For X32, TLSDESC generates 1. -maddress-mode=short lea foo@TLSDESC(%rip), %eax call *foo@TLSCALL(%eax) 2. -maddress-mode=long lea foo@TLSDESC(%rip), %rax call *foo@TLSCALL(%rax) Currently, I got [hjl@gnu-cfl-2 tls-x32]$ cat main.c #include <stdlib.h> #include <stdio.h> __thread int bar = 301; extern int *test1 (int); extern int *test2 (int); int main () { int *p; p = test1 (30); if (*p != 30) abort (); *p = 40; test1 (40); p = test2 (301); if (*p != 301) abort (); if (p != &bar) abort (); *p = 40; test2 (40); puts ("PASS"); return 0; } [hjl@gnu-cfl-2 tls-x32]$ cat tlsdes.S .text .p2align 4 .globl test1 .type test1, @function test1: .cfi_startproc subl $8, %esp .cfi_def_cfa_offset 16 lea foo@TLSDESC(%rip), %eax call *foo@TLSCALL(%eax) addl %fs:0, %eax cmpl %edi, (%eax) jne .L5 addl $8, %esp .cfi_remember_state .cfi_def_cfa_offset 8 ret .L5: .cfi_restore_state call abort@PLT .cfi_endproc .size test1, .-test1 .p2align 4 .globl test2 .type test2, @function test2: .cfi_startproc subl $8, %esp .cfi_def_cfa_offset 16 lea bar@TLSDESC(%rip), %eax call *bar@TLSCALL(%eax) addl %fs:0, %eax cmpl %edi, (%eax) jne .L9 addl $8, %esp .cfi_remember_state .cfi_def_cfa_offset 8 ret .L9: .cfi_restore_state call abort@PLT .cfi_endproc .size test2, .-test2 .section .tdata,"awT",@progbits .align 4 .type foo, @object .size foo, 4 foo: .long 30 .section .note.GNU-stack,"",@progbits [hjl@gnu-cfl-2 tls-x32]$ gcc -mx32 tlsdes.S main.c /usr/local/bin/ld: /tmp/ccdQjj3W.o: TLS transition from R_X86_64_GOTPC32_TLSDESC to R_X86_64_TPOFF32 against `foo' at 0x5 in section `.text' failed collect2: error: ld returned 1 exit status [hjl@gnu-cfl-2 tls-x32]$
For x32, we must encode "lea x@TLSDESC(%rip), %eax" with a REX prefix even if it isn't required. Otherwise linker can’t safely perform GDesc -> LE optimization.
The master branch has been updated by H.J. Lu <hjl@sourceware.org>: https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=14470f0755dbc942aa684ed647df978ddfc7cff2 commit 14470f0755dbc942aa684ed647df978ddfc7cff2 Author: H.J. Lu <hjl.tools@gmail.com> Date: Mon Jan 20 06:58:51 2020 -0800 x86-64: Fix TLSDESC relaxation for x32 For x32, we must encode "lea x@TLSDESC(%rip), %reg" with a REX prefix even if it isn't required. Otherwise linker can’t safely perform GDesc -> IE/LE optimization. X32 TLSDESC sequences can be: 40 8d 05 00 00 00 00 rex lea x@TLSDESC(%rip), %reg ... 67 ff 10 call *x@TLSCALL(%eax) or the same sequence as LP64: 48 8d 05 00 00 00 00 lea foo@TLSDESC(%rip), %reg ... ff 10 call *foo@TLSCALL(%rax) We need to support both sequences for x32. For both GDesc -> IE/LE transitions, 67 ff 10 call *x@TLSCALL(%eax) should relaxed to 0f 1f 00 nopl (%rax) For GDesc -> LE transition, 40 8d 05 00 00 00 00 rex lea x@TLSDESC(%rip), %reg should relaxed to 40 c7 c0 fc ff ff ff rex movl $x@tpoff, %reg For GDesc -> IE transition, 40 8d 05 00 00 00 00 rex lea x@TLSDESC(%rip), %reg should relaxed to 40 8b 05 00 00 00 00 rex movl x@gottpoff(%rip), %eax bfd/ PR ld/25416 * elf64-x86-64.c (elf_x86_64_check_tls_transition): Support "rex leal x@tlsdesc(%rip), %reg" and "call *x@tlsdesc(%eax)" in X32 mode. (elf_x86_64_relocate_section): In x32 mode, for GDesc -> LE transition, relax "rex leal x@tlsdesc(%rip), %reg" to "rex movl $x@tpoff, %reg", for GDesc -> IE transition, relax "rex leal x@tlsdesc(%rip), %reg" to "rex movl x@gottpoff(%rip), %eax". For both transitions, relax "call *(%eax)" to "nopl (%rax)". gas/ PR ld/25416 * config/tc-i386.c (output_insn): Add a dummy REX_OPCODE prefix for lea with R_X86_64_GOTPC32_TLSDESC relocation when generating x32 object. * testsuite/gas/i386/ilp32/x32-tls.d: Updated. * testsuite/gas/i386/ilp32/x32-tls.s: Add tests for lea with R_X86_64_GOTPC32_TLSDESC relocation. ld/ PR ld/25416 * testsuite/ld-x86-64/pr25416-1.s: New file * testsuite/ld-x86-64/pr25416-1a.d: Likewise. * testsuite/ld-x86-64/pr25416-1b.d: Likewise. * testsuite/ld-x86-64/pr25416-1.s: Likewise. * testsuite/ld-x86-64/pr25416-2.s: Likewise. * testsuite/ld-x86-64/pr25416-2a.d: Likewise. * testsuite/ld-x86-64/pr25416-2b.d: Likewise. * testsuite/ld-x86-64/pr25416-3.d: Likewise. * testsuite/ld-x86-64/pr25416-3.s: Likewise. * testsuite/ld-x86-64/pr25416-4.d: Likewise. * testsuite/ld-x86-64/pr25416-4.s: Likewise. * testsuite/ld-x86-64/pr25416-5a.c: Likewise. * testsuite/ld-x86-64/pr25416-5b.s: Likewise. * testsuite/ld-x86-64/pr25416-5c.s: Likewise. * testsuite/ld-x86-64/pr25416-5d.s: Likewise. * testsuite/ld-x86-64/pr25416-5e.s: Likewise. * testsuite/ld-x86-64/x86-64.exp: Run PR ld/25416 tests.
Fixed for 2.35.
The master branch has been updated by H.J. Lu <hjl@sourceware.org>: https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=6a462ad49ec13bec4659f2a8fb2301cac9742da9 commit 6a462ad49ec13bec4659f2a8fb2301cac9742da9 Author: H.J. Lu <hjl.tools@gmail.com> Date: Wed Jan 22 06:22:41 2020 -0800 x86-64: Skip GNU2 TLS tests only without compiler support After fixing: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93319 https://sourceware.org/bugzilla/show_bug.cgi?id=25416 -mtls-dialect=gnu2 is working for x32 with GCC 10. Skip GNU2 TLS tests only if compiler doesn't support it. PR ld/25416 * testsuite/ld-x86-64/tls.exp: Skip GNU2 TLS tests only without compiler support.