[PATCH] Make InstCombine aware of TargetTransformInfo when optimize extension

[Jingyue Wu]

Hi David,

The thread you cited is talking about DAGTypeLegalizer::isTypeLegal which
looks separate from DataLayout::isLegalInteger we've been talking about
here. Maybe LLVM assumes they should be consistent?

I see two approaches here.

1. Making 64-bit illegal in NVPTX's data layout. This will solve the
problem this patch originates from, because ShouldChangeType already
disallows conversion to data-layout-illegal types. I also suspect this
won't break the codegen, because DAGTypeLegalizer still considers 64-bit a
legal type and is able to emit 64-bit PTX instructions such as add.s64.
However, this will cause performance issues in LSR (which is being
discussed in Bjarke's thread
http://lists.cs.uiuc.edu/pipermail/llvmdev/2015-June/087234.html) and maybe
other passes. Currently, LSR only promotes variables with legal bitwidths
to indvars. Pointers in NVPTX64 are 64-bit so LSR, so LSR won't consider
them at all. Therefore, if we make 64-bit data-layout-illegal, we will have
to fix some callers of DataLayout::isLegalInteger to handle non-native
integer/pointer types (using TTI's instruction cost model e.g.) to get good
performance.

2. (the approach taken in this patch) Keeping 64-bit legal in NVPTX's data
layout, and let the callers of isLegalInteger (InstCombineCasts in this
case) be aware of cost difference in legal types.

I don't see a clear winner. Should 64-bit be legal or illegal in NVPTX's
data layout? NVPTX emits PTX ISA in which 64-bit is native, but 64-bit in
not native in the more low-level machine code. Any thoughts?

Jingyue


On Wed, Jul 1, 2015 at 11:44 AM, David Majnemer <david.majnemer at gmail.com>
wrote:

>
>
> On Tue, Jun 30, 2015 at 9:34 AM, Justin Holewinski <jholewinski at nvidia.com
> > wrote:
>
>>
>> > On Jun 26, 2015, at 4:20 AM, Chandler Carruth <chandlerc at gmail.com>
>> wrote:
>> >
>> > This doesn't seem like the right approach at all.
>> >
>> > I don't think we want to use cost heuristics to dictate the canonical
>> form in the IR. It makes the canonicalization *much* more complex and hard
>> to manage.
>>
>> I agree.  I would prefer not to mix the concepts of legality and
>> performance here.
>>
>> >
>> > I think that NVPTX should either change the datalayout to make i64
>> illegal (and that seems reasonable for instcombine to respect, but that
>> would for example mean it would be an *absolute* decision, not a cost
>> decision) or you'll need to do a late transform to narrow the operations as
>> much as possible.
>>
>> I don’t see how we could remove i64 as a legal type in NVPTX.  Our
>> pointer size is 64-bit, so we need a container for pointers.
>>
>
> I think other folks are doing (or would like to be doing) this sort of
> thing: http://lists.cs.uiuc.edu/pipermail/llvmdev/2015-June/087248.html
>
>
>>
>> >
>> >
>> > http://reviews.llvm.org/D10750
>> >
>> > EMAIL PREFERENCES
>> >  http://reviews.llvm.org/settings/panel/emailpreferences/
>> >
>> >
>>
>>
>>
>> -----------------------------------------------------------------------------------
>> This email message is for the sole use of the intended recipient(s) and
>> may contain
>> confidential information.  Any unauthorized review, use, disclosure or
>> distribution
>> is prohibited.  If you are not the intended recipient, please contact the
>> sender by
>> reply email and destroy all copies of the original message.
>>
>> -----------------------------------------------------------------------------------
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150701/5a9db9d3/attachment.html>