IPRA, interprocedural register allocation, question

Hello Peter,

Vivek,

I am looking into these function attributes in the clang docs

Preserve_most

Preserve_all

They are not available in the 3.6.2 that I am currently using, but I hope they exist in 3.8

These should provide enough info to solve my problem,

at the MC level calls to functions with these attributes

with be code-gen’ed through different “calling conventions”,

and CALL instructions to them should have different register USE and DEF info,

This CALL instruction register USE and DEF info should already be useful

to the intra-procedural register allocator (allowing values live across these

calls to be in what are otherwise caller-save registers),

at least that’s how I read the MC dumps, every call instruction seems to have

every caller-save register flagged as “imp-def”, IE implicitly-defined by the instruction,

and hopefully what is considered a caller-save register at a call-site is defined by the callee.

And this should be the information that IPRA takes advantage of in its bottom-up analysis.

Which leads me to this question, when compiling an entire whole program at one time,

so there is no linking and no LTO, will there ever be IPRA that works within LLC for this scenario,

and is this an objective of your project, or are you focusing only on LTO ?

I know this is not the typical “linux” scenario (dynamic linking of not only standard libraries,

but also sometimes even application libraries, and lots of static linking because of program

size), but it is a typical “embedded” scenario, which is where I am currently.

Other thoughts or comments ?

–Peter Lawrence.

Vivek,

             I am looking into these function attributes in the clang docs

                Preserve_most

                Preserve_all

They are not available in the 3.6.2 that I am currently using, but I hope
they exist in 3.8

These should provide enough info to solve my problem,

at the MC level calls to functions with these attributes

with be code-gen’ed through different “calling conventions”,

and CALL instructions to them should have different register USE and DEF
info,

Yes I believe that preserve_most or preserve_all should help you even with
out IPRA. But just to note IPRA can even help further for example on X86
preserve_most cc will not preserve R11 (this can be verified from
X86CallingConv.td and X86RegisterInfo.cpp) how ever IPAR calculates regmask
based on the actual register usage and if procedure with preserve_most cc
does not use R11 and none callsite inside of function body then IPRA will
mark R11 as preserved. Also IPRA produces RegMask which is super set of
RegMask due to calling convention.

This CALL instruction register USE and DEF info should already be useful

to the intra-procedural register allocator (allowing values live across
these

calls to be in what are otherwise caller-save registers),

at least that’s how I read the MC dumps, every call instruction seems to
have

every caller-save register flagged as “imp-def”, IE implicitly-defined by
the instruction,

and hopefully what is considered a caller-save register at a call-site is
defined by the callee.

And this should be the information that IPRA takes advantage of in its
bottom-up analysis.

Yes that is expected help from IPRA.

Which leads me to this question, when compiling an entire whole program at
one time,

so there is no linking and no LTO, will there ever be IPRA that works
within LLC for this scenario,

and is this an objective of your project, or are you focusing only on LTO ?

The current IPRA infrastructure works at compile time so it's scope of

optimization is restricted to a compilation unit. So IPRA can only
construct correct register usage information if the procedure's code is
generated by same compiler instance that means we can't optimize library
calls or procedure defined in other module. This is because we can't keep
register usage information data across two different compiler instance.

Now if we consider LTO, it eliminates above limitation by making a large IR
module from smaller modules before generating code and thus we can have
register usage information (at lest) for procedure which was previously
defined in other module, because now with LTO every thing is in one module.
So that also clarifies that IPRA does not do anything at link time.

Now coming to LLC, it can use IPRA and optimize for functions defined in
current module. So yes while compiling whole program ( a single huge .bc
file) IPRA can be used with LLC. Also just note that if a software is
written in separate files per module (which is very common) and still you
want to maximize benefits of IPRA, then we can use llvm-link tool to
combine several .bc files to produce a huge .bc file and use that with LLC
to get maximum benefits.

