GHC Causes opt to crash under LLVM 3.1 due to non const memcpy offset

[Axman6]

Under LLVM 3.1, opt is crashing when compiling code produced by GHC which makes calls to memcpy. After talking to the guys in #llvm they informed me that memcpy needs its alignment argument to be a const value, not from a value in a register. An example of the crash log:

alignment argument of memory intrinsics must be a constant int
  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %lnfVC, i8* %lnfVE, i64 %lnfVF, i32 %lnfVG, i1 false)
alignment argument of memory intrinsics must be a constant int
  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %lnfWc, i8* %lnfWe, i64 %lnfWf, i32 %lnfWg, i1 false)
alignment argument of memory intrinsics must be a constant int
  call void @llvm.memmove.p0i8.p0i8.i64(i8* %lnfWk, i8* %lnfWm, i64 %lnfWn, i32 %lnfWo, i1 false)
alignment argument of memory intrinsics must be a constant int
  call void @llvm.memmove.p0i8.p0i8.i64(i8* %lnfWJ, i8* %lnfWL, i64 %lnfWM, i32 %lnfWN, i1 false)
Broken module found, compilation aborted!
0  opt               0x000000010cec4be7 llvm::sys::ThreadLocal<llvm::PrettyStackTraceEntry const>::~ThreadLocal() + 6113
1  opt               0x000000010cec4eee llvm::sys::ThreadLocal<llvm::PrettyStackTraceEntry const>::~ThreadLocal() + 6888
2  libsystem_c.dylib 0x00007fff8dabacfa _sigtramp + 26
3  libsystem_c.dylib 000000000000000000 _sigtramp + 18446603338139325216
4  opt               0x000000010cec4dfc llvm::sys::ThreadLocal<llvm::PrettyStackTraceEntry const>::~ThreadLocal() + 6646
5  opt               0x000000010ce9865a llvm::StringMap<llvm::Value*, llvm::MallocAllocator>::clear() + 8942
6  opt               0x000000010ce975cd llvm::StringMap<llvm::Value*, llvm::MallocAllocator>::clear() + 4705
7  opt               0x000000010ce8b50b llvm::PassRegistrationListener::passEnumerate(llvm::PassInfo const*) + 12531
8  opt               0x000000010ce8b704 llvm::PassRegistrationListener::passEnumerate(llvm::PassInfo const*) + 13036
9  opt               0x000000010ce8b848 llvm::PassRegistrationListener::passEnumerate(llvm::PassInfo const*) + 13360
10 opt               0x000000010ce8bb28 llvm::PassRegistrationListener::passEnumerate(llvm::PassInfo const*) + 14096
11 opt               0x000000010ce8bf27 llvm::PassRegistrationListener::passEnumerate(llvm::PassInfo const*) + 15119
12 opt               0x000000010cc7a708 llvm::DenseMap<llvm::BasicBlock*, unsigned int, llvm::DenseMapInfo<llvm::BasicBlock*>, llvm::DenseMapInfo<unsigned int> >::init(unsigned int) + 4866
13 opt               0x000000010cc76364
Stack dump:
0.	Program arguments: opt mmulttest.ll -o /var/folders/0n/bjp22m948xlgtl006s6kwyj00000gv/T/ghc87286_0/ghc87286_0.bc -O3 -std-compile-opts -time-passes -loop-unroll 
1.	Running pass 'Function Pass Manager' on module 'mmulttest.ll'.
2.	Running pass 'Module Verifier' on function '@seo9_info'

[Axman6]

A minimal test case which produces the problem

[Axman6]

LLVM IR from compiling testcase.hs compiled with ghc -O2 -fllvm testcase.hs

[simonmar]

David, any chance you could look at this please?

[dterei]

Yes, not sure when but hopefully soon.

[davidterei@…]

commit 41b37a233e92b3f7df07f1d8d4240e80643dff96

Author: David Terei <davidterei@gmail.com>
Date:   Mon Jun 25 00:51:47 2012 -0700

    Fix #6158. LLVM 3.1 doesn't like certain constructions that 3.0 and
    earlier did, so we avoid them.

 compiler/llvmGen/LlvmCodeGen/CodeGen.hs |   18 +++++++++++++++---
 1 files changed, 15 insertions(+), 3 deletions(-)

[dterei]

OK fixed. Damn LLVM doesn't like this code anymore:

%a = i32 1
call ccc llvm.memcpy.p0i8.p0i8.i64( i8* %x, i8* %y, i64 %z, i32 %a, i1 0)

As while the alignment (a) is constant the fact it's in a variable throws LLVM off in 3.1. A work around for now is to run LLVM's constant propigation pass at the very start to inline the variable. e.g do this:

$ ghc --make testcase.hs -fllvm -O2 -optlo-constprop -optlo-O3

and it should compile fine.