* MIPS gas problem
@ 2000-04-20 2:13 Hiroyuki Machida
2000-04-20 10:16 ` Florian Lohoff
2000-04-20 12:07 ` Koundinya.K
0 siblings, 2 replies; 8+ messages in thread
From: Hiroyuki Machida @ 2000-04-20 2:13 UTC (permalink / raw)
To: binutils, linux, debian-mips
Hi
I found the problem "__attribute__ ((aligned(xx))" doesn't work
properly on MIPS/Linux. Please try to execute the attached test.
I think this problem can be reproduced on any ELF/MIPS box except
EMBEDED system which has OS name "elf".
I tracked down and finaly found gas/config/t-mips.c:s_change_sec(sec)
sets always ".rodata" section-alignment to 2**4. This should be set
to the maximum rodata object's alignment value.
% cc -c rotest.c -o rotest.o
% objdump -h rotest.o
rotest.o: file format elf32-littlemips
Sections:
Idx Name Size VMA LMA File off Algn
0 .text 00000330 0000000000000000 0000000000000000 00000040 2**4
CONTENTS, ALLOC, LOAD, RELOC, READONLY, CODE
1 .data 00000010 0000000000000000 0000000000000000 00000370 2**4
CONTENTS, ALLOC, LOAD, DATA
2 .bss 00000000 0000000000000000 0000000000000000 00000380 2**4
ALLOC
3 .reginfo 00000018 0000000000000000 0000000000000000 00000380 2**2
CONTENTS, ALLOC, LOAD, READONLY, DATA,LINK_ONCE_SAME_SIZE
4 .mdebug 000002f8 0000000000000000 0000000000000000 00000398 2**2
CONTENTS, READONLY, DEBUGGING
5 .rodata 00002060 0000000000000000 0000000000000000 00000690 2**4
CONTENTS, ALLOC, LOAD, READONLY, DATA
The sample fix for binutils 2.9.1 is following. (This can be
apply to current CVS version, I suppose.)
Does anyone know the reason why gas/config/t-mips.c set to "rodata'
section alignment to 2**4 and ignore the maximum rodata object
alignment.? We have to know why this restriction is made, anyway.
---
Hiroyuki Machida
Creative Station SCE Inc./Sony Corp.
Index: tc-mips.c
===================================================================
RCS file: /usr/cvsroot/caesar/src/compiler/binutils/gas/config/tc-mips.c,v
retrieving revision 1.4
retrieving revision 1.5
diff -U10 -p -r1.4 -r1.5
--- tc-mips.c 1999/07/07 01:03:55 1.4
+++ tc-mips.c 2000/04/12 06:50:19 1.5
@@ -9727,21 +9727,28 @@ s_change_sec (sec)
(subsegT) get_absolute_expression ());
if (OUTPUT_FLAVOR == bfd_target_elf_flavour)
{
bfd_set_section_flags (stdoutput, seg,
(SEC_ALLOC
| SEC_LOAD
| SEC_READONLY
| SEC_RELOC
| SEC_DATA));
if (strcmp (TARGET_OS, "elf") != 0)
- bfd_set_section_alignment (stdoutput, seg, 4);
+ {
+ static char first = 1;
+ if (first)
+ {
+ bfd_set_section_alignment (stdoutput, seg, 4);
+ first=0;
+ }
+ }
}
demand_empty_rest_of_line ();
}
else
{
as_bad ("No read only data section in this object file format");
demand_empty_rest_of_line ();
return;
}
break;
@@ -9749,21 +9756,28 @@ s_change_sec (sec)
case 's':
if (USE_GLOBAL_POINTER_OPT)
{
seg = subseg_new (".sdata", (subsegT) get_absolute_expression ());
if (OUTPUT_FLAVOR == bfd_target_elf_flavour)
{
bfd_set_section_flags (stdoutput, seg,
SEC_ALLOC | SEC_LOAD | SEC_RELOC
| SEC_DATA);
if (strcmp (TARGET_OS, "elf") != 0)
- bfd_set_section_alignment (stdoutput, seg, 4);
+ {
+ static char first = 1;
+ if (first)
+ {
+ bfd_set_section_alignment (stdoutput, seg, 4);
+ first=0;
+ }
+ }
}
demand_empty_rest_of_line ();
break;
}
else
{
as_bad ("Global pointers not supported; recompile -G 0");
demand_empty_rest_of_line ();
return;
}
--- end of patch
/*
* rotest.c - rodata alignment test.
