llvm dev - Sep 2017 - Errors linking with LLVM 5.0 - dump() missing

On 09/25/2017 06:19 PM, Matthias Braun wrote:
> 
>> On Sep 25, 2017, at 6:03 PM, David Keaton via llvm-dev <llvm-dev at
lists.llvm.org> wrote:
>>
>> On 09/25/2017 02:53 PM, Matthias Braun via llvm-dev wrote:
>>> The dump methods can be included in the release builds anyway by
enabling LLVM_ENABLE_DUMP.
>>
>>      Is there a way to pass LLVM_ENABLE_DUMP into cmake from outside
without changing any of the configuration files?  With Chapel's use of LLVM,
we try to avoid modifying any of the source files that come from the LLVM
distribution.  Our own Makefile that builds Chapel also invokes cmake to build
LLVM.  I have tried setting environment variables and passing command-line
arguments, and nothing seems to work.
> Looks like nobody got around adding a flag. (this really should be fixed).
> 
> In the meantime `cmake -DCMAKE_CXX_FLAGS="-DLLVM_ENABLE_DUMP"`
should do the trick.
      Thank you.  That was the flag we needed.

      Unfortunately, when I tried this just now, I found that the 
declaration of llvm::MachineRegisterInfo::dumpUses() in 
include/llvm/CodeGen/MachineRegisterInfo.h is missing the check against 
LLVM_ENABLE_DUMP.  It has only

#ifndef NDEBUG

which causes the build to fail because the definition in 
lib/CodeGen/MachineRegisterInfo.cpp is guarded by this.

#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)

					David

On 09/25/2017 06:47 PM, David Keaton via llvm-dev wrote:
> On 09/25/2017 06:19 PM, Matthias Braun wrote:
>>
>>> On Sep 25, 2017, at 6:03 PM, David Keaton via llvm-dev 
>>> <llvm-dev at lists.llvm.org> wrote:
>>>
>>> On 09/25/2017 02:53 PM, Matthias Braun via llvm-dev wrote:
>> In the meantime `cmake
-DCMAKE_CXX_FLAGS="-DLLVM_ENABLE_DUMP"` should
>> do the trick.
> 
>       Thank you.  That was the flag we needed.
> 
>       Unfortunately, when I tried this just now, I found that the 
> declaration of llvm::MachineRegisterInfo::dumpUses() in 
> include/llvm/CodeGen/MachineRegisterInfo.h is missing the check against 
> LLVM_ENABLE_DUMP.  It has only
> 
> #ifndef NDEBUG
> 
> which causes the build to fail because the definition in 
> lib/CodeGen/MachineRegisterInfo.cpp is guarded by this.
> 
> #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
      Unfortunately, this was only the beginning.  I tried updating the 
header file to test LLVM_ENABLE_DUMP as well, and when I did that, the 
build hit another such case, which I tried fixing, and then it hit 
another, and another.  Here is what I have found so far that would need 
to be fixed under the current arrangement.

LLVM_ENABLE_DUMP added to:

include/llvm/CodeGen/MachineRegisterInfo.h
	dumpUses()

include/llvm/CodeGen/MachineScheduler.h
	dumpScheduledState()

include/llvm/CodeGen/TargetSchedule.h
	getResourceName()

include/llvm/MC/MCSchedule.h
	Name

utils/TableGen/SubtargetEmitter.cpp
	emitted code inside EmitSchedModel()

Unknown location in AArch64
	This is not the end.  I just lost track of what to change at this point.

      Perhaps it would be easier to revert the change that hid the 
dump() methods.  That would have another advantage as well.  Before LLVM 
5.0.0, Chapel had the ability to link with an existing pre-built LLVM if 
it was installed on the system.  With LLVM 5.0.0, we cannot do that any 
longer because we need to compile with LLVM_ENABLE_DUMP defined.  It 
takes 10x longer to compile LLVM than it does to compile the rest of 
Chapel, so it is quite a burden to require a recompilation of LLVM to 
get the dump() methods.

					David

> On Sep 26, 2017, at 7:04 AM, David Keaton via llvm-dev <llvm-dev at
lists.llvm.org> wrote:
> 
> On 09/25/2017 06:47 PM, David Keaton via llvm-dev wrote:
>> On 09/25/2017 06:19 PM, Matthias Braun wrote:
>>> 
>>>> On Sep 25, 2017, at 6:03 PM, David Keaton via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
>>>> 
>>>> On 09/25/2017 02:53 PM, Matthias Braun via llvm-dev wrote:
>>> In the meantime `cmake
-DCMAKE_CXX_FLAGS="-DLLVM_ENABLE_DUMP"` should do the trick.
>>      Thank you.  That was the flag we needed.
>>      Unfortunately, when I tried this just now, I found that the
declaration of llvm::MachineRegisterInfo::dumpUses() in
include/llvm/CodeGen/MachineRegisterInfo.h is missing the check against
LLVM_ENABLE_DUMP.  It has only
>> #ifndef NDEBUG
>> which causes the build to fail because the definition in
lib/CodeGen/MachineRegisterInfo.cpp is guarded by this.
>> #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
> 
>     Unfortunately, this was only the beginning.  I tried updating the
header file to test LLVM_ENABLE_DUMP as well, and when I did that, the build hit
another such case, which I tried fixing, and then it hit another, and another. 
Here is what I have found so far that would need to be fixed under the current
arrangement.
> 
> LLVM_ENABLE_DUMP added to:
> 
> include/llvm/CodeGen/MachineRegisterInfo.h
> 	dumpUses()
> 
> include/llvm/CodeGen/MachineScheduler.h
> 	dumpScheduledState()
> 
> include/llvm/CodeGen/TargetSchedule.h
> 	getResourceName()
> 
> include/llvm/MC/MCSchedule.h
> 	Name
> 
> utils/TableGen/SubtargetEmitter.cpp
> 	emitted code inside EmitSchedModel()
> 
> Unknown location in AArch64
> 	This is not the end.  I just lost track of what to change at this point.
I guess that was to be expected with an option that didn't even have a cmake
flag. I would propose to just check for LLVM_ENABLE_DUMP in the sourcecode and
move the logic that debug builds have dump() methods to the cmake side of things
to simplify things. But I think that needs agreement and someone to write a
patch.

+CC some of the people who I believe made the decision for the current style.
> 
>     Perhaps it would be easier to revert the change that hid the dump()
methods.  That would have another advantage as well.  Before LLVM 5.0.0, Chapel
had the ability to link with an existing pre-built LLVM if it was installed on
the system.  With LLVM 5.0.0, we cannot do that any longer because we need to
compile with LLVM_ENABLE_DUMP defined.  It takes 10x longer to compile LLVM than
it does to compile the rest of Chapel, so it is quite a burden to require a
recompilation of LLVM to get the dump() methods.
Again: I consider client code calling dump() to be a misuse.

If you want to dump IR, can't you just switch your code from
    Value.dump();  to Value.print(dbgs(), true); dbgs() << '\n';
?

- Matthias