[llvm-dev] should we have IR intrinsics for integer min/max?

[Sanjay Patel via llvm-dev]

Hi -

The answer to this question may help to resolve larger questions about intrinsics and vectorization that were discussed at the dev mtg last week, but let's start with the basics:

Which, if any, of these is the canonical IR?

; ret = x < y ? 0 : x-y

define i32 @max1(i32 %x, i32 %y) {
  %sub = sub nsw i32 %x, %y
  %cmp = icmp slt i32 %x, %y ; cmp is independent of sub
  %sel = select i1 %cmp, i32 0, i32 %sub
  ret i32 %sel
}

; ret = (x-y) < 0 ? 0 : x-y

define i32 @max2(i32 %x, i32 %y) {
  %sub = sub nsw i32 %x, %y
  %cmp = icmp slt i32 %sub, 0 ; cmp depends on sub, but this looks more like a max?
  %sel = select i1 %cmp, i32 0, i32 %sub
  ret i32 %sel
}

; ret = (x-y) > 0 ? x-y : 0

define i32 @max3(i32 %x, i32 %y) {
  %sub = sub nsw i32 %x, %y
  %cmp = icmp sgt i32 %sub, 0 ; canonicalize cmp+sel - looks even more like a max?
  %sel = select i1 %cmp, i32 %sub, i32 0
  ret i32 %sel
}

define i32 @max4(i32 %x, i32 %y) {
  %sub = sub nsw i32 %x, %y

  %max = llvm.smax.i32(i32 %sub, i32 0) ; this intrinsic doesn't exist today

  ret i32 %max
}

FWIW, InstCombine doesn't canonicalize any of the first 3 options currently. Codegen suffers because of that (depending on the target machine and data types). Regardless of the IR choice, some backend fixes are needed.

Another possible consideration is the structure/accuracy of the cost models used by the vectorizers and other passes. I don't think they ever special-case the cmp+sel pair as a possibly unified (and therefore cheaper than the sum of the parts) operation.

Note that we added FP variants for min/max ops with:
https://reviews.llvm.org/rL220341

[Matt Arsenault via llvm-dev]

On 11/07/2016 11:01 AM, Sanjay Patel via llvm-dev wrote:
>
> FWIW, InstCombine doesn't canonicalize any of the first 3 options
> currently. Codegen suffers because of that (depending on the target
> machine and data types). Regardless of the IR choice, some backend
> fixes are needed.
>
> Another possible consideration is the structure/accuracy of the cost
> models used by the vectorizers and other passes. I don't think they
> ever special-case the cmp+sel pair as a possibly unified (and
> therefore cheaper than the sum of the parts) operation.
>
> Note that we added FP variants for min/max ops with:
> https://reviews.llvm.org/rL220341

FP min/max is different and more complicated due to the special NaN
handling behavior. Integer min/max is representable with only a compare
and select, so I think it would be preferable to just canonicalize to
using those two instructions

-Matt
_______________________________________________
LLVM Developers mailing list
llvm...@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

[Manuel Jacob via llvm-dev]

On 2016-11-07 20:01, Sanjay Patel via llvm-dev wrote:
> FWIW, InstCombine doesn't canonicalize any of the first 3 options
> currently. Codegen suffers because of that (depending on the target
> machine
> and data types). Regardless of the IR choice, some backend fixes are
> needed.

I'm missing context here. Can you describe in more detail how the IR
choice affects the code generation? In case the target has special
integer min / max instructions, why is matching all three variants
difficult?

-Manuel

[Sanjay Patel via llvm-dev]

Codegen is not the primary motivation here, so maybe I shouldn't have even mentioned that. However, you can find more context in:
https://reviews.llvm.org/D26091
https://reviews.llvm.org/D26096 (note how the optimizer can regress codegen)

The main concern is that we should choose a canonical form for IR that is easiest to reason about, and then we should transform all IR to that form. The backend shouldn't have to pattern match all of these variants - that's what IR is for.

