在 2025/8/26 下午6:52, Richard Biener 写道:
On Tue, 26 Aug 2025, Zhou Zhao wrote:
在 2025/8/26 下午3:37, Richard Biener 写道:
On Thu, 21 Aug 2025, Zhou Zhao wrote:
This patch is a respond of the patch posted at
https://gcc.gnu.org/pipermail/gcc-patches/2025-January/673051.html
as some suggestion by Richard Biener, I have adopted these suggestions
and regenerated the patch.
In the 538.imagick_r benchmark of Spec2017, I find these pattern from
MagickRound function. This patch implements these pattern in match.pd
for 4 rules under -funsafe-math-optimizations:
1) (x-floor(x)) < (ceil(x)-x) ? floor(x) : ceil(x) -> floor(x+0.5)
2) (x-floor(x)) >= (ceil(x)-x) ? ceil(x) : floor(x) -> floor(x+0.5)
3) (ceil(x)-x) > (x-floor(x)) ? floor(x) : ceil(x) -> floor(x+0.5)
4) (ceil(x)-x) <= (x-floor(x)) ? ceil(x) : floor(x) -> floor(x+0.5)
The patch implements floor(x+0.5) operation to replace these pattern
that semantics of round(x) function.
Can you say why you particularly chose floor (x + 0.5) as result
while describing it to have the semantics of round (x)? The reasonable
other choice is round(x) itself?
What exact differences prompt you do gate this with
-funsafe-math-optimizations? I can see signed zeros have
different behavior, for x == -0.0 all forms in original form
return -0.0 while the simplification will return 0.0. The behavior
for Infs and NaNs looks unchanged. 0.5 and -0.5 seem to compute
to the same value when using floor(x+0.5) as simplification (unless
I made a mistake). floor or ceil do not raise IEEE exceptions,
so wouldn't -fno-signed-zeros be enough as a gate?
Thanks,
Richard.
Thank you for your reply. The time interval since the last patch
submission might be too long, so I will re-describe our discussion on
the above issues:
1. I consider the round functions are round(x) and rint(x). In round
halfway cases, round(x) away from zero, rint(x) rounds x to the nearest
even integer. When the pattern input is x=-2.5, return -2.0, but
round(-2.5) return -3.0. When the pattern input is x=2.5, it return
3.0, but rint(2.5) return 2.0. Therefore, using floor(x + 0.5) is the
best matches expression I think. Do you have any other functions with
semantics of round that could be used to represent this pattern?
No, I think that covers it. rint() also is affected by the rounding
mode so I think cannot be used here.
2. As you mentioned, I need to add the -funsafe-math-optimizations
option to protect the (+0.5) operation. With the -Ofast option, which
enables -fno-signed-zeros, I observed that when x = -0.4, the pattern
returns -0.0 on aarch64-linux-gnu but returns 0.0 on x86_64-linux-gnu.
floor(x + 0.5) will return 0.0 on all the above targets. Additionally,
the pattern behaves the same for all double values, including INFs
and NaNs.
3. Indeed, I cannot guarantee that (+0.5) will always yield the
expected value, so I use -funsafe-math-optimizations for protection.
I think this is better than checking HONOR_NANS/INFS/SIGNED_ZEROS,
because when the true result of x + 0.5 cannot be exactly represented
in the target floating-point format, and happens to be halfway between
two adjacent floating-point numbers, different rounding rules will
make different choices, leading to different final results.
I think 0.5 can be always exactly represented, but there might be
a special 'x' where + 0.5 triggers a one-ulp difference depending
on rounding mode.
That said, HONOR_SIGN_DEPENDENT_ROUNDING might be also an issue
because of that.
But other than that the transform should be value-preserving?
yes, I think so.
So I'd gate with !HONOR_SIGNED_ZEROS && !HONOR_SIGN_DEPENDENT_ROUNDING
instead?
These conditions can indeed cover all the case I have considered,
I'm no issues with it.
If no other issues, I will make the revisions and then resubmit
[patch v3] afterward.
Thanks,
Zhou Zhao.
Richard.
Thanks,
Zhou Zhao.
The patch was regtested on aarch64-linux-gnu and x86_64-linux-gnu,
SPEC 2017 and SPEC 2006 were run:
As for SPEC 2017, 538.imagick_r benchmark performance increased by 3%+
in base test of ratio mode.
As for SPEC 2006, while the transform does not seem to be triggered,
we also see no non-noise impact on performance.
OK for mainline?
gcc/ChangeLog:
* match.pd: Add new pattern for round.
gcc/testsuite/ChangeLog:
* gcc.dg/fold-round-1.c: New test.
