gas generates incorrect ia64 unwind rlen values
Another gas bug with ia64 unwind data in binutils-2.11.90.0.8-12.
arch/ia64/lib/memset.S generates unwind data of
<memset>: [0xe0020000005e07c0-0xe0020000005e0be0), info at +0x6aa5e8
v1, flags=0x0 ( ), len=8 bytes
R1:prologue(rlen=1)
R3:body(rlen4)
B3:epilogue(t=1,ecount890)An epilogue count of 3890? I think not. But wait, it gets worse! The raw info data is
e0020000006aa5e8 01000000 00000100 0161c201 ea01b29e
0100 version 1 0000 no flags 01000000 ulen=1 (8 bytes), followed by 8 bytes 0161c201 ea01b29e
01 Prologue R1, rlen=1 61c201 Body R3, rlen4 ea According to IA-64 Software Conventions and Runtime
Architecture Guide August 2000 (24535802.pdf), Appendix B, ea is not a valid code, B3 records should be e0! Why is gas generating ea?
01 t=1
b29e uleb128 number with x80 set in both bytes! We have reached
the end of the data area (ulen=1) but uleb128 says there is more data to come.
Sorry for the surfeit of exclamation marks but the ia64 unwind code is driving me up the wall. I want kdb to backtrace correctly, all this garbage data just gets in the way.
Quoted 4 lines · Jim Wilson
Quoted 1 line
RH 7.2-ia64, Dec 15 2001.
The tools for this release were based on Summer 2000 FSF sources with patches. So they are about 2.5 years old now.
I know, but AFAICT it is the latest that is available on the RH site for ia64. If there is a more recent binutils ia64.rpm, point me at it and I will try it.
Quoted 3 lines
Are you sure this is an assembler problem? If the compiler is emitting bad unwind directives, then it could be a compiler problem. You didn't mention anything about looking at assembly code.
See my follow up mail about memset.S, that is pure assembler. But just to be sure, here is traps.c::ia64_fault assembler output. No spurious unwind directives from gcc but unwind is still wrong.
.text
.align 16 .align 32 .global ia64_fault# .proc ia64_fault# ia64_fault: .prologue 12, 37 .save ar.pfs, r38 alloc r38 = ar.pfs, 4, 3, 5, 0 .fframe 256 adds r12 = -256, r12 movl r14 = 34359738383 .save rp, r37 mov r37 = b0 ;; .body and r14 = r14, r33 movl r15 = 34359738372 [454 lines omitted for brevity] .L2933: mov r39 = r36 mov r40 = r33 br.call.sptk.many b0 = ia32_intercept# ;; cmp4.ne p6, p7 = 0, r8 (p7) br.cond.dpnt .L2880 addl r39 = @ltoff(.LC39), gp ;; ld8 r39 = [r39] br.call.sptk.many b0 = printk# ;; adds r14 = 8, r36 addl r39 = @ltoff(.LC40), gp mov r41 = r34 ;; ld8 r40 = [r14] ld8 r39 = [r39] mov r42 = r33 mov r43 = r35 br.call.sptk.many b0 = printk# [unwind tables in the object say that the body ends here] ;; addl r39 = 11, r0 mov r40 = r13 br.call.sptk.many b0 = force_sig# ;; br .L2880 .L2935: adds r34 = 144, r12 addl r40 = @ltoff(.LC41), gp shr.u r41 = r33, 16 ;; mov r39 = r34 ld8 r40 = [r40] br.call.sptk.many b0 = sprintf# ;; br .L2882 .L2936: adds r34 = 144, r12 addl r40 = @ltoff(.LC42), gp mov r41 = r32 ;; mov r39 = r34 ld8 r40 = [r40] br.call.sptk.many b0 = sprintf# ;; .L2882: mov r39 = r34 mov r40 = r36 mov r41 = r33 br.call.sptk.many b0 = die_if_kernel# ;; addl r39 = 4, r0 mov r40 = r13 br.call.sptk.many b0 = force_sig# ;; .L2880: mov ar.pfs = r38 mov b0 = r37 .restore sp adds r12 = 256, r12 br.ret.sptk.many b0 .endp ia64_fault#
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Keith Owens
Keith> Sorry for the surfeit of exclamation marks but the ia64 unwind code is Keith> driving me up the wall. I want kdb to backtrace correctly, all this Keith> garbage data just gets in the way.
Then use gcc-3.2 (or backport the fixes).
--david
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Keith Owens
Keith> On 16 Dec 2002 16:25:15 -0500,
Quoted 4 lines · Keith> Jim Wilson <[email protected]> wrote:
Quoted 4 lines
Quoted 1 line
RH 7.2-ia64, Dec 15 2001.
The tools for this release were based on Summer 2000 FSF sources with patches. So they are about 2.5 years old now.
