RFC: Separate machine IR from lib/CodeGen into lib/MIR

hyp · May 26, 2015, 4:46pm

Hi all,

The CodeGen library is a big bag of interdependent bits. This caused
a circular dependency in the MIR serialization commit (r237954), which got
reverted in r238007.

I propose separating the machine IR out of CodeGen and into its own
MIR library, living at lib/MIR. This touches every target but it’s mostly a
mechanical change that renames the header files, although a couple of pre-rename
commits are needed to decouple several things in CodeGen. I’ve attached a proof
of concept patch that renames the files and applies cleanly to r238038.

The MIR serialization code will go under the new MIR library and not under
CodeGen which will break the dependency. The CodeGen library will link MIR
automatically so the build files for tools that link with CodeGen won’t have to
be changed.

Thanks for reading,
Alex

mir-move.patch (286 KB)

dexonsmith · May 26, 2015, 8:37pm

FWIW, I'm in favour of this. I'd be interested to hear if there's
a reason *not* to do this.

Here's some more context:

http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20150518/278031.html

resistor · May 26, 2015, 8:46pm

I have no objections, though I think lib/MachineIR would make a prettier bikeshed

—Owen

Chris_Lattner · May 27, 2015, 4:28am

+1. Codegen should also be renamed to something else that starts with “Machine” as well.

-Chris

qcolombet · May 27, 2015, 6:20am

+1.

Could those two be subdirectories of one “Machine-Related-Stuff” directory?
E.g.,
MachineStuff/IR
MachineStuff/CodeGen

Where MachineStuff is something meaningful :).

That way, they keep a logic bound, more formal than the naming convention.

My 2c.

Q.

Chris_Lattner · May 27, 2015, 3:08pm

+1.

Could those two be subdirectories of one “Machine-Related-Stuff” directory?
E.g.,
MachineStuff/IR
MachineStuff/CodeGen

Where MachineStuff is something meaningful :).

That way, they keep a logic bound, more formal than the naming convention.

Something like?

lib/Machine/IR
lib/Machine/Passes

Unless there will be many subdirectories, it seems slightly better to flatten out the layer.

Also, if we’re getting crazy here, CodeGen in clang should be renamed to IRGen, AsmPrinter should be renamed to MCGen, and SelectionDAG should be replaced

-Chris

rnk · May 27, 2015, 5:13pm

+1 for this. I think you probably want to limit the scope of your change to
just splitting out MachineIR and not trying to rename CodeGen, AsmPrinter,
etc.

dexonsmith · May 27, 2015, 5:22pm

Agreed :). The rest seems out-of-scope for Alex's internship.

Chandler_Carruth · May 27, 2015, 5:24pm

+1.

Could those two be subdirectories of one “Machine-Related-Stuff” directory?
E.g.,
MachineStuff/IR
MachineStuff/CodeGen

Where MachineStuff is something meaningful :).

That way, they keep a logic bound, more formal than the naming convention.

Something like?

lib/Machine/IR
lib/Machine/Passes

Unless there will be many subdirectories, it seems slightly better to flatten out the layer.

I strongly prefer breaking it out into subdirectories. There are a bunch of reasons:

It will grow.
Without it, we cannot have separate libraries, which will lose some options for shrinking the size of libraries.
Without separate libraries we can’t as easily enforce the layering separations between these things. For example, making this split will be extremely hard currently because there is a lot of inappropriate dependencies that will block splitting things out.

However, “IR” and “Passes” cover only two of the things in CodeGen. There is also the implementation of a lot of common infrastructure used by targets’ code generators.

My initial suggestion would be to just sink CodeGen into Machine/CodeGen, add the …/IR and …/Passes directories, and then start extracting things from CodeGen into the two more narrow directories. I think there is likely some stuff that should continue to live in a “code generator” infrastructure directory as it is neither part of the machine IR, nor is it part of any particular pass.

My suggested layering would be:

Passes depend on IR, CodeGen depends on both Passes and IR. The idea is that anything passes require should be embedded into the IR.

However, this won’t currently work. There are things that seem to be parallel but independent of the machine IR and are used by any machine passes. There are also things that clearly use the machine passes. Currently, I’m not sure how to cleanly divide this library up without really significant refactoring of every part of the code generator.

While I would like to see this happen, is it really a good idea to put this in the critical path of getting MIR serialized and deserialized?

Also, if we’re getting crazy here, CodeGen in clang should be renamed to IRGen, AsmPrinter should be renamed to MCGen, and SelectionDAG should be replaced

I’m happy to actually do the CodeGen → IRGen rename. I actually built the change but didn’t submit it because of the concerns of some out-of-tree users of Clang. I still have all the perl scripts and such I used sitting around.

