In my previous patch enabling some of the built-in functions for _Float<N> and
_Float<N>X datatypes, I missed making fminf128 and fmaxf128 generate the
minimum and maximum inline code when -ffast-math is used. This patch to
match.pd enables the code generation using if-then-else if the machine does not
have an appropriate min/max instruction (the power9 hardware does not have
min/max instructions for quad floating point).
I have done builds on x86-64, little endian power8, and little endian power9
prototype systems. All builds bootstrapped and had no regressions. Can I
check this patch into the trunk?
[gcc]
2017-11-09 Michael Meissner <meiss...@linux.vnet.ibm.com>
* match.pd: Convert fminf<N>, fminf<N>x, fmax<N>, and fmax<N>x
into the min/max operations for _Float<N> and _Float<N>X types.
[gcc/testsuite]
2017-11-09 Michael Meissner <meiss...@linux.vnet.ibm.com>
* gcc.target/powerpc/float128-minmax.c: New test.
--
Michael Meissner, IBM
IBM, M/S 2506R, 550 King Street, Littleton, MA 01460-6245, USA
email: meiss...@linux.vnet.ibm.com, phone: +1 (978) 899-4797
Index: gcc/match.pd
===================================================================
--- gcc/match.pd (revision 254473)
+++ gcc/match.pd (working copy)
@@ -1723,7 +1723,7 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
/* Simplifications of MIN_EXPR, MAX_EXPR, fmin() and fmax(). */
-(for minmax (min max FMIN FMAX)
+(for minmax (min max FMIN FMIN_FN FMAX FMAX_FN)
(simplify
(minmax @0 @0)
@0))
@@ -1801,7 +1801,7 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
&& TYPE_PRECISION (TREE_TYPE (@0)) > TYPE_PRECISION (type))
(minmax @1 (convert @2)))))
-(for minmax (FMIN FMAX)
+(for minmax (FMIN FMIN_FN FMAX FMAX_FN)
/* If either argument is NaN, return the other one. Avoid the
transformation if we get (and honor) a signalling NaN. */
(simplify
@@ -1819,11 +1819,17 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
(FMIN @0 @1)
(min @0 @1))
(simplify
+ (FMIN_FN @0 @1)
+ (min @0 @1))
+ (simplify
(FMAX @0 @1)
+ (max @0 @1))
+ (simplify
+ (FMAX_FN @0 @1)
(max @0 @1)))
/* min (-A, -B) -> -max (A, B) */
-(for minmax (min max FMIN FMAX)
- maxmin (max min FMAX FMIN)
+(for minmax (min max FMIN FMIN_FN FMAX FMAX_FN)
+ maxmin (max min FMAX FMAX_FN FMIN FMAX_FN)
(simplify
(minmax (negate:s@2 @0) (negate:s@3 @1))
(if (FLOAT_TYPE_P (TREE_TYPE (@0))
Index: gcc/testsuite/gcc.target/powerpc/float128-minmax.c
===================================================================
--- gcc/testsuite/gcc.target/powerpc/float128-minmax.c (revision 0)
+++ gcc/testsuite/gcc.target/powerpc/float128-minmax.c (revision 0)
@@ -0,0 +1,15 @@
+/* { dg-do compile { target { powerpc*-*-* && lp64 } } } */
+/* { dg-require-effective-target powerpc_p9vector_ok } */
+/* { dg-options "-mpower9-vector -O2 -ffast-math" } */
+
+#ifndef TYPE
+#define TYPE _Float128
+#endif
+
+/* Test that the fminf128/fmaxf128 functions generate if/then/else and not a
+ call. */
+TYPE f128_min (TYPE a, TYPE b) { return __builtin_fminf128 (a, b); }
+TYPE f128_max (TYPE a, TYPE b) { return __builtin_fmaxf128 (a, b); }
+
+/* { dg-final { scan-assembler {\mxscmpuqp\M} } } */
+/* { dg-final { scan-assembler-not {\mbl\M} } } */
