Incompatibility in Inline TLS Assembly on Alpine 3.22 with `zend_jit_ir.c`

[thecaliskan]

Description

When building PHP on Alpine Linux 3.22 with musl and targeting 32-bit or 64-bit x86 platforms, a regression is observed related to inline assembly in ext/opcache/jit/zend_jit_ir.c, specifically:

This error does not occur on Alpine 3.21 using the same compilation toolchain. The issue stems from the @TLSGD usage which implicitly expects %rdi on x86_64 or %eax on i386, depending on the TLS model.

ext/opcache/jit/zend_jit_ir.c:3459: Error: @TLSGD operator requires %rdi' as dest register`

Environment

• Alpine Linux 3.22
• GCC (via build-base)
• musl libc
• Target: x86_64 or i386
• Affected file: ext/opcache/jit/zend_jit_ir.c

---

Related PR / Reference

• Add Alpine 3.22 (remove Alpine 3.20) docker-library/php#1580

PHP Version

8.4.7-zts
8.3.21-zts
8.2.28-zts
8.1.32-zts

Operating System

Alpine 3.22

[arnaud-lb]

Thank you for reporting this. This is the same root cause as #15074 (see #15074 (comment)), and will be fixed by #18660 (specifically ba04798).

[nielsdos]

@arnaud-lb But this can be fixed on lower branches, and I think we should.
For this particular issue the following suffices (on PHP-8.3) (but would need to generalise to other platforms too ofc as we would need to change other code paths with the same assembly too):

diff --git a/ext/opcache/jit/zend_jit_x86.dasc b/ext/opcache/jit/zend_jit_x86.dasc
index 1f1abb59a1c..6ea8521ba6d 100644
--- a/ext/opcache/jit/zend_jit_x86.dasc
+++ b/ext/opcache/jit/zend_jit_x86.dasc
@@ -2910,7 +2910,7 @@ static int zend_jit_setup(void)
 
 		__asm__(
 			"leaq _tsrm_ls_cache@tlsgd(%%rip), %0\n"
-			: "=a" (ti));
+			: "=D" (ti));
 		tsrm_tls_offset = ti[1];
 		tsrm_tls_index = ti[0] * 8;
 #elif defined(__FreeBSD__)

[arnaud-lb]

@nielsdos I agree, and it's great that there is a simpler fix, as ba04798 felt a bit risky for release branches.

The change you propose looks good to me. We don't need to use the entire code sequence as in ba04798, because gas is only checking a single instruction, and ld will only check the entire sequence when building opcache statically, which is not allowed in release branches.

Glibc seems unaffected as we use @gottpoff, for which gas doesn't check the destination register: https://github.com/bminor/binutils-gdb/blob/ec181e1710e37007a8d95c284609bfaa5868d086/gas/config/tc-i386.c#L6869

Do you want to take care of this or shall I?

[nielsdos]

Thanks for the link, I'm on it.