On Mon, 24 Aug 2026 15:16:47 GMT, Boris Ulasevich <[email protected]>
wrote:
>> These fixes address a series of failures reproducible when building arm32
>> natively on an arm32 platform (not cross-compilation), in both release and
>> fastdebug modes.
>>
>> 1. jniFastGetField_arm.cpp: use offset_shift
>>
>> After JEP 401 was integrated, jfieldIDWorkaround's offset_shift changed from
>> 2 to 3. ARM32 is the only port where the shift was hardcoded as the literal
>> 2 in JNI_FastGetField::generate_fast_get_int_field0. Other architectures
>> take this value from the jfieldIDWorkaround::offset_shift constant and
>> picked up the change automatically. This caused java.io.FileDescriptor's int
>> fd field to be read incorrectly in the native fdval() function (IOUtil.c,
>> libnio.so), which made the JVM hang during module loading at the
>> COMPILE_CREATE_SYMBOLS step of the JDK build (fd was incorrectly decoded as
>> 0 -- a blocking read from stdin). Fix: use the
>> jfieldIDWorkaround::offset_shift constant instead of the hardcoded literal.
>>
>> 2. oops/symbol.hpp: class alignas(8) Symbol
>>
>> JEP 401 widened VerificationType::TypeMask from 0x3 to 0x7, reserving a
>> third bit for inline types. As a result is_reference() now requires a
>> Symbol* to be 8-byte aligned. On 32-bit a Symbol sits at offset 4 inside a
>> SymbolTable node, so is_reference() is always false and verification of any
>> class fails with VerifyError. This only happens in release builds: in debug
>> builds the node carries an extra DEBUG_ONLY field, which pushes the Symbol
>> to offset 8. Fix: alignas(8) makes Symbol 8-byte aligned. On 64-bit the
>> layout is unchanged.
>>
>> 3. make/autoconf/jdk-options.m4: disable cds-archive-preview for 32-bit
>> systems
>>
>> The build unconditionally generates a preview CDS archive, but value objects
>> are not supported on 32-bit: InlineKlass::InlineKlass hits assert "Should
>> not be called in 32 bit mode" (markWord.hpp), and the jdk-image step fails
>> while dumping classes_preview.jsa. Fix: make the cds-archive-preview option
>> unavailable on non-64-bit platforms.
>>
>> 4. cds/archiveUtils.cpp: additional check for minimal gap size
>>
>> This fixes a fastdebug build failure at the generate-link-opt-data step:
>> assert(is_aligned(gap_bytes(), SharedSpaceObjectAlignment)) failed. It is
>> related to ee90f00b3b3 (8376822, UseCompactObjectHeaders: fill Klass
>> alignment gaps), not to JEP 401. That change started inserting an alignment
>> gap in front of a buffered Klass and recording it for later reuse. On arm32
>> the gap is always 4 bytes -- nothing can ever fit into it, and AllocGap
>> fails its precondition. Fix: onl...
>
> Boris Ulasevich has updated the pull request with a new target base due to a
> merge or a rebase. The incremental webrev excludes the unrelated changes
> brought in by the merge/rebase. The pull request contains four additional
> commits since the last revision:
>
> - C1: emit array access and acmp type profiles only when such profiling is
> enabled
> - allocate gap_bytes: move out of condition
> - Use CHECK_AVAILABLE to explain why cds-archive-preview is unavailable
> - 8390159: [ARM32] Native ARM32 build hangs in COMPILE_CREATE_SYMBOLS
Looks good to me otherwise. I'll run some testing and report back once it
passed.
src/hotspot/share/c1/c1_GraphBuilder.cpp line 1393:
> 1391: }
> 1392: if ((stream()->cur_bc() == Bytecodes::_if_acmpeq ||
> stream()->cur_bc() == Bytecodes::_if_acmpne) &&
> 1393: is_profiling() && profile_branches() &&
> MethodData::profile_acmp()) {
I think this should also go through `Compilation` and incorporate the
`is_profiling() && profile_branches()` checks.
-------------
Marked as reviewed by thartmann (Reviewer).
PR Review: https://git.openjdk.org/jdk/pull/32402#pullrequestreview-5039979040
PR Review Comment: https://git.openjdk.org/jdk/pull/32402#discussion_r3871065085
