https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108038
Bug ID: 108038
Summary: GCC generates poor code for select of consecutive
constants
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: law at gcc dot gnu.org
CC: rzinsly at ventanamicro dot com
Target Milestone: ---
Target: riscv-64
This testcase:
typedef signed long int __int64_t;
#define X(OP, OPN) __int64_t test##OPN(__int64_t x, __int64_t y) { __int64_t
cmp; cmp = (x OP y) ? 2 : 3; return cmp; }
X(>,GT)
X(<,LT)
typedef unsigned long int __uint64_t;
#define XU(OP, OPN) __uint64_t test##OPN(__uint64_t x, __uint64_t y) {
__uint64_t cmp; cmp = (x OP y) ? 2 : 3; return cmp; }
X(>,GTU)
X(<,LTU)
Compiled for rv64 with -O2 -march=rv64gc_zba gets this code for the first test.
The others are similar in having the extraneous xori.
testGT:
sgt a0,a0,a1 # 25 [c=4 l=4] *sgt_didi
xori a0,a0,1 # 27 [c=4 l=4] xordi3/1
addi a0,a0,2 # 16 [c=4 l=4] adddi3/1
ret # 35 [c=0 l=4] simple_return
But ISTM we could generate this instead:
testGT:
sgt a0,a1,a0 # 24 [c=4 l=4] *riscv.md:2535
addi a0,a0,2 # 16 [c=4 l=4] adddi3/1
ret # 32 [c=0 l=4] simple_return
I think the key is this:
Trying 25 -> 27:
25: r139:DI=r140:DI>r141:DI
REG_DEAD r141:DI
REG_DEAD r140:DI
27: r134:DI=r139:DI^0x1
REG_DEAD r139:DI
Failed to match this instruction:
(set (reg:DI 134 [ <retval> ])
(le:DI (reg:DI 140)
(reg:DI 141)))
We can invert the condition and swap the operands with something like this:
(define_insn ""
[(set (match_operand:GPR 0 "register_operand" "=r")
(any_le:GPR (match_operand:X 1 "register_operand" " r")
(match_operand:X 2 "register_operand" "r")))]
""
"sgt<u>\t%0,%2,%1"
[(set_attr "type" "slt")
(set_attr "mode" "<X:MODE>")])
We probably need another pattern to handle the LT/LTU cases which want to
match:
rying 24 -> 25:
24: r139:DI=r140:DI<r141:DI
REG_DEAD r141:DI
REG_DEAD r140:DI
25: r138:SI=r139:DI==0
REG_DEAD r139:DI
Failed to match this instruction:
(set (reg:SI 138)
(ge:SI (reg:DI 140)
(reg:DI 141)))
Again, I think by inverting the condition and swapping the operands we should
be good to go.