On Mon, 30 Mar 2026 14:29:11 GMT, Aleksey Shipilev <[email protected]> wrote:
> And are we sure on AArch64 it does not? I don't see a similar change there.
I'm sure. AArch64 uses `cbz`.
> I see at least this in `LIR_Assembler::increment_profile_ctr` on AArch64:
>
> ```
> if (!step->is_constant()) {
> // If step is 0, make sure the stub check below always fails
> __ cmp(step->as_register(), (u1)0);
> __ mov(rscratch1, InvocationCounter::count_increment *
> ProfileCaptureRatio);
> __ csel(dest, dest, rscratch1, __ NE);
> }
> ```
That snippet above is executed only for counters that may overflow.
> Also ugh, doing two tests just to restore the flags. But that one seems
> unavoidable.
I think so. There seems to be no better way to do it.
> src/hotspot/cpu/arm/c1_MacroAssembler_arm.cpp line 256:
>
>> 254:
>> 255: void C1_MacroAssembler::step_random(Register state, Register temp,
>> Register data) {
>> 256: // if (VM_Version::supports_crc32()) {
>
> Can be just deleted, IMO.
I guess so. There seems no way to test the feature flags on Arm 32.
> src/hotspot/share/c1/c1_LIR.cpp line 1082:
>
>> 1080: (_step, _result, _freq_op,
>> 1081: _md_reg, _md_op, _md_offset_op, _overflow_stub);
>> 1082: if (overflow_stub()) {
>
> `overflow_stub() != nullptr`
Yuck. OK. :-)
> src/hotspot/share/compiler/compiler_globals.hpp line 391:
>
>> 389: "information at the bailout point")
>> \
>> 390:
>> \
>> 391: product(int, ProfileCaptureRatio, 1, EXPERIMENTAL,
>> \
>
> Put it at `64`, if you believe this is a right tradeoff.
This is an experimental patch. I don't think I should turn an experiment on by
default.
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/28541#discussion_r3010816929
PR Review Comment: https://git.openjdk.org/jdk/pull/28541#discussion_r3010818885
PR Review Comment: https://git.openjdk.org/jdk/pull/28541#discussion_r3010823337
PR Review Comment: https://git.openjdk.org/jdk/pull/28541#discussion_r3010828593