Chandler_Carruth · May 27, 2015, 5:38pm

I think even this is going to be out-of-scope unless the existing code happens to have a clean layering separation. I don’t see one currently, but I might be missing it.

Specifically, I’m strongly opposed to splitting this out while there are circular references between any MI implementation and any CodeGen implementation. Not all of our linkers or platforms even allow this, but the errors tend to be brittle and sporadic rather than consistent.

dexonsmith · May 27, 2015, 5:41pm

+1.

Could those two be subdirectories of one “Machine-Related-Stuff” directory?
E.g.,
MachineStuff/IR
MachineStuff/CodeGen

Where MachineStuff is something meaningful :).

That way, they keep a logic bound, more formal than the naming convention.

Something like?

lib/Machine/IR
lib/Machine/Passes

Unless there will be many subdirectories, it seems slightly better to flatten out the layer.

There are two different layers we can flatten. We already have two
layers in lib/CodeGen. If we flatten the second layer and go with
Quentin's idea, we get something like this:

lib/Machine/MIR
lib/Machine/Passes
lib/Machine/AsmPrinter (or MCGen)
lib/Machine/SelectionDAG (or MIRGen?)

(IMO, lib/Machine/IR is too close to lib/IR.)

If we flatten the first, we get this:

lib/MachineIR
lib/CodeGen (or MachineIRGen)
lib/CodeGen/AsmPrinter (or MCGen)
lib/CodeGen/SelectionDAG (or ...)

Both seem pretty reasonable to me, but the first one seems a little
cleaner.

Also, if we’re getting crazy here, CodeGen in clang should be renamed to IRGen, AsmPrinter should be renamed to MCGen, and SelectionDAG should be replaced

+1

dexonsmith · May 27, 2015, 5:59pm

+1.

Could those two be subdirectories of one “Machine-Related-Stuff” directory?
E.g.,
MachineStuff/IR
MachineStuff/CodeGen

Where MachineStuff is something meaningful :).

That way, they keep a logic bound, more formal than the naming convention.

Something like?

lib/Machine/IR
lib/Machine/Passes

Unless there will be many subdirectories, it seems slightly better to flatten out the layer.

I strongly prefer breaking it out into subdirectories. There are a bunch of reasons:

1) It will grow.
2) Without it, we cannot have separate libraries, which will lose some options for shrinking the size of libraries.
3) Without separate libraries we can't as easily enforce the layering separations between these things. For example, making this split will be *extremely* hard currently because there is a lot of inappropriate dependencies that will block splitting things out.

However, "IR" and "Passes" cover only two of the things in CodeGen. There is also the implementation of a lot of common infrastructure used by targets' code generators.

My initial suggestion would be to just sink CodeGen into Machine/CodeGen, add the .../IR and .../Passes directories, and then start extracting things from CodeGen into the two more narrow directories. I think there is likely some stuff that should continue to live in a "code generator" infrastructure directory as it is neither part of the machine IR, nor is it part of any particular pass.

My suggested layering would be:

Passes depend on IR, CodeGen depends on both Passes and IR. The idea is that anything passes require should be embedded into the IR.

(Oops, missed this until after I sent my own response.

One thing I'd add to this from my email is that I think
lib/Machine/IR is likely to get confused with lib/IR for the same
reasons that lib/CodeGen is confusing between LLVM and Clang. IMO
lib/Machine/MIR is "safer".)

However, this won't currently work. There are things that seem to be parallel but independent of the machine IR and are used by any machine passes. There are also things that clearly use the machine passes. Currently, I'm not sure how to cleanly divide this library up without really significant refactoring of every part of the code generator.

While I would like to see this happen, is it really a good idea to put this in the critical path of getting MIR serialized and deserialized?

Not if it's as hard as you're saying. My impression was that Alex
was able to move the IR stuff he needed into a separately library
pretty trivially (based on the P.O.C. patch he posted), but if it's
not trivial he should just move on.

Also, if we’re getting crazy here, CodeGen in clang should be renamed to IRGen, AsmPrinter should be renamed to MCGen, and SelectionDAG should be replaced

I'm happy to actually do the CodeGen -> IRGen rename. I actually built the change but didn't submit it because of the concerns of some out-of-tree users of Clang. I still have all the perl scripts and such I used sitting around.

I'm a really big fan of this. If you supply the perl scripts
somehow (attached to a PR that you reference in the commit?) then
out-of-tree users shouldn't have a problem. Unless I'm missing
something?