Keith> I know, but AFAICT it is the latest that is available on the RH site Keith> for ia64. If there is a more recent binutils ia64.rpm, point me at it Keith> and I will try it.
Not Red Hat, but while Debian (unfortunately) still defaults to gcc-2.9x at the moment, you can do "apt-get install gcc-3.2" to get something reasonably modern.
--david
Quoted 3 lines
I know, but AFAICT it is the latest that is available on the RH site for ia64. If there is a more recent binutils ia64.rpm, point me at it and I will try it.
Yes, this is the latest one that we give away free of charge to non-customers. I think we stopped maintaining these tools a while ago when we started recommending RH AS 2.[01] instead of RHL 7.[012] for IA-64 customers.
If you want to try something newer, the FSF binutils sources are easily accessible at
http://sources.redhat.com/binutils
Quoted 1 line
See my follow up mail about
memset.S, that is pure assembler.
Right, I didn't look closely enough at the file name. There is unfortunately no testcase though. I'll have to grab a copy of the kernel sources in order to get a testcase.
Quoted 1 line
[454 lines omitted for brevity]
And this isn't usable as a testcase either.
Jim
Quoted 3 lines · Jim Wilson
Yes, this is the latest one that we give away free of charge to non-customers. I think we stopped maintaining these tools a while ago when we started recommending RH AS 2.[01] instead of RHL 7.[012] for IA-64 customers.
I built binutils-2.13.90.0.2-2 (from rh 8.0) on ia64. I extracted the generated assembler for ia64_fault as traps.S. The bug still exists in 2.13.90.0.2. traps.S has been sent to Jim Wilson as a test case.
/usr/src/redhat/BUILD/binutils-2.13.90.0.2/binutils/tmpdir/gas/as -x -mconstant-gp traps.S -o traps.o /usr/src/redhat/BUILD/binutils-2.13.90.0.2/binutils/readelf -u traps.o
Unwind section '.IA_64.unwind' at offset 0xb40 contains 1 entries:
<ia64_fault>: [0x0-0xad0), info at +0x0
v1, flags=0x0 ( ), len bytes
R2:prologue_gr(mask=[rp,ar.pfs],grsave=r37,rlen=8)
P7:pfs_when(t=0)
P7:mem_stack_f(t=2,size%6)
P7:rp_when(t=7)
R3:body(rlenG5) <= should be 511
B2:epilogue(t=2,ecount=0)
R1:prologue(rlen=0)0000000000000000 <ia64_fault>:
0: 0c 30 31 0e 80 05 [MFI] alloc r38=ar.pfs,12,7,0 6: 00 00 00 02 00 80 nop.f 0x0 c: 01 60 f8 8c adds r12=-256,r12 10: 04 00 00 00 01 00 [MLX] nop.m 0x0 16: 00 08 00 00 00 c0 movl r14=0x80000000f 1c: f1 00 00 60 20: 02 00 00 00 01 00 [MII] nop.m 0x0 26: 50 02 00 62 00 c0 mov r37°;; 2c: e1 08 31 80 and r14=r14,r33 ..... ab0: 00 00 00 00 01 00 [MII] nop.m 0x0 ab6: 00 30 01 55 00 00 mov.i ar.pfs=r38 abc: 50 0a 00 07 mov b0=r37 ac0: 1d 60 00 18 02 21 [MFB] adds r12%6,r12 ac6: 00 00 00 02 00 80 nop.f 0x0 acc: 08 00 84 00 br.ret.sptk.many b0;; ad0: 0c 00 00 00 01 00 [MFI] nop.m 0x0 <== end of code at slot 519 ad6: 00 00 00 02 00 00 nop.f 0x0 adc: 00 00 04 00 nop.i 0x0
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Keith Owens
Keith> I built binutils-2.13.90.0.2-2 (from rh 8.0) on ia64. I Keith> extracted the generated assembler for ia64_fault as traps.S. Keith> The bug still exists in 2.13.90.0.2. traps.S has been sent Keith> to Jim Wilson as a test case.
Which compiler did you use? With gcc-3.0, I get the info attached below. It looks correct to me (I only checked the total length, not the individual regions).
