offlineasm generates suboptimal code for load/store on ARM64

• Component: JavaScriptCore
• Status: RESOLVED FIXED

[Akos Kiss]

For every load/store instruction, where the address is in the form of base + offset, and the offset is out of range -255..4095, the ARM64 backend applies the standard risc lowering. That is, as the documentation of riscLowerMalformedAddresses describes:

loadp 10000[foo], bar

will become

move 10000, tmp
addp foo, tmp
loadp 0[tmp], bar

Which, in ARM64, means that, e.g.,

storep cfr, VM::topVMEntryFrame[vm]

becomes

movz x17, #45424, lsl #0
add x17, x17, x1
str x29, [x17, #0]

However, ARM64 can do better:

movz x17, #45424, lsl #0
str x29, [x1, x17, lsl #0]

This would achieve the same result with less instructions.

[Akos Kiss]

Created attachment 237682 [details]
Proposed patch.

[Michael Saboff]

Comment on attachment 237682 [details]
Proposed patch.

View in context: https://bugs.webkit.org/attachment.cgi?id=237682&action=review

r=me

> Source/JavaScriptCore/ChangeLog:10
> +        base to the temporary, and the change the load/store to use the

"the change" -> "then change"

> Source/JavaScriptCore/ChangeLog:13
> +        perform explicit register additons but it is enough to change load/store

Typo: "additions"

[Michael Saboff]

Comment on attachment 237682 [details]
Proposed patch.

r=me

[Akos Kiss]

Created attachment 237911 [details]
Updated patch

Typos fixed.

[Michael Saboff]

Comment on attachment 237911 [details]
Updated patch

r=me.

[WebKit Commit Bot]

Comment on attachment 237911 [details]
Updated patch

Clearing flags on attachment: 237911

Committed r173497: <http://trac.webkit.org/changeset/173497>

[WebKit Commit Bot]

All reviewed patches have been landed.  Closing bug.