Re: LLVM 3.2 failure

Glad to know I'm not crazy... There is already a ticket #7694 about the failure bootstrapping with LLVM 3.2. Also, Jan, could you send a fingerprint of your build tree? You can use utils/fingerprint/fingerprint.py to generate one. If your tree works and ours doesn't, that might help us narrow down the bug. Thanks, Geoff On 03/14/2013 02:40 PM, Austin Seipp wrote:

I was able to reproduce Geoffrey's failure on Mac OS X 10.8, with LLVM 3.2. The stage2 compiler eventually segfaults ("Segmentation Fault 11") during the build process after being compiled successfully with stage1.

Something recently happened, because I was bootstrapping fine with LLVM 3.2 recently after David's fixes landed (I filed a small raft of tickets.) It's times like these I really wish we had a reliable 'git bisect'...

I unfortunately haven't had time to dig into this, but I'll file a ticket to track it this morning. I can also reproduce this on my ARM/Linux machine. Previously, I got it to at least get done with stage2, and fail later in DPH. Now it seems to fail earlier in the same way the OS X build does.

OS's:

- 32bit ARM/Linux, Ubuntu 12.10 Linaro derivative; GCC 4.6.3. Bootstrapping compiler is GHC 7.4.1, LLVM 3.2. - 64bit OS X 10.8 Mountain Lion, llvm-gcc (XCode 4.6.) Bootstrapping compiler is GHC 7.6.2, LLVM 3.2.

Also, @Jan: the warnings during the build process probably come from your bootstrap compiler. The built compiler (stage1/stage2) both support LLVM 3.2 directly and have correct version checks, but the bootstrap compiler you're using won't. In practice this mismatch never proved a problem in the past; just weeks ago the entire tree validated with the LLVM build on OS X with no failures and I was working on ARM things.

I'll file a ticket and dig in soon when I get a chance.

On Thu, Mar 14, 2013 at 9:29 AM, Geoffrey Mainland wrote:

...

...

...

Hm, you're sure that LLVM 3.2 is in your path when you configure GHC? I removed LLVM 3.0 from my system so there's no possibility of mistaking 3.2 with 3.0. I'm also getting lots of compilation warnings about untested LLVM version -

On 03/14/2013 02:15 PM, Jan Stolarek wrote: this didn't happen with 3.0.

...

...

This is HEAD? Yes. Commit 56353e3da9d5718dfd25e25ccf61c78b25deefe8

...

Linux x64? Yes:

[killy@xerxes : ~] uname -a Linux xerxes.discovery 2.6.37.6-24-desktop #1 SMP PREEMPT 2012-10-18

22:36:08 +0200 x86_64 x86_64

...

x86_64 GNU/Linux

...

What is in your build.mk file? The relevant parts are:

BuildFlavour = quick-llvm

ifeq "$(BuildFlavour)" "quick-llvm"

SRC_HC_OPTS = -H64m -O0 -fllvm GhcStage1HcOpts = -O -fllvm GhcStage2HcOpts = -O0 -fllvm GhcLibHcOpts = -O -fllvm SplitObjs = NO HADDOCK_DOCS = NO BUILD_DOCBOOK_HTML = NO BUILD_DOCBOOK_PS = NO BUILD_DOCBOOK_PDF = NO

endif

As you can see I'm not building the documentation. It's because it fails to build on my system and I don't care much about resolving that.

Janek

You don't have the following line?

GhcLibWays = $(if $(filter $(DYNAMIC_BY_DEFAULT),YES),dyn,v)

I ask because I am using a stock build.mk copied from build.mk.sample with BuildFlavour = quick-llvm, GHC HEAD, and LLVM 3.2, and my stage2 compiler crashes. It would be good to know *exactly* what the contents of your build.mk are.

What version of GHC are you using to perform the build? Are you using parallel make?

Thanks, Geoff