LLVM 3.2 failure

Geoffrey Mainland mainland at apeiron.net
Thu Mar 14 22:48:34 CET 2013


On 03/14/2013 09:36 PM, Austin Seipp wrote:
> My stage2 compiler got built and also fails on any compilation, no
> matter how trivial. After linking stage2, my build fails with:
>
> $ make
> ===--- building phase 0
> make -r --no-print-directory -f ghc.mk phase=0 phase_0_builds
> make[1]: Nothing to be done for `phase_0_builds'.
> ===--- building phase 1
> make -r --no-print-directory -f ghc.mk phase=1 phase_1_builds
> make[1]: Nothing to be done for `phase_1_builds'.
> ===--- building final phase
> make -r --no-print-directory -f ghc.mk phase=final all
> "inplace/bin/ghc-stage2" -static -H64m -O0 -fllvm -package-name
old-time-1.1.0.1 -hide-all-packages -i -ilibraries/old-time/.
-ilibraries/old-time/dist-install/build
-ilibraries/old-time/dist-install/build/autogen
-Ilibraries/old-time/dist-install/build
-Ilibraries/old-time/dist-install/build/autogen
-Ilibraries/old-time/include -optP-include
-optPlibraries/old-time/dist-install/build/autogen/cabal_macros.h
-package base-4.7.0.0 -package old-locale-1.0.0.5 -XHaskell98 -XCPP
-XForeignFunctionInterface -O -fllvm -no-user-package-db -rtsopts -odir
libraries/old-time/dist-install/build -hidir
libraries/old-time/dist-install/build -stubdir
libraries/old-time/dist-install/build -hisuf hi -osuf o -hcsuf hc -c
libraries/old-time/dist-install/build/System/Time.hs -o
libraries/old-time/dist-install/build/System/Time.o
> make[1]: *** [libraries/old-time/dist-install/build/System/Time.o]
Segmentation fault: 11
>
>
> Just compiling 'hello world' is enough to trigger it, however:
>
> $ cat hi.hs
> main = putStrLn "hello world"
> $ ./inplace/bin/ghc-stage2 -v3 -fforce-recomp hi.hs ⏎
> Glasgow Haskell Compiler, Version 7.7.20130313, stage 2 booted by GHC
version 7.6.2
> Using binary package database:
/Users/a/ghc/ghc-pristine/inplace/lib/package.conf.d/package.cache
> wired-in package ghc-prim mapped to ghc-prim-0.3.1.0-inplace
> wired-in package integer-gmp mapped to integer-gmp-0.5.1.0-inplace
> wired-in package base mapped to base-4.7.0.0-inplace
> wired-in package rts mapped to builtin_rts
> wired-in package template-haskell mapped to
template-haskell-2.9.0.0-inplace
> wired-in package dph-seq not found.
> wired-in package dph-par not found.
> Hsc static flags:
> wired-in package ghc-prim mapped to ghc-prim-0.3.1.0-inplace
> wired-in package integer-gmp mapped to integer-gmp-0.5.1.0-inplace
> wired-in package base mapped to base-4.7.0.0-inplace
> wired-in package rts mapped to builtin_rts
> wired-in package template-haskell mapped to
template-haskell-2.9.0.0-inplace
> wired-in package dph-seq not found.
> wired-in package dph-par not found.
> *** Chasing dependencies:
> Chasing modules from: *hi.hs
> Stable obj: []
> Stable BCO: []
> Ready for upsweep
> [NONREC
> ModSummary {
> ms_hs_date = 2013-03-14 19:58:30 UTC
> ms_mod = main:Main,
> ms_textual_imps = [import (implicit) Prelude]
> ms_srcimps = []
> }]
> *** Deleting temp files:
> Deleting:
> compile: input file hi.hs
