Preserve device debug info for gdb-oneapi (#1004)

[pvelesko]

Adds llvm-patches/llvm/0007 for the configure_llvm.sh build path. The HIPSPV toolchain force-disabled device debug info (adjustDebugInfoKind -> NoDebugInfo), so -g/-O0 produced SPIR-V with no debug metadata and gdb-oneapi could not resolve locals in kernels. The historical blocker (translator aborting on DW_OP_LLVM_convert) is gone. The patch honors -g and, when set, emits the NonSemantic.Shader.DebugInfo form that Intel IGC/gdb-oneapi consume.

The same change is upstreamed in llvm/llvm-project#210504 and applied to the fork branch in CHIP-SPV/llvm-project#6.

Refs #1004

Hardware validation on Aurora (PVC, i915, oneAPI 2025.3, LLVM 22)

Full in-kernel debugging works, including reading arguments and locals per SIMD lane:

addKernel (a=0xff00fffffffe0000, b=0xff00fffffffd0000, c=0xff00fffffffc0000, n=256)
    at dbgtest.hip:9
n = 256   idx = 64   lhs = 64   rhs = 128
print a[idx]  ->  $4 = 64
thread :3     ->  idx = 70, lhs = 70

Getting there needs three changes. This PR is one of them; the other two are chipStar-side and are not yet written.

1. This patch: honor -g and request nonsemantic-shader-200 + SPV_KHR_non_semantic_info.
2. Add +SPV_INTEL_optnone to the same extension list. At -O0 clang marks kernels optnone; without the extension the translator silently drops it, so IGC optimizes the kernel and every variable reads . Needed only on the Triple::ChipStar branch upstream, whose allowlist excludes it; the generic branch already uses --spirv-ext=+all.
3. chipStar must stop discarding optnone and must pass -cl-opt-disable to zeModuleCreate:
   • llvm_passes/HipPasses.cpp RemoveNoInlineOptNoneAttrsPass strips optnone from every function unconditionally, before the translator runs. It should skip that when the module carries debug info (llvm.dbg.cu).
   • CHIPBackendLevel0::getDefaultJitFlags() returns "". It should yield -cl-opt-disable when the module has debug info.

Measured combinations, all with -g -O0 on PVC:

optnone in IR	SPV_INTEL_optnone	L0 build flags	locals
stripped	no	-cl-opt-disable -g	optimized out
stripped	yes	-cl-opt-disable -g	optimized out
preserved	yes	(empty, current default)	optimized out
preserved	yes	-g	optimized out
preserved	yes	-cl-opt-disable	readable

So -cl-opt-disable alone is not enough, and preserving optnone alone is not enough; both are required. -g in the L0 build flags is not needed.

Also confirmed: nonsemantic-shader-200 is the right choice. Testing nonsemantic-shader-100 changed nothing, and icpx -fsycl -g passes -spirv-debug-info-version=nonsemantic-shader-200 itself.

The -g binary runs correctly on PVC and there is no change when -g is absent.

Using gdb-oneapi with chipStar

Two prerequisites live outside the compiler entirely.

1. Enable EU debug on the node. It is off by default but writable by a normal user inside a job; no root needed:

   for f in /sys/class/drm/card*/prelim_enable_eu_debug; do echo 1 > "$f"; done
   

   Without this, Level Zero enumerates zero devices under ZET_ENABLE_PROGRAM_DEBUGGING=1 and chipStar fails with hipErrorNotInitialized.

2. Redirect gdb-oneapi's auto-attach hook. Its intelgt_auto_attach.py plants the trigger breakpoint on zeContextCreate, but chipStar calls zeContextCreateEx (src/backend/Level0/CHIPBackendLevel0.cc:2368). Different address, so the hook never fires, gdbserver-ze never starts, and device breakpoints stay pending while the program runs to completion. The extension provides an official override:

   export INTELGT_AUTO_ATTACH_HOOK="-qualified zeContextCreateEx"
   

Full session:

for f in /sys/class/drm/card*/prelim_enable_eu_debug; do echo 1 > "$f"; done
export ZET_ENABLE_PROGRAM_DEBUGGING=1
export INTELGT_AUTO_ATTACH_HOOK="-qualified zeContextCreateEx"
export CHIP_BE=level0
export CHIP_JIT_FLAGS_OVERRIDE="-cl-opt-disable"   # until item 3 above lands
hipcc -g -O0 test.hip -o test
gdb-oneapi -ex "set breakpoint pending on" -ex "break test.hip:42" -ex run ./test

Add INTELGT_AUTO_ATTACH_VERBOSE_LOG=1 to trace auto-attach; a successful run logs "gdbserver-ze started for process N".

Item 2 deserves its own fix: chipStar could call zeContextCreate when it is not subsetting devices, and Intel should hook both entry points, since any Level Zero application using the Ex form is invisible to auto-attach today.

[pvelesko]

/run-aurora-ci