I know this is not the typical “linux” scenario (dynamic linking of not

only standard libraries,

but also sometimes even application libraries, and lots of static linking
because of program

size), but it is a typical “embedded” scenario, which is where I am
currently.

I don't understand this use case but we can have further improvement in
IPRA for example if you have several libraries which has already compiled
and codegen, but you are able to provide information of register usage for
the functions of that libraries than we can think about an approach were we
can store register usage information into a file (which will obviously
increase compile time) and use that information across different compiler
instances so that we can provide register usage information with out having
actual code while compiling.

Other thoughts or comments ?

I am looking for some ideas that can improve current IPRA. So if you feel
anything relevant please let me know we can discuss and implement feasible
ideas.

Thanks,
Vivek

Vivek,

             I am looking into these function attributes in the clang docs

                Preserve_most

                Preserve_all

They are not available in the 3.6.2 that I am currently using, but I hope
they exist in 3.8

These should provide enough info to solve my problem,

at the MC level calls to functions with these attributes

with be code-gen’ed through different “calling conventions”,

and CALL instructions to them should have different register USE and DEF
info,

Yes I believe that preserve_most or preserve_all should help you even with
out IPRA. But just to note IPRA can even help further for example on X86
preserve_most cc will not preserve R11 (this can be verified from
X86CallingConv.td and X86RegisterInfo.cpp) how ever IPAR calculates regmask
based on the actual register usage and if procedure with preserve_most cc
does not use R11 and none callsite inside of function body then IPRA will
mark R11 as preserved. Also IPRA produces RegMask which is super set of
RegMask due to calling convention.

I believe that __attribute__ ((registermask = ....)) can provide
more flexibility compare to preserve_all or preserve_most CC in some case.
So believe that we should try it out.

-Vivek

Vivek,
             I am looking into these function attributes in the clang docs
                Preserve_most
                Preserve_all
They are not available in the 3.6.2 that I am currently using, but I hope they exist in 3.8

These should provide enough info to solve my problem,
at the MC level calls to functions with these attributes
with be code-gen’ed through different “calling conventions”,
and CALL instructions to them should have different register USE and DEF info,

This CALL instruction register USE and DEF info should already be useful
to the intra-procedural register allocator (allowing values live across these
calls to be in what are otherwise caller-save registers),
at least that’s how I read the MC dumps, every call instruction seems to have
every caller-save register flagged as “imp-def”, IE implicitly-defined by the instruction,
and hopefully what is considered a caller-save register at a call-site is defined by the callee.
And this should be the information that IPRA takes advantage of in its bottom-up analysis.

The idea of IPRA is to *produce* more accurate list of clobbered register by a functions, so that at call site the caller needs to only save/restore the minimum amount of registers across the call.

Which leads me to this question, when compiling an entire whole program at one time,
so there is no linking and no LTO, will there ever be IPRA that works within LLC for this scenario,
and is this an objective of your project, or are you focusing only on LTO ?

LTO is just a way of exposing more to the analysis: IPRA can only “optimize” calls to function that are codegen’d during the same compilation. With LTO since you codegen the full program at once you can basically optimize “every” call.

So IPRA works well without LTO, but will be able to operate only on calls to function that are part of the current compilation.

I know this is not the typical “linux” scenario (dynamic linking of not only standard libraries,
but also sometimes even application libraries, and lots of static linking because of program
size), but it is a typical “embedded” scenario, which is where I am currently.

Other thoughts or comments ?

Any reason *not* to use LTO in your case?

Vivek,

IIUC it seems that we need two pieces of information to do IPRA,

  1. what registers the callee clobbers

  2. what the callee does to the call-graph

And it is #2 that we are missing when we define an external function,

Even when we declare it with a preserves or a regmask attribute,

So what I / we need is another attribute that says this is a leaf function,

At least in my case all I’m really concerned with are leaf functions

Thoughts ?

–Peter Lawrence.

