| Summary: | ARM gcc 4.1 optimization bug | ||
|---|---|---|---|
| Product: | gcc | Reporter: | yfw <yfw.debian> |
| Component: | target | Assignee: | Paul Brook <pbrook> |
| Status: | RESOLVED FIXED | ||
| Severity: | major | CC: | dberlin, dirk.behme, enrico.scholz+bugsgcc, gcc-bugs, m.k.edwards, s_j_newbury, tbm |
| Priority: | P3 | ||
| Version: | 4.1.0 | ||
| Target Milestone: | 4.1.2 | ||
| Host: | i386 | Target: | arm xscale iwmmxt |
| Build: | i386 | Known to work: | |
| Known to fail: | Last reconfirmed: | 2006-07-18 21:42:28 | |
| Attachments: |
.i file of pcm_native.c and .s files for -Os, -O1 and -O2
Fix. |
||
We really need a self contianed example but I think this was already fixed for 4.1.1. Subject: Re: ARM gcc 4.1 optimization bug Hi pinskia, I tried to make a simple test example for this bug. But If I put the code from ALSA subsystem of Linux kernel to a test.c file, the gcc will product correct assembly code. :(. So I put the Linux kernel assembly code(with -Os and withoud -O) to bug reporter. I will try the 4.1.1 later. Thanks & Regards yfw On 30 Apr 2006 05:01:19 -0000, pinskia at gcc dot gnu dot org <gcc-bugzilla@gcc.gnu.org> wrote: > > > ------- Comment #1 from pinskia at gcc dot gnu dot org 2006-04-30 05:01 ------- > We really need a self contianed example but I think this was already fixed for > 4.1.1. > > > -- > > pinskia at gcc dot gnu dot org changed: > > What |Removed |Added > ---------------------------------------------------------------------------- > Status|UNCONFIRMED |WAITING > Component|c |target > > > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27363 > > ------- You are receiving this mail because: ------- > You reported the bug, or are watching the reporter. > Where can I get gcc 4.1.1? From the ftp site, the latest gcc 4.1 release is 4.1.0. Thanks. I tried the gcc 4.1.1 snapshot 20060421. The bug still there. The assembly code producted with -Os option is the same as gcc 4.1.0. (In reply to comment #4) > I tried the gcc 4.1.1 snapshot 20060421. The bug still there. The assembly code > producted with -Os option is the same as gcc 4.1.0. > Have you got anywhere with this? I wonder if I'm hitting it too. All the ALSA modules load but no devices become available. Kernels 2.6.16 - 2.6.17-rc4 + ALSA SoC GCC 4.1.0 & 4.1.1 Target arm-iwmmxt-linux-gnueabi (-mabi=linux-aapcs, -march=armv5te, -mtune=iwmmxt, -Os) Created attachment 11628 [details]
.i file of pcm_native.c and .s files for -Os, -O1 and -O2
Attached the .i and .s for -Os & -O2 (failing) and -O1 (working) (generated with --save-temps).
Used compile options (from Linux kernel 2.6.17-rc5):
arm-linux-gcc -Wp,-MD,sound/core/.pcm_native.o.d -nostdinc -isystem /usr/arm/arm-linux_4_1_0/bin/../lib/gcc/arm-linux/4.1.0/include -D__KERNEL__ -Iinclude -include include/linux/autoconf.h -mlittle-endian -Wall -Wundef -Wstrict-prototypes -Wno-trigraphs -fno-strict-aliasing -fno-common -O1 -fno-omit-frame-pointer -fno-optimize-sibling-calls -fno-omit-frame-pointer -mapcs -mno-sched-prolog -mabi=apcs-gnu -mno-thumb-interwork -D__LINUX_ARM_ARCH__=5 -march=armv5te -mtune=arm9tdmi -msoft-float -Uarm -Wdeclaration-after-statement -Wno-pointer-sign -DMODULE -D"KBUILD_STR(s)=#s" -D"KBUILD_BASENAME=KBUILD_STR(pcm_native)" -D"KBUILD_MODNAME=KBUILD_STR(snd_pcm)" --save-temps -c -o sound/core/pcm_native.o sound/core/pcm_native.c
Only -O1 & -O2 & -Os were modified.
Small note:
While compiling only pcm_native.c with -O1 and using the resulting modules (4 modules depend on it) while using -Os for all other modules make the error I observed go away. However, I get no sound output like if I compile all sound modules with -O1. Seems that some other files of kernels sound system have (the same?) optimization issue as well. But I think to start we should concentrate on this file.