Finkel_Hal_J · May 27, 2015, 6:11pm

From: "Duncan P. N. Exon Smith" <dexonsmith@apple.com>
To: "Chandler Carruth" <chandlerc@google.com>
Cc: "LLVM Developers Mailing List" <llvmdev@cs.uiuc.edu>
Sent: Wednesday, May 27, 2015 12:59:23 PM
Subject: Re: [LLVMdev] RFC: Separate machine IR from lib/CodeGen into lib/MIR

>
>>>
>>> +1.
>>>
>>> Could those two be subdirectories of one “Machine-Related-Stuff”
>>> directory?
>>> E.g.,
>>> MachineStuff/IR
>>> MachineStuff/CodeGen
>>>
>>> Where MachineStuff is something meaningful :).
>>>
>>> That way, they keep a logic bound, more formal than the naming
>>> convention.
>>
>> Something like?
>>
>> lib/Machine/IR
>> lib/Machine/Passes
>>
>> Unless there will be many subdirectories, it seems slightly better
>> to flatten out the layer.
>
> I strongly prefer breaking it out into subdirectories. There are a
> bunch of reasons:
>
> 1) It will grow.
> 2) Without it, we cannot have separate libraries, which will lose
> some options for shrinking the size of libraries.
> 3) Without separate libraries we can't as easily enforce the
> layering separations between these things. For example, making
> this split will be *extremely* hard currently because there is a
> lot of inappropriate dependencies that will block splitting things
> out.
>
> However, "IR" and "Passes" cover only two of the things in CodeGen.
> There is also the implementation of a lot of common infrastructure
> used by targets' code generators.
>
> My initial suggestion would be to just sink CodeGen into
> Machine/CodeGen, add the .../IR and .../Passes directories, and
> then start extracting things from CodeGen into the two more narrow
> directories. I think there is likely some stuff that should
> continue to live in a "code generator" infrastructure directory as
> it is neither part of the machine IR, nor is it part of any
> particular pass.
>
> My suggested layering would be:
>
> Passes depend on IR, CodeGen depends on both Passes and IR. The
> idea is that anything passes require should be embedded into the
> IR.
>

(Oops, missed this until after I sent my own response.

One thing I'd add to this from my email is that I think
lib/Machine/IR is likely to get confused with lib/IR for the same
reasons that lib/CodeGen is confusing between LLVM and Clang. IMO
lib/Machine/MIR is "safer".)

+1 -- Please only use IR to refer to LLVM IR. We should use some different abbreviation for any other IR we have in the backend.

> However, this won't currently work. There are things that seem to
> be parallel but independent of the machine IR and are used by any
> machine passes. There are also things that clearly use the machine
> passes. Currently, I'm not sure how to cleanly divide this library
> up without really significant refactoring of every part of the
> code generator.
>
> While I would like to see this happen, is it really a good idea to
> put this in the critical path of getting MIR serialized and
> deserialized?

Not if it's as hard as you're saying. My impression was that Alex
was able to move the IR stuff he needed into a separately library
pretty trivially (based on the P.O.C. patch he posted), but if it's
not trivial he should just move on.

>> Also, if we’re getting crazy here, CodeGen in clang should be
>> renamed to IRGen, AsmPrinter should be renamed to MCGen, and
>> SelectionDAG should be replaced
>
> I'm happy to actually do the CodeGen -> IRGen rename. I actually
> built the change but didn't submit it because of the concerns of
> some out-of-tree users of Clang. I still have all the perl scripts
> and such I used sitting around.

I'm a really big fan of this.

+1 -- Me too.

-Hal

If you supply the perl scripts

hyp · May 27, 2015, 9:01pm

>
>>>
>>> +1.
>>>
>>> Could those two be subdirectories of one “Machine-Related-Stuff”
directory?
>>> E.g.,
>>> MachineStuff/IR
>>> MachineStuff/CodeGen
>>>
>>> Where MachineStuff is something meaningful :).
>>>
>>> That way, they keep a logic bound, more formal than the naming
convention.
>>
>> Something like?
>>
>> lib/Machine/IR
>> lib/Machine/Passes
>>
>> Unless there will be many subdirectories, it seems slightly better to
flatten out the layer.
>
> I strongly prefer breaking it out into subdirectories. There are a bunch
of reasons:
>
> 1) It will grow.
> 2) Without it, we cannot have separate libraries, which will lose some
options for shrinking the size of libraries.
> 3) Without separate libraries we can't as easily enforce the layering
separations between these things. For example, making this split will be
*extremely* hard currently because there is a lot of inappropriate
dependencies that will block splitting things out.
>
> However, "IR" and "Passes" cover only two of the things in CodeGen.
There is also the implementation of a lot of common infrastructure used by
targets' code generators.
>
> My initial suggestion would be to just sink CodeGen into
Machine/CodeGen, add the .../IR and .../Passes directories, and then start
extracting things from CodeGen into the two more narrow directories. I
think there is likely some stuff that should continue to live in a "code
generator" infrastructure directory as it is neither part of the machine
IR, nor is it part of any particular pass.
>
> My suggested layering would be:
>
> Passes depend on IR, CodeGen depends on both Passes and IR. The idea is
that anything passes require should be embedded into the IR.
>

(Oops, missed this until after I sent my own response.

