gdb/testsuite: support optimized code in gdb.rocm testsuite

[spatrang]

Summary

This series makes the gdb.rocm testsuite pass when the test programs are
compiled with optimization (-O1/-O2/-O3, -flto) passed via
GDB_TESTCASE_OPTIONS, rather than the implicit -O0 the suite assumes today.

Three commits, split by concern:

1. Update gdb.rocm framework to support optimized code — adds helpers in
   gdb/testsuite/lib/rocm.exp:

   • check_for_any_flags {flags} — detects whether GDB_TESTCASE_OPTIONS
     contains any of the listed flags.
   • gdb_caching_proc no_optimized_code — predicate for -O1/-O2/-O3.
   • gdb_caching_proc no_lto — predicate for -flto.

2. Update some gdb.rocm tests to support optimized code (22 files) —
   two patterns applied surgically:

   • In .cpp/.c: add volatile on locals the test inspects, and
     __attribute__((optnone)) on helper functions the test breakpoints into.
     Add a trivial volatile int statement to empty "break here" functions
     so their body is not elided.
   • In .exp: relax line-number arithmetic, backtrace PC formatting,
     expected source strings after step, and add a bounded retry loop for
     extra ticks under optimization.

3. Skip unsupported tests in case of optimized code (19 files) — for tests
   whose semantics cannot be reconciled with optimized code, add
   require no_optimized_code at the top; mi-aspace.exp also gets
   require no_lto. register-watchpoint.exp soft-skips the tail via
   check_for_any_flags after the earlier portion has run.

Diffstat: 42 files changed, +146 / -46.

Context

This continues the work started in internal PR
AMD-ROCm-Internal/ROCgdb#6,
rebased onto the current amd-staging of the public repo. The commits are
content-identical to the internal PR, preserving the original authors
(Aleksandar Rikalo, Andrey Kasaurov) with their Co-Authored-By: trailers
and adding my Signed-off-by.

[spatrang]

Hi @lancesix

This PR is a re-submission of the optimized-code series that was originally
opened in the internal rocgdb repository. I'm summarising your comments and
how they've been addressed in the latest push:

1. gdb.rocm/deep-stack.exp — "This part of the change seems odd. Is
   it really necessary?"
   Reverted. Frames #10..#12 now use the same ".* <name>" regex
   style as #0..#9; the dropped space was unnecessary.

2. gdb.rocm/device-attach.cpp — "Is code still valid without this
   line?"
   No, it isn't. Restored the __global__ void qualifier on
   bit_extract_kernel; without it the function isn't a HIP kernel
   and hipLaunchKernelGGL would fail to compile.

3. gdb.rocm/hip-builtin-variables.cpp — "Not a fan of the implicit
   int … const volatile reads off … Looks like it could just be
   volatile int."
   Agreed. All const volatile X = …; declarators have been replaced
   with volatile int X = …;.

4. gdb.rocm/precise-memory-exec.c — "Why volatile int sometimes,
   attribute((optnone)) other times?"
   Good point. second() now uses __attribute__((optnone)) instead
   of an artificial volatile int i = 0;, matching the convention
   used in corefile.cpp, names.cpp, ocp_mx.cpp, and
   step-schedlock-spurious-waves.cpp (touched in the same series).

5. lib/rocm.exp — "missing '.'" (×2) and "indentation" (×2)
   on no_optimized_code and no_lto.
   Both comments were added a trailing period, and the body of the
   if-blocks in both procs is now tab-indented to match the
   surrounding style.

The corresponding fix-up commit is on top of this branch. Please take
another look when you have a moment

cc: @lumachad

[lumachad]

pre-commit seems to be grumpy.

[lumachad]

As stated elsewhere, the series only handles GDB_TESTCASE_OPTIONS. If someone passes CFLAGS directly to the compiler the optimization level checks won't work properly. Do we want to fix that as well? Maybe as a follow up if not in this one?

[lumachad]

There's a pre-commit failure.

Other than that, I'm still seeing FAIL's for gdb.rocm/watchpoint-basic.exp when running the testsuite with -O1 or -O3. It passes when running with -O0. Could you please check that you're running the testsuite correctly?

[lumachad]

Did the latest force-pushes provide anything that would fix gdb.rocm/watchpoint-basic.exp? I can't tell.

[lumachad]

Just small issues.

[lumachad]

For -lto runs, it looks like gdb.rocm/hip-builtin-variables.exp and gdb.rocm/hip-lang-detect.exp are still failing.

FAIL: gdb.rocm/hip-lang-detect.exp: language detection (dwarfdump execution)
FAIL: gdb.rocm/hip-lang-detect.exp: expected compiler language (found "unknown")

[spatrang]

For -lto runs, it looks like gdb.rocm/hip-builtin-variables.exp and gdb.rocm/hip-lang-detect.exp are still failing.

