[llvm] [SelectionDAG] Optimize 32-bit udiv with 33-bit magic constants on 64-bit targets (PR #181288)

[Craig Topper via llvm-commits]

topperc wrote:

> Oh, when I assembled that code with the original llc-18, it output the same code as the llc I proposed this time.
> 
> ```
> % llc-18 propose.ll -O3 -mattr=bmi2  -o -
>         .text
>         .file   "propose.ll"
>         .globl  udiv_by_19                      # -- Begin function udiv_by_19
>         .p2align        4, 0x90
>         .type   udiv_by_19, at function
> udiv_by_19:                             # @udiv_by_19
> # %bb.0:
>         movl    %edi, %edx
>         movabsq $970881267037344822, %rax       # imm = 0xD79435E50D79436
>         mulxq   %rax, %rax, %rax
>                                         # kill: def $eax killed $eax killed $rax
>         retq
> ```
> 
> However, the next code remained the same as the previous slow code.
> 
> ```c++
> uint32_t div7ext(uint32_t x)
> {
>         return uint32_t(uint64_t(x) / 7);
> }
> ```
> 
> ```shell
> % clang++-18 -O3 -march=native -S -masm=intel t.cpp -o -
> _Z7div7extj:
>         .cfi_startproc
> # %bb.0:
>         mov     eax, edi
>         imul    rax, rax, 613566757
>         shr     rax, 32
>         sub     edi, eax
>         shr     edi
>         add     eax, edi
>         shr     eax, 2
>                                         # kill: def $eax killed $eax killed $rax
>         ret
> ```

The middle end optimizations would remove the casts.

I think the constants between your approach and the zext have a similar magnitude and upper bits but differ in the lower bits.

https://github.com/llvm/llvm-project/pull/181288