One thing I'd add to this from my email is that I think
lib/Machine/IR is likely to get confused with lib/IR for the same
reasons that lib/CodeGen is confusing between LLVM and Clang. IMO
lib/Machine/MIR is "safer".)

> However, this won't currently work. There are things that seem to be
parallel but independent of the machine IR and are used by any machine
passes. There are also things that clearly use the machine passes.
Currently, I'm not sure how to cleanly divide this library up without
really significant refactoring of every part of the code generator.
>
> While I would like to see this happen, is it really a good idea to put
this in the critical path of getting MIR serialized and deserialized?

Not if it's as hard as you're saying. My impression was that Alex
was able to move the IR stuff he needed into a separately library
pretty trivially (based on the P.O.C. patch he posted), but if it's
not trivial he should just move on.

While it's true that there are things that prevent a straightforward
library division, I don't think that they will pose such a big problem.
I have a brief list of things that have to be done before I can move the
Machine IR stuff from CodeGen to some other library:
- Move the SplitCriticalEdge method from MachineBasicBlock (
⚙ D10064 Separate Machine IR from CodeGen: Move SplitCriticalEdge out of MachineBasicBlock).
- Move the UnpackMachineBundles and FinalizeMachineBundles passes from
MachineInstrBundle.cpp. (⚙ D10070 Separate Machine IR from CodeGen: Move UnpackMachineBundles pass from MachineInstrBundle.cpp + 1 upcoming patch).
- Refactor SlotIndexes.h: keep the SlotIndexes pass and move the rest to
SlotIndex.h. Introduce a new container class in SlotIndex.h that will
extract the map
between machine instructions and slot indexes from the SlotIndexes pass
and remove the dependency on the pass from MachineBasicBlock's print method.
- WinEHFuncInfo - Either move parseEHInfo from WinEHPrepare.cpp to new file
WinEHFuncInfo.cpp or move the declaration of parseEHInfo from
WinEHFuncInfo.h
to a new header WinEHPrepare.h
- Get rid of several unused #includes in MachineIR cpp files that include
passes.

After that it should be possible to move the Machine IR files out of
CodeGen without them actually including any CodeGen headers.
And all the Machine IR passes like MachineFunctionPass,
MachineFunctionAnalysis, etc. will remain in CodeGen.

Alex.

dexonsmith · May 27, 2015, 9:18pm

>
>>>
>>> +1.
>>>
>>> Could those two be subdirectories of one “Machine-Related-Stuff” directory?
>>> E.g.,
>>> MachineStuff/IR
>>> MachineStuff/CodeGen
>>>
>>> Where MachineStuff is something meaningful :).
>>>
>>> That way, they keep a logic bound, more formal than the naming convention.
>>
>> Something like?
>>
>> lib/Machine/IR
>> lib/Machine/Passes
>>
>> Unless there will be many subdirectories, it seems slightly better to flatten out the layer.
>
> I strongly prefer breaking it out into subdirectories. There are a bunch of reasons:
>
> 1) It will grow.
> 2) Without it, we cannot have separate libraries, which will lose some options for shrinking the size of libraries.
> 3) Without separate libraries we can't as easily enforce the layering separations between these things. For example, making this split will be *extremely* hard currently because there is a lot of inappropriate dependencies that will block splitting things out.
>
> However, "IR" and "Passes" cover only two of the things in CodeGen. There is also the implementation of a lot of common infrastructure used by targets' code generators.
>
> My initial suggestion would be to just sink CodeGen into Machine/CodeGen, add the .../IR and .../Passes directories, and then start extracting things from CodeGen into the two more narrow directories. I think there is likely some stuff that should continue to live in a "code generator" infrastructure directory as it is neither part of the machine IR, nor is it part of any particular pass.
>
> My suggested layering would be:
>
> Passes depend on IR, CodeGen depends on both Passes and IR. The idea is that anything passes require should be embedded into the IR.
>

(Oops, missed this until after I sent my own response.

One thing I'd add to this from my email is that I think
lib/Machine/IR is likely to get confused with lib/IR for the same
reasons that lib/CodeGen is confusing between LLVM and Clang. IMO
lib/Machine/MIR is "safer".)

> However, this won't currently work. There are things that seem to be parallel but independent of the machine IR and are used by any machine passes. There are also things that clearly use the machine passes. Currently, I'm not sure how to cleanly divide this library up without really significant refactoring of every part of the code generator.
>
> While I would like to see this happen, is it really a good idea to put this in the critical path of getting MIR serialized and deserialized?

