move sincos after pre (was: Re: [Ada,FYI] revamp ada.numerics.aux)

[Alexandre Oliva]

On Oct 22, 2020, Alexandre Oliva <oliva@adacore.com> wrote:

> On Oct 18, 2020, Alexandre Oliva <oliva@adacore.com> wrote:
>> The option is provided by default, but there is an alternate version
>> that doesn't, that is used for vxworks targets.

> vxworks float EFs not precise enough -> use long float

> From: Alexandre Oliva <oliva@adacore.com>

> Some acats-4 tests that check the precision of Float elementary
> functions fail with vxworks 7.2's implementations of single-precision
> math functions.

> This patch arranges for us to bypass the single-precision functions,
> and use the Aux_Long_Float implementation, based on the double-typed
> calls from the C library, for Float and Short_Float.


On platforms in which Aux_[Real_Type] involves non-NOP conversions
(e.g., between single- and double-precision, or between short float
and float), the conversions before the calls are CSEd too late for
sincos to combine calls.

This patch moves sincos after PRE, where the conversions are unified
at -O2.

I'm regstrapping this on x86_64-linux-gnu and powerpc64-linux-gnu, and
also testing it on affected platforms.  Another way to go, that would
take a little more effort, would be to extend sincos to take equivalent
conversions as the same operand, but I doubt I'll be able to undertake
that any time soon, so...  Is this one ok to install?


for  gcc/ChangeLog

	* passes.def: Move sincos after pre.

for  gcc/testsuite/ChangeLog

	* gnat.dg/sin_cos.ads: New.
	* gnat.dg/sin_cos.adb: New.
	* gcc.dg/sin_cos.c: New.
---
 gcc/passes.def                    |    2 +-
 gcc/testsuite/gcc.dg/sin_cos.c    |   41 +++++++++++++++++++++++++++++++++++++
 gcc/testsuite/gnat.dg/sin_cos.adb |   14 +++++++++++++
 gcc/testsuite/gnat.dg/sin_cos.ads |    4 ++++
 4 files changed, 60 insertions(+), 1 deletion(-)
 create mode 100644 gcc/testsuite/gcc.dg/sin_cos.c
 create mode 100644 gcc/testsuite/gnat.dg/sin_cos.adb
 create mode 100644 gcc/testsuite/gnat.dg/sin_cos.ads

diff --git a/gcc/passes.def b/gcc/passes.def
index cf15d8e..2743506 100644
--- a/gcc/passes.def
+++ b/gcc/passes.def
@@ -242,12 +242,12 @@ along with GCC; see the file COPYING3.  If not see
       NEXT_PASS (pass_ccp, true /* nonzero_p */);
       /* After CCP we rewrite no longer addressed locals into SSA
 	 form if possible.  */
-      NEXT_PASS (pass_cse_sincos);
       NEXT_PASS (pass_optimize_bswap);
       NEXT_PASS (pass_laddress);
       NEXT_PASS (pass_lim);
       NEXT_PASS (pass_walloca, false);
       NEXT_PASS (pass_pre);
+      NEXT_PASS (pass_cse_sincos);
       NEXT_PASS (pass_sink_code);
       NEXT_PASS (pass_sancov);
       NEXT_PASS (pass_asan);
diff --git a/gcc/testsuite/gcc.dg/sin_cos.c b/gcc/testsuite/gcc.dg/sin_cos.c
new file mode 100644
index 00000000..a4a7727
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/sin_cos.c
@@ -0,0 +1,41 @@
+/* { dg-do compile } */
+/* { dg-options "-O2" } */
+
+/* This maps to essentially the same gimple that is generated for
+   gnat.dg/sin_cos.adb, on platforms that use the wraplf variant of
+   Ada.Numerics.Aux_Float.  The value of EPSILON is not relevant to
+   the test, but the test must be there to keep the conversions in
+   different BBs long enough to trigger the problem that prevented the
+   sincos optimization, because the arguments passed to sin and cos
+   didn't get unified into a single SSA_NAME in time for sincos.  */
+
+#include <math.h>
+
+#define EPSILON 3.4526697709225118160247802734375e-4
+
+static float my_sinf(float x) {
+  return (float) sin ((double) x);
+}
+
+static float wrap_sinf(float x) {
+  if (fabs (x) < EPSILON)
+    return 0;
+  return my_sinf (x);
+}
+
+static float my_cosf(float x) {
+  return (float) cos ((double) x);
+}
+
+static float wrap_cosf(float x) {
+  if (fabs (x) < EPSILON)
+    return 1;
+  return my_cosf (x);
+}
+
+float my_sin_cos(float x, float *s, float *c) {
+  *s = wrap_sinf (x);
+  *c = wrap_cosf (x);
+}
+
+/* { dg-final { scan-assembler "sincos\|cexp" { target *-linux-gnu* *-w64-mingw* powerpc*-*-* } } } */
diff --git a/gcc/testsuite/gnat.dg/sin_cos.adb b/gcc/testsuite/gnat.dg/sin_cos.adb
new file mode 100644
index 00000000..e72f521
--- /dev/null
+++ b/gcc/testsuite/gnat.dg/sin_cos.adb
@@ -0,0 +1,14 @@
+--  { dg-do compile }
+--  { dg-options "-O2 -gnatn" }
+
+with Ada.Numerics.Elementary_Functions;
+use Ada.Numerics.Elementary_Functions;
+package body Sin_Cos is
+   procedure Sin_Cos (Angle : T; SinA, CosA : out T) is
+   begin
+      SinA := Sin (Angle);
+      CosA := Cos (Angle);
+   end;
+end Sin_Cos;
+
+--  { dg-final { scan-assembler "sincos\|cexp" { target *-linux-gnu* *-w64-mingw* powerpc*-*-* } } }
diff --git a/gcc/testsuite/gnat.dg/sin_cos.ads b/gcc/testsuite/gnat.dg/sin_cos.ads
new file mode 100644
index 00000000..a0eff3d
--- /dev/null
+++ b/gcc/testsuite/gnat.dg/sin_cos.ads
@@ -0,0 +1,4 @@
+package Sin_Cos is
+   subtype T is Float;
+   procedure Sin_Cos (Angle : T; SinA, CosA : out T);
+end Sin_Cos;


-- 
Alexandre Oliva, happy hacker
https://FSFLA.org/blogs/lxo/
Free Software Activist
GNU Toolchain Engineer