--david
<ia64_fault>: [0xe00000000442cd00-0xe00000000442d840], info at +0x545c80
v1, flags=0x0 (), len$ bytes
R2:prologue_gr(mask=[rp,ar.pfs],grsave=r37,rlen)
P7:pfs_when(t=0)
P7:mem_stack_f(t=1,size%6)
P7:rp_when(t)
R3:body(rlenU)
B1:label_state(label=1)
B2:epilogue(t=2,ecount=0)
R3:body(rlenG1)
B1:copy_state(label=1)e00000000442cd00 <ia64_fault>:e00000000442cd00: 00 30 31 0e 80 05 [MII] alloc r38=ar.pfs,12,
7,0
e00000000442cd06: c0 00 30 7c 46 00 adds r12=-256,r12
e00000000442cd0c: 00 00 04 00 nop.i 0x0
:
e00000000442d820: 00 70 c0 03 96 26 [MII] addl r14=-2077840,r1
e00000000442d826: 90 02 85 5e 29 60 shr.u r41=r33,16
e00000000442d82c: 04 61 04 84 adds r354,r12
e00000000442d830: 1d 00 00 00 01 00 [MFB] nop.m 0x0
e00000000442d836: 00 00 00 02 00 00 nop.f 0x0
e00000000442d83c: c0 fe ff 48 br.few e000000004426f0 <ia64_fault+0x9f0>;;David> Which compiler did you use? With gcc-3.0, I get the info David> attached below. It looks correct to me (I only checked the David> total length, not the individual regions).
Typo-alert: make that gcc-3.1 (not gcc-3.0).
--david
Quoted 8 lines · David Mosberger
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Keith Owens
Keith> I built binutils-2.13.90.0.2-2 (from rh 8.0) on ia64. I Keith> extracted the generated assembler for
ia64_faultastraps.S. Keith> The bug still exists in 2.13.90.0.2.traps.Shas been sent Keith> to Jim Wilson as a test case.Which compiler did you use?
I used gcc 2.96 20000731 (Red Hat Linux 7.1 2.96-101) to convert traps.c to traps.S, then extracted the headers and ia64_fault for testing against gas. traps.S only has one body. I wanted to avoid dumping 530+ lines to the list but ...
Interesting that a different gcc which generates two bodies gets valid unwind data, but this code with only one body generates invalid unwind data. David, what does your unwind data look like after assembling this code?
.file "traps.c" .pred.safe_across_calls p1-p5,p16-p63 .text .align 16 .align 32 .global ia64_fault# .proc ia64_fault# ia64_fault: .prologue 12, 37 .save ar.pfs, r38 alloc r38 = ar.pfs, 4, 3, 5, 0 .fframe 256 adds r12 = -256, r12 movl r14 = 34359738383 .save rp, r37 mov r37 = b0 ;; .body and r14 = r14, r33 movl r15 = 34359738372 ;; cmp.ne p6, p7 = r15, r14 adds r14 = 272, r12 ;; (p7) movl r15 = 8796093022208 mov r36 = r14 (p7) ld8 r14 = [r14] ;; (p7) or r14 = r15, r14 (p7) mov r15 = r36 ;; (p7) st8 [r15] = r14 (p7) br.cond.dpnt .L2880 adds r15 = -24, r32 ;; cmp.ltu p6, p7 = 23, r15 (p6) br.cond.dptk .L2936 addl r14 = @ltoff(.L2937), gp ;; ld8 r14 = [r14] ;; shladd r16 = r15, 3, r14 ;; ld8 r15 = [r16] ;; add r15 = r15, r14 ;; mov b6 = r15 br b6 .align 8 .L2937: data8 .L2883-.L2937 data8 .L2889-.L2937 data8 .L2903-.L2937 data8 .L2936-.L2937 data8 .L2936-.L2937 data8 .L2915-.L2937 data8 .L2936-.L2937 data8 .L2911-.L2937 data8 .L2926-.L2937 data8 .L2926-.L2937 data8 .L2929-.L2937 data8 .L2915-.L2937 data8 .L2915-.L2937 data8 .L2936-.L2937 data8 .L2936-.L2937 data8 .L2936-.L2937 data8 .L2936-.L2937 data8 .L2936-.L2937 data8 .L2936-.L2937 data8 .L2936-.L2937 data8 .L2936-.L2937 data8 .L2931-.L2937 data8 .L2933-.L2937 data8 .L2935-.L2937 .align 16 .L2883: extr.u r35 = r33, 4, 4 addl r14 = @ltoff(reason.3#), gp adds r34 = 144, r12 ;; cmp.ne p6, p7 = 3, r35 ld8 r14 = [r14] mov r39 = r34 ;; shladd r14 = r35, 3, r14 (p6) addl r42 = @ltoff(.LC31), gp ;; ld8 r41 = [r14] (p6) ld8 r42 = [r42] (p6) br.cond.dptk .L2885 tbit.z p6, p7 = r33, 37 ;; (p7) addl r42 = @ltoff(.LC29), gp ;; (p7) ld8 r42 = [r42] (p6) addl r42 = @ltoff(.LC30), gp ;; (p6) ld8 r42 = [r42] .L2885: addl r40 = @ltoff(.LC28), gp ;; ld8 r40 = [r40] br.call.sptk.many b0 = sprintf# ;; cmp.ne p6, p7 = 8, r35 (p6) br.cond.dptk .L2882 adds r15 = 64, r36 adds r14 = 8, r36 addl r39 = @ltoff(.LC32), gp ;; ld8 r40 = [r14] ld8 r41 = [r15] ld8 r39 = [r39] br.call.sptk.many b0 = printk# ;; br .L2880 .L2889: tbit.z p6, p7 = r33, 1 (p6) br.cond.dpnt .L2890 ld8 r14 = [r36] addl r15 = -524289, r0 ;; and r14 = r15, r14 ;; st8 [r36] = r14 #APP mov ar.k5=r13 #NO_APP adds r14 = 1496, r13 ;; ld4 r15 = [r14] ;; tbit.z p6, p7 = r15, 0 (p6) br.cond.dpnt .L2901 adds r39 = 2272, r13 br.call.sptk.many b0 = __ia64_load_fpu# ;; ld8 r14 = [r36] ;; and r14 = -33, r14 ;; st8 [r36] = r14 br .L2880 .L2901: br.call.sptk.many b0 = __ia64_init_fpu# ;; ld8 r14 = [r36] ;; or r14 = 32, r14 ;; st8 [r36] = r14 br .L2880 .L2890: adds r34 = 144, r12 addl r40 = @ltoff(.LC33), gp br .L2938 .L2903: ld8 r17 = [r36] ;; mov r15 = r17 ;; extr.u r14 = r15, 32, 2 ;; cmp4.eq p6, p7 = 0, r14 (p6) br.cond.dpnt .L2904 extr.u r14 = r33, 4, 4 adds r17 = 8, r36 extr.u r15 = r15, 41, 2 ;; cmp.ne p6, p7 = 2, r14 adds r16 = 20, r12 adds r18 = 44, r12 ;; (p7) addl r39 = 11, r0 adds r14 = 16, r12 (p7) addl r35 = 196610, r0 (p6) addl r39 = 4, r0 (p6) addl r35 = 196610, r0 mov r41 = r13 ;; st4 [r14] = r39 adds r14 = 24, r12 ;; st4 [r14] = r35 ld8 r14 = [r17] st4 [r16] = r0 ;; add r14 = r14, r15 adds r17 = 32, r12 ;; st8 [r17] = r14 adds r15 = 40, r12 addl r16 = 1, r0 ;; st4 [r15] = r32 adds r14 = 48, r12 adds r15 = 16, r12 st4 [r18] = r16 ;; mov r40 = r15 st8 [r14] = r33 br.call.sptk.many b0 = force_sig_info# ;; br .L2880 .L2904: adds r14 = 8, r36 ;; ld8 r39 = [r14] br.call.sptk.many b0 = search_exception_table# ;; cmp.eq p6, p7 = 0, r8 ;; (p6) mov r14 = r0 (p6) br.cond.dpnt .L2910 mov r40 = r8 mov r39 = r36 br.call.sptk.many b0 = handle_exception# ;; addl r14 = 1, r0 ;; .L2910: cmp4.eq p6, p7 = 0, r14 (p7) br.cond.dptk .L2880 adds r34 = 144, r12 addl r40 = @ltoff(.LC34), gp br .L2938 .L2911: ld8 r17 = [r36] ;; extr.u r14 = r17, 32, 2 ;; cmp4.eq p6, p7 = 0, r14 (p6) br.cond.dpnt .L2912 adds r14 = 16, r12 addl r39 = 4, r0 adds r16 = 24, r12 ;; st4 [r14] = r39 adds r18 = 8, r36 extr.u r17 = r17, 41, 2 addl r14 = 196610, r0 mov r41 = r13 ;; st4 [r16] = r14 ld8 r15 = [r18] adds r14 = 20, r12 adds r16 = 40, r12 ;; st4 [r14] = r0 add r15 = r15, r17 adds r14 = 32, r12 ;; st8 [r14] = r15 st4 [r16] = r32 adds r15 = 44, r12 addl r14 = 1, r0 ;; st4 [r15] = r14 adds r16 = 48, r12 adds r15 = 16, r12 ;; st8 [r16] = r33 br .L2939 .L2912: adds r34 = 144, r12 addl r40 = @ltoff(.LC35), gp br .L2938 .L2915: cmp.eq p6, p7 = 35, r32 ;; (p6) adds r15 = 24, r12 (p6) addl r14 = 196611, r0 (p6) mov r34 = r0 ;; (p6) st4 [r15] = r14 (p6) ld8 r17 = [r36] (p6) br.cond.dpnt .L2916 cmp.ltu p6, p7 = 35, r32 (p6) br.cond.dptk .L2923 ;; cmp.eq p6, p7 = 29, r32 (p6) br.cond.dpnt .L2917 ld8 r17 = [r36] br .L2916 ;; .L2923: cmp.eq p6, p7 = 36, r32 ;; (p6) adds r15 = 24, r12 (p6) addl r14 = 196610, r0 (p6) mov r34 = r0 ;; (p6) st4 [r15] = r14 (p6) ld8 r17 = [r36] (p6) br.cond.dpnt .L2916 ;; ld8 r17 = [r36] br .L2916 .L2917: addl r14 = 196612, r0 ;; ld8 r17 = [r36] adds r15 = 24, r12 ;; st4 [r15] = r14 tbit.nz p6, p7 = r17, 34 ;; (p7) adds r14 = 8, r36 ;; (p7) ld8 r34 = [r14] .L2916: extr.u r14 = r17, 32, 2 ;; cmp4.ne p6, p7 = 0, r14 (p6) br.cond.dptk .L2924 