Not if it's as hard as you're saying. My impression was that Alex
was able to move the IR stuff he needed into a separately library
pretty trivially (based on the P.O.C. patch he posted), but if it's
not trivial he should just move on.

While it's true that there are things that prevent a straightforward library division, I don't think that they will pose such a big problem.
I have a brief list of things that have to be done before I can move the Machine IR stuff from CodeGen to some other library:
- Move the SplitCriticalEdge method from MachineBasicBlock (⚙ D10064 Separate Machine IR from CodeGen: Move SplitCriticalEdge out of MachineBasicBlock).
- Move the UnpackMachineBundles and FinalizeMachineBundles passes from MachineInstrBundle.cpp. (⚙ D10070 Separate Machine IR from CodeGen: Move UnpackMachineBundles pass from MachineInstrBundle.cpp + 1 upcoming patch).
- Refactor SlotIndexes.h: keep the SlotIndexes pass and move the rest to SlotIndex.h. Introduce a new container class in SlotIndex.h that will extract the map
between machine instructions and slot indexes from the SlotIndexes pass and remove the dependency on the pass from MachineBasicBlock's print method.
- WinEHFuncInfo - Either move parseEHInfo from WinEHPrepare.cpp to new file WinEHFuncInfo.cpp or move the declaration of parseEHInfo from WinEHFuncInfo.h
to a new header WinEHPrepare.h
- Get rid of several unused #includes in MachineIR cpp files that include passes.

After that it should be possible to move the Machine IR files out of CodeGen without them actually including any CodeGen headers.
And all the Machine IR passes like MachineFunctionPass, MachineFunctionAnalysis, etc. will remain in CodeGen.

Alex.

Something Chandler pointed out on IRC is that #include dependencies
aren't enough -- we also need to root out any symbol dependencies.
Have you checked for symbols (mostly, functions) that are declared
in a header you're planning to move, but defined in a source file
you're *not* planning to move?

(Apparently there have been failed attempts in the past to bring
sanity here that I wasn't aware of...)

hyp · May 27, 2015, 9:25pm

I didn’t specifically check for them, but the WinEHFuncInfo header file problem that I described above
is an example of a symbol dependency that can be resolved in a fairly straight forward manner.

The remaining files seem pretty sane, but there might be some symbol dependencies that I haven’t
caught yet.

Alex.

majnemer · May 27, 2015, 10:40pm

> From: "Duncan P. N. Exon Smith" <dexonsmith@apple.com>
> To: "Chandler Carruth" <chandlerc@google.com>
> Cc: "LLVM Developers Mailing List" <llvmdev@cs.uiuc.edu>
> Sent: Wednesday, May 27, 2015 12:59:23 PM
> Subject: Re: [LLVMdev] RFC: Separate machine IR from lib/CodeGen into
lib/MIR
>
> >
> >>>
> >>> +1.
> >>>
> >>> Could those two be subdirectories of one “Machine-Related-Stuff”
> >>> directory?
> >>> E.g.,
> >>> MachineStuff/IR
> >>> MachineStuff/CodeGen
> >>>
> >>> Where MachineStuff is something meaningful :).
> >>>
> >>> That way, they keep a logic bound, more formal than the naming
> >>> convention.
> >>
> >> Something like?
> >>
> >> lib/Machine/IR
> >> lib/Machine/Passes
> >>
> >> Unless there will be many subdirectories, it seems slightly better
> >> to flatten out the layer.
> >
> > I strongly prefer breaking it out into subdirectories. There are a
> > bunch of reasons:
> >
> > 1) It will grow.
> > 2) Without it, we cannot have separate libraries, which will lose
> > some options for shrinking the size of libraries.
> > 3) Without separate libraries we can't as easily enforce the
> > layering separations between these things. For example, making
> > this split will be *extremely* hard currently because there is a
> > lot of inappropriate dependencies that will block splitting things
> > out.
> >
> > However, "IR" and "Passes" cover only two of the things in CodeGen.
> > There is also the implementation of a lot of common infrastructure
> > used by targets' code generators.
> >
> > My initial suggestion would be to just sink CodeGen into
> > Machine/CodeGen, add the .../IR and .../Passes directories, and
> > then start extracting things from CodeGen into the two more narrow
> > directories. I think there is likely some stuff that should
> > continue to live in a "code generator" infrastructure directory as
> > it is neither part of the machine IR, nor is it part of any
> > particular pass.
> >
> > My suggested layering would be:
> >
> > Passes depend on IR, CodeGen depends on both Passes and IR. The
> > idea is that anything passes require should be embedded into the
> > IR.
> >
>
> (Oops, missed this until after I sent my own response.
>
> One thing I'd add to this from my email is that I think
> lib/Machine/IR is likely to get confused with lib/IR for the same
> reasons that lib/CodeGen is confusing between LLVM and Clang. IMO
> lib/Machine/MIR is "safer".)