Vivek,

           IIUC it seems that we need two pieces of information to do IPRA,

1. what registers the callee clobbers

2. what the callee does to the call-graph

Yes I think this is enough, but in your case we don't require #2

And it is #2 that we are missing when we define an external function,

Even when we declare it with a preserves or a regmask attribute,

Because I think once we have effect of attribute at IR/MI level then we
can just parse it and populate register usage information vector for
declared function and then we can propagate reg mask on each call site
encountered.
But I am not user will it be easy to get new attribute working or we may
need to hack clang for that too.

I would also like to have thoughts from my mentors (Mehdi Amini and Hal
Finkel) about this.

So what I / we need is another attribute that says this is a leaf function,

At least in my case all I’m really concerned with are leaf functions

I am stating with a simple function declaration which have a custom
attribute.

-Vivek

Dear Peter, Hal and Mehdi,

I did some hack around clang so that I can attach a string attribute to function declaration.
So I think instead of adding new regmask attribute it would be better to use existing annotate attribute for example we can use it as follows:

extern void foo() attribute((annotate(“REGMASK:R11,R8”))); // here R11, R8 are clobbered regs

this will add string REGMASK:R11,R8 into llvm.metadata section and it will be tied to function foo via llvm.global.annotations. ( This currently works with function definitions only, work is needed to make this work with function declaration ) . The llvm.metadata should be accessed at IR level and then it can be parsed to create a regmask out of it.

The parsing will need to access Module object, and I hope when parsing for all such function reconnecting global annotation for function and string value would be simple.

An other approach would be adding a new attribute regmask and while codegen to IR this attribute should get lowered to corresponding string attribute in LLVM IR ( which should also be added) and then a pass would iterate through all such function which has such an attribute and populate register usage container.

But any idea to simplify is welcomed. Please share your views.

I have cced clang mailing list so that clang developers can correct me if I have make any mistake in context of clang.

Sincerely,
Vivek

Vivek,

Here’s the way I see it, let me know if you agree or disagree,

You cannot optimize a function’s calling convention (register-usage) unless

You can see and change every caller, and you only know this for non-static functions

if you know that all calls to external functions cannot call back into the current

compilation unit.

#1 gives you the info necessary to change the call-site to the external function

So you don’t need #2 to do RA around the call-site to the external function, instead

You need #2 before you can change any non-static function’s calling convention

within the current compilation unit, assuming you have this information for all

external functions.

To be more concrete, let foo() be a non-static function in the current compilation

Unit, any calls to foo() from external functions will have to use the “default”

Calling convention, so foo’s calling convention cannot be changed. We have to

Know that none of the external functions can call-back to the compilation unit

(they are “leaf” functions relative to the compilation unit) before we can change

Foo()’s calling convention.

Also, the issue of escaping-pointer-to-function is made clear by the example

Of the atexit() and exit() library functions, IE even static functions can end up

Being called by external functions. So exit() can never be declared “leaf”, and

To get the benefit of IPRA it needs to be within the compilation unit, either

By whole-program compilation or by LTO, if it is used.

–Peter.

Mehdi,

The external functions I need to call are all hand-written assembly language,

How would/could LTO handle that ?

–Peter Lawrence.

I thought about inline asm function, not pure .s files.

Vivek,
          Here’s the way I see it, let me know if you agree or disagree,

You cannot optimize a function’s calling convention (register-usage) unless
You can see and change every caller,

That’s true only if you want to “downgrade” the guarantees, i.e. if you want to reduce the callee-saved registers.
You can freely provide more information to limit the amount of caller-saved registers to a partial list of call-sites, which is in practice changing the “local" calling convention while keeping it compatible with the public one.

and you only know this for non-static functions
if you know that all calls to external functions cannot call back into the current
compilation unit.

I’m not sure why you consider calls to external functions and call back? If you don’t see main() (the common case) you don’t need a call to an external function to have a possible call to an externally visible function in the current module.

#1 gives you the info necessary to change the call-site to the external function

So you don’t need #2 to do RA around the call-site to the external function, instead
You need #2 before you can change any non-static function’s calling convention
within the current compilation unit, assuming you have this information for all
external functions.

