1710087 - ARM64: Use the zero register for zero constants when possible

Status: NEW

(Core :: JavaScript: WebAssembly, enhancement, P3)

Description

[Lars T Hansen [:lth]]

This may be tricky to do except in limited contexts because of the risk of confusing the zero register with the stack pointer, and how often is the zero value used for anything but initialization anyway, but in principle when a zero value is needed we should use the zero register.

Consider

  (func $f1 (param $p i32)
    (i32.store (local.get $p) (i32.const 0)))

The code for this is:

0x13cf8746f034  52800001  mov     w1, #0x0
0x13cf8746f038  b8206aa1  str     w1, [x21, x0]

where the first instruction is not necessary, the second could just store wzr.

Similarly, in bug 1709863 the zero value shows up in an addressing instruction, and though that addressing code should be fixed in other ways and the particular issue will go away, the situation is the same: a zero literal should not need a dedicated register.

Comment 1

[Lars T Hansen [:lth]]

Integer negate has a redundant zero when expressed as 0-n instead of n*-1:

0x1e381ef2b084  52800001  mov     w1, #0x0
0x1e381ef2b088  4b000020  sub     w0, w1, w0

We probably want to handle this case during lowering and generate an LNegI node (we should do this cross-platform; x86/x64 have the same problem), and we want to check that a redundant zero is not being generated in this case.

(Addressed in bug 1710403.)