Fix AMX tests failing with SIGILL on newer Linux kernels (#3802)

* Fix #3782: Fix AMX tests failing with SIGILL on newer Linux kernels

On Linux, applications must explicitly request permission from the kernel
via arch_prctl to use Intel AMX tile data registers (XFEATURE_XTILEDATA)
before executing any AMX instructions, otherwise the kernel kills the
process with SIGILL. Add the syscall at test harness startup and warn
on failure (e.g., when the CPU/kernel doesn't support AMX).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* Fix #3782: Silence arch_prctl warning that corrupted print test output

The warning printed to stdout on every test process, which polluted the
stdout comparison used by print_* functional tests, causing spurious
failures. Suppress the return value silently instead.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* Fix #3782: Add error message and compile flag

* Fix #3782: Remove TEST_REQUIRES_AMX from Windows builds

The AMX syscall (arch_prctl) is Linux-specific and not needed on Windows.
Windows handles AMX permissions automatically without requiring a syscall.
Removing the TEST_REQUIRES_AMX flag from Windows compilation fixes all
failing Windows test jobs.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

* Fix #3782: Fix macOS bitcode generation race condition and add validation

This commit addresses intermittent test failures on macOS-15-intel by:

1. **Serialize stdlib generation** (prevents race condition):
   - Add JOB_POOLS to limit stdlib bitcode generation to 1 job at a time
   - Prevents multiple parallel jobs from reading/writing the same .bc file
   - Fixes "file too small to contain bitcode header" errors

2. **Add pipeline failure detection** (fail fast on build errors):
   - Wrap all piped commands (m4|llvm-as, clang++|llvm-as, etc.) with pipefail
   - Validate output files are non-empty with "test -s"
   - Ensures build failures are caught early, not hidden in corrupted artifacts

The race condition occurred when cmake --build -j4 caused multiple jobs to
simultaneously access stdlib bitcode files during generation. The job pool
serializes these operations while allowing other build steps to parallelize.

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

* Revert "Fix #3782: Fix macOS bitcode generation race condition and add validation"

This reverts commit caad76905565efa2f25114b48b336e197af6581e.

---------

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

scripts/run_tests.py

@@ -686,6 +686,9 @@ def run_test(testname, host, target, jit_lib_path=None):
                 cc_cmd += ' -fPIE'
                 if should_fail:
                     cc_cmd += " -DEXPECT_FAILURE"
+                # Enable AMX syscall for AMX-capable targets (avx512spr and later)
+                if target.target.startswith('avx512spr') or target.target.startswith('avx512gnr'):
+                    cc_cmd += " -DTEST_REQUIRES_AMX"
 
                 if target.is_xe():
                     exe_name = "%s.run" % os.path.basename(testname)

tests/test_static.cpp

@@ -1,5 +1,5 @@
 /*
-  Copyright (c) 2010-2024, Intel Corporation
+  Copyright (c) 2010-2026, Intel Corporation
 
   SPDX-License-Identifier: BSD-3-Clause
 */
@@ -31,6 +31,16 @@
 #if defined ISPC_IS_LINUX || defined ISPC_IS_WASM
 #include <malloc.h>
 #endif
+#if defined(__linux__) && defined(__x86_64__)
+#include <sys/syscall.h>
+#include <unistd.h>
+#ifndef ARCH_REQ_XCOMP_PERM
+#define ARCH_REQ_XCOMP_PERM 0x1023
+#endif
+#ifndef XFEATURE_XTILEDATA
+#define XFEATURE_XTILEDATA 18
+#endif
+#endif
 
 #if (TEST_SIG == 7)
 #define varying_f_sz f_sz
@@ -53,7 +63,7 @@
 #else
 // __vectorcall calling convention is off by default.
 #define CALLINGCONV //__vectorcall
-#endif // VECTORCALL_CONV
+#endif              // VECTORCALL_CONV
 #else
 #define CALLINGCONV
 #endif // ISPC_IS_WINDOWS64
@@ -123,6 +133,15 @@ void *ISPCAlloc(void **handle, int64_t size, int32_t alignment) {
 #endif
 
 int main(int argc, char *argv[]) {
+#if defined(__linux__) && defined(__x86_64__) && defined(TEST_REQUIRES_AMX)
+    // Request permission to use AMX tile data
+    long result = syscall(SYS_arch_prctl, ARCH_REQ_XCOMP_PERM, XFEATURE_XTILEDATA);
+    if (result != 0) {
+        fprintf(stderr, "Error: Failed to enable AMX support (syscall returned %ld). "
+                "AMX instructions require kernel permission on Linux 5.16+.\n", result);
+        return 1;
+    }
+#endif
     int w = width();
     assert(w <= 64);
 

tests/test_static_jit.cpp

@@ -1,5 +1,5 @@
 /*
-  Copyright (c) 2025, Intel Corporation
+  Copyright (c) 2025-2026, Intel Corporation
 
   SPDX-License-Identifier: BSD-3-Clause
 */
@@ -35,6 +35,16 @@
 #if defined ISPC_IS_LINUX || defined ISPC_IS_WASM
 #include <malloc.h>
 #endif
+#if defined(__linux__) && defined(__x86_64__)
+#include <sys/syscall.h>
+#include <unistd.h>
+#ifndef ARCH_REQ_XCOMP_PERM
+#define ARCH_REQ_XCOMP_PERM 0x1023
+#endif
+#ifndef XFEATURE_XTILEDATA
+#define XFEATURE_XTILEDATA 18
+#endif
+#endif
 
 #include "ispc/ispc.h"
 
@@ -129,6 +139,15 @@ bool createTemporaryISPCFile(const std::string &filename, const std::string &con
 }
 
 int main(int argc, char *argv[]) {
+#if defined(__linux__) && defined(__x86_64__) && defined(TEST_REQUIRES_AMX)
+    // Request permission to use AMX tile data
+    long result = syscall(SYS_arch_prctl, ARCH_REQ_XCOMP_PERM, XFEATURE_XTILEDATA);
+    if (result != 0) {
+        fprintf(stderr, "Error: Failed to enable AMX support (syscall returned %ld). "
+                "AMX instructions require kernel permission on Linux 5.16+.\n", result);
+        return 1;
+    }
+#endif
     if (argc < 2) {
         printf("Usage: %s <ispc_source_file> [target]\n", argv[0]);
         return 1;
@@ -174,9 +193,9 @@ int main(int argc, char *argv[]) {
     }
 
     // Set the runtime functions that JIT will use
-    if (!engine->SetJitRuntimeFunction("ISPCLaunch", (void*)ISPCLaunch) ||
+    if (!engine->SetJitRuntimeFunction("ISPCLaunch", (void *)ISPCLaunch) ||
-        !engine->SetJitRuntimeFunction("ISPCSync", (void*)ISPCSync) ||
+        !engine->SetJitRuntimeFunction("ISPCSync", (void *)ISPCSync) ||
-        !engine->SetJitRuntimeFunction("ISPCAlloc", (void*)ISPCAlloc)) {
+        !engine->SetJitRuntimeFunction("ISPCAlloc", (void *)ISPCAlloc)) {
         printf("Failed to set JIT runtime functions\n");
         ispc::Shutdown();
         return 1;