FAIL: gdb.rocm/hip-lang-detect.exp: language detection (dwarfdump execution)
FAIL: gdb.rocm/hip-lang-detect.exp: expected compiler language (found "unknown")

Both hip-builtin-variables.exp and hip-lang-detect.exp now require no_lto. LTO can re-inline across TUs (defeating noinline/optnone boundaries) and rewrite the per-CU DW_LANG attribute, so the test premises don't hold at -flto. Matches the existing require no_lto pattern in watchpoint-at-end-of-shader.exp and mi-aspace.exp.

[lumachad]

Only had a minor comment. Otherwise this LGTM. I'd like @lancesix to eyeball this as well.

[lumachad]

Things look clean from -O1 through -O3. I do see a FAIL with -flto though:

FAIL: gdb.rocm/hip-builtin-completions.exp: language detection (dwarfdump execution)

Since we claim to be clean on -flto runs as well, we should fix that or skip it.

[lumachad]

Confirmed a clean run for -flto as well. Thanks.

[lumachad]

Since we're doing a new pass. Please cleanup the following:

Inconsistent spacing around attribute across the PR.

• gdb/testsuite/gdb.rocm/hip-builtin-variables.cpp:37,45,52,71 — attribute((optnone))
• gdb/testsuite/gdb.rocm/lane-info.cpp:38,49,53 — attribute((noinline)), attribute((noinline, disable_tail_calls))
• gdb/testsuite/gdb.rocm/line-breakpoint-in-kernel.cpp:34 — attribute((optnone))
• gdb/testsuite/gdb.rocm/multi-inferior-fork.cpp:39 — attribute((optnone))
• gdb/testsuite/gdb.rocm/multi-inferior-gpu.cpp:35 — attribute((optnone))
• gdb/testsuite/gdb.rocm/names.cpp:127 — attribute((optnone))
• gdb/testsuite/gdb.rocm/ocp_mx.cpp — five occurrences of attribute((optnone))
• gdb/testsuite/gdb.rocm/scheduler-locking.cpp:36 — attribute((optnone))
• gdb/testsuite/gdb.rocm/static-global.cpp:40 — attribute((optnone))
• gdb/testsuite/gdb.rocm/step-schedlock-spurious-waves.cpp:27 — attribute((optnone))
• gdb/testsuite/gdb.rocm/unaligned-memory-access.cpp:52,57 — attribute((noinline)), attribute((optnone))
• gdb/testsuite/gdb.rocm/watch-gpu-global-from-host.cpp:27 — attribute((optnone))
• gdb/testsuite/gdb.rocm/finish.cpp:225 — new attribute((optnone)) (still missing space, only on the kernel declaration)
• Fix: normalise to attribute ((...)) everywhere added by the patch.

Typos / Comments

• Spelling: inconsistent British/American mix introduced by the PR. New comments use optimiser, optimisation, initialisation, optimised in some files (alu-exceptions.cpp, fork-exec-.cpp, precise-memory-.c, deref-scoped-pointer.exp, hip-builtin-variables.exp,
  watchpoint-basic.exp) and optimization/optimized in others (the new lib/rocm.exp predicates, static-global.cpp, several require no_optimized_code skips). Existing gdb tree leans American (optimize); pick one for the whole PR.
• gdb/testsuite/gdb.rocm/static-global.cpp:37–39 — the new comment explains why optnone is used but doesn't explain the unrelated change on the same line from static to non-static. The companion file deref-scoped-pointer.cpp:35–37 does call out "Non-static so a line
  breakpoint…". Add the same rationale here for consistency.

Bugs

• gdb/testsuite/gdb.rocm/watch-gpu-global-from-host.cpp:34 — silently downgrades a C++ cast to a C-style cast:

  -  int *devGlobal;
  -  if (hipGetSymbolAddress (reinterpret_cast<void **> (&devGlobal), global))
  +  volatile int *devGlobal;
  +  if (hipGetSymbolAddress ((void **) &devGlobal, global))

