[LLVMdev] Unrolling an arithmetic expression inside a loop

[Duncan Sands]

Hi Juliusz,

> 2) In exec1 however it fails to recognize that the temporary variables
> are not reused anywhere and fails to simplify the arithmetic expression,
> producing:
>      %scevgep = getelementptr i32* %X, i64 %indvar
>      %scevgep4 = getelementptr i32* %Y, i64 %indvar
>      %scevgep5 = getelementptr i32* %res, i64 %indvar
>      %3 = load i32* %scevgep, align 4, !tbaa !0
>      %4 = load i32* %scevgep4, align 4, !tbaa !0
>      %5 = add nsw i32 %4, %3
>      %6 = mul nsw i32 %5, %4
>      %7 = mul nsw i32 %4, %4
>      %8 = sub i32 %6, %7
>      store i32 %8, i32* %scevgep5, align 4, !tbaa !0
>
> It is interesting, that I get better compilation output when inputting
> clang with worse input source (creating redundant temporary tables on
> stack).

this is now fixed in the development branch of LLVM.

Ciao,

Duncan.