gh-128883: Store ELF perf trampoline templates in read-only data

CPython copies the trampoline templates into executable memory. Keep the
ELF source in .rodata so BOLT preserves its bytes and boundaries.

Author: pablogsal

Lib/test/test_perf_profiler.py

@@ -62,7 +62,6 @@ def tearDown(self) -> None:
         for file in files_to_delete:
             file.unlink()
 
-    @unittest.skipIf(support.check_bolt_optimized(), "fails on BOLT instrumented binaries")
     def test_trampoline_works(self):
         code = """if 1:
                 def foo():
@@ -113,7 +112,6 @@ def baz():
                 "Address should contain only hex characters",
             )
 
-    @unittest.skipIf(support.check_bolt_optimized(), "fails on BOLT instrumented binaries")
     def test_trampoline_with_unencodable_name(self):
         code = """if 1:
                 import sys
@@ -123,7 +121,6 @@ def test_trampoline_with_unencodable_name(self):
                 """
         assert_python_ok("-c", code, PYTHON_JIT="0")
 
-    @unittest.skipIf(support.check_bolt_optimized(), "fails on BOLT instrumented binaries")
     def test_trampoline_works_with_forks(self):
         code = """if 1:
                 import os, sys
@@ -193,7 +190,6 @@ def baz():
         self.assertNotIn(f"py::bar:{script}", child_perf_file_contents)
         self.assertNotIn(f"py::baz:{script}", child_perf_file_contents)
 
-    @unittest.skipIf(support.check_bolt_optimized(), "fails on BOLT instrumented binaries")
     def test_trampoline_works_after_fork_with_many_code_objects(self):
         code = """if 1:
                 import gc, os, sys, signal
@@ -233,7 +229,6 @@ def test_trampoline_works_after_fork_with_many_code_objects(self):
         self.assertEqual(process.returncode, 0, stderr)
         self.assertEqual(stderr, "")
 
-    @unittest.skipIf(support.check_bolt_optimized(), "fails on BOLT instrumented binaries")
     def test_sys_api(self):
         for define_eval_hook in (False, True):
             code = """if 1:

Misc/NEWS.d/next/Build/2026-10-21-18-00-00.gh-issue-128883.perf-trampoline.rst

@@ -1,2 +1,2 @@
-Add a byte after the ELF perf trampoline end marker so it does not share
-the following function's address, which BOLT may otherwise move.
+Fix crashes with ``-X perf`` and ``-X perf_jit`` in BOLT builds.
+The ELF perf trampoline templates now use read-only data sections.

Python/asm_trampoline_aarch64.S

@@ -35,7 +35,12 @@
   #define GNU_PROPERTY_AARCH64_GCS 0
 #endif
 
+/* CPython copies this template into executable memory. */
+#if defined(__ELF__)
+    .section .rodata
+#else
     .text
+#endif
 #if defined(__aarch64__) && defined(__AARCH64EL__) && !defined(__ILP32__)
 #if defined(__APPLE__)
     .globl	__Py_trampoline_func_start
@@ -57,7 +62,6 @@ __Py_trampoline_func_end:
 #else
     .globl	_Py_trampoline_func_end
 _Py_trampoline_func_end:
-    .byte 0
     .section        .note.GNU-stack,"",@progbits
 #endif
 

Python/asm_trampoline_riscv64.S

@@ -1,4 +1,5 @@
-    .text
+/* CPython copies this template into executable memory. */
+    .section .rodata
     .globl	_Py_trampoline_func_start
 _Py_trampoline_func_start:
     addi    sp,sp,-16
@@ -9,5 +10,4 @@ _Py_trampoline_func_start:
     jr      ra
     .globl	_Py_trampoline_func_end
 _Py_trampoline_func_end:
-    .byte 0
     .section        .note.GNU-stack,"",@progbits

Python/asm_trampoline_x86_64.S

@@ -1,4 +1,9 @@
+/* CPython copies this template into executable memory. */
+#if defined(__ELF__)
+    .section .rodata
+#else
     .text
+#endif
 #ifdef __x86_64__
 #if defined(__APPLE__)
     .globl	__Py_trampoline_func_start
@@ -21,7 +26,6 @@ __Py_trampoline_func_end:
 #else
     .globl	_Py_trampoline_func_end
 _Py_trampoline_func_end:
-    .byte 0
     .section        .note.GNU-stack,"",@progbits
 #endif
 # Note for indicating the assembly code supports CET