• The cast change is presumably because reinterpret_cast<void**> won't strip the new volatile qualifier — but the C-style cast hides exactly that fact. This is the kind of cast that the GNU/ROCgdb style explicitly discourages in C++ code. Better: keep an explicit const_cast/reinterpret_cast pair so the volatile strip is visible, or store devGlobal as non-volatile and only mark the access path volatile.
• gdb/testsuite/gdb.rocm/lane-info.cpp:51–55 — the trailing asm volatile ("" : : : "memory"); in bar() is unreachable because sleep_forever() is an endless loop (volatile bool keep_going = true; while (keep_going) …). The author's stated rationale ("volatile asm prevents the call from being constant-folded away at the call site") refers to foo(), where it makes sense. In bar() the disable_tail_calls attribute is the load-bearing piece; the asm is dead. Either drop the asm in bar() and rely on disable_tail_calls, or annotate sleep_forever as attribute ((noreturn)) and remove the dead asm.
• gdb/testsuite/lib/rocm.exp check_for_any_flags (lines ~165–200) — uses lsearch -exact $src $flag, which relies on Tcl's string→list coercion of $src. If a future caller puts a flag with {} or unbalanced quotes into GDB_TESTCASE_OPTIONS/CFLAGS_FOR_TARGET, the implicit list conversion throws. Defensive form: if {[lsearch -exact [split $src] $flag] != -1} — split with no separators yields whitespace-tokenised words and avoids the auto-list coercion. Same logic, harder to break later.

[lumachad]

Sent offline comments.

[lumachad]

Just one nit this time:

• gdb/testsuite/gdb.rocm/until-tests.exp:76-79 — the test was: until lands on exactly breakpoint_loc+1. After this patch it accepts +1 or +2 unconditionally, even at -O0. That weakens the assertion for the non-optimized path it used to cover. Gate the loose pattern on
  [check_for_any_flags {-O1 -O2 -O3}] and keep the strict +1 check for the unoptimized build.

I'd give Lancelot a chance to go through it again.

[spatrang]

Just one nit this time:

• gdb/testsuite/gdb.rocm/until-tests.exp:76-79 — the test was: until lands on exactly breakpoint_loc+1. After this patch it accepts +1 or +2 unconditionally, even at -O0. That weakens the assertion for the non-optimized path it used to cover. Gate the loose pattern on
  [check_for_any_flags {-O1 -O2 -O3}] and keep the strict +1 check for the unoptimized build.

I'd give Lancelot a chance to go through it again.

Thanks! Pushed 86c2e23: gated the (+1|+2) alternation on check_for_any_flags {-O1 -O2 -O3}, restored the strict +1 check at -O0. until-tests.exp 28/28 PASS at -O0 and -O2.

[lancesix]

Note that this time I only looked at the overall diff, but having individual commits is still important, I'll check this on next round.

[aktemur]

I'm having some difficulty following the changes introduced by patches. Some patches should've been fixups -- they modify previous changes introduced by patches in the same PR. Can we please split, re-order, re-organize the patches so that the changes are focused?

[lancesix]

Hi,

I did not really re-read evertything, but it seems you push commits to this branch to fix previous commits of the branch still under review (gdb/testsuite/gdb.rocm: drop redundant test decorations for example).

We do not squash commits in GDB, but rather craft series and refine patches in the series as needed. The bits you change in that last patch have to be integrated into previous patches.

[spatrang]

Commit refactor (force-pushed). Squashed the series into two single-purpose commits:

1. gdb/testsuite: update gdb.rocm framework to support optimized code -- the lib/rocm.exp helpers (check_for_any_flags, no_optimized_code, no_lto).
2. gdb/testsuite: update gdb.rocm tests to support optimized code -- all test source/expectation adjustments plus the require no_optimized_code / require no_lto gates.

Folded in this review round's feedback:

• Commit titles now carry the gdb/testsuite: label (@aktemur).
• check_for_any_flags uses [llength $flags] == 0 instead of a boolean test.
• mi-aspace.exp combines the two predicates into require no_optimized_code no_lto.
• no_optimized_code now also catches -O, -Os, -Ofast, -Og (whole-token match).
• Trailing periods added to the verbose -log strings.

The numerous "this hunk should've been a fixup of the originating patch" comments are addressed by the squash itself -- the intra-file churn from the old multi-commit history is collapsed into these two logical commits.

Verified deep-stack.exp passes at -O3 (full #0..#12 backtrace).

One open question: should the Change-Id trailers be kept (amd-staging / Gerrit convention) or dropped? Happy to do either.

[spatrang]

Rebased onto the latest amd-staging.

The only conflicts were in the require lines of the gated .exp tests: amd-staging renamed allow_hipcc_tests -> allow_hip_tests. Resolved by taking the new proc name and keeping the require no_optimized_code / require no_lto gates this series adds.

Re-tested the full gdb.rocm/*.exp suite on the rebased branch (freshly built rocgdb), no unexpected failures at any level:

Level	pass	fail	unsupported
-O0	1891	0	8
-O1	1456	0	27
-O2	1456	0	27
-O3	1456	0	27

The lower pass count at -O1+ is the expected effect of the require no_optimized_code / require no_lto gates. Remaining XFAIL/UNTESTED entries (finish, runtime-core, ocp_mx) are pre-existing environment limitations unrelated to this series.