On Thursday 24 April 2014, 14:49:18, Carter Schonwald wrote:

What OS? Is this on a vm?

Oops, sorry. openSuSE 12.3 (64 bit), nothing virtual.

I7-avx is the instruction family. I5 will be ok.

Okay, then it's probably that what openSuSE calls 3.2 is not what GHC thinks 3.2 is.

(And note I base this assumption on the principle that e.g. ARM is currently working relatively well with the latest versions of LLVM, where that's the only option for the code generator. But we've had other bugs exposed here before, or exposed LLVM bugs, so it may just be whack-a-mole...) On Thu, Apr 24, 2014 at 2:04 PM, Austin Seipp wrote:

GHC 7.6 did not support LLVM 3.2 - it was only tested with up-to LLVM 3.1.

The odd version number for LLVM 3.2 was a known problem with their tarball, they forgot to remove the 'svn' suffix. This shouldn't be a problem for GHC, it should correctly parse the right thing anyway.

Can you try another LLVM version? LLVM 3.2 has been known to be particularly problematic - I believe me and David looked into this a while back, and we couldn't get even get the compiler to bootstrap with it at one point, let alone figure out what the hell was going on (I'll see if I can find the ticket related to this). LLVM 3.3 or LLVM 3.4 should work just fine, and you can put them on the front of your $PATH before running the testsuite to check this.

On Thu, Apr 24, 2014 at 1:57 PM, Daniel Fischer wrote:

...

On Thursday 24 April 2014, 14:49:18, Carter Schonwald wrote:

...

What OS? Is this on a vm?

Oops, sorry. openSuSE 12.3 (64 bit), nothing virtual.

...

I7-avx is the instruction family. I5 will be ok.

Okay, then it's probably that what openSuSE calls 3.2 is not what GHC thinks 3.2 is.

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://www.haskell.org/mailman/listinfo/ghc-devs

-- Regards,

Austin Seipp, Haskell Consultant Well-Typed LLP, http://www.well-typed.com/

-- Regards, Austin Seipp, Haskell Consultant Well-Typed LLP, http://www.well-typed.com/

We should reject LLVM 3.2 if it's known to be bad - I remember running into this on the RPi too. LlvmCodeGen already checks the version number and emits warnings if it's too old or too new. On 24/04/2014 20:04, Austin Seipp wrote:

GHC 7.6 did not support LLVM 3.2 - it was only tested with up-to LLVM 3.1.

The odd version number for LLVM 3.2 was a known problem with their tarball, they forgot to remove the 'svn' suffix. This shouldn't be a problem for GHC, it should correctly parse the right thing anyway.

Can you try another LLVM version? LLVM 3.2 has been known to be particularly problematic - I believe me and David looked into this a while back, and we couldn't get even get the compiler to bootstrap with it at one point, let alone figure out what the hell was going on (I'll see if I can find the ticket related to this). LLVM 3.3 or LLVM 3.4 should work just fine, and you can put them on the front of your $PATH before running the testsuite to check this.

On Thu, Apr 24, 2014 at 1:57 PM, Daniel Fischer wrote:

...

On Thursday 24 April 2014, 14:49:18, Carter Schonwald wrote:

...

What OS? Is this on a vm?

Oops, sorry. openSuSE 12.3 (64 bit), nothing virtual.

...

I7-avx is the instruction family. I5 will be ok.

Okay, then it's probably that what openSuSE calls 3.2 is not what GHC thinks 3.2 is.

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://www.haskell.org/mailman/listinfo/ghc-devs