Until a fix for this bug is found, there are two possible workarounds: - Compile kernels sound system as modules and compile these modules with -O1 instead of default -Os or -O2 (e.g. by changing main Makefile). - Use this patch from Fengwei Yin <yfw.debian@gmail.com> (then -Os or -O2 should work): -- linux/include/sound/pcm_params.h 2005-03-02 09:31:53.000000000 +0800 +++ linux-ok/include/sound/pcm_params.h 2006-06-08 09:57:11.000000000 +0800 @@ -196,6 +196,11 @@ INLINE int snd_mask_refine(snd_mask_t *m snd_mask_t old; assert(!snd_mask_empty(mask)); snd_mask_copy(&old, mask); + /* + * add the barrier to fix the optimization + * error of GCC 4.1 + */ + mb(); snd_mask_intersect(mask, v); if (snd_mask_empty(mask)) return -EINVAL; Note that this is only a workaround. It will drop the Alsa subsystem performance. One thing we have is some extra virtual operands from CCP: before: <bb 2>: mask_5 = &old; v_7 = mask_6; # SFT.2_33 = V_MAY_DEF <SFT.2_32>; *mask_5 = *v_7; mask_8 = mask_6; v_10 = v_9; i_11 = 0; goto <bb 4> (<L4>); after: <bb 2>: mask_5 = &old; v_7 = mask_6; # SFT.2 = V_MUST_DEF <SFT.2>; # VUSE <SFT.2>; old = *v_7; mask_8 = mask_6; v_10 = v_9; i_11 = 0; goto <bb 4> (<L4>); all of the SFT.2 are actually the same tree object. *** Bug 28362 has been marked as a duplicate of this bug. *** Bug #28362 contains a self contained example. Basically, it needs only | some_struct = *some_other_struct; to trigger this bug. I wonder for how much other segfaults/brokeness this bug is responsible for. Btw, happens for every gcc release (3.4.6, 4.0.3, 4.1.1). I'm working on this. Looks like a CSE bug. Subject: Bug 27363 Author: pbrook Date: Thu Jul 20 13:57:31 2006 New Revision: 115614 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=115614 Log: 2006-07-20 Paul Brook <paul@codesourcery.com> PR 27363 gcc/ * cse.c (cse_insn): Add destination addresses to hash table. Check if they are invalidated by this instruction. gcc/testsuite/ * gcc.dg/pr27363.c: New test. Added: trunk/gcc/testsuite/gcc.dg/pr27363.c Modified: trunk/gcc/ChangeLog trunk/gcc/cse.c trunk/gcc/testsuite/ChangeLog Subject: Bug 27363 Author: pbrook Date: Thu Jul 20 13:59:22 2006 New Revision: 115616 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=115616 Log: Backport from mainline. PR 27363 gcc/ * cse.c (cse_insn): Add destination addresses to hash table. Check if they are invalidated by this instruction. gcc/testsuite/ * gcc.dg/pr27363.c: New test. Added: branches/csl/sourcerygxx-4_1/gcc/testsuite/gcc.dg/pr27363.c Modified: branches/csl/sourcerygxx-4_1/ChangeLog.csl branches/csl/sourcerygxx-4_1/gcc/cse.c Subject: Bug 27363 Author: pbrook Date: Thu Jul 20 15:07:25 2006 New Revision: 115620 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=115620 Log: 2006-07-20 Paul Brook <paul@codesourcery.com> Backport from mainline. PR 27363 gcc/ * cse.c (cse_insn): Add destination addresses to hash table. Check if they are invalidated by this instruction. gcc/testsuite/ * gcc.dg/pr27363.c: New test. Added: branches/gcc-4_1-branch/gcc/testsuite/gcc.dg/pr27363.c Modified: branches/gcc-4_1-branch/gcc/ChangeLog branches/gcc-4_1-branch/gcc/cse.c branches/gcc-4_1-branch/gcc/testsuite/ChangeLog FIxed. Coooool. pbrook. I will try the latest gcc snapshot. Thanks a lot. Regards Yin, Fengwei Subject: Re: ARM gcc 4.1 optimization bug
yfw dot debian at gmail dot com wrote:
> ------- Comment #16 from yfw dot debian at gmail dot com 2006-07-22 06:21 -------
> Coooool. pbrook.
> I will try the latest gcc snapshot.
>
> Thanks a lot.
I tested it using crosstool and applying the patch manually
against GCC 4.1.0 and it works.
