Stream: git-wasmtime

Topic: wasmtime / issue #14699 Cranelift: missed optimization Im...

Wasmtime GitHub notifications bot (Oct 13 2026 at 13:29):

Takashiidobe opened issue #14699:

Feature

Optimization of unsigned constant-remainder tests:

Take a look at: https://godbolt.org/z/MPfdhbc3q (divide x by constant, in this case 7, and then test it).

for signed/unsigned x, and for == and !=, rustc on llvm emits a short sequence, e.g. for the != case

is_multiple_of_7:
        imul    eax, edi, -1227133513
        cmp     eax, 613566757
        setae   al
        ret

The corresponding cranelift side is much longer:

is_multiple_of_7:
     mov    rax, rdi
     mul    dword ptr [rip + 0x2b]  # 0x38 <is_multiple_of_7+0x38>
     mov    r10, rdi
     sub    r10d, edx
     shr    r10d
     add    r10d, edx
     shr    r10d, 0x2
     imul   r11d, r10d, 0x7
     sub    edi, r11d
     test   edi, edi
     setne  al
     movzx  rax, al

Benefit

llvm-mca says this is 8 cycles for llvm or 20 for cranelift on a zen4. There is a similar disparity for arm64 and riscv64.

Implementation

Not sure. LLVM implements this in target independent SDAG, maybe there's a similar target independent ISLE fold that would be a good place for it?

extra note: make sure to handle signed vs unsigned, ne and eq, and odd and even numbers.

Alternatives

• We could leave this as is. I think this is a mildly popular pattern, like for byte aligning so I'd like to see it improved.
• Would need to look into cranelift more for other places to put this optimization. It could be target dependent but I assume this optimization should work the same for any target since it's a strength reduction.
  

Wasmtime GitHub notifications bot (Oct 13 2026 at 14:19):

alexcrichton added the cranelift label to Issue #14699.

Wasmtime GitHub notifications bot (Oct 13 2026 at 14:19):

alexcrichton added the performance label to Issue #14699.

Wasmtime GitHub notifications bot (Oct 13 2026 at 14:19):

alexcrichton added the cranelift:goal:optimize-speed label to Issue #14699.

Wasmtime GitHub notifications bot (Oct 13 2026 at 14:19):

alexcrichton removed the performance label from Issue #14699.

Wasmtime GitHub notifications bot (Oct 16 2026 at 20:28):

fitzgen commented on issue #14699:

It is possible that https://github.com/bytecodealliance/wasmtime/blob/ce2794bbf28c4ae1d2afb6620b2de17ab8e7b740/cranelift/codegen/src/opts/div_const.rs is missing some cases or that our cost function for egraph extraction is not weighing things correctly. Needs more investigation.

Last updated: Oct 31 2026 at 23:03 UTC