[cfe-dev] Unaligned loads

[John McCall]

On Mar 30, 2011, at 12:10 PM, Damien Vincent wrote:
> I generated the llvm code for the following simple function (based on clang/llvm 2.8) using:
> clang -S -O2 -emit-llvm memory_unalignedint.c
> 
> int f(char *ptr)
> {
>   int r = *((int*)(ptr+2));
>   return r;
> }
> 
> And I was wondering why the alignment for the load is set to 4 although the load is not aligned ? (%tmp1 = load i32* %0, align 4)

This load is required to be 4-byte aligned;  the fact that the address is computed with pointer arithmetic doesn't matter.  The standards-compliant way of loading an unaligned object is memcpy, which LLVM will happily optimize to an unaligned load/store if possible.

We also support the __attribute__((aligned(N))) GCC extension to do this more naturally;  this would look something like:

  typedef int unaligned_int __attribute__((aligned(1)));
  int r = *((unaligned_int*) (ptr+2));

However, we have a pretty large number of bugs in this area, and in particular the load here does not seem to get emitted with align 1.

John.