> Created temporary directory:
/var/folders/f6/rjtvxfp92j3ffvm3zs7hv7vh0000gn/T/ghc39205_0
> *** Checking old interface for main:Main:
> [1 of 1] Compiling Main ( hi.hs, hi.o )
> *** Parser:
> *** Renamer/typechecker:
> *** Desugar:
> Result size of Desugar (after optimization)
> = {terms: 7, types: 5, coercions: 0}
> *** Simplifier:
> [1] 39205 segmentation fault ./inplace/bin/ghc-stage2 -v3
-fforce-recomp hi.hs
>
> It says the segfault occurs near the simplifier, but I'm a little
skeptical this is the actual cause. Digging further with LLDB on my Mac
OS X machine, I can see this:
>
> $ lldb /Users/a/ghc/ghc-pristine/inplace/lib/bin/ghc-stage2 --
-B/Users/a/ghc/ghc-pristine/inplace/lib -v3 -fforce-recomp hi.hs
>
> Current executable set to
'/Users/a/ghc/ghc-pristine/inplace/lib/bin/ghc-stage2' (x86_64).
> (lldb) r
> Process 39787 launched:
'/Users/a/ghc/ghc-pristine/inplace/lib/bin/ghc-stage2' (x86_64)
> Glasgow Haskell Compiler, Version 7.7.20130313, stage 2 booted by GHC
version 7.6.2
> Using binary package database:
/Users/a/ghc/ghc-pristine/inplace/lib/package.conf.d/package.cache
> wired-in package ghc-prim mapped to ghc-prim-0.3.1.0-inplace
> wired-in package integer-gmp mapped to integer-gmp-0.5.1.0-inplace
> wired-in package base mapped to base-4.7.0.0-inplace
> wired-in package rts mapped to builtin_rts
> wired-in package template-haskell mapped to
template-haskell-2.9.0.0-inplace
> wired-in package dph-seq not found.
> wired-in package dph-par not found.
> Hsc static flags:
> wired-in package ghc-prim mapped to ghc-prim-0.3.1.0-inplace
> wired-in package integer-gmp mapped to integer-gmp-0.5.1.0-inplace
> wired-in package base mapped to base-4.7.0.0-inplace
> wired-in package rts mapped to builtin_rts
> wired-in package template-haskell mapped to
template-haskell-2.9.0.0-inplace
> wired-in package dph-seq not found.
> wired-in package dph-par not found.
> *** Chasing dependencies:
> Chasing modules from: *hi.hs
> Stable obj: []
> Stable BCO: []
> Ready for upsweep
> [NONREC
> ModSummary {
> ms_hs_date = 2013-03-14 19:58:30 UTC
> ms_mod = main:Main,
> ms_textual_imps = [import (implicit) Prelude]
> ms_srcimps = []
> }]
> *** Deleting temp files:
> Deleting:
> compile: input file hi.hs
> Created temporary directory:
/var/folders/f6/rjtvxfp92j3ffvm3zs7hv7vh0000gn/T/ghc39787_0
> *** Checking old interface for main:Main:
> [1 of 1] Compiling Main ( hi.hs, hi.o )
> *** Parser:
> *** Renamer/typechecker:
> *** Desugar:
> Result size of Desugar (after optimization)
> = {terms: 7, types: 5, coercions: 0}
> *** Simplifier:
> Process 39787 stopped
> * thread #1: tid = 0x1c03, 0x0000000101f9a153 ghc-stage2`threadPaused
+ 163, stop reason = EXC_BAD_ACCESS (code=1, address=0xfffffffffffffff8)
> frame #0: 0x0000000101f9a153 ghc-stage2`threadPaused + 163
> ghc-stage2`threadPaused + 163:
> -> 0x101f9a153: movl -8(%rax), %ecx
> 0x101f9a156: addl $-31, %ecx
> 0x101f9a159: cmpl $7, %ecx
> 0x101f9a15c: ja 0x101f9a2ed ; threadPaused + 573
>
> So the crash is actually happening in the RTS. Looking at the
