On Wed, May 22, 2024 at 3:17 AM <pan2...@intel.com> wrote:
>
> From: Pan Li <pan2...@intel.com>
>
> This patch would like to support the __builtin_add_overflow branch form for
> unsigned SAT_ADD. For example as below:
>
> uint64_t
> sat_add (uint64_t x, uint64_t y)
> {
> uint64_t ret;
> return __builtin_add_overflow (x, y, &ret) ? -1 : ret;
> }
>
> Different to the branchless version, we leverage the simplify to
> convert the branch version of SAT_ADD into branchless if and only
> if the backend has supported the IFN_SAT_ADD. Thus, the backend has
> the ability to choose branch or branchless implementation of .SAT_ADD.
> For example, some target can take care of branches code more optimally.
>
> When the target implement the IFN_SAT_ADD for unsigned and before this
> patch:
>
> uint64_t sat_add (uint64_t x, uint64_t y)
> {
> long unsigned int _1;
> long unsigned int _2;
> uint64_t _3;
> __complex__ long unsigned int _6;
>
> ;; basic block 2, loop depth 0
> ;; pred: ENTRY
> _6 = .ADD_OVERFLOW (x_4(D), y_5(D));
> _2 = IMAGPART_EXPR <_6>;
> if (_2 != 0)
> goto <bb 4>; [35.00%]
> else
> goto <bb 3>; [65.00%]
> ;; succ: 4
> ;; 3
>
> ;; basic block 3, loop depth 0
> ;; pred: 2
> _1 = REALPART_EXPR <_6>;
> ;; succ: 4
>
> ;; basic block 4, loop depth 0
> ;; pred: 3
> ;; 2
> # _3 = PHI <_1(3), 18446744073709551615(2)>
> return _3;
> ;; succ: EXIT
> }
>
> After this patch:
> uint64_t sat_add (uint64_t x, uint64_t y)
> {
> long unsigned int _12;
>
> ;; basic block 2, loop depth 0
> ;; pred: ENTRY
> _12 = .SAT_ADD (x_4(D), y_5(D)); [tail call]
> return _12;
> ;; succ: EXIT
> }
>
> The below test suites are passed for this patch:
> * The x86 bootstrap test.
> * The x86 fully regression test.
> * The riscv fully regression test.
I'm not convinced we should match this during early if-conversion, should we?
The middle-end doesn't really know .SAT_ADD but some handling of
.ADD_OVERFLOW is present.
But please add a comment before the new pattern, esp. since it's
non-obvious that this is an improvent.
I suspect you rely on this form being recognized as .SAT_ADD later but
what prevents us from breaking this? Why not convert it to .SAT_ADD
immediately? If this is because the ISEL pass (or the widen-mult pass)
cannot handle PHIs then I would suggest to split out enough parts of
tree-ssa-phiopt.cc to be able to query match.pd for COND_EXPRs.
> gcc/ChangeLog:
>
> * match.pd: Add new simplify to convert branch SAT_ADD into
> branchless, if and only if backend implement the IFN.
>
> Signed-off-by: Pan Li <pan2...@intel.com>
> ---
> gcc/match.pd | 11 +++++++++++
> 1 file changed, 11 insertions(+)
>
> diff --git a/gcc/match.pd b/gcc/match.pd
> index cff67c84498..2dc77a46e67 100644
> --- a/gcc/match.pd
> +++ b/gcc/match.pd
> @@ -3080,6 +3080,17 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> (match (unsigned_integer_sat_add @0 @1)
> (bit_ior:c (usadd_left_part_2 @0 @1) (usadd_right_part_2 @0 @1)))
>
> +#if GIMPLE
> +
> +(simplify
> + (cond (ne (imagpart (IFN_ADD_OVERFLOW@2 @0 @1)) integer_zerop)
> + integer_minus_onep (realpart @2))
> + (if (ternary_integer_types_match_p (type, @0, @1) && TYPE_UNSIGNED (type)
> + && direct_internal_fn_supported_p (IFN_SAT_ADD, type,
> OPTIMIZE_FOR_BOTH))
> + (bit_ior (plus@3 @0 @1) (negate (convert (lt @3 @0))))))
> +
> +#endif
> +
> /* x > y && x != XXX_MIN --> x > y
> x > y && x == XXX_MIN --> false . */
> (for eqne (eq ne)
> --
> 2.34.1
>
