Re: [PATCH] vect: Tweak comparisons with existing epilogue loops

2021-08-04 Thread Richard Biener via Gcc-patches
On Tue, Aug 3, 2021 at 3:52 PM Richard Sandiford via Gcc-patches
 wrote:
>
> This patch uses a more accurate scalar iteration estimate when
> comparing the epilogue of a constant-iteration loop with a candidate
> replacement epilogue.
>
> In the testcase, the patch prevents a 1-to-3-element SVE epilogue
> from seeming better than a 64-bit Advanced SIMD epilogue.
>
> Tested on aarch64-linux-gnu and x86_64-linux-gnu.  OK to install?

OK.

Richard.

> Richard
>
>
> gcc/
> * tree-vect-loop.c (vect_better_loop_vinfo_p): Detect cases in
> which old_loop_vinfo is an epilogue loop that handles a constant
> number of iterations.
>
> gcc/testsuite/
> * gcc.target/aarch64/sve/cost_model_12.c: New test.
> ---
>  .../gcc.target/aarch64/sve/cost_model_12.c| 19 +++
>  gcc/tree-vect-loop.c  | 10 +-
>  2 files changed, 28 insertions(+), 1 deletion(-)
>  create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/cost_model_12.c
>
> diff --git a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c
> index 0009d0964af..0a5b65adb04 100644
> --- a/gcc/tree-vect-loop.c
> +++ b/gcc/tree-vect-loop.c
> @@ -2778,7 +2778,15 @@ vect_better_loop_vinfo_p (loop_vec_info new_loop_vinfo,
>
>/* Limit the VFs to what is likely to be the maximum number of iterations,
>   to handle cases in which at least one loop_vinfo is fully-masked.  */
> -  HOST_WIDE_INT estimated_max_niter = likely_max_stmt_executions_int (loop);
> +  HOST_WIDE_INT estimated_max_niter;
> +  loop_vec_info main_loop = LOOP_VINFO_ORIG_LOOP_INFO (old_loop_vinfo);
> +  unsigned HOST_WIDE_INT main_vf;
> +  if (main_loop
> +  && LOOP_VINFO_NITERS_KNOWN_P (main_loop)
> +  && LOOP_VINFO_VECT_FACTOR (main_loop).is_constant (_vf))
> +estimated_max_niter = LOOP_VINFO_INT_NITERS (main_loop) % main_vf;
> +  else
> +estimated_max_niter = likely_max_stmt_executions_int (loop);
>if (estimated_max_niter != -1)
>  {
>if (known_le (estimated_max_niter, new_vf))
> diff --git a/gcc/testsuite/gcc.target/aarch64/sve/cost_model_12.c 
> b/gcc/testsuite/gcc.target/aarch64/sve/cost_model_12.c
> new file mode 100644
> index 000..4c5226e05de
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sve/cost_model_12.c
> @@ -0,0 +1,19 @@
> +/* { dg-options "-O3 -mtune=neoverse-512tvb" } */
> +
> +void
> +f (float x[restrict 10][1024],
> +   float y[restrict 10][1024], float z)
> +{
> +  for (int i = 0; i < 10; ++i)
> +{
> +#pragma GCC unroll 10
> +  for (int j = 0; j < 10; ++j)
> +   x[j][i] = y[j][i] * z;
> +}
> +}
> +
> +/* We should unroll the outer loop, with 2x 16-byte vectors and 1x
> +   8-byte vectors.  */
> +/* { dg-final { scan-assembler-not {\tptrue\t} } } */
> +/* { dg-final { scan-assembler {\tv[0-9]+\.4s,} } } */
> +/* { dg-final { scan-assembler {\tv[0-9]+\.2s,} } } */


[PATCH] vect: Tweak comparisons with existing epilogue loops

2021-08-03 Thread Richard Sandiford via Gcc-patches
This patch uses a more accurate scalar iteration estimate when
comparing the epilogue of a constant-iteration loop with a candidate
replacement epilogue.

In the testcase, the patch prevents a 1-to-3-element SVE epilogue
from seeming better than a 64-bit Advanced SIMD epilogue.

Tested on aarch64-linux-gnu and x86_64-linux-gnu.  OK to install?

Richard


gcc/
* tree-vect-loop.c (vect_better_loop_vinfo_p): Detect cases in
which old_loop_vinfo is an epilogue loop that handles a constant
number of iterations.

gcc/testsuite/
* gcc.target/aarch64/sve/cost_model_12.c: New test.
---
 .../gcc.target/aarch64/sve/cost_model_12.c| 19 +++
 gcc/tree-vect-loop.c  | 10 +-
 2 files changed, 28 insertions(+), 1 deletion(-)
 create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/cost_model_12.c

diff --git a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c
index 0009d0964af..0a5b65adb04 100644
--- a/gcc/tree-vect-loop.c
+++ b/gcc/tree-vect-loop.c
@@ -2778,7 +2778,15 @@ vect_better_loop_vinfo_p (loop_vec_info new_loop_vinfo,
 
   /* Limit the VFs to what is likely to be the maximum number of iterations,
  to handle cases in which at least one loop_vinfo is fully-masked.  */
-  HOST_WIDE_INT estimated_max_niter = likely_max_stmt_executions_int (loop);
+  HOST_WIDE_INT estimated_max_niter;
+  loop_vec_info main_loop = LOOP_VINFO_ORIG_LOOP_INFO (old_loop_vinfo);
+  unsigned HOST_WIDE_INT main_vf;
+  if (main_loop
+  && LOOP_VINFO_NITERS_KNOWN_P (main_loop)
+  && LOOP_VINFO_VECT_FACTOR (main_loop).is_constant (_vf))
+estimated_max_niter = LOOP_VINFO_INT_NITERS (main_loop) % main_vf;
+  else
+estimated_max_niter = likely_max_stmt_executions_int (loop);
   if (estimated_max_niter != -1)
 {
   if (known_le (estimated_max_niter, new_vf))
diff --git a/gcc/testsuite/gcc.target/aarch64/sve/cost_model_12.c 
b/gcc/testsuite/gcc.target/aarch64/sve/cost_model_12.c
new file mode 100644
index 000..4c5226e05de
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/sve/cost_model_12.c
@@ -0,0 +1,19 @@
+/* { dg-options "-O3 -mtune=neoverse-512tvb" } */
+
+void
+f (float x[restrict 10][1024],
+   float y[restrict 10][1024], float z)
+{
+  for (int i = 0; i < 10; ++i)
+{
+#pragma GCC unroll 10
+  for (int j = 0; j < 10; ++j)
+   x[j][i] = y[j][i] * z;
+}
+}
+
+/* We should unroll the outer loop, with 2x 16-byte vectors and 1x
+   8-byte vectors.  */
+/* { dg-final { scan-assembler-not {\tptrue\t} } } */
+/* { dg-final { scan-assembler {\tv[0-9]+\.4s,} } } */
+/* { dg-final { scan-assembler {\tv[0-9]+\.2s,} } } */