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:
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.
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.
gcc/ChangeLog:
* match.pd: Add new pattern for round.
gcc/testsuite/ChangeLog:
* gcc.dg/fold-round-1.c: New test.
---
gcc/match.pd | 21 +++++++++++
gcc/testsuite/gcc.dg/fold-round-1.c | 56 +++++++++++++++++++++++++++++
2 files changed, 77 insertions(+)
create mode 100644 gcc/testsuite/gcc.dg/fold-round-1.c
diff --git a/gcc/match.pd b/gcc/match.pd
index a4248a521cf..f7e074eaa56 100644
--- a/gcc/match.pd
+++ b/gcc/match.pd
@@ -794,6 +794,27 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
(rdiv @0 (negate @1))
(rdiv (negate @0) @1))
+/* 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 lt lt lt ge ge ge ge)
+ bt (BUILT_IN_FLOORF BUILT_IN_FLOOR BUILT_IN_FLOORL IFN_FLOOR
+ BUILT_IN_CEILF BUILT_IN_CEIL BUILT_IN_CEILL IFN_CEIL)
+ bf (BUILT_IN_CEILF BUILT_IN_CEIL BUILT_IN_CEILL IFN_CEIL
+ BUILT_IN_FLOORF BUILT_IN_FLOOR BUILT_IN_FLOORL IFN_FLOOR)
+ floor (BUILT_IN_FLOORF BUILT_IN_FLOOR BUILT_IN_FLOORL IFN_FLOOR
+ BUILT_IN_FLOORF BUILT_IN_FLOOR BUILT_IN_FLOORL IFN_FLOOR)
+ ceil (BUILT_IN_CEILF BUILT_IN_CEIL BUILT_IN_CEILL IFN_CEIL
+ BUILT_IN_CEILF BUILT_IN_CEIL BUILT_IN_CEILL IFN_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))
+ (if (!HONOR_SIGNED_ZEROS (type) && !HONOR_SIGN_DEPENDENT_ROUNDING (type))
+ (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..0d7f95429ac
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/fold-round-1.c
@@ -0,0 +1,56 @@
+/* { dg-do compile } */
+/* { dg-options "-Ofast" } */
+
+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" } } */
--
2.20.1
