[LLVMdev] Alignment of vectors

[Eli Friedman]

On Fri, Jul 18, 2008 at 6:45 AM, Benedict Gaster
<benedict.gaster at amd.com> wrote:
> Consider the following C code:
>
> typedef __attribute__(( ext_vector_type(2) )) float float2;
> typedef __attribute__(( ext_vector_type(2) )) __attribute__(( aligned(4) ))

AFAIK, the aligned attribute doesn't do anything on a typedef of
anything other than a struct/union type in either gcc or clang.  It
would be possible to implement something like this, but someone would
need to spec out exactly what it would do in various situations and
send the proposal past the gcc devs.

> float float2_align2;
>
> void foo(void)
> {
>   const float * p;
>   size_t offset;
>   float2 tmp = *((float2_align2 *)(p+offset));
> }
>
> When compiled with clang —emit-llvm I get:
>
> define void @foo() {
> entry:
>     %p = alloca float*, align 4        ; <float**> [#uses=1]
>     %offset = alloca i32, align 4        ; <i32*> [#uses=1]
>     %tmp = alloca <2 x float>, align 8        ; <<2 x float>*> [#uses=1]
>     %tmp1 = load float** %p        ; <float*> [#uses=1]
>     %tmp2 = load i32* %offset        ; <i32> [#uses=1]
>     %add.ptr = getelementptr float* %tmp1, i32 %tmp2        ; <float*>
> [#uses=1]
>     %conv = bitcast float* %add.ptr to <2 x float>*        ; <<2 x float>*>
> [#uses=1]
>     %tmp3 = load <2 x float>* %conv        ; <<2 x float>> [#uses=1]
>     store <2 x float> %tmp3, <2 x float>* %tmp
>     ret void
> }
>
> The problem is that the load into tmp3 seems to have lost any information
> that %conv should not be aligned to 8 bytes but rather 4. Of course, GCC
> only states that the alignment attribute will try and enforce a minimal
> alignment and so the above code generated by clang is valid

Right.

> but what about
> if the following code had been generated:
>
> define void @foo() {
> entry:
>     %p = alloca float*, align 4        ; <float**> [#uses=1]
>     %offset = alloca i32, align 4        ; <i32*> [#uses=1]
>     %tmp = alloca <2 x float>, align 4        ; <<2 x float>*> [#uses=1]
>     %tmp1 = load float** %p        ; <float*> [#uses=1]
>     %tmp2 = load i32* %offset        ; <i32> [#uses=1]
>     %add.ptr = getelementptr float* %tmp1, i32 %tmp2        ; <float*>
> [#uses=1]
>     %conv = bitcast float* %add.ptr to <2 x float>*        ; <<2 x float>*>
> [#uses=1]
>     %tmp3 = load <2 x float>* %conv        ; <<2 x float>> [#uses=1]
>     store <2 x float> %tmp3, <2 x float>* %tmp
>     ret void
> }

clang shouldn't be generating code like that.  Per the current gcc
documentation, the aligned attribute isn't allowed to decrease
alignment.  What would happen if the code subsequently passed a
pointer to the variable with the aligned attribute to another
function?

> Of course, it is now up to the backend to assure that an unaligned load is
> handled correctly in hardware, but assuming this is ok, then is this
> correct?

Unaligned loads should codegen correctly in LLVM; if they don't,
please file a bug.

-Eli