On 7/6/24 07:20, Jeff Law wrote:
On 7/3/24 3:16 PM, Patrick O'Neill wrote:
Regarding the amocas.q follow-up patch:
I'm having trouble with matching any TImode compare-and-swap
patterns. Here's the RTL I'm trying:
(define_mode_iterator SUPERGPR [SI DI TI])
(define_insn "zacas_atomic_cas_value<mode>"
[(set (match_operand:SUPERGPR 0 "register_operand" "=&r")
(match_operand:SUPERGPR 1 "memory_operand" "+A"))
(set (match_dup 1)
(unspec_volatile:SUPERGPR [(match_operand:SUPERGPR 2
"reg_or_0_operand" "rJ")
(match_operand:SUPERGPR 3 "reg_or_0_operand" "rJ")
(match_operand:SI 4 "const_int_operand") ;; mod_s
(match_operand:SI 5 "const_int_operand")] ;; mod_f
I'd check if it's the mode on operands 4 and 5. CONST_INTs do not
have modes.
I gave it a shot with no dice. Same behavior where SI/DI mode match but
TImode falls over.
https://gcc.gnu.org/onlinedocs/gccint/Machine-Independent-Predicates.html
| Function: const_int_operand
| This predicate allows any CONST_INT expression that fits in mode.
It is an appropriate choice for an immediate operand that does not allow
a symbol or label.
If I'm reading this right the const_int mode constrains the const_int to
a given mode.
Not super relevant here since mod_s/mod_f is a 6 state enum but it looks
like const_int is at least aware of the mode?
{
emit_insn (gen_zacas_atomic_cas_value<mode> (operands[1],
operands[2],
operands[3], operands[4],
operands[6], operands[7]));
Don't you need to pass operands[1] twice? Once for operands[0] in the
final pattern, then again for operands[1] in that final pattern?
I don't think so - operands[1] is a register while operands[2] is the
memory location.
expand pattern insn pattern
operands[1] register -> operands[0] register "=&r"
operands[2] memory -> operands[1] memory "+A"
operands[3] reg_0 -> operands[2] reg_0 "rJ"
operands[4] reg_0 -> operands[3] reg_0 "rJ"
operands[6] const_int -> operands[4] const_int
operands[7] const_int -> operands[5] const_int
Thanks for taking a look - I'll spend some time digging into
insn-recog.cc to get to the bottom of this.
Patrick
jeff