+1 -- Please only use IR to refer to LLVM IR. We should use some different
abbreviation for any other IR we have in the backend.

>
> > However, this won't currently work. There are things that seem to
> > be parallel but independent of the machine IR and are used by any
> > machine passes. There are also things that clearly use the machine
> > passes. Currently, I'm not sure how to cleanly divide this library
> > up without really significant refactoring of every part of the
> > code generator.
> >
> > While I would like to see this happen, is it really a good idea to
> > put this in the critical path of getting MIR serialized and
> > deserialized?
>
> Not if it's as hard as you're saying. My impression was that Alex
> was able to move the IR stuff he needed into a separately library
> pretty trivially (based on the P.O.C. patch he posted), but if it's
> not trivial he should just move on.
>
> >> Also, if we’re getting crazy here, CodeGen in clang should be
> >> renamed to IRGen, AsmPrinter should be renamed to MCGen, and
> >> SelectionDAG should be replaced
> >
> > I'm happy to actually do the CodeGen -> IRGen rename. I actually
> > built the change but didn't submit it because of the concerns of
> > some out-of-tree users of Clang. I still have all the perl scripts
> > and such I used sitting around.
>
> I'm a really big fan of this.

+1 -- Me too.

Renaming CodeGen to IRGen will make it easier for newcomers to grok clang,
big +1 from me.

pete · May 27, 2015, 11:53pm

+1.

Could those two be subdirectories of one “Machine-Related-Stuff” directory?
E.g.,
MachineStuff/IR
MachineStuff/CodeGen

Where MachineStuff is something meaningful :).

That way, they keep a logic bound, more formal than the naming convention.

Something like?

lib/Machine/IR
lib/Machine/Passes

Unless there will be many subdirectories, it seems slightly better to flatten out the layer.

I strongly prefer breaking it out into subdirectories. There are a bunch of reasons:

1) It will grow.
2) Without it, we cannot have separate libraries, which will lose some options for shrinking the size of libraries.
3) Without separate libraries we can't as easily enforce the layering separations between these things. For example, making this split will be *extremely* hard currently because there is a lot of inappropriate dependencies that will block splitting things out.

However, "IR" and "Passes" cover only two of the things in CodeGen. There is also the implementation of a lot of common infrastructure used by targets' code generators.

My initial suggestion would be to just sink CodeGen into Machine/CodeGen, add the .../IR and .../Passes directories, and then start extracting things from CodeGen into the two more narrow directories. I think there is likely some stuff that should continue to live in a "code generator" infrastructure directory as it is neither part of the machine IR, nor is it part of any particular pass.

My suggested layering would be:

Passes depend on IR, CodeGen depends on both Passes and IR. The idea is that anything passes require should be embedded into the IR.

(Oops, missed this until after I sent my own response.

One thing I'd add to this from my email is that I think
lib/Machine/IR is likely to get confused with lib/IR for the same
reasons that lib/CodeGen is confusing between LLVM and Clang. IMO
lib/Machine/MIR is "safer".)

However, this won't currently work. There are things that seem to be parallel but independent of the machine IR and are used by any machine passes. There are also things that clearly use the machine passes. Currently, I'm not sure how to cleanly divide this library up without really significant refactoring of every part of the code generator.

While I would like to see this happen, is it really a good idea to put this in the critical path of getting MIR serialized and deserialized?

Not if it's as hard as you're saying. My impression was that Alex
was able to move the IR stuff he needed into a separately library
pretty trivially (based on the P.O.C. patch he posted), but if it's
not trivial he should just move on.

While it's true that there are things that prevent a straightforward library division, I don't think that they will pose such a big problem.
I have a brief list of things that have to be done before I can move the Machine IR stuff from CodeGen to some other library:
- Move the SplitCriticalEdge method from MachineBasicBlock (⚙ D10064 Separate Machine IR from CodeGen: Move SplitCriticalEdge out of MachineBasicBlock).
- Move the UnpackMachineBundles and FinalizeMachineBundles passes from MachineInstrBundle.cpp. (⚙ D10070 Separate Machine IR from CodeGen: Move UnpackMachineBundles pass from MachineInstrBundle.cpp + 1 upcoming patch).
- Refactor SlotIndexes.h: keep the SlotIndexes pass and move the rest to SlotIndex.h. Introduce a new container class in SlotIndex.h that will extract the map
between machine instructions and slot indexes from the SlotIndexes pass and remove the dependency on the pass from MachineBasicBlock's print method.
- WinEHFuncInfo - Either move parseEHInfo from WinEHPrepare.cpp to new file WinEHFuncInfo.cpp or move the declaration of parseEHInfo from WinEHFuncInfo.h
to a new header WinEHPrepare.h
- Get rid of several unused #includes in MachineIR cpp files that include passes.

