From: Pan Li <pan2...@intel.com>
There are sorts of forms for the unsigned SAT_ADD. Some of them are
complicated while others are cheap. This patch would like to simplify
the complicated form into the cheap ones. For example as below:
>From the form 4 (branch):
SUM = ADD_OVERFLOW (X, Y)
SAT_U_ADD = IMAGPART_EXPR (SUM) == 0 ? REALPART_EXPR (SUM) : -1
To (branchless):
SAT_U_ADD = (X + Y) | - ((X + Y) < X).
#define T uint8_t
T sat_add_u_1 (T x, T y)
{
T ret;
return __builtin_add_overflow (x, y, &ret) == 0 ? ret : - 1;
}
Before this patch in phiopt2:
4 │ uint8_t sat_u_add_uint8_t_9 (uint8_t x, uint8_t y)
5 │ {
6 │ unsigned char _1;
7 │ unsigned char _2;
8 │ uint8_t _3;
9 │ __complex__ unsigned char _6;
10 │
11 │ <bb 2> [local count: 1073741824]:
12 │ _6 = .ADD_OVERFLOW (x_4(D), y_5(D));
13 │ _1 = REALPART_EXPR <_6>;
14 │ _2 = IMAGPART_EXPR <_6>;
15 │ if (_2 != 0)
16 │ goto <bb 3>; [35.00%]
17 │ else
18 │ goto <bb 4>; [65.00%]
19 │
20 │ <bb 3> [local count: 375809640]:
21 │
22 │ <bb 4> [local count: 1073741824]:
23 │ # _3 = PHI <255(3), _1(2)>
24 │ return _3;
25 │
26 │ }
After this patch
14 │ uint8_t sat_u_add_uint8_t_9 (uint8_t x, uint8_t y)
15 │ {
16 │ unsigned char _1;
17 │ __complex__ unsigned char _6;
18 │ unsigned char _8;
19 │ _Bool _9;
20 │ unsigned char _10;
21 │ unsigned char _11;
22 │ unsigned char _12;
23 │
24 │ <bb 2> [local count: 1073741824]:
25 │ _6 = .ADD_OVERFLOW (x_4(D), y_5(D)); // Dead code
26 │ _1 = REALPART_EXPR <_6>; // Ditto
27 │ _8 = x_4(D) + y_5(D);
28 │ _9 = x_4(D) > _8;
29 │ _10 = (unsigned char) _9;
30 │ _11 = -_10;
31 │ _12 = _8 | _11;
32 │ return _12;
33 │
34 │ }
This patch also add 2 more imm matching for unsigned SAT_ADD due to
remove the add_overflow match.
The below test suites are passed for this patch.
* The rv64gcv fully regression test.
* The x86 bootstrap test.
* The x86 fully regression test.
gcc/ChangeLog:
* match.pd: Remove unsigned branch form 4 and simplify to the
branchless of unsigned SAT_ADD, add 2 new form for imm.
Signed-off-by: Pan Li <pan2...@intel.com>
---
gcc/match.pd | 45 +++++++++++++++++++++++++++++++++++++--------
1 file changed, 37 insertions(+), 8 deletions(-)
diff --git a/gcc/match.pd b/gcc/match.pd
index f6f55491b89..8ed08b95bc0 100644
--- a/gcc/match.pd
+++ b/gcc/match.pd
@@ -3160,14 +3160,13 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
/* From SAT_U_ADD = (X + Y) < x ? -1 : (X + Y). */
(simplify (cond (lt (plus:c@2 @0 @1) @0) integer_minus_onep @2)
(if (types_match (type, @0, @1))
- (bit_ior (plus@2 @0 @1) (negate (convert (lt @2 @0)))))))
-
-/* Unsigned saturation add, case 5 (branch with eq .ADD_OVERFLOW):
- SUM = ADD_OVERFLOW (X, Y)
- SAT_U_ADD = IMAGPART_EXPR (SUM) == 0 ? REALPART_EXPR (SUM) : -1. */
-(match (unsigned_integer_sat_add @0 @1)
- (cond^ (eq (imagpart (IFN_ADD_OVERFLOW:c @0 @1)) integer_zerop)
- (usadd_left_part_2 @0 @1) integer_minus_onep))
+ (bit_ior (plus@2 @0 @1) (negate (convert (lt @2 @0))))))
+ /* From SUM = ADD_OVERFLOW (X, Y)
+ SAT_U_ADD = IMAGPART_EXPR (SUM) == 0 ? REALPART_EXPR (SUM) : -1 */
+ (simplify (cond (eq (imagpart (IFN_ADD_OVERFLOW:c@2 @0 @1)) integer_zerop)
+ (realpart @2) integer_minus_onep)
+ (if (types_match (type, @0, @1))
+ (bit_ior (plus@3 @0 @1) (negate (convert (lt @3 @0)))))))
/* Unsigned saturation add, case 6 (branch with ne .ADD_OVERFLOW):
SUM = ADD_OVERFLOW (X, Y)
@@ -3201,6 +3200,36 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
(if (INTEGRAL_TYPE_P (type) && TYPE_UNSIGNED (type)
&& types_match (type, @0) && int_fits_type_p (@1, type))))
+/* Unsigned saturation add, case 11 (one op is imm):
+ SAT_U_ADD = (X + 3) | -(X > (umax - 3)). */
+(match (unsigned_integer_sat_add @0 @1)
+ (bit_ior:c (plus:c @0 INTEGER_CST@1) (negate (convert (gt @0 INTEGER_CST@2))))
+ (if (INTEGRAL_TYPE_P (type) && TYPE_UNSIGNED (type)
+ && types_match (type, @0, @1))
+ (with
+ {
+ unsigned precision = TYPE_PRECISION (type);
+ wide_int cst_1 = wi::to_wide (@1);
+ wide_int cst_2 = wi::to_wide (@2);
+ wide_int max = wi::mask (precision, false, precision);
+ wide_int sum = wi::add (cst_1, cst_2);
+ }
+ (if (wi::eq_p (max, sum))))))
+
+/* Unsigned saturation add, case 12 (one op is imm). Almost the same form
+ as the case 11 but the gimple changes when INTEGER_CST@1 is umax. */
+(match (unsigned_integer_sat_add @0 @1)
+ (bit_ior:c (plus:c @0 INTEGER_CST@1) (negate (convert (ne @0 integer_zerop))))
+ (if (INTEGRAL_TYPE_P (type) && TYPE_UNSIGNED (type)
+ && types_match (type, @0, @1))
+ (with
+ {
+ unsigned precision = TYPE_PRECISION (type);
+ wide_int cst_1 = wi::to_wide (@1);
+ wide_int max = wi::mask (precision, false, precision);
+ }
+ (if (wi::eq_p (max, cst_1))))))
+
/* Signed saturation add, case 1:
T sum = (T)((UT)X + (UT)Y)
SAT_S_ADD = (X ^ sum) & !(X ^ Y) < 0 ? (-(T)(X < 0) ^ MAX) : sum;
--
2.43.0
