> -----Original Message-----
> From: Jakub Jelinek <[email protected]>
> Sent: Tuesday, September 3, 2024 2:56 AM
> To: Andrew Pinski <[email protected]>
> Cc: Jiang, Haochen <[email protected]>; Richard Biener
> <[email protected]>; [email protected]; Liu, Hongtao
> <[email protected]>; [email protected]; Hu, Lin1 <[email protected]>
> Subject: Re: [PATCH 2/8] i386: Optimize ordered and nonequal
>
> On Mon, Sep 02, 2024 at 11:25:36AM -0700, Andrew Pinski wrote:
> > On Mon, Sep 2, 2024 at 11:20 AM Jakub Jelinek <[email protected]> wrote:
> > >
> > > On Mon, Aug 26, 2024 at 02:42:31PM +0800, Haochen Jiang wrote:
> > > > * match.pd: Optimize (and ordered non-equal) to
> > > > (not (or unordered equal))
> > > >
> > > > gcc/testsuite/ChangeLog:
> > > >
> > > > * gcc.target/i386/optimize_one.c: New test.
> > >
> > > The testcase FAILs on i686-linux, because it uses -mfpmath=sse
> > > without enabling -msse2.
> > >
> > > I've committed the following fix as obvious to fix that.
> > >
> > > > --- a/gcc/match.pd
> > > > +++ b/gcc/match.pd
> > > > @@ -6636,6 +6636,9 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> > > > (ltgt @0 @0)
> > > > (if (!flag_trapping_math || !tree_expr_maybe_nan_p (@0))
> > > > { constant_boolean_node (false, type); }))
> > > > +(simplify
> > > > + (bit_and (ordered @0 @1) (ne @0 @1)) (bit_not (uneq @0 @1)))
> > >
> > > I wonder whether there shouldn't be some :c (e.g. on bit_and and
> > > maybe ne too), because ordered is commutative and so is ne and so is
> > > bit_and, and perhaps you want to match also (bit_and (ne @0 @1)
> > > (ordered @1 @0)) etc. What about negation of this (bit_ior (unordered @0
> @1) (eq @0 @1))?
> >
> > The :c is needed for bit_and for sure. BUT should not needed for
> > ordered/ne though because the canonicalization of the operations
> > should have the operands in the same order as `a ordered b` is the
> > same as `b ordered a`.
>
> Maybe. Also, isn't (bit_not (uneq @0 @1)) (ltgt @0 @1) ?
>
> Jakub
I add some tests like:
6 int is_ordered_and_nonequal_sh_1 (float a, float b)
7 {
8 return !__builtin_isunordered (a, b) && (a != b);
9 }
10
11 int is_ordered_and_nonequal_sh_2 (float a, float b)
12 {
13 return !__builtin_isunordered (a, b) && (b != a);
14 }
15
16 int is_ordered_and_nonequal_sh_3 (float a, float b)
17 {
18 return (b != a) && !__builtin_isunordered (a, b);
19 }
20
21 int is_ordered_and_nonequal_sh_4 (float a, float b)
22 {
23 return !__builtin_isunordered (a, b) && !(a == b);
24 }
25
26 int is_ordered_and_nonequal_sh_5 (float a, float b)
27 {
28 return !__builtin_isunordered (a, b) && !(b == a);
29 }
30
31 int is_ordered_and_nonequal_sh_6 (float a, float b)
32 {
33 return !(b == a) && !__builtin_isunordered (a, b);
34 }
35
36 int is_unordered_or_equal_sh_1 (float a, float b)
37 {
38 return __builtin_isunordered (a, b) || (a == b);
39 }
40
41 int is_unordered_or_equal_sh_2 (float a, float b)
42 {
43 return __builtin_isunordered (a, b) || (b == a);
44 }
45
46 int is_unordered_or_equal_sh_3 (float a, float b)
47 {
48 return (a == b) || __builtin_isunordered (b, a);
49 }
50
51 int is_ordered_and_ltgt_sh_1 (float a, float b)
52 {
53 return !__builtin_isunordered (a, b) && ((a < b) || (a > b));
54 }
55
56 int is_ordered_and_ltgt_sh_2 (float a, float b)
57 {
58 return !__builtin_isunordered (a, b) && ((b > a) || (a > b));
59 }
60
61 int is_ordered_and_ltgt_sh_3 (float a, float b)
62 {
63 return ((b > a) || (a > b)) && !__builtin_isunordered (a, b);
64 }
From the result, I get some conclusions or quesitions:
1. We need to add :c to bit_and.
2. We need to add (ltgt @0 @1).
But, in 006t.gimple !__builtin_isunordered (a, b) && ((b > a) || (a > b)) will
be
126
127 if (a ord b) goto <D.2813>; else goto <D.2811>;
128 <D.2813>:
129 if (b <> a) goto <D.2814>; else goto <D.2811>;
130 <D.2814>:
131 iftmp.1 = 1;
GCC don't use bit_and like other testcases, looks like I can't just add a
simplify in match.pd, do you have any idea to solve this question in the
frontend (maybe frontend)?
3. After add these testcases, I consider to change the testcase's name to
opt-ucomi-1.c.
