[Issue]: `set(ENV{LD_PRELOAD})` in `test/integration/CMakeLists.txt` poisons cmake configure on incremental builds

[davidd-amd]

Problem Description

Summary

test/integration/CMakeLists.txt sets the process-wide LD_PRELOAD environment variable at configure time, which contaminates all subsequent execute_process() calls and post-generate ninja invocations. On incremental asan builds (where libintercept.so already exists from a prior build), this causes rpm, ninja, and other system tools to crash with:

rpm: error while loading shared libraries: libclang_rt.asan-x86_64.so: cannot open shared object file: No such file or directory

/usr/local/bin/ninja: error while loading shared libraries: libclang_rt.asan-x86_64.so: cannot open shared object file: No such file or directory

CMake Generate step failed.  Build files cannot be regenerated correctly.

Environment

• Build system: TheRock (ROCm super-build) with --preset linux-release-asan
• Container image: ghcr.io/rocm/therock_build_manylinux_x86_64:latest (manylinux_2_28, RHEL 8 based)
• aqlprofile sanitizer: OFF (aqlprofile_SANITIZER=OFF)
• Dependent libraries: libhsa-runtime64.so built with ASAN (ROCR-Runtime_SANITIZER=ASAN)

Root Cause

In test/integration/CMakeLists.txt (around line 66):

# Add a PRELOAD environment with libintercept
set(ENV{LD_PRELOAD} "$ENV{LD_PRELOAD}:${CMAKE_CURRENT_BINARY_DIR}/libintercept.so")

add_test(NAME testv2 COMMAND testv2)
set_tests_properties(testv2 PROPERTIES ENVIRONMENT "${LD_PRELOAD}" TIMEOUT 45 LABELS "unittests" FAIL_REGULAR_EXPRESSION "${AQLPROFILE_DEFAULT_FAIL_REGEX}")

set(ENV{LD_PRELOAD} ...) modifies the cmake process environment, not a cmake variable. This environment persists for:

• All subsequent execute_process() calls during configure (e.g., rpm --eval %{?dist} for CPACK)
• The post-generate ninja -t recompact and ninja -t restat operations

Why it works on clean builds (and in CI)

On a first configure, libintercept.so does not yet exist in CMAKE_CURRENT_BINARY_DIR. The dynamic linker encounters a nonexistent file in LD_PRELOAD, prints a warning, and continues. System tools like rpm and ninja run successfully.

CI always starts with a clean build directory, so this path is always taken.

Why it fails on incremental builds

On a reconfigure (e.g., after a reboot, stamp file deletion, or any re-run of the configure step), libintercept.so already exists from the prior build. The dynamic linker loads it via LD_PRELOAD, which triggers its transitive dependency chain:

LD_PRELOAD: libintercept.so
  → DT_NEEDED: libhsa-amd-aqlprofile64.so
  → DT_NEEDED: libhsa-runtime64.so.1  (built with ASAN)
    → DT_NEEDED: libclang_rt.asan-x86_64.so  →  NOT FOUND  →  FATAL

The ASAN runtime is not discoverable because:

• aqlprofile has SANITIZER=OFF, so no ASAN-related LD_LIBRARY_PATH or RPATH is configured
• The ASAN runtime is not in the system ldconfig cache
• No LD_LIBRARY_PATH points to the clang resource directory

Every system tool invoked by cmake (rpm, ninja, gzip, date) inherits the poisoned LD_PRELOAD and crashes.

Reproduction

1. Configure and build aqlprofile in an ASAN super-build (first time — succeeds)
2. Delete the configure stamp: rm -f <build>/profiler/aqlprofile/stamp/configure.stamp
3. Re-run configure: cmake --build <build> --target aqlprofile+configure
4. Observe failure: rpm: error while loading shared libraries: libclang_rt.asan-x86_64.so

Additional Issue

The set_tests_properties call on the next line uses ${LD_PRELOAD} (a cmake variable), not $ENV{LD_PRELOAD} (the environment variable that was just set). Unless a cmake variable named LD_PRELOAD is defined elsewhere, this test property is likely empty/incorrect — meaning the set(ENV{...}) call is both harmful (pollutes configure) and ineffective for its intended purpose (setting the test environment).

Suggested Fix

Remove the set(ENV{LD_PRELOAD}) and set the preload directly via test properties:

# BEFORE (buggy):
set(ENV{LD_PRELOAD} "$ENV{LD_PRELOAD}:${CMAKE_CURRENT_BINARY_DIR}/libintercept.so")
add_test(NAME testv2 COMMAND testv2)
set_tests_properties(testv2 PROPERTIES ENVIRONMENT "${LD_PRELOAD}" TIMEOUT 45 LABELS "unittests" FAIL_REGULAR_EXPRESSION "${AQLPROFILE_DEFAULT_FAIL_REGEX}")

# AFTER (fixed):
add_test(NAME testv2 COMMAND testv2)
set_tests_properties(testv2 PROPERTIES
  ENVIRONMENT "LD_PRELOAD=${CMAKE_CURRENT_BINARY_DIR}/libintercept.so"
  TIMEOUT 45
  LABELS "unittests"
  FAIL_REGULAR_EXPRESSION "${AQLPROFILE_DEFAULT_FAIL_REGEX}")

This scopes LD_PRELOAD to the test execution only, preventing configure-time environment pollution.

Workaround

Delete the stale libintercept.so before reconfiguring:

rm -f <build>/profiler/aqlprofile/build/test/integration/libintercept.so

Operating System

manylinux

CPU

NA

GPU

NA

ROCm Version

7.11

ROCm Component

No response

Steps to Reproduce

No response

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

No response