[WASM] Suboptimal `memcmp` expansion for odd lengths

[MaxGraey]

Example (Godbolt)

Calls to memcmp(a, b, n) == 0, where n is odd is suboptimal for wasm32 target where such call didn't lower to expansion and calls memcmp routine even for small lengths (n < 16)

Here codegen for memcmp(a, "abc", 3) == 0:

x64

movzx   eax, word ptr [rdi]
xor     eax, 25185
movzx   ecx, byte ptr [rdi + 2]
xor     ecx, 99
or      cx, ax
sete    al
ret

wasm32 (actual)

local.get  0
i32.const  .L.str.2
i32.const  3
call       memcmp
i32.eqz

wasm32 (expected)

local.get    0
i32.load16_u 0     ;;align=1 ???
i32.const    25185
i32.xor 
local.get    0
i32.load8_u  2
i32.const    99
i32.xor 
i32.or  
i32.eqz

[llvmbot]

@llvm/issue-subscribers-backend-webassembly

[junparser]

we shall implement TTI hook enableMemCmpExpansion for wasm which enables mergeicmps and expandmemcmp pass

[ParkHanbum]

Taking note of what @junparser said, I implemented the code as follows.

WebAssemblyTTIImpl::TTI::MemCmpExpansionOptions
WebAssemblyTTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const {
  TTI::MemCmpExpansionOptions Options;
  Options.LoadSizes = {8, 4, 2, 1};
  Options.MaxNumLoads = TLI->getMaxExpandSizeMemcmp(OptSize);
  return Options;
}

Is it okay to submit this as is?

[badumbatish]

hi @ParkHanbum, have you submitted the snippet as a PR to the issue? if you haven't, and if you don't mind, I'd like to take up the issue

[ParkHanbum]

@badumbatish sure, I'm not on work