[llvm-dev] how experimental are the llvm.experimental.vector.reduce.* functions?

Sat Feb 9 09:37:07 PST 2019

On Sat, Feb 9, 2019 at 6:25 PM Simon Pilgrim <llvm-dev at redking.me.uk> wrote:

> The add/sub (+mul) overflow intrinsics are being updated to support
> vectors to match the related add/sub saturation intrinsics. We haven't
> updated the docs yet as legalization, vectorization and various minor bits
> of plumbing still need to be finished before it can be officially supported
> (Nikita Popov has been looking at the legalization recently).
>
> Regarding the reduction functions - I think the integer intrinsics at
> least are relatively stable and we can probably investigate dropping the
> experimental tag before the next release (assuming someone has the time to
> take on the work) - it'd be nice to have the SLP vectorizer emit reduction
> intrinsics directly for these.
>
> The floating point intrinsics are trickier as they (may) have stricter
> ordering constraints that is still causing issues and may need tweaking
> (e.g. see PR36734).
>
The vector reduction intrinsics still need quite a lot of work. Apart from
SplitVecOp, all legalizations are currently missing. This is only
noticeable on AArch64 right now, because all other targets expand vector
reductions prior to codegen.

Nikita

> On 09/02/2019 16:17, Sanjay Patel wrote:
>
> The IR update to allow vector types was here:
> https://reviews.llvm.org/D57090
> ...we didn't update the docs at that time because it was not clear what
> the backend would do with that, but that might've changed with some of the
> more recent patches.
>
> On Sat, Feb 9, 2019 at 1:42 AM Craig Topper via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
>
>> I don't think I understand your pseudocode using
>> llvm.experimental.vector.reduce.umax. All of the types you showed are
>> scalar, but that intrinsic doesn't work on scalars so I'm having a hard
>> time understanding what you're trying to do with it.
>> llvm.experimental.vector.reduce.umax takes a vector input and returns a
>> scalar result. Are you wanting to find if any of the additions overflowed
>> or a mask of which addition overflowed?
>>
>> The sadd.with.overflow intrinsics are in the process of gaining vector
>> support if not already complete. Simon Pilgrim made some commits recently.
>> I know the documentation in the LangRef hasn't been updated. It will return
>> a <X x i1> vector for overflow instead i1 when vectors are used.
>>
>> ~Craig
>>
>>
>> On Fri, Feb 8, 2019 at 11:03 PM Andrew Kelley via llvm-dev <
>> llvm-dev at lists.llvm.org> wrote:
>>
>>> I'm interested in using @llvm.experimental.vector.reduce.smax/umax to
>>> implement runtime overflow checking for vectors. Here's an example
>>> checked addition, without vectors, and then I'll follow the example with
>>> what I would do for checked addition with vectors.
>>>
>>> Frontend code (zig):
>>>
>>> export fn entry() void {
>>>     var a: i32 = 1;
>>>     var b: i32 = 2;
>>>     var x = a + b;
>>> }
>>>
>>> LLVM IR code:
>>>
>>> define void @entry() #2 !dbg !41 {
>>> Entry:
>>>   %a = alloca i32, align 4
>>>   %b = alloca i32, align 4
>>>   %x = alloca i32, align 4
>>>   store i32 1, i32* %a, align 4, !dbg !52
>>>   call void @llvm.dbg.declare(metadata i32* %a, metadata !45, metadata
>>> !DIExpression()), !dbg !52
>>>   store i32 2, i32* %b, align 4, !dbg !53
>>>   call void @llvm.dbg.declare(metadata i32* %b, metadata !48, metadata
>>> !DIExpression()), !dbg !53
>>>   %0 = load i32, i32* %a, align 4, !dbg !54
>>>   %1 = load i32, i32* %b, align 4, !dbg !55
>>>   %2 = call { i32, i1 } @llvm.sadd.with.overflow.i32(i32 %0, i32 %1),
>>> !dbg !56
>>>   %3 = extractvalue { i32, i1 } %2, 0, !dbg !56
>>>   %4 = extractvalue { i32, i1 } %2, 1, !dbg !56
>>>   br i1 %4, label %OverflowFail, label %OverflowOk, !dbg !56
>>>
>>> OverflowFail:                                     ; preds = %Entry
>>>   tail call fastcc void @panic(%"[]u8"* @2, %StackTrace* null), !dbg !56
>>>   unreachable, !dbg !56
>>>
>>> OverflowOk:                                       ; preds = %Entry
>>>   store i32 %3, i32* %x, align 4, !dbg !57
>>>   call void @llvm.dbg.declare(metadata i32* %x, metadata !50, metadata
>>> !DIExpression()), !dbg !57
>>>   ret void, !dbg !58
>>> }
>>>
>>> You can see this takes advantage of @llvm.sadd.with.overflow, which is
>>> not available with vectors. So here is a different approach (pseudocode):
>>>
>>> %a_zext = zext %a to i33 # 1 more bit
>>> %b_zext = zext %b to i33 # 1 more bit
>>> %result_zext = add %a_zext, %b_zext
>>> %max_result = @llvm.experimental.vector.reduce.umax(%result_zext)
>>> %overflow = icmp %max_result > @max_i32_value
>>> %result = trunc %result_zext to i32
>>>
>>> You can imagine how this would work for signed integers, replacing zext
>>> with sext and umax with smax.
>>>
>>> This depends on an "experimental" API. Can anyone advise on depending on
>>> this API? Is it a bad idea? Is it about to be promoted to
>>> non-experimental soon? Can anyone advise on how to best achieve my goal?
>>>
>>> Kind regards,
>>> Andrew
>>>
>>> _______________________________________________
>>> LLVM Developers mailing list
>>> llvm-dev at lists.llvm.org
>>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>>
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20190209/73af8193/attachment.html>