If I understand the case you have in mind, it is only when you see the main() function in the current module and you’re trying to prove that an externally visible function could not be called from outside the module basically?

It seems to me that this is a bit orthogonal to IPRA: multiple optimizations (IPRA included) work best when functions are deduced local, non-recursive, are not tail called (for IPRA in particular), and don’t have their address taken.
The “infer-func-attr” and “globalopt” passes try to do their best to make this happen, especially during LTO.

The attribute case that Vivek is adding seems more murky though.

Mehdi,

I’m compiling embedded applications which are small enough to do

whole-program-compilation. There’s no advantage in breaking them up into

separate compilation pieces and linking them, even though in source form

they are composed of a couple of separate source files.

So for me the compilation unit is always the entire program (and includes main())

Except for some hand-coded-assembly-language support functions that are “external”

to the compilation unit and in my case never call back into the compilation unit,

IE they are always “leaf” functions from the point of view of the compilation unit’s call-graph.

Hence I would like a clang function attribute that says this function is “leaf”

So that IPRA can know that none of the functions it is compiling is ever called

From outside this compilation unit.

And I apologize to everyone for confusingly using the term “compilation unit”

When I meant “whole program”.

Yes I am aware of the fact that if you change a function’s calling convention

By converting some scratch regs into save regs (for example because they aren’t even touched)

Then you are safe to call it from either the default calling convention or the

Optimized calling convention. This is the safe thing to do, and is why I will

Only use “preserves_most” and “preserves_all” optimized calling conventions,

As those will have been implemented by a back-end writer who is aware of

All these compilations (as opposed to the “registermask=” calling convention

Which is much less safe)

I do however feel that IPRA in the whole-program case should not be restricted to

Only scratch-becoming-save changes, I don’t have any data to support the notion,

But it begs to be investigated, unless someone can somehow prove that it can’t help

Performance.

–Peter.

Mehdi,

I’m compiling embedded applications which are small enough to do

whole-program-compilation. There’s no advantage in breaking them up into

separate compilation pieces and linking them, even though in source form

they are composed of a couple of separate source files.

Sent from my iPhone

Mehdi,

            I’m compiling embedded applications which are small enough to
do

whole-program-compilation. There’s no advantage in breaking them up into

separate compilation pieces and linking them, even though in source form

they are composed of a couple of separate source files.

Ok, so LTO case basically.

So for me the compilation unit is always the entire program (and includes
main())

Except for some hand-coded-assembly-language support functions that are
“external”

to the compilation unit and in my case never call back into the
compilation unit,

IE they are always “leaf” functions from the point of view of the
compilation unit’s call-graph.

Hence I would like a clang function attribute that says this function is
“leaf”

So that IPRA can know that none of the functions it is compiling is ever
called

From outside this compilation unit.

@Peter How do you want this information to be used by IPRA, i.e if a

function (probably function declaration) is marked as "leaf" then how it
should impact IPRA from your point of view?

I believe the usual (and best way from the compiler point of view) way to
address your particular scenario is to have a proper export list and use
LTO.
For instance if you never call into the program from one of your
hand-coded assembly routines, LTO should be able to turn every global
functions/variables into local ones.

I am agree with Mehdi here. Also as he has summarized with adding

attribute regmask we don't want to change calling convention for callee
(i.e reducing callee saved registers) but just it can help intra-procedural
register allocators at call sites inside the caller of particular hand
written assembly function. If you have very frequent uses of such assembly
functions and preserve_all or preserve_most is not able help you describing
exact register usage in that case using regmask attribute will help a lot.
I believe that propagating actual register usage is totally safe provided
user has provided correct information ( compiler can only help to detect
semantical error i.e for example R45 is specified but such register is not
available on given platform )

-Vivek

Mehdi,

I am looking for an understanding of 1) IPRA in general, 2) IPRA in LLVM.

Whether I want to use LTO or not is a separate issue.

