Sent from my Verizon Wireless 4G LTE DROID On Nov 19, 2016 10:26 AM, Sanjay Patel <spatel@rotateright.com> wrote: > > If I have my FMA intrinsics story straight now (thanks for the explanation, Hal!), I think it raises another question about IR canonicalization (and may affect the proposed revision to IR FMF):
No, I think that we specifically don’t want to canonicalize to fmuladd at the IR level at all. If the backend has the freedom to form FMAs as it sees fit, then we should delay the decision until whenever the backend finds most appropriate. Some backends, for example, form FMAs using the MachineCombiner pass which considers critical path, latency, throughputs, etc. in order to find the best fusion opportunities. We only use fmuladd when required to restrict the backend to certain choices due to source-language semantics.
Thanks again, Hal
> > define float @foo(float %a, float %b, float %c) { > %mul = fmul fast float %a, %b ; using ‘fast’ because there is no ‘fma’ flag > %add = fadd fast float %mul, %c > ret float %add > } > > Should this be: > > define float @goo(float %a, float %b, float %c) { > %maybe.fma= call fast float @llvm.fmuladd.f32(float %a, float %b, float %c) > ret float %maybe.fma > } > declare float @llvm.fmuladd.f32(float %a, float %b, float %c) > >
From: "Hal J. via llvm-dev Finkel" <llvm-dev@lists.llvm.org>
To: "Sanjay Patel" <spatel@rotateright.com>
Cc: "llvm-dev" <llvm-dev@lists.llvm.org>
Sent: Saturday, November 19, 2016 10:58:27 AM
Subject: Re: [llvm-dev] FMA canonicalization in IR
Sent from my Verizon Wireless 4G LTE DROID
>
> If I have my FMA intrinsics story straight now (thanks for the
> explanation, Hal!), I think it raises another question about IR
> canonicalization (and may affect the proposed revision to IR FMF):
No, I think that we specifically don't want to canonicalize to
fmuladd at the IR level at all. If the backend has the freedom to
form FMAs as it sees fit, then we should delay the decision until
whenever the backend finds most appropriate. Some backends, for
example, form FMAs using the MachineCombiner pass which considers
critical path, latency, throughputs, etc. in order to find the best
fusion opportunities. We only use fmuladd when required to restrict
the backend to certain choices due to source-language semantics.
I'll also add that, in general, we canonicalize in order to enable other transformations (and reduce the number of input forms those transformations need to match in order to be effective). Forming @llvm.fmulall at the IR level does not seem to further this goal. Did you have something in mind that this canonicalization would help?
The potential advantage I was considering would be more accurate cost modeling in the vectorizer, inliner, etc. Like min/max, this is another case where the sum of the IR parts is greater than the actual cost.
Beyond that, it seems odd to me that we’d choose the longer IR expression of something that could be represented in a minimal form. I know we make practical concessions in IR based on backend deficiencies, but in this case I think the fix would be easy - if we’re in contract=fast mode, just split all of these intrinsics at DAG creation time and let the DAG or other passes behave exactly like they do today to fuse them back together again?
From: "Sanjay Patel" <spatel@rotateright.com>
To: "Hal Finkel" <hfinkel@anl.gov>
Cc: "llvm-dev" <llvm-dev@lists.llvm.org>
Sent: Saturday, November 19, 2016 10:40:27 PM
Subject: Re: [llvm-dev] FMA canonicalization in IR
The potential advantage I was considering would be more accurate cost
modeling in the vectorizer, inliner, etc. Like min/max, this is
another case where the sum of the IR parts is greater than the
actual cost.
This is indeed a problem, but is a much larger problem than just FMAs (as you note). Our cost-modeling interfaces should be extended to handle instruction patterns -- I don't see any other way of solving this in general.
Beyond that, it seems odd to me that we'd choose the longer IR
expression of something that could be represented in a minimal form.
My fear is that, by forming the FMAs earlier than necessary, you'll just end up limiting opportunities for CSE, reassociation, etc. without any corresponding benefit.
I know we make practical concessions in IR based on backend
deficiencies, but in this case I think the fix would be easy - if
we're in contract=fast mode, just split all of these intrinsics at
DAG creation time and let the DAG or other passes behave exactly
like they do today to fuse them back together again?
This is a good point; we could do this in fp-contract=fast mode.
Hi Mehdi,
I can’t think of any (and I’m away from my dev machine, so I can’t check). If you’re concerned about inhibiting transforms by introducing intrinsics (as Hal also mentioned), I agree.
However, I see fmuladd as a special case - we already use these intrinsics in contract=on mode, so we should already be required to handle these as “first class” ops in the cost model and other passes. If we’re not, I think that would be a bug.
From: "Sanjay Patel" <spatel@rotateright.com>
To: "Mehdi Amini" <mehdi.amini@apple.com>
Cc: "Hal Finkel" <hfinkel@anl.gov>, "llvm-dev" <llvm-dev@lists.llvm.org>
Sent: Sunday, November 20, 2016 9:01:36 AM
Subject: Re: [llvm-dev] FMA canonicalization in IR
Hi Mehdi,
I can't think of any (and I'm away from my dev machine, so I can't
check). If you're concerned about inhibiting transforms by
introducing intrinsics (as Hal also mentioned), I agree.
However, I see fmuladd as a special case - we already use these
intrinsics in contract=on mode, so we should already be required to
handle these as "first class" ops in the cost model and other
passes. If we're not, I think that would be a bug.
I don't think it is a matter of handling them as "first class" or otherwise. The intrinsics specifically represent a tradeoff: Specific add/multiply pairs we're permitted to fuse by source-language rules. Should I perform a CSE, reassociation, etc. that would require splitting apart a fmuladd intrinsic? I suspect not. Doing so would lose information. On the other hand, if we're free to fuse later as we see fit, then we probably should perform the CSE and then fuse, if possible, later. So the question is really: do we want to teach IR-level optimizations to split apart the fmuladd intrinsics when they'd otherwise block transformations? I suspect the answer is no, and I think that's the assumption embedded in the use of an intrinsic in the design (although I'm certainly open to being convinced otherwise).
From: "Sanjay Patel" <spatel@rotateright.com>
To: "Hal Finkel" <hfinkel@anl.gov>
Cc: "llvm-dev" <llvm-dev@lists.llvm.org>
Sent: Saturday, November 19, 2016 10:40:27 PM
Subject: Re: [llvm-dev] FMA canonicalization in IR
The potential advantage I was considering would be more accurate cost
modeling in the vectorizer, inliner, etc. Like min/max, this is
another case where the sum of the IR parts is greater than the
actual cost.
This is indeed a problem, but is a much larger problem than just FMAs (as you note). Our cost-modeling interfaces should be extended to handle instruction patterns -- I don't see any other way of solving this in general.
This proposal - cost model instruction patterns, not just instructions - keeps coming up in a number of contexts. We've seen a number of proposals recently to add intrinsics at various places in the pipeline to get around this limitation. Investing in infrastructure to solve this problem via the cost model seems like a generally useful path forward which would benefit far more than FMA.