mov r40 = r32 mov r41 = r36 addl r39 = 4, r0 br.call.sptk.many b0 = kdb# ;; cmp4.eq p6, p7 = 0, r8 (p7) br.cond.dptk .L2880 .L2924: adds r16 = 16, r12 addl r39 = 5, r0 adds r15 = 44, r12 ;; mov r14 = r16 adds r17 = 48, r12 adds r18 = 32, r12 ;; st4 [r14] = r39, 4 adds r16 = 40, r12 mov r41 = r13 ;; st4 [r14] = r0 st4 [r15] = r0 adds r14 = 16, r12 st8 [r17] = r0 ;; mov r40 = r14 st8 [r18] = r34 br .L2940 .L2926: cmp.eq p6, p7 = 32, r32 ;; mov r40 = r36 mov r41 = r33 (p6) addl r39 = 1, r0 (p7) mov r39 = r0 br.call.sptk.many b0 = handle_fpu_swa# ;; cmp4.gt p6, p7 = r0, r8 (p6) br.cond.dpnt .L2928 adds r14 = 1496, r13 ;; ld8 r15 = [r14] ;; tbit.nz p6, p7 = r15, 7 (p7) br.cond.dpnt .L2880 .L2928: adds r15 = 16, r12 addl r39 = 8, r0 adds r17 = 8, r36 ;; mov r14 = r15 adds r18 = 24, r12 mov r41 = r13 ld8 r15 = [r36] ;; st4 [r14] = r39, 4 extr.u r15 = r15, 41, 2 ;; st4 [r14] = r0 ld8 r16 = [r17] addl r14 = 196615, r0 ;; st4 [r18] = r14 add r16 = r16, r15 adds r17 = 32, r12 ;; st8 [r17] = r16 addl r14 = 1, r0 adds r15 = 44, r12 ;; st4 [r15] = r14 adds r18 = 48, r12 adds r16 = 40, r12 adds r14 = 16, r12 ;; st8 [r18] = r33 mov r40 = r14 .L2940: st4 [r16] = r0 br.call.sptk.many b0 = force_sig_info# ;; br .L2880 .L2929: ld8 r20 = [r36] ;; extr.u r14 = r20, 32, 2 ;; cmp4.eq p6, p7 = 0, r14 (p6) br.cond.dpnt .L2930 adds r15 = 16, r12 addl r39 = 4, r0 adds r16 = 24, r12 ;; st4 [r15] = r39 addl r14 = 196617, r0 adds r17 = 20, r12 ;; st4 [r16] = r14 adds r15 = 8, r36 adds r18 = 44, r12 st4 [r17] = r0 adds r19 = 48, r12 extr.u r16 = r20, 41, 2 ;; ld8 r14 = [r15] mov r41 = r13 st4 [r18] = r0 adds r15 = 40, r12 ;; add r14 = r14, r16 st8 [r19] = r0 adds r16 = 32, r12 st4 [r15] = r0 adds r15 = 16, r12 ;; st8 [r16] = r14 .L2939: mov r40 = r15 br.call.sptk.many b0 = force_sig_info# ;; br .L2880 .L2930: adds r34 = 144, r12 addl r40 = @ltoff(.LC36), gp ;; .L2938: mov r39 = r34 ld8 r40 = [r40] br.call.sptk.many b0 = sprintf# ;; br .L2882 .L2931: mov r39 = r36 mov r40 = r33 br.call.sptk.many b0 = ia32_exception# ;; cmp4.ne p6, p7 = 0, r8 (p7) br.cond.dpnt .L2880 addl r39 = @ltoff(.LC37), gp ;; ld8 r39 = [r39] br.call.sptk.many b0 = printk# ;; addl r39 = @ltoff(.LC38), gp adds r14 = 8, r36 mov r41 = r34 ;; ld8 r40 = [r14] ld8 r39 = [r39] mov r42 = r33 br.call.sptk.many b0 = printk# ;; addl r39 = 11, r0 mov r40 = r13 br.call.sptk.many b0 = force_sig# ;; adds r34 = 144, r12 br .L2882 .L2933: mov r39 = r36 mov r40 = r33 br.call.sptk.many b0 = ia32_intercept# ;; cmp4.ne p6, p7 = 0, r8 (p7) br.cond.dpnt .L2880 addl r39 = @ltoff(.LC39), gp ;; ld8 r39 = [r39] br.call.sptk.many b0 = printk# ;; adds r14 = 8, r36 addl r39 = @ltoff(.LC40), gp mov r41 = r34 ;; ld8 r40 = [r14] ld8 r39 = [r39] mov r42 = r33 mov r43 = r35 br.call.sptk.many b0 = printk# ;; addl r39 = 11, r0 mov r40 = r13 br.call.sptk.many b0 = force_sig# ;; br .L2880 .L2935: adds r34 = 144, r12 addl r40 = @ltoff(.LC41), gp shr.u r41 = r33, 16 ;; mov r39 = r34 ld8 r40 = [r40] br.call.sptk.many b0 = sprintf# ;; br .L2882 .L2936: adds r34 = 144, r12 addl r40 = @ltoff(.LC42), gp mov r41 = r32 ;; mov r39 = r34 ld8 r40 = [r40] br.call.sptk.many b0 = sprintf# ;; .L2882: mov r39 = r34 mov r40 = r36 mov r41 = r33 br.call.sptk.many b0 = die_if_kernel# ;; addl r39 = 4, r0 mov r40 = r13 br.call.sptk.many b0 = force_sig# ;; .L2880: mov ar.pfs = r38 mov b0 = r37 .restore sp adds r12 = 256, r12 br.ret.sptk.many b0 .endp ia64_fault#