*/
#define AL1 1024
#define AL2 4096
const char globalc='a';
const int global1 __attribute__ ((aligned(AL1))) =0 ;
const int global2 __attribute__ ((aligned(AL2))) =1;
const int local1 __attribute__ ((aligned(AL1))) =0 ;
const int local2 __attribute__ ((aligned(AL2))) =1;
int total_ng=0;
void check(char *addr, unsigned long al)
{
unsigned long mask = (al -1);
if ((unsigned long) addr & mask){
total_ng ++;
printf("err:%x expected:%x\n", addr,
(unsigned long) addr & ~mask);
}
}
int
main(void)
{
printf("* readonly local/gloabl\n");
printf("chcking align:%x\n",AL1);
check((char *)&local1,AL1); check((char *)&global1,AL1);
printf("%x:%x\n", &local1,&global1);
printf("chcking align:%x\n",4096);
check((char *)&global1,AL2); check((char *)&global2,AL2);
printf("%x:%x\n", &global1,&global2);
printf("\n");
if (total_ng) {
printf("NG:%d\n",total_ng);
} else {
printf("OK\n");
}
printf("\n");
return total_ng;
}
--- end of test
^ permalink raw reply [flat|nested] 8+ messages in thread* Re: MIPS gas problem 2000-04-20 2:13 MIPS gas problem Hiroyuki Machida @ 2000-04-20 10:16 ` Florian Lohoff 2000-04-20 12:07 ` Koundinya.K 1 sibling, 0 replies; 8+ messages in thread From: Florian Lohoff @ 2000-04-20 10:16 UTC (permalink / raw) To: Hiroyuki Machida; +Cc: linux On Thu, Apr 20, 2000 at 11:13:20AM +0900, Hiroyuki Machida wrote: > > I found the problem "__attribute__ ((aligned(xx))" doesn't work > properly on MIPS/Linux. Please try to execute the attached test. > I think this problem can be reproduced on any ELF/MIPS box except > EMBEDED system which has OS name "elf". > > I tracked down and finaly found gas/config/t-mips.c:s_change_sec(sec) > sets always ".rodata" section-alignment to 2**4. This should be set > to the maximum rodata object's alignment value. Hmm, if i understand that correctly i am seeing a different behaviour: (egcs 1.0.3a + binutils 2.8.1) [flo@repeat flo]$ gcc -save-temps -o rotest rotest.c rotest.c:8: warning: alignment of `global1' is greater than maximum object file alignment rotest.c:9: warning: alignment of `global2' is greater than maximum object file alignment rotest.c:11: warning: alignment of `local1' is greater than maximum object file alignment rotest.c:12: warning: alignment of `local2' is greater than maximum object file alignment And i see only a max alignment 3 in the rotest.s, and as expected the test fails: [flo@repeat flo]$ ./rotest * readonly local/gloabl chcking align:400 err:4007a8 expected:400400 err:400798 expected:400400 4007a8:400798 chcking align:1000 err:400798 expected:400000 err:4007a0 expected:400000 400798:4007a0 NG:4 Flo -- Florian Lohoff flo@rfc822.org +49-subject-2-change "Technology is a constant battle between manufacturers producing bigger and more idiot-proof systems and nature producing bigger and better idiots." ^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: MIPS gas problem 2000-04-20 2:13 MIPS gas problem Hiroyuki Machida 2000-04-20 10:16 ` Florian Lohoff @ 2000-04-20 12:07 ` Koundinya.K [not found] ` <20000420164812.17210.qmail@daffy.airs.com> 1 sibling, 1 reply; 8+ messages in thread From: Koundinya.K @ 2000-04-20 12:07 UTC (permalink / raw) To: Hiroyuki Machida; +Cc: binutils, linux, debian-mips machida@sm.sony.co.jp said: -> Hi -> I found the problem "__attribute__ ((aligned(xx))" doesn't work -> properly on MIPS/Linux. Please try to execute the attached test. I -> think this problem can be reproduced on any ELF/MIPS box except -> EMBEDED system which has OS name "elf". -> I tracked down and finaly found gas/config/t-mips.c:s_change_sec(sec) -> sets always ".rodata" section-alignment to 2**4. This should be set -> to the maximum rodata object's