[Hal Finkel via llvm-dev]

Noting that all of the above use the same number of IR instructions, I prefer this third option:

 1. It uses fewer values in the icmp/select, so the live range of the x and y, individually, is shorter. This seems like a reasonable metric for simplicity.
 2. Using a comparison of (x-y) against zero likely makes it easier for computing known bits to simply the answer (you only need to compute the sign bit).
 3. The constant of the select, 0, is the second argument (which seems to reflect our general canonical choice).

define i32 @max4(i32 %x, i32 %y) {
  %sub = sub nsw i32 %x, %y

  %max = llvm.smax.i32(i32 %sub, i32 0) ; this intrinsic doesn't exist today

  ret i32 %max
}

I don't currently see the need for a new intrinsic.

FWIW, InstCombine doesn't canonicalize any of the first 3 options currently. Codegen suffers because of that (depending on the target machine and data types). Regardless of the IR choice, some backend fixes are needed.

Another possible consideration is the structure/accuracy of the cost models used by the vectorizers and other passes. I don't think they ever special-case the cmp+sel pair as a possibly unified (and therefore cheaper than the sum of the parts) operation.

We don't have a facility currently for the target to provide a cost for combined operations. We should, but there's design work to be done.

 -Hal

Note that we added FP variants for min/max ops with:
https://reviews.llvm.org/rL220341

_______________________________________________
LLVM Developers mailing list
llvm...@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

--

Hal Finkel
Lead, Compiler Technology and Programming Languages
Leadership Computing Facility
Argonne National Laboratory

[Sanjay Patel via llvm-dev]

Thanks, Hal and Matt for the feedback. As usual, my instincts about canonicalization were probably wrong. :)

I thought that @max1 vs. @max3 would be viewed as an unknowable trade-off between reducing the dependency chain and the pseudo-canonical min/max form, so we'd add intrinsics, and defer that decision to the backend.

I'll wait to see if there are any other arguments presented.

@max2 vs. @max3 is a straightforward commute that we should have been doing anyway, so I can start there. Assuming we go with @max3, we need to add something to DAGCombine to turn that back into @max1 (PPC w/ isel and AArch64 do better with @max1; x86 is the same).

 

[Zaks, Ayal via llvm-dev]

Hal> Noting that all of the above use the same number of IR instructions, I prefer this third option:

 

as does RecurrenceDescriptor::isMinMaxSelectCmpPattern(), which looks for “Select(ICmp(X, Y), X, Y)”.

 

 

Sanjay> Another possible consideration is the structure/accuracy of the cost models used by the vectorizers and other passes. I don't think they ever special-case the cmp+sel pair as a possibly unified (and therefore cheaper than the sum of the parts) operation.

 

They call the above to identify min/max reductions; and use TTI::getCmpSelInstrCost(). Special-casing may be redundant if the costs of scalar-vs-vector min/max correspond to the costs of scalar-vs-vector cmp&sel.

 

Ayal.

---------------------------------------------------------------------
Intel Israel (74) Limited

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.

[Sanjay Patel via llvm-dev]

Thanks, Ayal. I had not seen that API until now.

So...it's a bit of a mess right now. We have:
1. RecurrenceDescriptor::isMinMaxSelectCmpPattern()

2. ValueTracking's llvm::matchSelectPattern()

3. At least 3 places in InstCombine that check for select(icmp(x,y) x, y) OR select(icmp(x, y), y, x); grep for:
"instruction is used exclusively by a select as" ...
  // part of a minimum or maximum operation. If so, refrain from doing
  // any other folding. This helps out other analyses which understand
  // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
  // and CodeGen.

I posted a patch for a first step of improving canonicalization here:
https://reviews.llvm.org/D26525

And I've enhanced ValueTracking in these commits:
https://reviews.llvm.org/rL285499
https://reviews.llvm.org/rL286318

https://reviews.llvm.org/rL286776

Should everyone be using the ValueTracking function as the point-of-truth about whether something is min/max? Even with that, the InstCombine bail-out logic seems shaky to me - we probably need to refine where exactly we want to *not* do a transform in order to prevent breaking the min/max idiom.

Does this change anyone's opinion about whether we need min/max intrinsics?