Keith> I used gcc 2.96 20000731 (Red Hat Linux 7.1 2.96-101) to Keith> convert traps.c to traps.S, then extracted the headers and Keith> ia64_fault for testing against gas. traps.S only has one Keith> body. I wanted to avoid dumping 530+ lines to the list but Keith> ...
Keith> Interesting that a different gcc which generates two bodies Keith> gets valid unwind data, but this code with only one body Keith> generates invalid unwind data. David, what does your unwind Keith> data look like after assembling this code?
Why do you insist on re-discovering old bugs? For the code in question, the problem is caused by the jump-table generated for the switch-statement. gcc 2.9x placed jump-tables in the text section, which violates the ia64 ABI and also confuses the assembler.
--david
Quoted 4 lines · David Mosberger
Why do you insist on re-discovering old bugs? For the code in question, the problem is caused by the jump-table generated for the switch-statement. gcc 2.9x placed jump-tables in the text section, which violates the ia64 ABI and also confuses the assembler.
Which version of gcc to use is not my decision, I have to use 2.96 :(. Thanks anyway.
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Mon, 16 Dec 2002 14:08:43 -0800, David Mosberger
Quoted 4 lines · <[email protected]> said:
There have been many unwind-related bug fixes to the toolchain. It's one reason why 2.9x is hopelessly obsolete. Distros should really switch to gcc-3.2.
--david
David:
I have a RHAS kernel compiled with *gcc3.2*. Using a script based on readelf/objdmp I found out that there are 7 instances in this kernel where 'rlen' may be wrong. The invariant the script is looking for is this:
Sum(rlen for various regions) = Number of slots in the code range.
The script found following violations of the invariant:
<ia64_trace_syscall>: [0xe00000000440e1a0-0xe00000000440e240), info at +0x54cdd8
lo = 440E1A0 hi = 440E240
sum_rlen = 28 no_slots = 30
*******ERROR ***********
sum_rlen: 28 != no_slots:30
<ia64_ret_from_clone>: [0xe00000000440e240-0xe00000000440e270), info at
+0x54ce08
lo = 440E240 hi = 440E270
sum_rlen = 7 no_slots = 9
*******ERROR ***********
sum_rlen: 7 != no_slots:9
<ia64_prepare_handle_unaligned>: [0xe00000000440e7a0-0xe00000000440e800),
info at +0x54cf70
lo = 440E7A0 hi = 440E800
sum_rlen = 17 no_slots = 18
*******ERROR ***********
sum_rlen: 17 != no_slots:18
<ia32_ret_from_clone>: [0xe0000000044506a0-0xe0000000044506d0), info at
+0x54fd00
lo = 44506A0 hi = 44506D0
sum_rlen = 7 no_slots = 9
*******ERROR ***********
sum_rlen: 7 != no_slots:9
<memset>: [0xe0000000049338a0-0xe000000004933cc0), info at +0x5803d0
lo = 49338A0 hi = 4933CC0
sum_rlen = 195 no_slots = 198
*******ERROR ***********
sum_rlen: 195 != no_slots:198
<memcpy>: [0xe0000000049365a0-0xe000000004936a40), info at +0x580618
lo = 49365A0 hi = 4936A40
sum_rlen = 219 no_slots = 222
*******ERROR ***********
sum_rlen: 219 != no_slots:222
code_range= 0xe000000004b18000-0xe000000004b182b0
lo = 4B18000 hi = 4B182B0
sum_rlen = 130 no_slots = 129
*******ERROR ***********
sum_rlen: 130 != no_slots:129-Harish
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Harish
Harish> I have a RHAS kernel compiled with *gcc3.2*. Using a script Harish> based on readelf/objdmp I found out that there are 7 Harish> instances in this kernel where 'rlen' may be wrong. The Harish> invariant the script is looking for is this: Sum(rlen for Harish> various regions) = Number of slots in the code range.