After that it should be possible to move the Machine IR files out of CodeGen without them actually including any CodeGen headers.
And all the Machine IR passes like MachineFunctionPass, MachineFunctionAnalysis, etc. will remain in CodeGen.

Alex.

Something Chandler pointed out on IRC is that #include dependencies
aren't enough -- we also need to root out any symbol dependencies.
Have you checked for symbols (mostly, functions) that are declared
in a header you're planning to move, but defined in a source file
you're *not* planning to move?

(Apparently there have been failed attempts in the past to bring
sanity here that I wasn't aware of…)

The solution to this is to have a tool which links only the parts we need. So for example, write a tool which only verifies machine IR. We parse it and verify it, and exit. That only needs the MIR and parser and will root out any dependencies on IR/CodeGen.

+1 for everything else BTW, not that we needed another +1.

Cheers,
Pete

_sean_silva · May 28, 2015, 1:16am

> From: "Duncan P. N. Exon Smith" <dexonsmith@apple.com>
> To: "Chandler Carruth" <chandlerc@google.com>
> Cc: "LLVM Developers Mailing List" <llvmdev@cs.uiuc.edu>
> Sent: Wednesday, May 27, 2015 12:59:23 PM
> Subject: Re: [LLVMdev] RFC: Separate machine IR from lib/CodeGen into
lib/MIR
>
> >
> >>>
> >>> +1.
> >>>
> >>> Could those two be subdirectories of one “Machine-Related-Stuff”
> >>> directory?
> >>> E.g.,
> >>> MachineStuff/IR
> >>> MachineStuff/CodeGen
> >>>
> >>> Where MachineStuff is something meaningful :).
> >>>
> >>> That way, they keep a logic bound, more formal than the naming
> >>> convention.
> >>
> >> Something like?
> >>
> >> lib/Machine/IR
> >> lib/Machine/Passes
> >>
> >> Unless there will be many subdirectories, it seems slightly better
> >> to flatten out the layer.
> >
> > I strongly prefer breaking it out into subdirectories. There are a
> > bunch of reasons:
> >
> > 1) It will grow.
> > 2) Without it, we cannot have separate libraries, which will lose
> > some options for shrinking the size of libraries.
> > 3) Without separate libraries we can't as easily enforce the
> > layering separations between these things. For example, making
> > this split will be *extremely* hard currently because there is a
> > lot of inappropriate dependencies that will block splitting things
> > out.
> >
> > However, "IR" and "Passes" cover only two of the things in CodeGen.
> > There is also the implementation of a lot of common infrastructure
> > used by targets' code generators.
> >
> > My initial suggestion would be to just sink CodeGen into
> > Machine/CodeGen, add the .../IR and .../Passes directories, and
> > then start extracting things from CodeGen into the two more narrow
> > directories. I think there is likely some stuff that should
> > continue to live in a "code generator" infrastructure directory as
> > it is neither part of the machine IR, nor is it part of any
> > particular pass.
> >
> > My suggested layering would be:
> >
> > Passes depend on IR, CodeGen depends on both Passes and IR. The
> > idea is that anything passes require should be embedded into the
> > IR.
> >
>
> (Oops, missed this until after I sent my own response.
>
> One thing I'd add to this from my email is that I think
> lib/Machine/IR is likely to get confused with lib/IR for the same
> reasons that lib/CodeGen is confusing between LLVM and Clang. IMO
> lib/Machine/MIR is "safer".)

+1 -- Please only use IR to refer to LLVM IR. We should use some different
abbreviation for any other IR we have in the backend.

Also, "MIR" is also widely used in the literature and compiler textbooks to
refer to what we call "IR", so "MIR" is confusing anyway.

