Fix __builtin_verbose_trap support for GPU and CPU

[amd-bfilipov]

Ticket: AIROCGDB-558

Summary

Implements complete support for __builtin_verbose_trap on both AMD GPU and CPU (x86_64), making verbose trap diagnostic messages visible in backtraces.

Problem

When __builtin_verbose_trap fired, the diagnostic inline frames were not visible in backtraces, making it impossible to see the trap category and message. Two issues prevented this:

1. PC pointing past trap instruction: After s_trap 2 on GPU, PC was not rewound to point at the trap instruction, causing GDB to miss the inlined DWARF frame containing the diagnostic info
2. Inline frames hidden by default: GDB's skip_inline_frames() intentionally hides inline frames to improve stepping experience, but this also hid the verbose trap frames

Solution

This PR provides a two-part fix:

1. PC Adjustment for s_trap 2 (GPU)

Rewind PC by 4 bytes when a wave stops with s_trap 2, similar to breakpoint handling. This ensures the PC remains in the inlined DWARF frame containing __builtin_verbose_trap debug info.

Files changed:

• gdb/amd-dbgapi-target.c - PC adjustment logic
• gdb/testsuite/gdb.rocm/builtin_verbose_trap.{cpp,exp} - Test coverage

2. gdbarch Hook for Inline Frame Control (GPU + CPU)

Adds architecture-specific control over inline frame visibility using GDB's gdbarch pattern. Architectures can now prevent skipping of compiler-generated inline frames containing diagnostic information.

Changes:

• gdb/gdbarch_components.py - Define show_verbose_trap_inline_frame method
• gdb/arch-utils.{c,h} - Default implementation (preserves current behavior)
• gdb/amdgpu-tdep.c - AMD GPU implementation (checks SIGABRT + name pattern)
• gdb/amd64-linux-tdep.c - CPU x86_64 implementation (checks SIGILL + name pattern)
• gdb/inline-frame.c - Call gdbarch hook in skip loop
• gdb/gdbarch-gen.{c,h} - Auto-generated

Detection logic:

• Only shows frame when both conditions met:
  1. Stop signal matches trap type (SIGABRT for GPU s_trap 2, SIGILL for CPU ud2)
  2. Symbol name starts with __clang_trap_msg$<category>$<message>
• Preserves normal stepping behavior (frames hidden when user steps into them)

Results

Before:

Thread 6 received signal SIGABRT, Aborted.
#0  test_trap_kernel () at test.cpp:24

After (GPU):

Thread 6 received signal SIGABRT, Aborted.
#0  __clang_trap_msg$check verbose$This is verbose trap! ()
    at test.cpp:23
#1  test_trap_kernel () at test.cpp:24

After (CPU):

Program received signal SIGILL, Illegal instruction.
#0  __clang_trap_msg$check negative$Value must be non-negative! ()
    at test.cpp:4
#1  check_value (value=-1) at test.cpp:5
#2  main () at test.cpp:12

[aktemur]

Please squash the second patch to the second, so that the test accompanies the fix.

[lancesix]

The test is now using __builtin_verbose_trap, this test is about that.

[lancesix]

The main thing we'd want to test here is wether the current frame gives us the trap, i.e. does your message show un bt or similar message.

The expectation is to see that in the message with the SIGABRT signal, saying where we stopped.

[lancesix]

The test should exercise the stepping as well. And since you add host support, I suppose this should have a test as well.

[amd-bfilipov]

The test should exercise the stepping as well. And since you add host support, I suppose this should have a test as well.

Ok. Added another inline function to the test, so the test steps over it and checks whether regular inline frames stay hidden.
Added a cpu version of the test. It's in a separate commit but we can rearrange that if needed.

[jhuber6]

I'm not intimately familiar with gdb, but the lldb output seems much more descriptive here. Any chance we can do something similar? From the example with -fsanitize-runtime=trap -fsanitize=undefined in Clang.

(lldb) run
Process 302532 launched: 'llvm-project/build/a.out' (x86_64)
Process 302532 stopped
* thread #1, name = 'a.out', stop reason = Undefined Behavior Sanitizer: signed integer addition overflow in 'x + 1'
       frame #1: 0x0000555555555750 a.out`signed_overflow(x=2147483647) at ubsan_test.c:4:39
   1     #include <limits.h>
   2     #include <stdio.h>
   3     
-> 4     int signed_overflow(int x) { return x + 1; }
   5     
   6     int shift_oob(int x, int amount) { return x << amount; }
   7     

[lancesix]

FYI, I am not sure this should be the way to go.

I have ROCm/rocm-systems#9776 to change the trap handler behaviour so this is not necessary. Such change would fix other issues reported independent to this PR.

[palves]

Re. the "gdb: Add gdbarch hook to show verbose trap inline frames" patch:

The commit log needs to be turned upside down.

This adds the gdbarch hook, yes, but it needs to begin by stating what the problem is, talk about __builtin_verbose_trap, show what goes wrong today without the fix.

I.e., establish rationale for why a change is needed.

Explain that this affects both CPU and GPU.

Then state how we're fixing this. The new gdbarch hook.