Harish> The script found following violations of the invariant:
Harish> <ia64_trace_syscall>:
Harish> <ia64_ret_from_clone>:
Harish> <ia64_prepare_handle_unaligned>:
Harish> <ia32_ret_from_clone>:
Harish> <memset>:
Harish> <memcpy>:
These are all handwritten assembly routines. I think I see what's wrong:
(a) for the first 4 routines, note how they all start with an empty prologue (e.g., PT_REGS_UNWIND_INFO(0)); these empty prologues seem to cause gas to not count the "nop"s at the beginning of the code. Or perhaps gas never counts the starting "nop"s and this is just one of the places where the first instruction cannot go into the first slot of a bundle.
(b) for memcpy and memset, I suspect the ".align 32" directives are confusing gas
(a) is probably a bug in gas. For (b), gas can't guarantee much because (in general) it can't know the alignment of the first instruction of a procedure (not beyond the minimum of 16 bytes). We should fix the assembly source here (use label_state/copy_state around the .align directives, or get rid of them alltogether). One thing we could do in gas, however, is add a warning so that a programmer will know when s/he accidentally uses a .align directive inside a region.
Harish> code_range= 0xe000000004b18000-0xe000000004b182b0
Harish> lo = 4B18000 hi = 4B182B0
Harish> sum_rlen = 130 no_slots = 129
Harish> *******ERROR ***********
Harish> sum_rlen: 130 != no_slots:129
Do you know what code this is? The addresses are not terribly useful because they'll vary from one kernel to another.
Is your script fully automatic? If so, perhaps we could add it to the kernel sources and run it as part of "make check". I'd love to have better unwind-info checking tools. Actually, it's even more important to use such tools to verify the correctness of the user-level unwind info.
Someone wants to volunteer to fix this stuff?
--david
RE: [Linux-ia64] Re: gas generates incorrect ia64 unwind rlen val
In reply to Keith OwensQuoted 11 lines
From: David Mosberger [mailto:[email protected]]
Sent: Thursday, December 19, 2002 3:19 PM
To: Patil, Harish
Cc: '[email protected]'; '[email protected]'
Subject: Re: [Linux-ia64] Re: gas generates incorrect ia64 unwind rlen valuesQuoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
On Tue, 17 Dec 2002 08:33:38 -0800, "Patil, Harish"
<[email protected]> said:
Quoted 7 lines
Harish> I have a RHAS kernel compiled with *gcc3.2*. Using a script Harish> based on readelf/objdmp I found out that there are 7 Harish> instances in this kernel where 'rlen' may be wrong. The Harish> invariant the script is looking for is this: Sum(rlen for Harish> various regions) = Number of slots in the code range.
Harish> The script found following violations of the invariant:
<stuff deleted>
Quoted 9 lines
Harish> code_range= 0xe000000004b18000-0xe000000004b182b0
Harish> lo = 4B18000 hi = 4B182B0
Harish>sum_rlen= 130no_slots= 129
Harish> *******ERROR ***********
Harish>sum_rlen: 130 != no_slots:129Do you know what code this is? The addresses are not terribly useful because they'll vary from one kernel to another.
Sorry, the procedure name somehow got lost.. I disassembled the image and the relevant procedure is: <ia64_sigtramp>
Quoted 5 lines
Is your script fully automatic? If so, perhaps we could add it to the kernel sources and run it as part of "make check". I'd love to have better unwind-info checking tools. Actually, it's even more important to use such tools to verify the correctness of the user-level unwind info.
Yes, the script is fully automatic and is included below (requires a 'readelf' that
understands "-u" option).
