Compiler warnings audit: 4,100+ warnings across ROCm components (gfx94X build)

[talumbau]

Summary

An audit of compiler warnings from the CI build run #25360331898 (Linux gfx94X-dcgpu, 2026-05-05) found 4,139 compiler warnings across ROCm-owned components. Some of these indicate potential real bugs (use of uninitialized memory, delete on non-heap objects, dead NULL checks).

Build logs were obtained from the TheRock S3 log archive:
https://therock-ci-artifacts-external.s3.amazonaws.com/ROCm-rocm-libraries/25360331898-linux/logs/gfx94X-dcgpu/index.html

Warning Counts by Component

Component	Owner Area	Warning Count	Top Warning Category
hip-tests	HIP	2,744	-Wc++17-extensions (inline vars) — 2,604 instances
hip-clr	HIP Runtime	427	-Winconsistent-missing-override — ~300+ instances
ROCR-Runtime	ROCr	250	-Wunused-variable, -Winconsistent-missing-override, -Wtypedef-redefinition
amdsmi	SMI	236	-Wsign-conversion — ~150+ instances
aqlprofile	Profiler	182	-Winconsistent-missing-override — 180 instances
amd-llvm	Compiler	163	-Wunused-but-set-parameter — 56 instances
rocprofiler-sdk	Profiler	90	-Winconsistent-missing-override, -Wuninitialized
rocprofiler-compute	Profiler	30	-Wunused-value (ignored nodiscard on hipError_t)
rocm_smi_lib	SMI	8	-Waddress (comparing stack variable address to NULL)
roctracer	Profiler	8	-Wunused-value (ignored nodiscard on hipError_t)
therock-flatbuffers	Third-party	2	-Wstringop-overflow
hipBLASLt	BLAS	0	Clean
rocBLAS	BLAS	0	Clean
rocRoller	BLAS	0	Clean
hipBLAS	BLAS	0	Clean

Third-party generated code (not actionable by ROCm teams):

• therock-host-blas64: 17,858 warnings (generated LAPACK/BLAS f2c code)
• therock-host-blas: 17,826 warnings (same, 32-bit variant)

Potential Bugs (High Priority)

amd-llvm: -Wfree-nonheap-object (4 instances)

Calling operator delete[] on a pointer to a non-heap object. This is likely a real bug that could cause crashes or memory corruption.

rocprofiler-sdk: -Wuninitialized (5 instances)

base class 'MemoryManager' is uninitialized when used here to access 'MemoryManager::agent' — accessing an uninitialized base class member during construction.

rocm_smi_lib: -Waddress (8 instances)

Comparing the address of stack/local variables (originalMemoryPartition, originalComputePartition, memory_partition, current_compute_partition) to NULL. These checks are always true — likely a logic error or dead code.

amdsmi: -Wsign-conversion (150+ instances)

Implicit signedness conversions between int/uint32_t/uint64_t/size_type. Potential truncation or sign-extension bugs.

Code Quality (Medium Priority)

Missing override keyword (~800 warnings across 4 components)

• hip-clr (427): ~30 virtual methods missing override, each appearing ~12 times
• aqlprofile (182): WaitIdle, Enable, Disable, BuildWriteConfigRegPacket, etc.
• rocprofiler-sdk (90): Same aqlprofile pattern (likely shared code)
• ROCR-Runtime (part of 250): PopulateMipmapSrd, PopulateImageSrd, SubmitCommand, etc.

This is a mechanical fix — add override to all virtual method overrides.

Ignored nodiscard return values (38 warnings)

• rocprofiler-compute (30): Ignoring hipError_t return values
• roctracer (8): Same issue

Should either check the return value or explicitly cast to (void).

Low Priority / Noisy

hip-tests: C++17 extension warnings (2,604)

Test code uses inline variables but appears to compile with a pre-C++17 standard. Either bump -std=c++17 or remove inline specifiers.

amd-llvm: Unused parameters/variables (80+)

Likely in upstream LLVM code or OpenMP runtime. May not be actionable in the ROCm fork.

ROCR-Runtime: Miscellaneous (remaining ~100)

Unused variables, typedef redefinitions, const-correctness issues, mismatched struct/class tags.