> disassembly of the current frame, we see:
>
> (lldb) disassemble -f
> ghc-stage2`threadPaused:
> 0x101f9a0b0: pushq %rbp
> 0x101f9a0b1: pushq %r15
> 0x101f9a0b3: pushq %r14
> 0x101f9a0b5: pushq %r13
> 0x101f9a0b7: pushq %r12
> 0x101f9a0b9: pushq %rbx
> 0x101f9a0ba: subq $56, %rsp
> 0x101f9a0be: movq %rsi, %rbx
> 0x101f9a0c1: movq %rdi, 48(%rsp)
> 0x101f9a0c6: movq %rbx, %rsi
> 0x101f9a0c9: callq 0x101f8f350 ; maybePerformBlockedException
> 0x101f9a0ce: cmpw $3, 32(%rbx)
> 0x101f9a0d3: je 0x101f9a540 ; threadPaused + 1168
> 0x101f9a0d9: movq 24(%rbx), %rax
> 0x101f9a0dd: movl 8(%rax), %ecx
> 0x101f9a0e0: leaq 24(%rax,%rcx,8), %rcx
> 0x101f9a0e5: movq %rcx, 40(%rsp)
> 0x101f9a0ea: movq 16(%rax), %r14
> 0x101f9a0ee: xorl %eax, %eax
> 0x101f9a0f0: movl %eax, 12(%rsp)
> 0x101f9a0f4: leaq 1109(%rip), %r15 ; threadPaused + 1184
> 0x101f9a0fb: leaq 102230(%rip), %r12 ; stg_WHITEHOLE_info
> 0x101f9a102: movl %eax, 8(%rsp)
> 0x101f9a106: movl %eax, 32(%rsp)
> 0x101f9a10a: movl %eax, %r13d
> 0x101f9a10d: jmpq 0x101f9a305 ; threadPaused + 597
> 0x101f9a112: movslq (%r15,%rcx,4), %rcx
> 0x101f9a116: addq %r15, %rcx
> 0x101f9a119: jmpq *%rcx
> 0x101f9a11b: cmpq 1097902(%rip), %rax ; (void *)0x0000000101fb4a28:
stg_marked_upd_frame_info
> 0x101f9a122: jne 0x101f9a164 ; threadPaused + 180
> 0x101f9a124: testl %r13d, %r13d
> 0x101f9a127: je 0x101f9a310 ; threadPaused + 608
> 0x101f9a12d: movl 32(%rsp), %r13d
> 0x101f9a132: addl %r13d, 8(%rsp)
> 0x101f9a137: addl $2, 12(%rsp)
> 0x101f9a13c: jmpq 0x101f9a310 ; threadPaused + 608
> 0x101f9a141: nopw %cs:(%rax,%rax)
> 0x101f9a150: movq (%r14), %rax
> -> 0x101f9a153: movl -8(%rax), %ecx
> 0x101f9a156: addl $-31, %ecx
> 0x101f9a159: cmpl $7, %ecx
> 0x101f9a15c: ja 0x101f9a2ed ; threadPaused + 573
> ... lots more code
>
> The segfaulting instruction attempts to load ECX from RAX, but RAX is
null, hence the access violation on 0xfffffffffffffff8
>
> (lldb) register read
> General Purpose Registers:
> rax = 0x0000000000000000
> rbx = 0x00000001061d4000
> rcx = 0x00000000fffffff0
> rdx = 0x00000001024e10a0 ghc-stage2`large_alloc_lim
> rdi = 0x00000001024d8540 ghc-stage2`MainCapability
> rsi = 0x00000001061d4000
> rbp = 0x00000001020c6760 ghczm7zi7zi20130313_Demand_cprProdSig_closure
+ 16
> rsp = 0x00007fff5fbfb3f0
> r8 = 0x00000001020c8789
ghczm7zi7zi20130313_IdInfo_MayHaveCafRefs_closure + 1
> r9 = 0x00000001020c8779 ghczm7zi7zi20130313_IdInfo_NoLBVarInfo_closure + 1
> r10 = 0x0000000100685920 ghc-stage2`s5or_info
> r11 = 0x00000001051e98c8
> r12 = 0x0000000101fb3058 ghc-stage2`stg_WHITEHOLE_info
> r13 = 0x0000000000000000
> r14 = 0x00000001020c67d8
ghczm7zi7zi20130313_Demand_worthSplittingThunk_closure + 8
> r15 = 0x0000000101f9a550 ghc-stage2`threadPaused + 1184
> rip = 0x0000000101f9a153 ghc-stage2`threadPaused + 163
> rflags = 0x0000000000010287
> cs = 0x000000000000002b
> fs = 0x0000000000000000
> gs = 0x0000000000000000
>
>
> So there's still something weird going on. Looking at
rts/ThreadPaused.c in threadPaused, we see some code like:
>
> while ((P_)frame < stack_end) {
