PHIOPT: Fix minmax_replacement for three way
Checks
Commit Message
So when diamond bb support was added to minmax_replacement in r13-1950-g9bb19e143cfe,
the code was not expecting the alt_middle_bb not to exist if it was empty (for threeway_p).
So when factor_out_conditional_conversion was used to factor out conversions, it turns out
the assumption for alt_middle_bb to be wrong and we ended up with threeway_p being true but
having middle_bb being empty but alt_middle_bb not being empty which causes wrong code in
many cases.
This patch fixes the issue by adding a test for the 2 cases where the assumption on
threeway_p case having the other bb being empty.
Note my plan for GCC 15 is remove minmax_replacement as match.pd will catch all cases
at that point.
OK? Bootstrapped and tested on x86_64-linux-gnu with no regressions.
PR tree-optimization/111469
gcc/ChangeLog:
* tree-ssa-phiopt.cc (minmax_replacement): Fix
the assumption for the `non-diamond` handling cases
of diamond code.
gcc/testsuite/ChangeLog:
* gcc.c-torture/execute/pr111469-1.c: New test.
---
.../gcc.c-torture/execute/pr111469-1.c | 38 +++++++++++++++++++
gcc/tree-ssa-phiopt.cc | 10 ++++-
2 files changed, 46 insertions(+), 2 deletions(-)
create mode 100644 gcc/testsuite/gcc.c-torture/execute/pr111469-1.c
Comments
On Thu, Sep 21, 2023 at 10:10 AM Andrew Pinski <apinski@marvell.com> wrote:
>
> So when diamond bb support was added to minmax_replacement in r13-1950-g9bb19e143cfe,
> the code was not expecting the alt_middle_bb not to exist if it was empty (for threeway_p).
> So when factor_out_conditional_conversion was used to factor out conversions, it turns out
> the assumption for alt_middle_bb to be wrong and we ended up with threeway_p being true but
> having middle_bb being empty but alt_middle_bb not being empty which causes wrong code in
> many cases.
>
> This patch fixes the issue by adding a test for the 2 cases where the assumption on
> threeway_p case having the other bb being empty.
>
> Note my plan for GCC 15 is remove minmax_replacement as match.pd will catch all cases
> at that point.
>
> OK? Bootstrapped and tested on x86_64-linux-gnu with no regressions.
OK.
> PR tree-optimization/111469
>
> gcc/ChangeLog:
>
> * tree-ssa-phiopt.cc (minmax_replacement): Fix
> the assumption for the `non-diamond` handling cases
> of diamond code.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.c-torture/execute/pr111469-1.c: New test.
> ---
> .../gcc.c-torture/execute/pr111469-1.c | 38 +++++++++++++++++++
> gcc/tree-ssa-phiopt.cc | 10 ++++-
> 2 files changed, 46 insertions(+), 2 deletions(-)
> create mode 100644 gcc/testsuite/gcc.c-torture/execute/pr111469-1.c
>
> diff --git a/gcc/testsuite/gcc.c-torture/execute/pr111469-1.c b/gcc/testsuite/gcc.c-torture/execute/pr111469-1.c
> new file mode 100644
> index 00000000000..b68d5989eac
> --- /dev/null
> +++ b/gcc/testsuite/gcc.c-torture/execute/pr111469-1.c
> @@ -0,0 +1,38 @@
> +/* PR tree-optimization/111469 */
> +
> +long f;
> +char *g;
> +__attribute__((noinline))
> +char o() {
> + char l;
> + while (f)
> + ;
> + l = *g;
> + return l;
> +}
> +
> +/* factor_out_conditional_conversion is able to remove the casts
> + from the 2 bbs (correctly)
> + but then minmax_replacement should not optimize this to a MIN_EXPR
> + as o has side effects. */
> +
> +__attribute__((noinline))
> +unsigned short gg(unsigned short a, unsigned short b)
> +{
> + short d;
> + if (a > b)
> + {
> + d= b;
> + }
> + else
> + {
> + o();
> + d = a;
> + }
> + return d;
> +}
> +
> +int main(void)
> +{
> + gg(3, 2);
> +}
> diff --git a/gcc/tree-ssa-phiopt.cc b/gcc/tree-ssa-phiopt.cc
> index 3835d25d08c..96901a40444 100644
> --- a/gcc/tree-ssa-phiopt.cc
> +++ b/gcc/tree-ssa-phiopt.cc
> @@ -1823,7 +1823,9 @@ minmax_replacement (basic_block cond_bb, basic_block middle_bb, basic_block alt_
> arg_false = arg0;
> }
>
> - if (empty_block_p (middle_bb))
> + if (empty_block_p (middle_bb)
> + && (!threeway_p
> + || empty_block_p (alt_middle_bb)))
> {
> if ((operand_equal_for_phi_arg_p (arg_true, smaller)
> || (alt_smaller
> @@ -2006,7 +2008,9 @@ minmax_replacement (basic_block cond_bb, basic_block middle_bb, basic_block alt_
>
> return true;
> }
> - else
> + else if (middle_bb == alt_middle_bb
> + && (!threeway_p
> + || empty_block_p (alt_middle_bb)))
> {
> /* Recognize the following case, assuming d <= u:
>
> @@ -2182,6 +2186,8 @@ minmax_replacement (basic_block cond_bb, basic_block middle_bb, basic_block alt_
> SSA_OP_DEF));
> gsi_move_before (&gsi_from, &gsi);
> }
> + else
> + return false;
>
> /* Emit the statement to compute min/max. */
> gimple_seq stmts = NULL;
> --
> 2.31.1
>
new file mode 100644
@@ -0,0 +1,38 @@
+/* PR tree-optimization/111469 */
+
+long f;
+char *g;
+__attribute__((noinline))
+char o() {
+ char l;
+ while (f)
+ ;
+ l = *g;
+ return l;
+}
+
+/* factor_out_conditional_conversion is able to remove the casts
+ from the 2 bbs (correctly)
+ but then minmax_replacement should not optimize this to a MIN_EXPR
+ as o has side effects. */
+
+__attribute__((noinline))
+unsigned short gg(unsigned short a, unsigned short b)
+{
+ short d;
+ if (a > b)
+ {
+ d= b;
+ }
+ else
+ {
+ o();
+ d = a;
+ }
+ return d;
+}
+
+int main(void)
+{
+ gg(3, 2);
+}
@@ -1823,7 +1823,9 @@ minmax_replacement (basic_block cond_bb, basic_block middle_bb, basic_block alt_
arg_false = arg0;
}
- if (empty_block_p (middle_bb))
+ if (empty_block_p (middle_bb)
+ && (!threeway_p
+ || empty_block_p (alt_middle_bb)))
{
if ((operand_equal_for_phi_arg_p (arg_true, smaller)
|| (alt_smaller
@@ -2006,7 +2008,9 @@ minmax_replacement (basic_block cond_bb, basic_block middle_bb, basic_block alt_
return true;
}
- else
+ else if (middle_bb == alt_middle_bb
+ && (!threeway_p
+ || empty_block_p (alt_middle_bb)))
{
/* Recognize the following case, assuming d <= u:
@@ -2182,6 +2186,8 @@ minmax_replacement (basic_block cond_bb, basic_block middle_bb, basic_block alt_
SSA_OP_DEF));
gsi_move_before (&gsi_from, &gsi);
}
+ else
+ return false;
/* Emit the statement to compute min/max. */
gimple_seq stmts = NULL;