Detailed Per-Component Breakdowns

hip-tests (2,744 warnings)

Count	Warning	Flag
2,604	inline variables are a C++17 extension	-Wc++17-extensions
56	variable 'expected_value' used uninitialized whenever switch default taken	-Wsometimes-uninitialized
28	variable 'src' used uninitialized whenever switch default taken	-Wsometimes-uninitialized
28	variable 'dst' used uninitialized whenever switch default taken	-Wsometimes-uninitialized
28	unused function 'test_hipMallocFromPoolAsync_MThread'	-Wunused-function

hip-clr (427 warnings)

Dominated by -Winconsistent-missing-override. ~30+ virtual method overrides missing the override keyword, each appearing ~12 times across translation units.

Top methods missing override: submitWriteMemory, submitVirtualMap, submitUnmapMemory, submitThreadTrace, submitThreadTraceMemObjects, submitSvmUnmapMemory, SubmitSvmPrefetchBatchAsync, submitSvmPrefetchAsync, submitSvmMapMemory, submitSvmFreeMemory, and many more.

ROCR-Runtime (250 warnings)

Count	Warning	Flag
24	unused variable 'err'	-Wunused-variable
18	redefinition of typedef 'HsaKFDContext'	-Wtypedef-redefinition
12	'PopulateMipmapSrd' missing override	-Winconsistent-missing-override
12	'PopulateImageSrd' missing override	-Winconsistent-missing-override
9	unused variable 'status'	-Wunused-variable
8	struct previously declared as class	-Wmismatched-tags
8	cast drops const qualifier	-Wcast-qual
6	'SubmitCommand' missing override	-Winconsistent-missing-override
6	'printSwizzleMode' missing override	-Winconsistent-missing-override
6	'printSRDDetailed' missing override	-Winconsistent-missing-override

amdsmi (236 warnings)

Count	Warning	Flag
63	private field 'm_current_offset' not used	-Wunused-private-field
32	implicit conversion: 'int' to 'uint32_t'	-Wsign-conversion
10	implicit conversion: 'int' to 'uint64_t'	-Wsign-conversion
8	implicit conversion: 'unsigned int' to 'int'	-Wsign-conversion
8	implicit conversion: 'rsmi_status_t' to 'int32_t'	-Wsign-conversion
8	implicit conversion: 'int' to 'size_type'	-Wsign-conversion
7	implicit conversion: 'int64_t' to 'uint64_t'	-Wsign-conversion
6	implicit conversion: 'streamsize' to 'size_t'	-Wsign-conversion
6	implicit conversion: 'long' to 'uint64_t'	-Wsign-conversion

aqlprofile (182 warnings)

All -Winconsistent-missing-override:

Count	Method
20	WaitIdle
20	Enable
20	Disable
16	BuildWriteConfigRegPacket
16	BuildThreadTraceEventFinish
16	BuildNopPacket
16	BuildIndirectBufferCmd
14	GetSpmSampleDelayMax
14	BuildWriteWaitIdlePacket
8	BuildWriteShRegPacket

amd-llvm (163 warnings)

Count	Warning	Flag
56	parameter 'P' set but not used	-Wunused-but-set-parameter
23	private field 'OmptInitialized' not used	-Wunused-private-field
9	field designators not in declaration order	-Wreorder-init-list
5	enumerated and non-enumerated type in conditional	-Wextra
4	operator delete[] called on non-heap object	-Wfree-nonheap-object
4	std::scoped_lock may not support CTAD	-Wctad-maybe-unsupported
3	comparison of integers of different signs	-Wsign-compare

rocprofiler-sdk (90 warnings)

Primarily -Winconsistent-missing-override (same aqlprofile pattern, likely shared code) plus:

• 5 instances of -Wuninitialized (base class 'MemoryManager' uninitialized when accessed)

Methodology

All *_build.log files were downloaded from the TheRock S3 log archive for the gfx94X-dcgpu build job. Warnings were extracted by matching warning: in each log file, then deduplicated and counted. Only compiler warnings (not CMake, pip, or git warnings) are included in the component counts.

---

Reference: ROCM-24273

[ScottTodd]

Want to close the older #47 in favor of this new analysis?