  1. I currently believe it is a true statement that:

If all external functions are known to not call back into the “whole-program”

Being compiled, then IPRA is free to do anything at all to the functions being

Compiled, not limited to only “upgrades” calling convention changes, but

Also allowing “downgrades” calling convention changes as well.

Do you think my current belief #1 is correct ?

  1. it seems that LLVM currently limits itself to “upgrades” calling convention changes,

The reason being so that not all call sites are required to be changed,

therefore calls through function pointers can use the default calling convention

If for example there is insufficient analysis to know for sure what functions can be

called from that site.

Is my understanding #2 of IPRA in LLVM correct ?

–Peter.

“whole-program” here is a misnomer since there are external functions, but I don’t

Have a better term for this.

“upgrades” means some scratch regs are converted to save

(the callee either doesn’t touch them at all, or does do save/restore)

“downgrades” means some save regs are converted to scratch

(the callee no longer does save/restore to some registers, and does clobber them)

Mehdi,
             I am looking for an understanding of 1) IPRA in general, 2) IPRA in LLVM.
Whether I want to use LTO or not is a separate issue.

1) I currently believe it is a true statement that:
                If all external functions are known to not call back into the “whole-program”
                Being compiled, then IPRA is free to do anything at all to the functions being
                Compiled, not limited to only “upgrades” calling convention changes, but
                Also allowing “downgrades” calling convention changes as well.

Do you think my current belief #1 is correct ?

Yes, with some extra assumptions (you don’t use dlsym for instance, and you won’t link to another file with a global initializer that can call any of these).

I expressed this earlier (which include the other issues I mentioned just before) as “we can turn the linkage of every function into local” (or private, or static, whatever denomination you prefer).

2) it seems that LLVM currently limits itself to “upgrades” calling convention changes,
The reason being so that not all call sites are required to be changed,
therefore calls through function pointers can use the default calling convention
If for example there is insufficient analysis to know for sure what functions can be
called from that site.

Is my understanding #2 of IPRA in LLVM correct ?

I don’t believe this is correct, currently IPRA will limit itself to this for function that can be called from another module.
I will freely change the calling convention, including downgrades, when it knows that it can see all call sites (+ extra conditions, like no recursion being involved I think).

“whole-program” here is a misnomer since there are external functions, but I don’t
Have a better term for this.

I believe you can talk about “main module”, i.e. the module defines the entry point for the program.
Note LLVM can’t make assumption about the lack of dlsym() or global initializer in other module for example, so the linkage type of functions is what tells us about the possibility to call back or not.

Mehdi,

In my mind at least, “whole program” means no dynamic libraries, so the only

external functions are simple runtime support, do you have a suggested term for that ?

–Peter.

Hello Peter,

Are you still interested in attribute(regmask) ?
I have done some hack ( both clang+IPRA) to get it working if you want to play around it I can send a patch by tomorrow.

Sincerely,
Vivek

Vivek,

I apologize if you took my original email as a request for implementation,

I meant to be asking what is already available, I think the answer to that

is the ‘preserves_most’ and ‘preserves_all’ attributes, but I will also

Use ‘regmask’ if those prove to be too sub-optimal.

I am still interested in figuring out the necessary and sufficient conditions

For LLC to do optimal IPRA when given a “whole program”

(as per my previous definition of “whole program”),

As opposed to how to accomplish this with LTO,

If you are open to having such discussions, even though your focus

IIUC is supposed to be LTO, then great. I think Mehdi is stuck trying

To convince me to use LTO, but given all the changes I’ve had to make

To CodeGen (IE outside of my Target sub-dir) for having separate Data and Address

register sets, I think using LTO is a long term solution that I can’t take

On just now (IE the svn branch merge problem)

As one of my old math professors used to say “don’t use a sledge hammer

To crush a pea”, to wit I am only compiling a single source file as an entire whole

Program and I don’t do any linking, why should I have to use a linker.

–Peter Lawrence