On Thu, 10 Nov 2022, Philipp Tomsich wrote:
> Our ifcombine pass combines 2 single-bit tests into a single test of
> the form "(a & T) == T".
>
> However, detection of the bit-tests does not work for sign-bits (of
> any mode) as the RTL presented to the pass uses a cast to a
> signed-type and an order-operator. E.g., the test for 'a & 0x80'
> presents as:
> _1 = *a_5(D);
> _2 = (signed char) _1;
> if (_2 < 0)
> goto <bb 3>;
> else
> goto <bb 5>;
>
> This adds detection logic to recognize_single_bit_test() for this case
> and reports it as a single-bit test.
>
> gcc/ChangeLog:
>
> * tree-ssa-ifcombine.cc (recognize_single_bit_test): Add
> detection for tests against the sign-bit of the relevant
> type as a single-bit test.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.dg/tree-ssa/ssa-ifcombine-14.c: New test.
>
> Signed-off-by: Philipp Tomsich <philipp.toms...@vrull.eu>
> ---
>
> .../gcc.dg/tree-ssa/ssa-ifcombine-14.c | 14 +++++++++++++
> gcc/tree-ssa-ifcombine.cc | 20 +++++++++++++++++++
> 2 files changed, 34 insertions(+)
> create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/ssa-ifcombine-14.c
>
> diff --git a/gcc/testsuite/gcc.dg/tree-ssa/ssa-ifcombine-14.c
> b/gcc/testsuite/gcc.dg/tree-ssa/ssa-ifcombine-14.c
> new file mode 100644
> index 00000000000..f9b1115a34a
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/tree-ssa/ssa-ifcombine-14.c
> @@ -0,0 +1,14 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O -fdump-tree-ifcombine-details-blocks" } */
> +
> +void sink();
> +
> +void same(unsigned char *a)
> +{
> + if (*a & 0x80)
> + if (*a & 0x40)
> + g();
> +}
> +
> +/* { dg-final { scan-tree-dump "optimizing double bit test" } } */
> +
> diff --git a/gcc/tree-ssa-ifcombine.cc b/gcc/tree-ssa-ifcombine.cc
> index 80c41c45489..cd6331f84db 100644
> --- a/gcc/tree-ssa-ifcombine.cc
> +++ b/gcc/tree-ssa-ifcombine.cc
> @@ -225,6 +225,26 @@ recognize_single_bit_test (gcond *cond, tree *name, tree
> *bit, bool inv)
> {
> gimple *stmt;
>
> + /* Handle the test for a sign-bit:
> + unsigned charD.15 _1;
> + _2 = (signed char) _1;
> + if (_2 < 0) */
> + if (TREE_CODE (gimple_cond_lhs (cond)) == SSA_NAME
> + && !TYPE_UNSIGNED (TREE_TYPE (gimple_cond_lhs (cond)))
> + && gimple_cond_code (cond) == (inv ? GE_EXPR : LT_EXPR)
> + && integer_zerop (gimple_cond_rhs (cond)))
Can you please keep the common early out
|| TREE_CODE (gimple_cond_lhs (cond)) != SSA_NAME
|| !integer_zerop (gimple_cond_rhs (cond)))
and the common
stmt = SSA_NAME_DEF_STMT (gimple_cond_lhs (cond));
if (!is_gimple_assign (stmt))
return false;
and instead add the sign bit case to the list of cases below? I'd
suggest assigning gimple_cond_code to a temporary and amending the
individual tests
> + {
> + tree type = TREE_TYPE (gimple_cond_lhs (cond));
> +
> + stmt = SSA_NAME_DEF_STMT (gimple_cond_lhs (cond));
> + if (!is_gimple_assign (stmt))
> + return false;
> + *name = gimple_assign_rhs1 (stmt);
> + *bit = build_int_cst (integer_type_node, TYPE_PRECISION (type) - 1);
> +
> + return true;
> + }
> +
> /* Get at the definition of the result of the bit test. */
> if (gimple_cond_code (cond) != (inv ? EQ_EXPR : NE_EXPR)
> || TREE_CODE (gimple_cond_lhs (cond)) != SSA_NAME
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg,
Germany; GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman;
HRB 36809 (AG Nuernberg)
