Hi all, While working with conditional execution and in particular with define_cond_exec, I would like to add an attribute to cond_exec patterns that are produced through define_cond_exec. This is done by adding an attributes vector to the define_cond_exec construct. When a pattern is then processed through define_cond_exec, the attributes vector is appended to the existing attributes. This would allow us to, for example, disable the predicable variant of an insn based on a non-constant variable.
For example, a setup like:
(define_attr "predicated" "yes,no" (const_string "no"))
(define_attr "control_attr" "yes,no" (const_string "yes"))
(define_attr "enabled" "no,yes"
(cond [(and (eq_attr "control_attr" "no")
(and (eq_attr "predicated" "yes")
(eq_attr "TARGET_RESTRICT_CE" "no")))
(const_string "no")]
(const_string "yes"))
(define_cond_exec
[(match_operator 0 "arm_comparison_operator"
[(match_operand 1 "cc_register" "")
(const_int 0)])]
""
""
[(set_attr "predicated" "yes")] // <<<< New field
)
(define_insn ...
[(set_attr "predicable" "yes")
(set_attr "control_attr" "yes,no")]) // <<<< disables predicated version of
the second alternative when TARGET_RESTRICT_CE is set
would let us disable the conditional variant of an insn by setting its
control_attr attribute to "no"
when the variable TARGET_RESTRICT_CE is true.
This new field for define_cond_exec is optional, and leaving it out should
not affect any backend that does not use it.
This would be a very useful feature for the arm backend to use, and without
it,
dynamically disabling conditional variants of insns
(i.e. dependent on a compiler flag)
involves a lot of pattern duplication and ugliness.
Patch has been bootstrapped on an arm-none-linux-gnueabihf target with
running on a Cortex-A15 and regtested.
Also, bootstrapped on x86_64-unknown-linux-gnu.
I've been working and testing on the arm backend, however I noticed that
ia64 and c6x use define_cond_exec as well,
so I'm cc'ing Steve. Would it be possible for you to confirm that this
doesn't break ia64? I compared the outputs of the gen*
programs with and without the patch and didn't see anything bad (the way
define_cond_exec is parsed changes due to the format change)
Any other feedback on the patch would, of course, be very appreciated.
Ok for trunk if no objections?
Thanks,
Kyrill
2013-05-20 Kyrylo Tkachov <kyrylo.tkac...@arm.com>
* rtl.def: Add extra third optional field to define_cond_exec.
* gensupport.c (process_one_cond_exec): Process attributes from
define_cond_exec.
* doc/md.texi: Document third field in define_cond_exec.
02-define_cond_exec.patch
Description: Binary data
