[prev in list] [next in list] [prev in thread] [next in thread] 

List:       gcc
Subject:    Re: Transformation of relational expression
From:       Richard Biener <richard.guenther () gmail ! com>
Date:       2016-02-23 11:53:51
Message-ID: CAFiYyc2oarqHJmH-Y0EyzPK9_pzsicE2famd4a7UBSrugQQ-qQ () mail ! gmail ! com
[Download RAW message or body]

On Tue, Feb 23, 2016 at 11:39 AM, Dmytro Sheyko <dmytro.sheyko@gmail.com> wrote:
> Hello,
>
> I am compiling code like this:
>
>     int test(int x, int eq, int lt, int gt) {
>         return x < 2000 ? lt : x > 2000 ? gt : eq;
>     }
>
> and expect that compiler would generate one CMP instruction for both
> comparisons:
>
>         cmpl    $2000, %edi
>         jl      .L37
>         jne     .L38
>     ...
>
> however result is quite different:
>
>         cmpl    $1999, %edi
>         jle     .L37
>         cmpl    $2000, %edi
>         jne     .L38
>
> For some reason, compiler transforms (x < 2000) to (x <= 1999). Though
> such transformation looks semantically legal, not sure it is really
> necessary here. Especially taking into account that it impedes
> subsequent optimizations.
>
> It seems that 'maybe_canonicalize_comparison' from gcc/fold-const.c is
> responsible for such transformation (I am using 5.3.0, as for trunk it
> seems that some part of such logic is moved to gcc/match.pd).
>
> I wonder why such transformation is necessary. What benefits does it offer?

Canonicalization with x <= 1999 and thus simplifying of x < 2000 && x <= 1999
for example.

I don't see how we can get a single compare at GIMPLE so it must be RTL
maybe in compare-elim.c doing the recognition.

Richard.

> Thank you,
> Dmytro
[prev in list] [next in list] [prev in thread] [next in thread]