Optimize R_386_GOT32/R_386_GOT32X only if addend is 0

Linker can't optimize R_386_GOT32 and R_386_GOT32X relocations if addend
isn't 0.  It isn't valid to convert

movl	foo@GOT+1(%ecx), %eax

to

leal	foo@GOTOFF+1(%ecx), %eax

nor to convert

movq	foo@GOTPCREL+1(%rip), %rax

to

leaq	foo(%rip), %rax

for x86-64.  We should check if addend is 0 before optimizing R_386_GOT32
and R_386_GOT32X relocations.  Testcases are added for i386 and x86-64.

bfd/

	* elf32-i386.c (elf_i386_convert_load): Skip if addend isn't 0.
	(elf_i386_relocate_section): Skip R_386_GOT32X optimization if
	addend isn't 0.

ld/testsuite/

	* ld-i386/i386.exp: Run mov2a, mov2b and mov3.
	* ld-i386/mov2.s: New file.
	* ld-i386/mov2a.d: Likewise.
	* ld-i386/mov2b.d: Likewise.
	* ld-i386/mov3.d: Likewise.
	* ld-i386/mov3.s: Likewise.
	* ld-x86-64/mov2.s: Likewise.
	* ld-x86-64/mov2a.d: Likewise.
	* ld-x86-64/mov2b.d: Likewise.
	* ld-x86-64/mov2c.d: Likewise.
	* ld-x86-64/mov2d.d: Likewise.
	* ld-x86-64/x86-64.exp: Run mov2a, mov2b, mov2c and mov2d.
This commit is contained in:
H.J. Lu 2015-12-04 08:43:45 -08:00
parent f27c5390b2
commit 02e2aef89b
15 changed files with 186 additions and 13 deletions

View file

@ -1,3 +1,9 @@
2015-12-04 H.J. Lu <hongjiu.lu@intel.com>
* elf32-i386.c (elf_i386_convert_load): Skip if addend isn't 0.
(elf_i386_relocate_section): Skip R_386_GOT32X optimization if
addend isn't 0.
2015-12-04 Tristan Gingold <gingold@adacore.com>
* elf.c (rewrite_elf_program_header): Remove useless loop.

View file

@ -2820,6 +2820,11 @@ elf_i386_convert_load (bfd *abfd, asection *sec,
if (roff < 2)
continue;
/* Addend for R_386_GOT32 and R_386_GOT32X relocations must be 0. */
addend = bfd_get_32 (abfd, contents + roff);
if (addend != 0)
continue;
modrm = bfd_get_8 (abfd, contents + roff - 1);
baseless = (modrm & 0xc7) == 0x5;
@ -2913,11 +2918,6 @@ elf_i386_convert_load (bfd *abfd, asection *sec,
{
/* The function is locally defined. */
convert_branch:
addend = bfd_get_32 (abfd, contents + roff);
/* Addend for R_386_GOT32X relocation must be 0. */
if (addend != 0)
continue;
/* Convert R_386_GOT32X to R_386_PC32. */
if (modrm == 0x15 || (modrm & 0xf8) == 0x90)
{
@ -3007,11 +3007,6 @@ convert_load:
}
else
{
/* Addend for R_386_GOT32X relocation must be 0. */
addend = bfd_get_32 (abfd, contents + roff);
if (addend != 0)
continue;
if (opcode == 0x85)
{
/* Convert "test %reg1, foo@GOT(%reg2)" to
@ -3998,8 +3993,11 @@ elf_i386_relocate_section (bfd *output_bfd,
branch to direct branch. It is OK to convert adc,
add, and, cmp, or, sbb, sub, test, xor only when PIC
is false. */
unsigned int opcode;
opcode = bfd_get_8 (abfd, contents + rel->r_offset - 2);
unsigned int opcode, addend;
addend = bfd_get_32 (input_bfd, contents + rel->r_offset);
if (addend != 0)
goto r_386_got32;
opcode = bfd_get_8 (input_bfd, contents + rel->r_offset - 2);
if (opcode != 0x8b && opcode != 0xff)
goto r_386_got32;
}
@ -4031,7 +4029,7 @@ elf_i386_relocate_section (bfd *output_bfd,
/* If not PIC, add the .got.plt section address for
baseless addressing. */
unsigned int modrm;
modrm = bfd_get_8 (abfd, contents + rel->r_offset - 1);
modrm = bfd_get_8 (input_bfd, contents + rel->r_offset - 1);
if ((modrm & 0xc7) == 0x5)
relocation += offplt;
}