> info = get_ret_itbl(frame);
>
> switch (info->i.type) {
>
> case UPDATE_FRAME:
>
> // If we've already marked this frame, then stop here.
> if (frame->header.info == (StgInfoTable *)&stg_marked_upd_frame_info) {
> if (prev_was_update_frame) {
> words_to_squeeze += sizeofW(StgUpdateFrame);
> weight += weight_pending;
> weight_pending = 0;
> }
> goto end;
> }
>
> SET_INFO(frame, (StgInfoTable *)&stg_marked_upd_frame_info);
>
> bh = ((StgUpdateFrame *)frame)->updatee;
> bh_info = bh->header.info;
>
> Which I believe roughly corresponds to this assembly:
>
> 0x101f9a11b: cmpq 1097902(%rip), %rax ; (void *)0x0000000101fb4a28:
stg_marked_upd_frame_info
> 0x101f9a122: jne 0x101f9a164 ; threadPaused + 180 # check if
frame->header.info = stg_marked_upd_frame_info
> 0x101f9a124: testl %r13d, %r13d ; Check if prev_was_update_frame == 0
> 0x101f9a127: je 0x101f9a310 ; threadPaused + 608; if
prev_was_update_frame == 0
> 0x101f9a12d: movl 32(%rsp), %r13d ; increment words_to_squeeze, etc
> 0x101f9a132: addl %r13d, 8(%rsp) ; same as above
> 0x101f9a137: addl $2, 12(%rsp) ; same as above
> 0x101f9a13c: jmpq 0x101f9a310 ; threadPaused + 608; exit if
frame->header.info == stg_marked_up_frame_info
> 0x101f9a141: nopw %cs:(%rax,%rax) ; Crap opcodes for alignment
> 0x101f9a150: movq (%r14), %rax ; Load info into RAX
> -> 0x101f9a153: movl -8(%rax), %ecx ; deref info->i.type
>
> Due to optimization settings, the code is rather reorganized and
> coalesced for being nice to the processor, but the segfault actually
> occurs on this line:
>
> while ((P_)frame < stack_end) {
> info = get_ret_itbl(frame);
>
> switch (info->i.type) { <- SEGFAULT HERE
>
> So this seems to be some interaction between the compiler and info
> table layout, possibly? If we rebuild the stage2 compiler with a debug
> RTS and disassemble with source, we see the same thing:
>
> $ lldb ...
> ...
> ...
> *** Desugar:
> Result size of Desugar (after optimization)
> = {terms: 7, types: 5, coercions: 0}
> *** Simplifier:
> Process 42406 stopped
> * thread #1: tid = 0x1c03, 0x0000000101fb79c1
ghc-stage2`threadPaused(cap=0x0000000102521980, tso=0x00000001060d0000)
+ 177 at ThreadPaused.c:223, stop reason = EXC_BAD_ACCESS (code=1,
address=0xfffffffffffffff8)
> frame #0: 0x0000000101fb79c1
ghc-stage2`threadPaused(cap=0x0000000102521980, tso=0x00000001060d0000)
+ 177 at ThreadPaused.c:223
> 220 while ((P_)frame < stack_end) {
> 221 info = get_ret_itbl(frame);
> 222
> -> 223 switch (info->i.type) {
> 224
> 225 case UPDATE_FRAME:
> 226
> (lldb) disassemble -f -m
> ghc-stage2`threadPaused + 157 at ThreadPaused.c:221
> 220 while ((P_)frame < stack_end) {
> 221 info = get_ret_itbl(frame);
> 222
> 0x101fb79ad: movq -32(%rbp), %rax
> 0x101fb79b1: movq %rax, %rdi
> 0x101fb79b4: callq 0x101fb7730 ; get_ret_itbl at ClosureMacros.h:88
> 0x101fb79b9: movq %rax, -40(%rbp)
> ghc-stage2`threadPaused + 173 at ThreadPaused.c:223
> 222
> 223 switch (info->i.type) {
> 224
> 0x101fb79bd: movq -40(%rbp), %rax
> -> 0x101fb79c1: movl 16(%rax), %eax
> 0x101fb79c4: leal -37(%rax), %ecx