alignment value. I am seeing something different (gcc 2.95.2 and binutils from snapshot 000213) on my mips based machine (mips-dde-sysv4.2MP) [~] gcc -c rotest.c -o rotest.o rotest.c:8: warning: alignment of `global1' is greater than maximum object file alignment. Using 8. rotest.c:9: warning: alignment of `global2' is greater than maximum object file alignment. Using 8. rotest.c:11: warning: alignment of `local1' is greater than maximum object file alignment. Using 8. rotest.c:12: warning: alignment of `local2' is greater than maximum object file alignment. Using 8. koundinya ^ permalink raw reply [flat|nested] 8+ messages in thread
[parent not found: <20000420164812.17210.qmail@daffy.airs.com>]
* Re: MIPS gas problem [not found] ` <20000420164812.17210.qmail@daffy.airs.com> @ 2000-04-21 1:52 ` Hiroyuki Machida 2000-04-24 5:06 ` Koundinya.K 2000-04-21 3:30 ` Hiroyuki Machida 1 sibling, 1 reply; 8+ messages in thread From: Hiroyuki Machida @ 2000-04-21 1:52 UTC (permalink / raw) To: kk, flo, ian; +Cc: binutils, linux, debian-mips I rewrite the test program which is not depending your gcc version or config. Please try this. #define P1 10 #define P2 12 #define P1STR "10" #define P2STR "12" #define AL1 (2<<(P1-1)) #define AL2 (2<<(P2-1)) asm( " .globl globalc;\n" " .rdata;\n" " .type globalc,@object;\n" " .size globalc,1;\n" "globalc:\n" " .byte 97;\n" ); asm( " .globl global1;\n" " .align " P1STR ";\n" " .type global1,@object;\n" " .size global1,4;\n" "global1:\n" " .word 0;\n" ); asm( " .globl global2;\n" " .align " P2STR ";\n" " .type global2,@object;\n" " .size global2,4;\n" "global2:\n" " .word 1;\n" ); int total_ng=0; void check(char *addr, unsigned long al) { unsigned long mask = (al -1); if ((unsigned long) addr & mask){ total_ng ++; printf("err:%x expected:%x\n", addr, (unsigned long) addr & ~mask); } } int main(void) { extern int global1, global2; extern char globalc; printf("* readonly gloabl\n"); printf("chcking align:%x\n",AL1); check((char *)&global1,AL1); printf("%x\n", &global1); printf("chcking align:%x\n",AL2); check((char *)&global2,AL2); printf("%x\n", &global2); printf("\n"); if (total_ng) { printf("NG:%d\n",total_ng); } else { printf("OK\n"); } printf("\n"); return total_ng; } ^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: MIPS gas problem 2000-04-21 1:52 ` Hiroyuki Machida @ 2000-04-24 5:06 ` Koundinya.K 0 siblings, 0 replies; 8+ messages in thread From: Koundinya.K @ 2000-04-24 5:06 UTC (permalink / raw) To: Hiroyuki Machida; +Cc: flo, ian, binutils, linux, debian-mips Hi, machida@sm.sony.co.jp said: -> I rewrite the test program which is not depending your gcc version or -> config. Please try this. -> #define P1 10 #define P2 12 Sorry for the delay in trying out this test case. I was over the edge for the past 3 days :-) O.K, There were no warnings with this test case unlike the previous one. [~] gcc -c rotest1.c -o rotest1.o [~] objdump -h objtest1.o otest1.o: file format elf32-tradbigmips Sections: Idx Name Size VMA LMA File off Algn 0 .text 00000320 00000000 00000000 00000040 2**4 CONTENTS, ALLOC, LOAD, RELOC, READONLY, CODE 1 .data 00000010 00000000 00000000 00000360 2**4 CONTENTS, ALLOC, LOAD, DATA 2 .bss 00000000 00000000 00000000 00000370 2**4 ALLOC 3 .reginfo 00000018 00000000 00000000 00000370 2**2 CONTENTS, ALLOC, LOAD, READONLY, DATA, LINK_ONCE_SAME_SIZE 4 .mdebug 00000240 00000000 00000000 00000388 2**2 CONTENTS, READONLY, DEBUGGING 5 .note 00000014 00000000 00000000 000005c8 2**0 CONTENTS, READONLY 6 .comment 00000035 00000000 00000000 000005dc 2**0 CONTENTS, READONLY 7 .rodata 00001060 00000000 00000000 00000620 2**4 CONTENTS, ALLOC, LOAD, READONLY, DATA