RFR: 8347422: Crash during safepoint handler execution with -XX:+UseAPX

[Vladimir Kozlov]

On Fri, 10 Jan 2025 13:24:21 GMT, Jatin Bhateja <jbhateja at openjdk.org> wrote:

> This bug fix patch fixes an internal error seen during safepoint handler execution.
> The problem occurs due to missing handling for REX2 prefixed polling test instruction.
> 
> Manually verified the patch with the -XX:+SafepointALot runtime flag.
> 
> Best Regards,
> Jatin
> PS: Patch will be opened for review after some validation.

I have few comments.

src/hotspot/cpu/x86/sharedRuntime_x86_64.cpp line 3129:

> 3127:     //  o  In case BASE register is one of the 32 extended GPR registers available only on targets
> 3128:     //     supporting Intel APX extension, then we need to emit two byte REX2 prefix to hold
> 3129:     //     most significant two bits of 5 bit register encoding.

Change `two byte` -> `two bytes` in two places.

src/hotspot/cpu/x86/sharedRuntime_x86_64.cpp line 3133:

> 3131:     if (VM_Version::supports_apx_f()) {
> 3132:       __ cmpb(Address(rbx, 0), Assembler::REX2);
> 3133:       __ jcc(Assembler::notEqual, check_rex_prefix);

Why we not using short branch `jccb` here and in following code? We skip only `addptr` instruction.

On other hand jccb(Assembler::notEqual, no_adjust) at line 3093 above may be far from target after new code is added. May be use jcc there.

-------------

PR Review: https://git.openjdk.org/jdk/pull/23035#pullrequestreview-2556850168
PR Review Comment: https://git.openjdk.org/jdk/pull/23035#discussion_r1918967338
PR Review Comment: https://git.openjdk.org/jdk/pull/23035#discussion_r1918968951