Bug 124721 – ARM64: Implement push/pop equivalents in LLInt

• Status: RESOLVED
• Resolution: FIXED
• Component: JavaScriptCore

[Michael Saboff]

ARM64 doesn't allow operation to the stack point that aren't 16 byte aligned.  Therefore push and pop need to be changed or abandoned.  One possibility is push and pop a pair of registers using the store pair (STP) and load pair (LDP) instructions.

[Geoffrey Garen]

Which push/pop instructions do you have in mind? 

Are you saying that the stack can't be misaligned even for a moment, such that two 8 byte pushes in a row would crash?

[Michael Saboff]

(In reply to comment #1)
> Which push/pop instructions do you have in mind? 
> 
> Are you saying that the stack can't be misaligned even for a moment, such that two 8 byte pushes in a row would crash?

ARM64 doesn't have push or pop instructions.  There are pre/post indexed memory ops.  The stack does need to be 16 byte aligned.  I don't know about whether temporary 8 byte alignment is fine.  There are load / store pair instructions which we use in the JSC macro assembler for pushToSave() and popToRestore().

There are LLInt push / pop macro ops.  They emit push and pop instructions, so I wondered if the clang assembler mapped those to the appropriate memory ops.  Turns out the it doesn't.  The current LLInt push / pop are currently unused, wrong and should result in a compile error.

When implementing the LLInt callToJavaScript(), I added {push,pop}CalleeSaves macro ops that does the right thing for each platform.  I followed after the push / pop macro op implementations.  Trying to work through the ARM64 compile issues I found there aren't ARM64 pseudo ops for push / pop.  I have reimplemented {push,pop}CalleeSaves using store / load pair and it appears to be working fine.

[Michael Saboff]

Created attachment 217628 [details]
Patch

[Michael Saboff]

Committed r159655: <http://trac.webkit.org/changeset/159655>

[Michael Saboff]

Reviewed in person by Phil Pizlo.

[Geoffrey Garen]

Does stp decrement the stack pointer register? Does ldp increment it? I'm worried that you're just storing/loading the same locations over and over again.

[Michael Saboff]

(In reply to comment #6)
> Does stp decrement the stack pointer register? Does ldp increment it? I'm worried that you're just storing/loading the same locations over and over again.

The stp is the pre-increment variety with an increment of -16.  That is what the "[sp, #-16]!" syntax means.
The ldp is the post-increment variety with an increment of 16.  The syntax for post increment has the increment amount after the closing ']': "[sp], #16".  A little confusing that's for sure.