[PATCH, rs6000] optimization for long long and double vec_reve [PR100868]

[PATCH, rs6000] optimization for long long and double vec_reve [PR100868]

[HAO CHEN GUI]

Hi

    The patch optimized expansion for long long or double vec_reve builtin.

      Bootstrapped and tested on powerpc64le-linux with no regressions. 
Is this okay for trunk? Any recommendations? Thanks a lot.

ChangeLog

2021-09-06 Haochen Gui <guihaoc@linux.ibm.com>

gcc/
     * config/rs6000/altivec.md (altivec_vreve<mode>2):
     Modify the expansion when number of elements is 2.

gcc/testsuite/
     * gcc.target/powerpc/vec_reve.c: New test.


patch.diff

diff --git a/gcc/config/rs6000/altivec.md b/gcc/config/rs6000/altivec.md
index 1351dafbc41..96e71979fbd 100644
--- a/gcc/config/rs6000/altivec.md
+++ b/gcc/config/rs6000/altivec.md
@@ -4063,6 +4063,13 @@ (define_expand "altivec_vreve<mode>2"
    size = GET_MODE_UNIT_SIZE (<MODE>mode);
    num_elements = GET_MODE_NUNITS (<MODE>mode);

+  if (num_elements == 2)
+    {
+      emit_insn (gen_vsx_xxpermdi_<mode> (operands[0], operands[1],
+                      operands[1], GEN_INT (2)));
+      DONE;
+    }
+
    for (j = 0; j < num_elements; j++)
      for (i = 0; i < size; i++)
        RTVEC_ELT (v, i + j * size)
diff --git a/gcc/testsuite/gcc.target/powerpc/vec_reve.c 
b/gcc/testsuite/gcc.target/powerpc/vec_reve.c
new file mode 100644
index 00000000000..d86ce11ba91
--- /dev/null
+++ b/gcc/testsuite/gcc.target/powerpc/vec_reve.c
@@ -0,0 +1,17 @@
+/* { dg-do compile { target { powerpc*-*-* } } } */
+/* { dg-require-effective-target powerpc_altivec_ok } */
+/* { dg-options "-O2 -maltivec" } */
+
+#include <altivec.h>
+
+vector double foo1 (vector double a)
+{
+   return vec_reve (a);
+}
+
+vector long long foo2 (vector long long a)
+{
+   return vec_reve (a);
+}
+
+/* { dg-final { scan-assembler-times {\mxxpermdi\M} 2 } } */

Re: [PATCH, rs6000] optimization for long long and double vec_reve [PR100868]

[Segher Boessenkool]

Hi!

On Mon, Sep 06, 2021 at 02:10:58PM +0800, HAO CHEN GUI wrote:
>    The patch optimized expansion for long long or double vec_reve builtin.
> 
>      Bootstrapped and tested on powerpc64le-linux with no regressions. 
> Is this okay for trunk? Any recommendations? Thanks a lot.

In the future, please say something about what you changed, something
about the new code?

"Use xxswapd instead of <whatever we did before>" or even "Use xxswapd".
It doesn't have to be long, but something to guide the reader helps a
lot :-)

>     * config/rs6000/altivec.md (altivec_vreve<mode>2):
>     Modify the expansion when number of elements is 2.

	* config/rs6000/altivec.md (altivec_vreve<mode>2): Modify the expansion
	when num_elements is 2.

But "Modify" does not say what you changed it to.  So something like

	* config/rs6000/altivec.md (altivec_vreve<mode>2): Use xxswapd if
	num_elements is 2.

But, there are multiple patterns with that name.  So this would be

	* config/rs6000/altivec.md (altivec_vreve<mode>2 for VEC_A): Use xxswapd
	if num_elements is 2.

That check is a bit nasty, so maybe you should just split this pattern
into one for VEC_K and one for VEC_64?

> --- a/gcc/config/rs6000/altivec.md
> +++ b/gcc/config/rs6000/altivec.md
> @@ -4063,6 +4063,13 @@ (define_expand "altivec_vreve<mode>2"
>    size = GET_MODE_UNIT_SIZE (<MODE>mode);
>    num_elements = GET_MODE_NUNITS (<MODE>mode);
> 
> +  if (num_elements == 2)
> +    {
> +      emit_insn (gen_vsx_xxpermdi_<mode> (operands[0], operands[1],
> +                      operands[1], GEN_INT (2)));
> +      DONE;
> +    }

emit_insn (gen_xxswapd_<mode> (operands[0], operands[1]));

If this doesn't compile like this, it will need some "parameterized
names" magic (for xxswapd), a bit more work (but one-time work, and it
is worth it).

> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/powerpc/vec_reve.c
> @@ -0,0 +1,17 @@
> +/* { dg-do compile { target { powerpc*-*-* } } } */

That is the only allowed target in gcc.target/powerpc anyway.  Just
write

/* { dg-do compile } */

and nothing more (this (compile) is the default as well, you can just
leave it out completely if you want).

Finally: should whatever the old code generated have been optimised
better?


Segher