there is no danger involved in changing it.

Ciao,

Duncan.

PS: I didn't check, but presumably this same problem occurs elsewhere:
we get a pointer from Emit and need alignment for the pointee.