View file

@ -1,3 +1,16 @@
2015-12-04 H.J. Lu <hongjiu.lu@intel.com>
* ld-i386/i386.exp: Run mov2a and mov2b.
* ld-i386/mov2.s: New file.
* ld-i386/mov2a.d: Likewise.
* ld-i386/mov2b.d: Likewise.
* ld-x86-64/mov2.s: Likewise.
* ld-x86-64/mov2a.d: Likewise.
* ld-x86-64/mov2b.d: Likewise.
* ld-x86-64/mov2c.d: Likewise.
* ld-x86-64/mov2d.d: Likewise.
* ld-x86-64/x86-64.exp: Run mov2a, mov2b, mov2c and mov2d.
2015-12-01 H.J. Lu <hongjiu.lu@intel.com>
PR ld/19319

View file

@ -293,6 +293,9 @@ run_dump_test "lea1e"
run_dump_test "lea1f"
run_dump_test "mov1a"
run_dump_test "mov1b"
run_dump_test "mov2a"
run_dump_test "mov2b"
run_dump_test "mov3"
run_dump_test "branch1"
run_dump_test "call1"
run_dump_test "call2"

View file

@ -0,0 +1,15 @@
.section my_section,"aw",@progbits
.long 0x12345678
.text
.globl foo
.type foo, @function
foo:
ret
.size foo, .-foo
.globl _start
.type _start, @function
_start:
movl foo@GOT+1(%ecx), %eax
movl __start_my_section@GOT+1(%ecx), %eax
movl __stop_my_section@GOT+1(%ecx), %eax
.size _start, .-_start

View file

@ -0,0 +1,15 @@
#source: mov2.s
#as: --32
#ld: -shared -melf_i386
#objdump: -dw
.*: +file format .*
Disassembly of section .text:
#...
[ ]*[a-f0-9]+: 8b 81 ([0-9a-f]{2} ){4} * mov -0x[a-f0-9]+\(%ecx\),%eax
[ ]*[a-f0-9]+: 8b 81 ([0-9a-f]{2} ){4} * mov -0x[a-f0-9]+\(%ecx\),%eax
[ ]*[a-f0-9]+: 8b 81 ([0-9a-f]{2} ){4} * mov -0x[a-f0-9]+\(%ecx\),%eax
#pass

View file

@ -0,0 +1,15 @@
#source: mov2.s
#as: --32
#ld: -pie -melf_i386
#objdump: -dw
.*: +file format .*
Disassembly of section .text:
#...
[ ]*[a-f0-9]+: 8b 81 ([0-9a-f]{2} ){4} * mov -0x[a-f0-9]+\(%ecx\),%eax
[ ]*[a-f0-9]+: 8b 81 ([0-9a-f]{2} ){4} * mov -0x[a-f0-9]+\(%ecx\),%eax
[ ]*[a-f0-9]+: 8b 81 ([0-9a-f]{2} ){4} * mov -0x[a-f0-9]+\(%ecx\),%eax
#pass

View file

@ -0,0 +1,14 @@
#as: --32
#ld: -melf_i386
#objdump: -dw
.*: +file format .*
Disassembly of section .text:
#...
[ ]*[a-f0-9]+: 8b 05 ([0-9a-f]{2} ){4} * mov 0x[a-f0-9]+\,%eax
[ ]*[a-f0-9]+: 8b 05 ([0-9a-f]{2} ){4} * mov 0x[a-f0-9]+\,%eax
[ ]*[a-f0-9]+: 8b 05 ([0-9a-f]{2} ){4} * mov 0x[a-f0-9]+\,%eax
#pass

View file

@ -0,0 +1,15 @@
.section my_section,"aw",@progbits
.long 0x12345678
.text
.globl foo
.type foo, @function
foo:
ret
.size foo, .-foo
.globl _start
.type _start, @function
_start:
movl foo@GOT+1, %eax
movl __start_my_section@GOT+1, %eax
movl __stop_my_section@GOT+1, %eax
.size _start, .-_start

View file

@ -0,0 +1,15 @@
.section my_section,"aw",@progbits
.long 0x12345678
.text
.globl foo
.type foo, @function
foo:
ret
.size foo, .-foo
.globl _start
.type _start, @function
_start:
movq foo@GOTPCREL+1(%rip), %rax
movq __start_my_section@GOTPCREL+1(%rip), %rax
movq __stop_my_section@GOTPCREL+1(%rip), %rax
.size _start, .-_start

View file

@ -0,0 +1,15 @@
#source: mov2.s
#as: --64
#ld: -shared -melf_x86_64
#objdump: -dw
.*: +file format .*
Disassembly of section .text:
#...
[ ]*[a-f0-9]+: 48 8b 05 ([0-9a-f]{2} ){4} * mov 0x[a-f0-9]+\(%rip\),%rax # [a-f0-9]+ <_DYNAMIC\+0x[a-f0-9]+>
[ ]*[a-f0-9]+: 48 8b 05 ([0-9a-f]{2} ){4} * mov 0x[a-f0-9]+\(%rip\),%rax # [a-f0-9]+ <_DYNAMIC\+0x[a-f0-9]+>
[ ]*[a-f0-9]+: 48 8b 05 ([0-9a-f]{2} ){4} * mov 0x[a-f0-9]+\(%rip\),%rax # [a-f0-9]+ <_DYNAMIC\+0x[a-f0-9]+>
#pass

View file

@ -0,0 +1,15 @@
#source: mov2.s
#as: --64
#ld: -pie -melf_x86_64
#objdump: -dw
.*: +file format .*
Disassembly of section .text:
#...
[ ]*[a-f0-9]+: 48 8b 05 ([0-9a-f]{2} ){4} * mov 0x[a-f0-9]+\(%rip\),%rax # [a-f0-9]+ <_DYNAMIC\+0x[a-f0-9]+>
[ ]*[a-f0-9]+: 48 8b 05 ([0-9a-f]{2} ){4} * mov 0x[a-f0-9]+\(%rip\),%rax # [a-f0-9]+ <_DYNAMIC\+0x[a-f0-9]+>
[ ]*[a-f0-9]+: 48 8b 05 ([0-9a-f]{2} ){4} * mov 0x[a-f0-9]+\(%rip\),%rax # [a-f0-9]+ <_DYNAMIC\+0x[a-f0-9]+>
#pass

View file

@ -0,0 +1,15 @@
#source: mov2.s
#as: --x32
#ld: -shared -melf32_x86_64
#objdump: -dw
.*: +file format .*
Disassembly of section .text:
#...
[ ]*[a-f0-9]+: 48 8b 05 ([0-9a-f]{2} ){4} * mov 0x[a-f0-9]+\(%rip\),%rax # [a-f0-9]+ <_DYNAMIC\+0x[a-f0-9]+>
[ ]*[a-f0-9]+: 48 8b 05 ([0-9a-f]{2} ){4} * mov 0x[a-f0-9]+\(%rip\),%rax # [a-f0-9]+ <_DYNAMIC\+0x[a-f0-9]+>
[ ]*[a-f0-9]+: 48 8b 05 ([0-9a-f]{2} ){4} * mov 0x[a-f0-9]+\(%rip\),%rax # [a-f0-9]+ <_DYNAMIC\+0x[a-f0-9]+>
#pass

View file

@ -0,0 +1,15 @@
#source: mov2.s
#as: --x32
#ld: -pie -melf32_x86_64
#objdump: -dw
.*: +file format .*
Disassembly of section .text:
#...
[ ]*[a-f0-9]+: 48 8b 05 ([0-9a-f]{2} ){4} * mov 0x[a-f0-9]+\(%rip\),%rax # [a-f0-9]+ <_DYNAMIC\+0x[a-f0-9]+>
[ ]*[a-f0-9]+: 48 8b 05 ([0-9a-f]{2} ){4} * mov 0x[a-f0-9]+\(%rip\),%rax # [a-f0-9]+ <_DYNAMIC\+0x[a-f0-9]+>
[ ]*[a-f0-9]+: 48 8b 05 ([0-9a-f]{2} ){4} * mov 0x[a-f0-9]+\(%rip\),%rax # [a-f0-9]+ <_DYNAMIC\+0x[a-f0-9]+>
#pass

View file

@ -321,6 +321,10 @@ run_dump_test "mov1a"
run_dump_test "mov1b"
run_dump_test "mov1c"
run_dump_test "mov1d"
run_dump_test "mov2a"
run_dump_test "mov2b"
run_dump_test "mov2c"
run_dump_test "mov2d"
run_dump_test "load1a"
run_dump_test "load1b"
run_dump_test "load1c"