koundinya ^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: MIPS gas problem [not found] ` <20000420164812.17210.qmail@daffy.airs.com> 2000-04-21 1:52 ` Hiroyuki Machida @ 2000-04-21 3:30 ` Hiroyuki Machida [not found] ` <20000421154859.3820.qmail@daffy.airs.com> 1 sibling, 1 reply; 8+ messages in thread From: Hiroyuki Machida @ 2000-04-21 3:30 UTC (permalink / raw) To: ian; +Cc: kk, binutils, linux, debian-mips Hi Ian, From: Ian Lance Taylor <ian@zembu.com> Subject: Re: MIPS gas problem Date: 20 Apr 2000 09:48:12 -0700 > [~] gcc -c rotest.c -o rotest.o > rotest.c:8: warning: alignment of `global1' is greater than maximum object > file alignment. Using 8. This message is a gcc issue, as you said. But, my reported plobelm is really binutils issue. Sorry to confuse you. I will describe the problem using only asm file, again. Assmbling and linking attached al.S, I got the following result. % cc -o al -save-temps al.S % nm al.o | grep global 0000000000000400 R global1 0000000000001000 R global2 0000000000000000 R globalc % nm al | grep global 0000000000400dc0 R global1 00000000004019c0 R global2 00000000004009c0 R globalc Alignments of global1 and global2 in "al.o" looks good. But in "al" not. % objdump -h al.o | grep rodata 5 .rodata 00001010 0000000000000000 0000000000000000 00000250 2**4 % objdump -h al | grep rodata 9 .rodata 00001010 00000000004009c0 00000000004009c0 000009c0 2**4 I think when ld gathers read only section, the alignment will be broken. Because .rodata section alignment of al.o is not the maximum rodata object's alignment value(in this case 2**12). I tracked down and found that s_change_sec():gas/config/tc-mips.c resets section alignment to 2**4. So, the maximum rodata object's alignment value will be lost in the following sequence. .rdata // reset section-alignment to 2**4 .algin 10 // update section-alignment to 2**10 .type global1,@object; .size global1,4; global1: .word 0; .rdata // reset section-alignment to 2**4, Why?? .align 2 // not update ( 2 <= current section-alignment) string: .ascii "test\n" I want to know the reason why s_change_sec() always reset section-alignment to 2**4. I think this is wrong. --- al.S .abicalls .globl globalc; .rdata; .type globalc,@object; .size globalc,1; globalc: .byte 97; // rodata aligned 10 (1024) .globl global1; .align 10; .type global1,@object; .size global1,4; global1: .word 0; // rodata aligned 12 (4096) .globl global2; .align 12; .type global2,@object; .size global2,4; global2: .word 1; // data .globl total_ng .data .align 2 .type total_ng,@object .size total_ng,4 total_ng: .word 0 // rodata again .rdata .align 2 string: .ascii "test\n" .text .align 3 .globl main .ent main main: nop; nop nop .end main --- endof al.S --- Hiroyuki Machida Creative Station SCE Inc./Sony Corp. ^ permalink raw reply [flat|nested] 8+ messages in thread
[parent not found: <20000421154859.3820.qmail@daffy.airs.com>]
* Re: MIPS gas problem [not found] ` <20000421154859.3820.qmail@daffy.airs.com> @ 2000-04-24 3:16 ` Hiroyuki Machida 2000-04-24 3:16 ` Hiroyuki Machida 0 siblings, 1 reply; 8+ messages in thread From: Hiroyuki Machida @ 2000-04-24 3:16 UTC (permalink / raw) Cc: linux, debian-mips Thank you Ian and Alan. From: Ian Lance Taylor <ian@zembu.com> Subject: Re: MIPS gas problem Date: 21 Apr 2000 08:48:59 -0700 > Date: Fri, 21 Apr 2000 12:30:14 +0900 > From: Hiroyuki Machida <machida@sm.sony.co.jp> > > This message is a gcc issue, as you said. > > But, my reported plobelm is really binutils issue. > > I know. I didn't mean to imply otherwise. Alan Modra's suggested > patch was the right one. > > Ian I summarize the problem and attach patch. Please install