---
gcc/match.pd | 17 +++++++++
gcc/testsuite/gcc.dg/fold-round-1.c | 56 +++++++++++++++++++++++++++++
2 files changed, 73 insertions(+)
create mode 100644 gcc/testsuite/gcc.dg/fold-round-1.c
diff --git a/gcc/match.pd b/gcc/match.pd
index 66e8a787449..94036603e08 100644
--- a/gcc/match.pd
+++ b/gcc/match.pd
@@ -794,6 +794,23 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
(rdiv @0 (negate @1))
(rdiv (negate @0) @1))
+(if (flag_unsafe_math_optimizations)
+/* convert semantics of round(x) function to floor(x+0.5). */
+/* (x-floor(x)) < (ceil(x)-x) ? floor(x) : ceil(x) --> floor(x+0.5). */
+/* (x-floor(x)) >= (ceil(x)-x) ? ceil(x) : floor(x) --> floor(x+0.5). */
+/* (ceil(x)-x) > (x-floor(x)) ? floor(x) : ceil(x) --> floor(x+0.5). */
+/* (ceil(x)-x) <= (x-floor(x)) ? ceil(x) : floor(x) --> floor(x+0.5). */
+(for op (lt ge)
+ bt (FLOOR CEIL)
+ bf (CEIL FLOOR)
+ floor (FLOOR FLOOR)
+ ceil (CEIL CEIL)
+ (simplify
+ (cond (op:c (minus:s SSA_NAME@0 (floor SSA_NAME@0))
+ (minus:s (ceil SSA_NAME@0) SSA_NAME@0))
+ (bt SSA_NAME@0) (bf SSA_NAME@0))
+ (floor (plus @0 { build_real (type, dconsthalf); })))))
+
(if (flag_unsafe_math_optimizations)
/* Simplify (C / x op 0.0) to x op 0.0 for C != 0, C != Inf/Nan.
Since C / x may underflow to zero, do this only for unsafe math. */
diff --git a/gcc/testsuite/gcc.dg/fold-round-1.c
b/gcc/testsuite/gcc.dg/fold-round-1.c
new file mode 100644
index 00000000000..845d6d2e475
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/fold-round-1.c
@@ -0,0 +1,56 @@
+/* { dg-do compile } */
+/* { dg-options "-Ofast -funsafe-math-optimizations" } */
+
+extern void link_error (void);
+
+#define TEST_ROUND(TYPE, FFLOOR, FCEIL)
\
+ void round_##FFLOOR##_1 (TYPE x)
\
+ {
\
+ TYPE t1 = 0;
\
+ TYPE t2 = __builtin_##FFLOOR (x + 0.5);
\
+ if ((x - __builtin_##FFLOOR (x)) < (__builtin_##FCEIL (x) - x))
\
+ t1 = __builtin_##FFLOOR (x);
\
+ else
\
+ t1 = __builtin_##FCEIL (x);
\
+ if (t1 != t2)
\
+ link_error ();
\
+ }
\
+ void round_##FFLOOR##_2 (TYPE x)
\
+ {
\
+ TYPE t1 = 0;
\
+ TYPE t2 = __builtin_##FFLOOR (x + 0.5);
\
+ if ((__builtin_##FCEIL (x) - x) > (x - __builtin_##FFLOOR (x)))
\
+ t1 = __builtin_##FFLOOR (x);
\
+ else
\
+ t1 = __builtin_##FCEIL (x);
\
+ if (t1 != t2)
\
+ link_error ();
\
+ }
\
+ void round_##FFLOOR##_3 (TYPE x)
\
+ {
\
+ TYPE t1 = 0;
\
+ TYPE t2 = __builtin_##FFLOOR (x + 0.5);
\
+ if ((__builtin_##FCEIL (x) - x) <= (x - __builtin_##FFLOOR (x)))
\
+ t1 = __builtin_##FCEIL (x);
\
+ else
\
+ t1 = __builtin_##FFLOOR (x);
\
+ if (t1 != t2)
\
+ link_error ();
\
+ }
\
+ void round_##FFLOOR##_4 (TYPE x)
\
+ {
\
+ TYPE t1 = 0;
\
+ TYPE t2 = __builtin_##FFLOOR (x + 0.5);
\
+ if ((x - __builtin_##FFLOOR (x)) >= (__builtin_##FCEIL (x) - x))
\
+ t1 = __builtin_##FCEIL (x);
\
+ else
\
+ t1 = __builtin_##FFLOOR (x);
\
+ if (t1 != t2)
\
+ link_error ();
\
+ }
+
+TEST_ROUND (float, floorf, ceilf)
+TEST_ROUND (double, floor, ceil)
+TEST_ROUND (long double, floorl, ceill)
+
+/* { dg-final { scan-assembler-not "link_error" } } */
