[llvm-dev] clarification needed for the constrained fp implementation.

[Kaylor, Andrew via llvm-dev]

Hi Brian,

I was just cleaning up my inbox and saw that I had never responded to your e-mail below. I apologize for that. It was not intentional. Did the subsequent discussion answer your questions?

In any event, let me address your specific question about the language reference, where it says "These intrinsics are used to provide special handling of floating point operations when specific rounding mode or floating point exception behavior is required" (your emphasis repeated for discussion). The idea here was that these intrinsics can be used to control the behavior of the optimizer, in order to allow non-default rounding mode and floating point exception behavior.  The "required" behavior must be expressed in the IR, as always, but the intrinsics only perform the operation they are documented to perform. I suppose I should clean up the language reference to make this less ambiguous.

It's unclear to me whether your processor supports directly encoding the rounding mode into floating point instructions. If it does not, then I don't see how having the semantics of the intrinsics force a given rounding mode would help, as the generated code would still require instructions that explicitly set the rounding mode before (and possibly after) each floating point instruction. If your processor does allow the rounding mode to be encoded in an instruction, then I can understand the desire to have a way to encode that into the IR, but for the IR to remain target-independent we would still need to support targets that do not have this capability. My earlier suggestion of intrinsics to establish a "local" rounding mode was intended to enable this sort of model.

Regarding your question about having to save and restore the rounding mode when calling a function, I think that depends on the language-specific semantics of the function you are calling. The C99 standard, for instance, requires that functions not change the rounding mode unless they are specifically documented as doing so and that any function called is assumed to require the default rounding mode unless it is specifically documented otherwise. (The FENV_ACCESS pragma can provide a sort of "documentation" for locally defined functions in this regard.) So, basically, it's up to the front end (or whatever other mechanism generates your IR) to ensure that the necessary conventions are correctly followed.

Does that help?

Thanks,
Andy


From: Sumner, Brian [mailto:Brian.Sumner at amd.com]
Sent: Friday, November 03, 2017 1:54 PM
To: Kaylor, Andrew <andrew.kaylor at intel.com>; Ding, Wei <Wei.Ding2 at amd.com>; Arsenault, Matthew <Matthew.Arsenault at amd.com>
Cc: LLVM Developers Mailing List (llvmdev at cs.uiuc.edu) <llvmdev at cs.uiuc.edu>
Subject: RE: clarification needed for the constrained fp implementation.

Hi Andy,

I'm having some trouble correlating your comments with the LLVM language reference manual which says: "These intrinsics are used to provide special handling of floating point operations when specific rounding mode or floating point exception behavior is required."  (Emphasis added by me.) We believed that this meant that we could use these intrinsics when we required our target to generate a floating point operation with a specific rounding mode.   We have been attempting to implement these semantics, which are somewhat different than "a way of telling the optimizer what it can and cannot assume about rounding mode and FP exception behavior".

We had hoped to avoid defining target specific intrinsics to request FP operations with specific rounding, and we prefer to attach the desired rounding to each specific operation.  The approach using something like llvm.get.roundingmode() and llvm.set.roundingmode() is not nearly as appealing to us.  For example, suppose we have library functions that require the default (to-nearest-even) rounding to achieve correct behavior.  Must we now start saving and restoring the rounding mode in every such function to protect it from a higher level caller possibly changing the rounding mode?

In short, I would prefer to continue to follow the semantics stated in the manual, and would be happy to see the ".experimental" dropped as well.

Thanks,
Brian



From: Kaylor, Andrew [mailto:andrew.kaylor at intel.com]
Sent: Friday, November 03, 2017 1:11 PM
To: Ding, Wei; Sumner, Brian; Arsenault, Matthew
Cc: LLVM Developers Mailing List (llvmdev at cs.uiuc.edu<mailto:llvmdev at cs.uiuc.edu>)
Subject: RE: clarification needed for the constrained fp implementation.

Hi Wei,

I've been meaning to write something up for discussion on the LLVM Dev list about this.  I hope you don't mind if I copy the list now to accomplish that while also answering your questions.  Eventually I create a document describing this in more detail and less formally than the language definition.

Basically, the "constraints" in the constrained FP intrinisics are constraints on the optimizer.  They are a way of telling the optimizer what it can and cannot assume about rounding mode and FP exception behavior.  By default, the optimizer assumes that the rounding mode is round-to-nearest and that FP exceptions are being ignored.  If the user code is going to do anything that invalidates these assumptions, then we need a way to make the optimizer stop assuming that.  That's what the intrinisics do.  Because most passes don't recognize the intrinisics, they can't do anything with the operations they represent and therefore can't make any assumption about them.

The intrinsics are not intended to do anything to change the rounding mode or FP exception handling state.  I have an idea in mind for some additional intrinsics that would provide a way to control the FP environment.  There are already some target-specific mechanisms for doing that, but I'd like to have something that's target independent.  I'll say more about this in a minute.

I mentioned in my review comments that my work on this has been motivated by the STDC pragmas, and I think if I explain that it might make the semantics of the intrinsics seem a little more natural.  The primary pragma I have in mind here is the "STDC FENV_ACCESS" pragma.  I believe this is part of the C99 standard, but compiler support for it is still mostly (if not entirely) missing.  For instance, if you try to use this pragma with clang you will get a message telling you that the pragma isn't supported and it will have no other effect.  We want to change that.