it, if you can. I don't know what procedure is required to install the patch to CVS tree. *Problem Description: Current tc-mips.c:s_change_sec() always sets '.rdata' and '.sdata' section alignments to 2**4, even if those contain over 2**4 aligned objects. *Solution: Tc-mips.c:s_change_sec() should use record_alignment(), not bfd_set_section_alignment(), for preventing reset section alignments. *Patch: 2000-04-24 Machida Hiroyuki <machida@sm.sony.co.jp> tc-mips.c (s_change_sec): Use record_alignment(), not bfd_set_section_alignment() to preserve section aligenments. --- tc-mips.c.ORG Mon Apr 24 11:22:39 2000 +++ tc-mips.c Mon Apr 24 11:24:48 2000 @@ -10045,7 +10045,7 @@ s_change_sec (sec) | SEC_RELOC | SEC_DATA)); if (strcmp (TARGET_OS, "elf") != 0) - bfd_set_section_alignment (stdoutput, seg, 4); + record_alignment (stdoutput, seg, 4); } demand_empty_rest_of_line (); } @@ -10067,7 +10067,7 @@ s_change_sec (sec) SEC_ALLOC | SEC_LOAD | SEC_RELOC | SEC_DATA); if (strcmp (TARGET_OS, "elf") != 0) - bfd_set_section_alignment (stdoutput, seg, 4); + record_alignment (stdoutput, seg, 4); } demand_empty_rest_of_line (); break; ^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: MIPS gas problem 2000-04-24 3:16 ` Hiroyuki Machida @ 2000-04-24 3:16 ` Hiroyuki Machida 0 siblings, 0 replies; 8+ messages in thread From: Hiroyuki Machida @ 2000-04-24 3:16 UTC (permalink / raw) Cc: linux, debian-mips Thank you Ian and Alan. From: Ian Lance Taylor <ian@zembu.com> Subject: Re: MIPS gas problem Date: 21 Apr 2000 08:48:59 -0700 > Date: Fri, 21 Apr 2000 12:30:14 +0900 > From: Hiroyuki Machida <machida@sm.sony.co.jp> > > This message is a gcc issue, as you said. > > But, my reported plobelm is really binutils issue. > > I know. I didn't mean to imply otherwise. Alan Modra's suggested > patch was the right one. > > Ian I summarize the problem and attach patch. Please install it, if you can. I don't know what procedure is required to install the patch to CVS tree. *Problem Description: Current tc-mips.c:s_change_sec() always sets '.rdata' and '.sdata' section alignments to 2**4, even if those contain over 2**4 aligned objects. *Solution: Tc-mips.c:s_change_sec() should use record_alignment(), not bfd_set_section_alignment(), for preventing reset section alignments. *Patch: 2000-04-24 Machida Hiroyuki <machida@sm.sony.co.jp> tc-mips.c (s_change_sec): Use record_alignment(), not bfd_set_section_alignment() to preserve section aligenments. --- tc-mips.c.ORG Mon Apr 24 11:22:39 2000 +++ tc-mips.c Mon Apr 24 11:24:48 2000 @@ -10045,7 +10045,7 @@ s_change_sec (sec) | SEC_RELOC | SEC_DATA)); if (strcmp (TARGET_OS, "elf") != 0) - bfd_set_section_alignment (stdoutput, seg, 4); + record_alignment (stdoutput, seg, 4); } demand_empty_rest_of_line (); } @@ -10067,7 +10067,7 @@ s_change_sec (sec) SEC_ALLOC | SEC_LOAD | SEC_RELOC | SEC_DATA); if (strcmp (TARGET_OS, "elf") != 0) - bfd_set_section_alignment (stdoutput, seg, 4); + record_alignment (stdoutput, seg, 4); } demand_empty_rest_of_line (); break; ^ permalink raw reply [flat|nested] 8+ messages in thread
end of thread, other threads:[~2000-04-24 5:18 UTC | newest]
Thread overview: 8+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2000-04-20 2:13 MIPS gas problem Hiroyuki Machida
2000-04-20 10:16 ` Florian Lohoff
2000-04-20 12:07 ` Koundinya.K
[not found] ` <20000420164812.17210.qmail@daffy.airs.com>
2000-04-21 1:52 ` Hiroyuki Machida
2000-04-24 5:06 ` Koundinya.K
2000-04-21 3:30 ` Hiroyuki Machida
[not found] ` <20000421154859.3820.qmail@daffy.airs.com>
2000-04-24 3:16 ` Hiroyuki Machida
2000-04-24 3:16 ` Hiroyuki Machida
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox