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 (&main_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 00000000000..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,} } } */