Basically, what the "STDC FENV_ACCESS" pragma does is provide programmers with a way to tell the compiler that the program might change the FP environment.  This pragma represents a setting that has only two states -- on and off.  The default setting of this state is documented as being implementation defined.  In clang the default state will be off.  The C99 standard states that accessing the FP environment (testing FP status flags, changing FP control modes, etc.) when FENV_ACCESS is off is undefined behavior.  The C99 standard provides library calls to access the environment (fesetround, fegetround, fetestexcept, etc.) but you can only safely use these if you have set FENV_ACCESS to the "on" state.  A typical usage might look like this:

#include <fenv.h>

double someFunc(double A, double B, bool ForceRoundUp) {
  #pragma STDC FENV_ACCESS ON
  double Result;
  if (ForceRoundUp) {
    int OldRM = fegetround();
    fesetround(FE_UPWARD);
    Result = A/B;
    fesetround(OldRM);
  } else {
    Result = A/B;
  }
  return Result;
}

So you see here that there are explicit calls to change the rounding mode.  If you were to do this in clang today, the generated IR would look like this:

define double @someFunc(double, double, i1) {
  br i1 %2, label %4, label %8

; <label>:4:                                      ; preds = %3
  %5 = tail call i32 @fegetround()
  %6 = tail call i32 @fesetround(i32 2048)
  %7 = tail call i32 @fesetround(i32 %5)
  br label %8

; <label>:8:                                      ; preds = %3, %4
  %9 = fdiv double %0, %1
  ret double %9
}

Notice that the fdiv got sunk outside of the calls to change the rounding mode.  Once we support the FENV_ACCESS pragma, the generated IR will look like this instead:

define double @someFunc(double, double, i1) {
  br i1 %2, label %4, label %8

; <label>:4:                                      ; preds = %3
  %5 = tail call i32 @fegetround()
  %6 = tail call i32 @fesetround(i32 2048)
  %7 = call double llvm.experimental.constrained.fdiv.f64(double %0, double%1, metadata "round.dynamic", metadata "fpexcept.strict")
  %8 = tail call i32 @fesetround(i32 %5)
  br label %11

; <label>:9:                                      ; preds = %3
  %10 = call double llvm.experimental.constrained.fdiv.f64(double %0, double%1, metadata "round.dynamic", metadata "fpexcept.strict")
  Br label %11

; <label>:11:                                      ; preds = %4, %9
  %12 = phi double [ %7, %4 ], [ %10, %9 ]
  ret double %12
}

Note that I've left the rounding mode as "round.dynamic" here.  In theory we could implement an optimization that recognizes the calls to fesetround and changes that argument to "round.upward", but initially it will be "round.dynamic" which indicates that the program is allowed to change the rounding mode at runtime and that's what I expect will always come out of the front end.

The key point here is that the pragma just gives the programmer permission to change things.  There are some other pragmas in draft standards that allow the programmer to specify what the rounding mode will be for a given scope, but it is my understanding that they do not actually change the rounding mode either.  They just tell the compiler what the programmer is promising the rounding mode will be at that point.

All of this is substantially background information with regard to LLVM since the LLVM optimizer is independent of any front end.  The key point I'm trying to convey is that the constrained intrinsics are meant to behave in a way that is analogous to these pragmas.

What I think we need now are a set of language-neutral and target-independent intrinsics that allow us to control the FP rounding mode.  For instance, suppose you are implementing some function and you want to be able to control the rounding mode for the entire scope of the function.  You typically will need to do that similarly to what I showed in my example above -- get the current rounding mode, set the rounding mode, perform some operations, and restore the rounding mode.  I don't think you would want the constrained intrinsics that we have to be responsible for controlling the rounding mode because that would mean that the backend would need to generate instructions to get-set-restore the rounding mode around every operation in your function (unless the target supports explicit rounding mode operands).

What I'm thinking is that we need something like this:

void llvm.set.roundingmode(i32 mode)
i32 lllvm.get.roundingmode()

These would then get translated during instruction selection to target-specific instructions, which is equivalent to what fesetround() and fegetround() do. But I think it would also be useful to have something like this:

void llvm.begin.local.roundingmode(i32 mode)
void llvm.end.local.roundingmode()

This could encapsulate to get-set-restore idiom.  My thinking then is that if a target does support explicit rounding mode operands, these intrinsics wouldn't need to result in any instructions that change the processors rounding mode and could instead just be used to determine what the rounding mode operand should be where applicable.

There are still some issues that need to be worked out here (such as the fact that a i32 here is far too general), but that's basically what I'm thinking.

Does that make sense?

-Andy



From: Ding, Wei [mailto:Wei.Ding2 at amd.com]
Sent: Friday, November 03, 2017 12:04 PM
To: Kaylor, Andrew <andrew.kaylor at intel.com<mailto:andrew.kaylor at intel.com>>; Sumner, Brian <Brian.Sumner at amd.com<mailto:Brian.Sumner at amd.com>>; Arsenault, Matthew <Matthew.Arsenault at amd.com<mailto:Matthew.Arsenault at amd.com>>
Subject: clarification needed for the constrained fp implementation.

Hi Andy,

Thanks a lot for your comments https://reviews.llvm.org/D38634. Actually, I don't think I got 100% clear about the way you are trying to implement for the constrained fps. If possible, could you please elaborate on? Especially I don't quite follow your comments like "I'm approaching this from the perspective of the STDC pragmas related to the FP environment. ".

Thank you so much!

Best regards,

Wei
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20171127/f406b125/attachment.html>