https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110089
Kewen Lin <linkw at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |segher at gcc dot gnu.org
--- Comment #7 from Kewen Lin <linkw at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #2)
> powerpc could use subf. and jgt? This is for the downward iterating IV for
> partial-vector-usage=2 loop vectorization.
Currently subf. exploitation only happens for CCmode and <MODE>mode == Pmode
(patterns *subf<mode>3_dot and *subf<mode>3_dot2), as the record form on all
fixed point insns perform "the first three bits of CR Field 0 are set by
**signed** comparison of the result to zero", it can't be used for CCUNSmode
and
<MODE>mode == Pmode. But it looks we can extend it for unsigned int on 64-bit.
Even if I did a hack to extend it with one *subf<mode>3_dot3 for UNSCC + SImode
+ Pmode==DImode, it still failed to match in combine, the pattern looks like:
(insn 24 23 25 4 (set (reg/v:DI 127 [ n ])
(reg/v:DI 131 [ n ])) 682 {*movdi_internal64}
(nil))
(insn 25 24 26 4 (set (reg:SI 142)
(minus:SI (subreg/s/v:SI (reg/v:DI 131 [ n ]) 0)
(subreg/s/v:SI (reg/v:DI 132 [ s ]) 0))) "test1.c":20:9 94
{*subfsi3}
(expr_list:REG_DEAD (reg/v:DI 131 [ n ])
(nil)))
(insn 26 25 28 4 (set (reg/v:DI 131 [ n ])
(zero_extend:DI (reg:SI 142))) "test1.c":20:9 16 {zero_extendsidi2}
(expr_list:REG_DEAD (reg:SI 142)
(nil)))
(insn 28 26 29 4 (set (reg:CCUNS 143)
(compare:CCUNS (subreg/s/v:SI (reg/v:DI 127 [ n ]) 0)
(subreg/s/v:SI (reg/v:DI 132 [ s ]) 0))) "test1.c":22:15 discrim
1 805 {*cmpsi_unsigned}
(expr_list:REG_DEAD (reg/v:DI 127 [ n ])
(nil)))
-------
At expand, we have
# n_9 = PHI <n_12(D)(2), n_19(3)>
# sum_10 = PHI <0(2), sum_18(3)>
_1 = MIN_EXPR <n_9, s_13(D)>;
len_14 = (int) _1;
_2 = (long unsigned int) len_14;
_3 = _2 * 4;
_4 = a_15(D) + _3;
_5 = *_4;
_6 = b_17(D) + _3;
_7 = *_6;
_8 = _5 + _7;
sum_18 = _8 + sum_10;
n_46 = n_9;
n_19 = n_9 - s_13(D); // different operand here and below,
if (n_46 > s_13(D)) // we don't need this n_46?
goto <bb 3>; [89.00%]
else
goto <bb 4>; [11.00%]
;; succ: 3
;; 4
---------
btw, I tried with a constant value for s like 16, it can exploit hardware loop
(bdnz) on Power:
#define TYPE unsigned int
#define MIN(a,b) ((a>b)?b:a)
int
foo (int *__restrict a, int *__restrict b, TYPE n)
{
if (n <= 0)
return 0;
int len;
int sum = 0;
TYPE oldn;
do
{
len = MIN (n, 16);
sum += a[len] + b[len];
oldn = n;
n = n - 16;
}
while (oldn > 16);
return sum;
}
.L3:
cmplwi 0,5,16
addi 8,5,-16
isel 9,7,5,1
rldicl 5,8,0,32
rldic 9,9,2,30
lwzx 8,3,9
lwzx 9,4,9
add 10,8,10
add 10,10,9
bdnz .L3