Many thanks from me as well,
Dirk
Created attachment 12464 [details] Fix. Copied from http://www.freaknet.org/martin/crosstool/patches/gcc-4.1.1/gcc-4.1.1-bugfix-27363.patch so that it linked to this report. Still generates bad code for snd_mask_refine in the gcc-4.1-20070115 snapshot. I have verified that the patch claimed to fix this bug is in this snapshot. My gcc is tuned for arm-926ejs, old ABI. -O1 works. |
The kernel is 2.6.14. When I build ALSA subsystem. I use following commandline: arm-iwmmxt-linux-gnueabi-gcc -Wp,-MD,sound/core/.pcm_native.o.d -nostdinc -isystem /usr/local/arm-iwmmxt-linux-gnueabi/bin/../lib/gcc/arm-iwmmxt-linux-gnueabi/4.1.0/include -D__KERNEL__ -Iinclude -include include/linux/autoconf.h -mlittle-endian -gdwarf-2 -Wall -Wundef -Wstrict-prototypes -Wno-trigraphs -fno-strict-aliasing -fno-common -ffreestanding -fno-omit-frame-pointer -fno-optimize-sibling-calls -gdwarf-2 -fno-omit-frame-pointer -mapcs -mno-sched-prolog -mabi=aapcs-linux -mno-thumb-interwork -D__LINUX_ARM_ARCH__=5 -march=armv5te -mtune=xscale -Wa,-mcpu=xscale -msoft-float -Uarm -Wdeclaration-after-statement -Wno-pointer-sign -gdwarf-2 -DKBUILD_BASENAME=pcm_native -DKBUILD_MODNAME=snd_pcm -Os -c -o sound/core/pcm_native.o sound/core/pcm_native.c And the function is like following (using arm-iwmmxt-linux-gnueabi-objdump -d pcm_nativ.o): 0000211c <snd_mask_refine>: 211c: e1a0c00d mov ip, sp 2120: e92dd8f0 stmdb sp!, {r4, r5, r6, r7, fp, ip, lr, pc} 2124: e24cb004 sub fp, ip, #4 ; 0x4 2128: e24dd020 sub sp, sp, #32 ; 0x20 212c: e5913000 ldr r3, [r1] 2130: e51b502c ldr r5, [fp, #-44] 2134: e24b603c sub r6, fp, #60 ; 0x3c 2138: e1a0c000 mov ip, r0 213c: e1a0e006 mov lr, r6 2140: e1a04000 mov r4, r0 2144: e0055003 and r5, r5, r3 2148: e1a07001 mov r7, r1 214c: e8bc000f ldmia ip!, {r0, r1, r2, r3} 2150: e8ae000f stmia lr!, {r0, r1, r2, r3} 2154: e89c000f ldmia ip, {r0, r1, r2, r3} 2158: e5845000 str r5, [r4] 215c: e597c004 ldr ip, [r7, #4] 2160: e3550000 cmp r5, #0 ; 0x0 2164: e88e000f stmia lr, {r0, r1, r2, r3} 2168: e001300c and r3, r1, ip /* r1 from 2154: e89c000f ldmia ip, {r0, r1, r2, r3} Using the wrong value. The r1 from this instruction should be used: 214c: ldmia ip!, {r0, r1, r2, r3} */ 216c: e1a00004 mov r0, r4 2170: e3a02008 mov r2, #8 ; 0x8 2174: e1a01006 mov r1, r6 2178: e5843004 str r3, [r4, #4] 217c: 1a000005 bne 2198 <snd_mask_refine+0x7c> 2180: e3530000 cmp r3, #0 ; 0x0 2184: e3e03015 mvn r3, #21 ; 0x15 2188: 1a000002 bne 2198 <snd_mask_refine+0x7c> 218c: e1a00003 mov r0, r3 2190: e24bd01c sub sp, fp, #28 ; 0x1c 2194: e89da8f0 ldmia sp, {r4, r5, r6, r7, fp, sp, pc} 2198: ebfffffe bl 0 <memcmp> 219c: e2503000 subs r3, r0, #0 ; 0x0 21a0: 13a03001 movne r3, #1 ; 0x1 21a4: eafffff8 b 218c <snd_mask_refine+0x70> The C code is like following: #define SNDRV_MASK_SIZE 2 struct mask_t { unsigned int bits[8]; }; typedef struct mask_t snd_mask_t; static inline int snd_mask_empty(const snd_mask_t *mask) { int i; for (i = 0; i < SNDRV_MASK_SIZE; i++) { if (mask->bits[i]) return 0; } return 1; } static inline void snd_mask_intersect(snd_mask_t *mask, const snd_mask_t *v) { int i; for (i = 0; i < SNDRV_MASK_SIZE; i++) mask->bits[i] &= v->bits[i]; } static inline void snd_mask_copy(snd_mask_t *mask, const snd_mask_t *v) { *mask = *v; } int snd_mask_refine(snd_mask_t *mask, const snd_mask_t *v) { snd_mask_t old; snd_mask_copy(&old, mask); snd_mask_intersect(mask, v); if (snd_mask_empty(mask)) return -1; return !snd_mask_eq(mask, &old); return 1; } When I remove the -O option, the ALSA works OK. the .s file is like following: 00000040 <snd_mask_refine>: 40: e1a0c00d mov ip, sp 44: e92dd800 stmdb sp!, {fp, ip, lr, pc} 48: e24cb004 sub fp, ip, #4 ; 0x4 4c: e24dd048 sub sp, sp, #72 ; 0x48 50: e50b0048 str r0, [fp, #-72] 54: e50b104c str r1, [fp, #-76] 58: e51b3048 ldr r3, [fp, #-72] 5c: e24be040 sub lr, fp, #64 ; 0x40 60: e1a0c003 mov ip, r3 64: e8bc000f ldmia ip!, {r0, r1, r2, r3} 68: e8ae000f stmia lr!, {r0, r1, r2, r3} 6c: e89c000f ldmia ip, {r0, r1, r2, r3} 70: e88e000f stmia lr, {r0, r1, r2, r3} 74: e3a03000 mov r3, #0 ; 0x0 78: e50b3020 str r3, [fp, #-32] 7c: ea00000c b b4 <snd_mask_refine+0x74> 80: e51b0020 ldr r0, [fp, #-32] 84: e51b2020 ldr r2, [fp, #-32] 88: e51b3048 ldr r3, [fp, #-72] 8c: e7931102 ldr r1, [r3, r2, lsl #2] 90: e51b2020 ldr r2, [fp, #-32] 94: e51b304c ldr r3, [fp, #-76] 98: e7933102 ldr r3, [r3, r2, lsl #2] 9c: e0012003 and r2, r1, r3 a0: e51b3048 ldr r3, [fp, #-72] a4: e7832100 str r2, [r3, r0, lsl #2] a8: e51b3020 ldr r3, [fp, #-32] ac: e2833001 add r3, r3, #1 ; 0x1 b0: e50b3020 str r3, [fp, #-32] b4: e51b3020 ldr r3, [fp, #-32] b8: e3530001 cmp r3, #1 ; 0x1 bc: daffffef ble 80 <snd_mask_refine+0x40> c0: e51b3048 ldr r3, [fp, #-72] c4: e50b3018 str r3, [fp, #-24] c8: e3a03000 mov r3, #0 ; 0x0 cc: e50b301c str r3, [fp, #-28] d0: ea00000a b 100 <snd_mask_refine+0xc0> d4: e51b301c ldr r3, [fp, #-28] d8: e51b2018 ldr r2, [fp, #-24] dc: e7923103 ldr r3, [r2, r3, lsl #2] e0: e3530000 cmp r3, #0 ; 0x0 e4: 0a000002 beq f4 <snd_mask_refine+0xb4> e8: e3a03000 mov r3, #0 ; 0x0 ec: e50b3050 str r3, [fp, #-80] f0: ea000007 b 114 <snd_mask_refine+0xd4> f4: e51b301c ldr r3, [fp, #-28] f8: e2833001 add r3, r3, #1 ; 0x1 fc: e50b301c str r3, [fp, #-28] 100: e51b301c ldr r3, [fp, #-28] 104: e3530001 cmp r3, #1 ; 0x1 108: dafffff1 ble d4 <snd_mask_refine+0x94> 10c: e3a03001 mov r3, #1 ; 0x1 110: e50b3050 str r3, [fp, #-80] 114: e51b3050 ldr r3, [fp, #-80] 118: e3530000 cmp r3, #0 ; 0x0 11c: 0a000002 beq 12c <snd_mask_refine+0xec> 120: e3e03015 mvn r3, #21 ; 0x15 124: e50b3054 str r3, [fp, #-84] 128: ea00000f b 16c <snd_mask_refine+0x12c> 12c: e51b3048 ldr r3, [fp, #-72] 130: e50b3010 str r3, [fp, #-16] 134: e24b3040 sub r3, fp, #64 ; 0x40 138: e50b3014 str r3, [fp, #-20] 13c: e51b0010 ldr r0, [fp, #-16] 140: e51b1014 ldr r1, [fp, #-20] 144: e3a02008 mov r2, #8 ; 0x8 148: ebfffffe bl 0 <memcmp> 14c: e1a03000 mov r3, r0 150: e3530000 cmp r3, #0 ; 0x0 154: 13a03000 movne r3, #0 ; 0x0 158: 03a03001 moveq r3, #1 ; 0x1 15c: e3530000 cmp r3, #0 ; 0x0 160: 13a03000 movne r3, #0 ; 0x0 164: 03a03001 moveq r3, #1 ; 0x1 168: e50b3054 str r3, [fp, #-84] 16c: e51b3054 ldr r3, [fp, #-84] 170: e1a00003 mov r0, r3 174: e24bd00c sub sp, fp, #12 ; 0xc 178: e89da800 ldmia sp, {fp, sp, pc}