[pushed] aarch64: Fix overly optimistic LDP/STP matching [PR104005]
Commit Message
In g:526e1639aa76b0a8496b0dc3a3ff2c450229544e I'd added support
for finding more consecutive MEMs. However, the check was too
eager, in that it matched MEM_REFs with the same base address
even if that base address was an arbitrary SSA name. This can
give wrong results if a MEM_REF from one loop iteration is
compared with a MEM_REF from another (e.g. after rtl unrolling).
In principle, we could still accept MEM_REFs based on the same
incoming SSA name, but there seems to be no out-of-the-box API
for doing that. Adding a new one at this stage in GCC 12 doesn't
feel like a good risk/reward trade-off.
This patch therefore restricts the MEM_EXPR comparison to base decls
only, excluding all MEM_REFs. It means we lose all the new STPs in
the PR testcase but keep the ones in the original stp_1.c testcase.
Tested on aarch64-linux-gnu and pushed.
Richard
gcc/
PR target/104005
* config/aarch64/aarch64.cc (aarch64_check_consecutive_mems):
When using MEM_EXPR, require the base to be a decl.
gcc/testsuite/
PR target/104005
* gcc.target/aarch64/pr104005.c: New test.
---
gcc/config/aarch64/aarch64.cc | 1 +
gcc/testsuite/gcc.target/aarch64/pr104005.c | 17 +++++++++++++++++
2 files changed, 18 insertions(+)
create mode 100644 gcc/testsuite/gcc.target/aarch64/pr104005.c
Comments
On Tue, Jan 18, 2022 at 1:22 PM Richard Sandiford via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> In g:526e1639aa76b0a8496b0dc3a3ff2c450229544e I'd added support
> for finding more consecutive MEMs. However, the check was too
> eager, in that it matched MEM_REFs with the same base address
> even if that base address was an arbitrary SSA name. This can
> give wrong results if a MEM_REF from one loop iteration is
> compared with a MEM_REF from another (e.g. after rtl unrolling).
Hmm, if it really happens that way it also poses an issue for other
must-alias analysis like DSE (though I don't remember RTL DSE
looking at MEM_EXPRs for this purpose). And may-alias as well
if say MEM[_1 + 0] and MEM[_1 + 4] are seen but _1 is assigned
_1 + 4 between both (RTL unrolling would basically violate the
single-definition guarantee of SSA).
> In principle, we could still accept MEM_REFs based on the same
> incoming SSA name, but there seems to be no out-of-the-box API
> for doing that. Adding a new one at this stage in GCC 12 doesn't
> feel like a good risk/reward trade-off.
It's also not exactly obvious how to do that.
> This patch therefore restricts the MEM_EXPR comparison to base decls
> only, excluding all MEM_REFs. It means we lose all the new STPs in
> the PR testcase but keep the ones in the original stp_1.c testcase.
>
> Tested on aarch64-linux-gnu and pushed.
>
> Richard
>
>
> gcc/
> PR target/104005
> * config/aarch64/aarch64.cc (aarch64_check_consecutive_mems):
> When using MEM_EXPR, require the base to be a decl.
>
> gcc/testsuite/
> PR target/104005
> * gcc.target/aarch64/pr104005.c: New test.
> ---
> gcc/config/aarch64/aarch64.cc | 1 +
> gcc/testsuite/gcc.target/aarch64/pr104005.c | 17 +++++++++++++++++
> 2 files changed, 18 insertions(+)
> create mode 100644 gcc/testsuite/gcc.target/aarch64/pr104005.c
>
> diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
> index fdf0c9bd5b8..296145e6008 100644
> --- a/gcc/config/aarch64/aarch64.cc
> +++ b/gcc/config/aarch64/aarch64.cc
> @@ -24747,6 +24747,7 @@ aarch64_check_consecutive_mems (rtx *mem1, rtx *mem2, bool *reversed)
> &expr_offset2);
> if (!expr_base1
> || !expr_base2
> + || !DECL_P (expr_base1)
> || !operand_equal_p (expr_base1, expr_base2, OEP_ADDRESS_OF))
> return false;
>
> diff --git a/gcc/testsuite/gcc.target/aarch64/pr104005.c b/gcc/testsuite/gcc.target/aarch64/pr104005.c
> new file mode 100644
> index 00000000000..09dd81910eb
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/pr104005.c
> @@ -0,0 +1,17 @@
> +/* { dg-options "-O2 -funroll-loops" } */
> +
> +typedef int v2 __attribute__((vector_size(8)));
> +
> +void f(void) {
> + v2 v[1024];
> + v2 *ptr = v;
> + for (int i = 0; i < 512; ++i)
> + {
> + ptr[0][0] = 0;
> + asm volatile ("":::"memory");
> + ptr[0][1] = 1;
> + ptr += 2;
> + }
> +}
> +
> +/* { dg-final { scan-assembler-not {\tstp\t} } } */
> --
> 2.25.1
>
@@ -24747,6 +24747,7 @@ aarch64_check_consecutive_mems (rtx *mem1, rtx *mem2, bool *reversed)
&expr_offset2);
if (!expr_base1
|| !expr_base2
+ || !DECL_P (expr_base1)
|| !operand_equal_p (expr_base1, expr_base2, OEP_ADDRESS_OF))
return false;
new file mode 100644
@@ -0,0 +1,17 @@
+/* { dg-options "-O2 -funroll-loops" } */
+
+typedef int v2 __attribute__((vector_size(8)));
+
+void f(void) {
+ v2 v[1024];
+ v2 *ptr = v;
+ for (int i = 0; i < 512; ++i)
+ {
+ ptr[0][0] = 0;
+ asm volatile ("":::"memory");
+ ptr[0][1] = 1;
+ ptr += 2;
+ }
+}
+
+/* { dg-final { scan-assembler-not {\tstp\t} } } */