https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115186
Bug ID: 115186
Summary: Suboptimal codes generated by rtl-expand for divmod
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: dizhao at os dot amperecomputing.com
Target Milestone: ---
For the code below:
typedef unsigned short int uint16_t;
typedef unsigned char uint8_t;
typedef long unsigned int size_t;
uint16_t fletcher16(const uint8_t *data, const size_t len)
{
register uint16_t sum1 = 0, sum2 = 0;
register size_t i;
for (i = 0; i < len; i++) {
sum1 = (sum1 + data[i]) % 255;
sum2 = (sum2 + sum1) % 255;
}
return ((uint16_t)(sum2 << 8)) | sum1;
}
, when compiled with "-O3 -mcpu=ampere1a", shift+add is used instead of div.
The dump file at 266r.expand shows:
20: NOTE_INSN_BASIC_BLOCK 4
21: debug i => r114:DI-r116:DI
22: debug sum2 => r111:SI#0
23: debug sum1 => r110:SI#0
24: debug begin stmt marker
25: r118:SI=zero_extend([r114:DI])
26: r119:SI=r118:SI+r110:SI
27: r120:DI=zero_extend(r119:SI)
28: r121:DI=r120:DI
29: r122:DI=r121:DI<<0x8
30: r123:DI=r122:DI+r120:DI
REG_EQUAL r120:DI*0x101
31: r124:DI=r123:DI<<0x10
32: r125:DI=r123:DI+r124:DI
REG_EQUAL r120:DI*0x1010101
33: r126:DI=r125:DI<<0x7
34: r127:DI=r126:DI+r120:DI
REG_EQUAL r120:DI*0x80808081
35: r128:DI=r127:DI 0>>0x20
36: r104:SI=r128:DI#0 0>>0x7
REG_EQUAL udiv(r119:SI,0xff)
...
However, using mult instruction is better, like (the result can be produce with
"-mtune=neoverse-n1"):
20: NOTE_INSN_BASIC_BLOCK 4
21: debug i => r114:DI-r116:DI
22: debug sum2 => r111:SI#0
23: debug sum1 => r110:SI#0
24: debug begin stmt marker
25: r118:SI=zero_extend([r114:DI])
26: r119:SI=r118:SI+r110:SI
27: r121:SI=0xffffffff80808081
28: r120:DI=zero_extend(r119:SI)*zero_extend(r121:SI)
29: r122:DI=r120:DI 0>>0x20
30: r104:SI=r122:DI#0 0>>0x7
REG_EQUAL udiv(r119:SI,0xff)
...
The problem is, in expmed.cc:expmed_mult_highpart, when the result here is 0:
/* See whether the specialized multiplication optabs are
cheaper than the shift/add version. */
tem = expmed_mult_highpart_optab (mode, op0, narrow_op1, target,
unsignedp,
, the "tem" produced by the code that follows (shift/add version) can be more
expensive than the result of:
return expmed_mult_highpart_optab (mode, op0, op1, target, unsignedp,
max_cost);
For -mcpu=ampere1a, the estimated cost of the former is 36, the cost of the
latter is 28.
