#7588 closed bug (fixed)

GHC HEAD built with LLVM on Mac OS X miscompiles RTS: SIGSEGV in stg_PAP_apply

Reported by: thoughtpolice Owned by:
Priority: normal Milestone:
Component: Compiler (LLVM) Version: 7.7
Keywords: sigsegv Cc: dterei, simonmar
Operating System: MacOS X Architecture: Unknown/Multiple
Type of failure: Runtime crash Difficulty:
Test Case: Blocked By: #7571, #7580
Blocking: #7590 Related Tickets:

Description

After fixing #7571 and #7580, with those two patches, I now get a working stage1 compiler that can produce binaries using the LLVM backend. But all of them segfault:

$ cat no-op.hs
main = return ()
$ ~/code/haskell/ghc/inplace/bin/ghc-stage1 -fforce-recomp -fllvm no-op.hs
[1 of 1] Compiling Main             ( no-op.hs, no-op.o )
Linking no-op ...
$ ./no-op
[1]    12434 segmentation fault  ./no-op
$  

This looks like an error in stg_PAP_apply:

gdb -q ./no-op                                                                                                                            ⏎
Reading symbols for shared libraries .... done
(gdb) r
Starting program: /Users/a/t/no-op 
Reading symbols for shared libraries +++............................. done

Program received signal EXC_BAD_ACCESS, Could not access memory.
Reason: 13 at address: 0x0000000000000000
0x00000001002917d4 in stg_PAP_apply ()
(gdb) 

I imagine this is due to some miscompilation of rts/Apply.cmm using LLVM.

I'll rebuild the stage1 compiler with debugging support for sanity, and also enable the debug RTS in the test, and report back soon.

(I imagine this failure is certainly possible *because* of my patches in the other tickets, although my intuition tells me those are strictly correctness fixes and something else is afoot here.)

Change History (10)

comment:1 Changed 15 months ago by thoughtpolice

The stage1 compiler and RTS were both compiled with LLVM 3.2, btw.

comment:2 Changed 15 months ago by dterei

Yes, when this occurs it has always been due to a miscompilation of a handwritten cmm file. Apply.cmm or Update.cmm are usual culprits. You could also check that the mangler is still working fine, that may be another issue.

comment:3 Changed 15 months ago by thoughtpolice

Great, thanks for the reaffirmation, David. I'll look into it later tonight and investigate further.

comment:4 Changed 15 months ago by dterei

  • Blocking 7589 added

comment:5 Changed 15 months ago by dterei

  • Blocked By 7590 added

comment:6 Changed 15 months ago by dterei

  • Blocked By 7590 removed
  • Blocking 7590 added

comment:7 Changed 15 months ago by dterei

  • Blocking 7589 removed

comment:8 Changed 15 months ago by dterei

  • Status changed from new to infoneeded

Austin, can you confirm if this occurs still with HEAD?

comment:9 Changed 15 months ago by thoughtpolice

Yes, I saw your patches go by. Thanks a lot! I'll ./validate with the latest HEAD and see how far I get.

comment:10 Changed 15 months ago by thoughtpolice

  • Resolution set to fixed
  • Status changed from infoneeded to closed

I just ran validate on the latest copy of HEAD. This is fixed and the stage2 compiler is running the testsuite now.

Note: See TracTickets for help on using tickets.