Re: [PATCH][tree-ssa-math-opts] Expand pow (x, CONST) using square roots when possible

[Kyrill Tkachov <kyrylo dot tkachov at foss dot arm dot com>]

Hi Richard,

On 13/05/15 12:27, Richard Biener wrote:
> > > I notice that we don't have a testuite check that the target has
> > > >>a hw sqrt instructions. Would you like me to add one? Or can I make
> > > >>the testcase aarch64-specific?
> Would be great to have a testsuite check for this.

I've committed the patch with r223167.

The attached patch adds a testsuite check for hardware sqrt instructions.
In this version I've included arm (on the condition that vfp is possible),
aarch64, x86_64 and powerpc with vsx.
Is this definition ok?

I'm particularly not familiar with the powerpc architectures.

With this check in place, I've migrated the pow synthesis test from gcc.target/aarch64 to gcc.dg.

This test passes on arm-none-eabi, aarch64-none-elf and x86_64-linux.

Ok?

2015-05-13  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

    * lib/target-supports.exp (check_effective_target_hw_sqrt): New check.
    * gcc.dg/pow-sqrt-synth-1.c: New test.
    * gcc.target/aarch64/pow-sqrt-synth-1.c: Delete.

commit e30889fa7024ccfd47731aafbaf2288646b65504
Author: Kyrylo Tkachov <kyrylo.tkachov@arm.com>
Date:   Wed May 13 16:08:03 2015 +0100

    Add testsuite check for hw sqrt. Add generic test for pow sqrt synthesis

diff --git a/gcc/testsuite/gcc.dg/pow-sqrt-synth-1.c b/gcc/testsuite/gcc.dg/pow-sqrt-synth-1.c
new file mode 100644
index 0000000..a65efeb
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/pow-sqrt-synth-1.c
@@ -0,0 +1,38 @@
+/* { dg-do compile { target hw_sqrt } } */
+/* { dg-options "-fdump-tree-sincos -Ofast --param max-pow-sqrt-depth=8" } */
+/* { dg-additional-options "-mfloat-abi=softfp -mfpu=neon-vfpv4" { target arm*-*-* } } */
+
+double
+foo (double a)
+{
+  return __builtin_pow (a, -5.875);
+}
+
+double
+foof (double a)
+{
+  return __builtin_pow (a, 0.75f);
+}
+
+double
+bar (double a)
+{
+  return __builtin_pow (a, 1.0 + 0.00390625);
+}
+
+double
+baz (double a)
+{
+  return __builtin_pow (a, -1.25) + __builtin_pow (a, 5.75) - __builtin_pow (a, 3.375);
+}
+
+#define N 256
+void
+vecfoo (double *a)
+{
+  for (int i = 0; i < N; i++)
+    a[i] = __builtin_pow (a[i], 1.25);
+}
+
+/* { dg-final { scan-tree-dump-times "synthesizing" 7 "sincos" } } */
+/* { dg-final { cleanup-tree-dump "sincos" } } */
diff --git a/gcc/testsuite/gcc.target/aarch64/pow-sqrt-synth-1.c b/gcc/testsuite/gcc.target/aarch64/pow-sqrt-synth-1.c
deleted file mode 100644
index 52514fb..0000000
--- a/gcc/testsuite/gcc.target/aarch64/pow-sqrt-synth-1.c
+++ /dev/null
@@ -1,38 +0,0 @@
-/* { dg-do compile } */
-/* { dg-options "-fdump-tree-sincos -Ofast --param max-pow-sqrt-depth=8" } */
-
-
-double
-foo (double a)
-{
-  return __builtin_pow (a, -5.875);
-}
-
-double
-foof (double a)
-{
-  return __builtin_pow (a, 0.75f);
-}
-
-double
-bar (double a)
-{
-  return __builtin_pow (a, 1.0 + 0.00390625);
-}
-
-double
-baz (double a)
-{
-  return __builtin_pow (a, -1.25) + __builtin_pow (a, 5.75) - __builtin_pow (a, 3.375);
-}
-
-#define N 256
-void
-vecfoo (double *a)
-{
-  for (int i = 0; i < N; i++)
-    a[i] = __builtin_pow (a[i], 1.25);
-}
-
-/* { dg-final { scan-tree-dump-times "synthesizing" 7 "sincos" } } */
-/* { dg-final { cleanup-tree-dump "sincos" } } */
\ No newline at end of file
diff --git a/gcc/testsuite/lib/target-supports.exp b/gcc/testsuite/lib/target-supports.exp
index 3728927..c8f20a4 100644
--- a/gcc/testsuite/lib/target-supports.exp
+++ b/gcc/testsuite/lib/target-supports.exp
@@ -4668,6 +4668,27 @@ proc check_effective_target_vect_call_copysignf { } {
     return $et_vect_call_copysignf_saved
 }
 
+# Return 1 if the target supports hardware square root instructions.
+
+proc check_effective_target_hw_sqrt { } {
+    global et_hw_sqrt_saved
+
+    if [info exists et_hw_sqrt_saved] {
+	verbose "check_effective_target_hw_sqrt: using cached result" 2
+    } else {
+	set et_hw_sqrt_saved 0
+	if { [istarget x86_64-*-*]
+	     || ([istarget powerpc*-*-*] && [check_vsx_hw_available])
+	     || [istarget aarch64*-*-*]
+	     || ([istarget arm*-*-*] && [check_effective_target_arm_vfp_ok]) } {
+	   set et_hw_sqrt_saved 1
+	}
+    }
+
+    verbose "check_effective_target_hw_sqrt: returning et_hw_sqrt_saved" 2
+    return $et_hw_sqrt_saved
+}
+
 # Return 1 if the target supports vector sqrtf calls.
 
 proc check_effective_target_vect_call_sqrtf { } {