-- Sean Silva

hyp · May 28, 2015, 9:33pm

>
>
>>
>>
>>
>> 2015-05-27 10:59 GMT-07:00 Duncan P. N. Exon Smith <
dexonsmith@apple.com>:
>>>
>>>>>
>>>>> +1.
>>>>>
>>>>> Could those two be subdirectories of one “Machine-Related-Stuff”
directory?
>>>>> E.g.,
>>>>> MachineStuff/IR
>>>>> MachineStuff/CodeGen
>>>>>
>>>>> Where MachineStuff is something meaningful :).
>>>>>
>>>>> That way, they keep a logic bound, more formal than the naming
convention.
>>>>
>>>> Something like?
>>>>
>>>> lib/Machine/IR
>>>> lib/Machine/Passes
>>>>
>>>> Unless there will be many subdirectories, it seems slightly better to
flatten out the layer.
>>>
>>> I strongly prefer breaking it out into subdirectories. There are a
bunch of reasons:
>>>
>>> 1) It will grow.
>>> 2) Without it, we cannot have separate libraries, which will lose some
options for shrinking the size of libraries.
>>> 3) Without separate libraries we can't as easily enforce the layering
separations between these things. For example, making this split will be
*extremely* hard currently because there is a lot of inappropriate
dependencies that will block splitting things out.
>>>
>>> However, "IR" and "Passes" cover only two of the things in CodeGen.
There is also the implementation of a lot of common infrastructure used by
targets' code generators.
>>>
>>> My initial suggestion would be to just sink CodeGen into
Machine/CodeGen, add the .../IR and .../Passes directories, and then start
extracting things from CodeGen into the two more narrow directories. I
think there is likely some stuff that should continue to live in a "code
generator" infrastructure directory as it is neither part of the machine
IR, nor is it part of any particular pass.
>>>
>>> My suggested layering would be:
>>>
>>> Passes depend on IR, CodeGen depends on both Passes and IR. The idea
is that anything passes require should be embedded into the IR.
>>>
>>
>> (Oops, missed this until after I sent my own response.
>>
>> One thing I'd add to this from my email is that I think
>> lib/Machine/IR is likely to get confused with lib/IR for the same
>> reasons that lib/CodeGen is confusing between LLVM and Clang. IMO
>> lib/Machine/MIR is "safer".)
>>
>>> However, this won't currently work. There are things that seem to be
parallel but independent of the machine IR and are used by any machine
passes. There are also things that clearly use the machine passes.
Currently, I'm not sure how to cleanly divide this library up without
really significant refactoring of every part of the code generator.
>>>
>>> While I would like to see this happen, is it really a good idea to put
this in the critical path of getting MIR serialized and deserialized?
>>
>> Not if it's as hard as you're saying. My impression was that Alex
>> was able to move the IR stuff he needed into a separately library
>> pretty trivially (based on the P.O.C. patch he posted), but if it's
>> not trivial he should just move on.
>>
>> While it's true that there are things that prevent a straightforward
library division, I don't think that they will pose such a big problem.
>> I have a brief list of things that have to be done before I can move
the Machine IR stuff from CodeGen to some other library:
>> - Move the SplitCriticalEdge method from MachineBasicBlock (
⚙ D10064 Separate Machine IR from CodeGen: Move SplitCriticalEdge out of MachineBasicBlock).
>> - Move the UnpackMachineBundles and FinalizeMachineBundles passes from
MachineInstrBundle.cpp. (⚙ D10070 Separate Machine IR from CodeGen: Move UnpackMachineBundles pass from MachineInstrBundle.cpp + 1 upcoming
patch).
>> - Refactor SlotIndexes.h: keep the SlotIndexes pass and move the rest
to SlotIndex.h. Introduce a new container class in SlotIndex.h that will
extract the map
>> between machine instructions and slot indexes from the SlotIndexes
pass and remove the dependency on the pass from MachineBasicBlock's print
method.
>> - WinEHFuncInfo - Either move parseEHInfo from WinEHPrepare.cpp to new
file WinEHFuncInfo.cpp or move the declaration of parseEHInfo from
WinEHFuncInfo.h
>> to a new header WinEHPrepare.h
>> - Get rid of several unused #includes in MachineIR cpp files that
include passes.
>>
>> After that it should be possible to move the Machine IR files out of
CodeGen without them actually including any CodeGen headers.
>> And all the Machine IR passes like MachineFunctionPass,
MachineFunctionAnalysis, etc. will remain in CodeGen.
>>
>> Alex.
>>
>
> Something Chandler pointed out on IRC is that #include dependencies
> aren't enough -- we also need to root out any symbol dependencies.
> Have you checked for symbols (mostly, functions) that are declared
> in a header you're planning to move, but defined in a source file
> you're *not* planning to move?
>
> (Apparently there have been failed attempts in the past to bring
> sanity here that I wasn't aware of…)
The solution to this is to have a tool which links only the parts we
need. So for example, write a tool which only verifies machine IR. We
parse it and verify it, and exit. That only needs the MIR and parser and
will root out any dependencies on IR/CodeGen.

This might be a bit of a problem as such a tool would need the
LLVMTargetMachine which is currently implemented in CodeGen.
As well as that, every target depends on CodeGen, thus we won't be able to
create any machine IR tools without linking in CodeGen given the current
proposal.
Such a tool would require much bigger changes IMHO.

Cheers,
Alex