-Harish
<----------- BEGIN unwcheck.sh
#!/bin/sh
# Usage: unwcheck.sh <executable_file_name>
# Pre-requisite: readelf [from Gnu binutils package]
# Purpose: Check the following invariant
# For each code range in the input binary:
# Sum[ lengths of unwind regions] = Number of slots in code range.
# Author : Harish Patil
# First version: January 2002
# Modified : 2/13/2002
# Modified : 3/15/2002: duplicate detection
readelf -u $1 | gawk '\
function todec(hexstr){
dec = 0;
l = length(hexstr);
for (i = 1; i <= l; i++)
{
c = substr(hexstr, i, 1);
if (c = "A")
dec = dec*16 + 10;
else if (c = "B")
dec = dec*16 + 11;
else if (c = "C")
dec = dec*16 + 12;
else if (c = "D")
dec = dec*16 + 13;
else if (c = "E")
dec = dec*16 + 14;
else if (c = "F")
dec = dec*16 + 15;
else
dec = dec*16 + c;
}
return dec; }
BEGIN { first = 1; sum_rlen = 0; no_slots = 0; errors=0; no_code_ranges=0;
}
{
if (NF=5 && $3="info")
{
no_code_ranges += 1;
if (first = 0)
{
if (sum_rlen != no_slots)
{
print full_code_range;
print " ", "lo = ", lo, " hi =", hi;
print " ", "sum_rlen = ", sum_rlen, "no_slots = "
no_slots;
print " "," ", "*******ERROR ***********";
print " "," ", "sum_rlen:", sum_rlen, " != no_slots:"
no_slots;
errors += 1;
}
sum_rlen = 0;
}
full_code_range = $0;
code_range = $2;
gsub("),", "", code_range);
gsub("^.", "", code_range);
split(code_range, addr, "-");
lo = toupper(addr[1]);
code_range_lo[no_code_ranges] = addr[1];
occurs[addr[1]] += 1;
full_range[addr[1]] = $0;
gsub("0X.[0]*", "", lo);
hi = toupper(addr[2]);
gsub("0X.[0]*", "", hi);
no_slots = (todec(hi) - todec(lo))/ 16*3
first = 0;
}
if (index($0,"rlen") > 0 )
{
rlen_str = substr($0, index($0,"rlen"));
rlen = rlen_str;
gsub("rlen=", "", rlen);
gsub(")", "", rlen);
sum_rlen = sum_rlen + rlen;
}
}
END {
if (first = 0)
{
if (sum_rlen != no_slots)
{
print "code_range=", code_range;
print " ", "lo = ", lo, " hi =", hi;
print " ", "sum_rlen = ", sum_rlen, "no_slots = "
no_slots;
print " "," ", "*******ERROR ***********";
print " "," ", "sum_rlen:", sum_rlen, " != no_slots:"
no_slots;
errors += 1;
}
}
no_duplicates = 0;
for (i=1; i<=no_code_ranges; i++)
{
cr = code_range_lo[i];
if (reported_cr[cr]=1) continue;
if ( occurs[cr] > 1)
{
reported_cr[cr] = 1;
print "Code range low ", code_range_lo[i], ":", full_range[cr],
" occurs: ", occurs[cr], " times.";
print " ";
no_duplicates++;
}
}
print "==================="
print "Total errors:", errors, "/", no_code_ranges, " duplicates:",
no_duplicates;
print "===================" }
'
---> END unwcheck.sh [note the "'" (single quote) on the last line ]
RE: [Linux-ia64] Re: gas generates incorrect ia64 unwind rlen val
In reply to Keith OwensQuoted 9 lines
Quoted 8 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line · Harish
From: David Mosberger [mailto:[email protected]]
Sent: Thursday, December 19, 2002 3:19 PM
To: Patil, Harish
Cc: '[email protected]'; '[email protected]'
Subject: Re: [Linux-ia64] Re: gas generates incorrect ia64 unwind rlen values
Quoted 5 lines
Quoted 5 lines
Is your script fully automatic? If so, perhaps we could add it to the kernel sources and run it as part of "make check". I'd love to have better unwind-info checking tools. Actually, it's even more important to use such tools to verify the correctness of the user-level unwind info.
Harish> Yes, the script is fully automatic and is included below (requires a Harish> 'readelf' that understands "-u" option).
OK, I checked this script into arch/ia64/scripts/unwcheck.sh now. I had to hack it a bit because readelf v2.13.90.0.10 generates the code-ranges in the format:
[0xe000000004408000-0xe000000004408230],
Which isn't what the script expected.
I also fixed the handful of problems it pointed out. I agree that those are really gas bugs (except perhaps for the unwind info in gate.S, which was a bit clunky). But for now, I fixed things by adding explicit nop's and commenting out .align directives. The removal of .align may cost us a cycle or two in some routines, but I don't it's noticable in practice.
Thanks,
--david
Bjorn Helgaas
AppliedRe: [Linux-ia64] Re: gas generates incorrect ia64 unwind rlen val ues
In reply to Keith OwensQuoted 8 lines · David Mosberger
OK, I checked this script into
arch/ia64/scripts/unwcheck.shnow.
...
I also fixed the handful of problems it pointed out. I agree that those are really gas bugs (except perhaps for the unwind info ingate.S, which was a bit clunky). But for now, I fixed things by adding explicit nop's and commenting out .align directives. The removal of .align may cost us a cycle or two in some routines, but I don't it's noticable in practice.
I applied these changes to 2.4 as well.