> 0x101fb79c7: cmpl $2, %ecx
> 0x101fb79ca: movl %eax, -96(%rbp)
> 0x101fb79cd: jb 0x101fb7c56 ; threadPaused + 838 at ThreadPaused.c:342
> 0x101fb79d3: movl -96(%rbp), %eax
> 0x101fb79d6: cmpl $35, %eax
> 0x101fb79d9: jne 0x101fb7c58 ; threadPaused + 840 at ThreadPaused.c:352
> ghc-stage2`threadPaused + 207 at ThreadPaused.c:228
> 227 // If we've already marked this frame, then stop here.
> 228 if (frame->header.info == (StgInfoTable
*)&stg_marked_upd_frame_info) {
> 229 if (prev_was_update_frame) {
> 0x101fb79df: movq -32(%rbp), %rax
> 0x101fb79e3: movq (%rax), %rax
> 0x101fb79e6: leaq 210779(%rip), %rcx ; stg_marked_upd_frame_info
> 0x101fb79ed: leaq (%rcx), %rcx
> 0x101fb79f0: cmpq %rcx, %rax
> 0x101fb79f3: jne 0x101fb7a1d ; threadPaused + 269 at ThreadPaused.c:237
> ...
> ...
> ...
>
>
> Geoffrey, if you have time, can you confirm this behavior on your
> Linux machine with LLVM 3.2? I think we should really fix this; it's
> rather unfortunate if we have to tell users to use some specific LLVM
> 3.3 SVN revision, or stay on 3.1 (and it's a pain to keep multiple
> LLVM installs synchronized.) On that note, we should test this with
> 3.1 as well possibly.
>
> You can rebuild a stage2 GHC with the debug RTS very easily - that'll
> give you RTS source and extra sanity checks, etc. Just run it under
> GDB instead and look at the trace. You can rebuild stage2 by saying
> (from the top-level source directory.)
>
> $ cd ghc
> $ make re2 GhcDebugged=YES
>
> And the new inplace/bin/ghc-stage2 compiler will have the debug
> runtime enabled. I don't have a lot of more time to dig at this exact
> moment. I'll look more tonight when I have time.
>
>
> On Thu, Mar 14, 2013 at 2:23 PM, Geoffrey Mainland
<mainland at apeiron.net> wrote:
>
> My stage 2 compiler was crashing the first time it was invoked.
>
> I just finished building GHC HEAD using LLVM compiled from HEAD, and that
> worked, so perhaps this was just a 3.2 bug. I have yet to run the
> testsuite though.
>
> Geoff
>
> On 03/14/2013 07:16 PM, Jan Stolarek wrote:
> > Goeff, Austin,
> >
> > do the build errors always happen with the same module or do
> > they occur randomly? If the former, which modules do you have
> > problems with?
> >
> > Janek
>
> --
> Regards,
> Austin

Yes, I'm seeing exactly this failure. I have been building stage2 with
the debug RTS. The current TSO's stack is getting corrupted, but I
haven't had time to dig in and find out why this is happening. I defined
the following macro and sprinkled it everywhere I saw the current TSO's
stack being modified:

#define CHECK_STACK(stackobj) \
    ASSERT((StgWord*) (stackobj)->sp >  (stackobj)->stack);                \
    ASSERT((StgWord*) (stackobj)->sp <= (stackobj)->stack +
(stackobj)->stack_size);

Make the following check the first statement in the function
threadPaused in ThreadPaused.c and I suspect you will get a failed
assertion quite quickly:

CHECK_STACK(tso->stackobj);

FWIW, LLVM 3.3 built from subversion seems fine...

Geoff




More information about the ghc-devs mailing list