1312138 - Reduce code size when zeroing out memory.

Status: NEW

(Core :: JavaScript Engine: JIT, defect, P5)

Description

[Michael Bebenita [:mbx]]

We emit a lot of code to zero out memory in ION (WebAssembly). For instance, in AngryBots, there are 45,161 instructions of the form:

mov dword ptr [r15 + reg + ##], 0

This is encoded as [41 c7 44 3f ## 00 00 00 00] which is 9 bytes, so 406,449 bytes or ~ 5% (out of 10M). Since there is no 8-bit immediate mov, perhaps we can zero out a temp register and use that instead to save 2 bytes.

xor rbx, rbx
mov dword ptr [r15 + reg, ##], rbx


Another common pattern is:

mov dword ptr [r15 + rax], 0
mov dword ptr [r15 + rax + 4], 0
mov dword ptr [r15 + rax + 8], 0

There are 1247 matches of this exact pattern in AngryBots. And lots more if you consider a more general pattern:

mov dword ptr [r15 + reg + (i    ) * 4], 0
mov dword ptr [r15 + reg + (i + 1) * 4], 0
mov dword ptr [r15 + reg + (i + 2) * 4], 0
mov dword ptr [r15 + reg + (i + n) * 4], 0

Can we optimize this?

Comment 1

[Luke Wagner [:luke]]

The former is an interesting question.  It'd be nice if, in Lowering, you could ask for a register that could be denied if it would cause register pressure.  Otherwise, it's hard to weigh the tradeoff of *always* requiring a register to zero or never.  Poking around, gcc and clang don't seem to do anything fancier here.  (Although clang does convert grouped stores of 0 to SIMD stores.)

Bug 897425 would help with the latter case so we could expose the redundant (HeapReg + reg) computation.