On Sat, 12 Dec 2020, Jakub Jelinek via Gcc-patches wrote:
> This patch adds the ~(X - Y) -> ~X + Y simplification requested
> in the PR (plus also ~(X + C) -> ~X + (-C) for constants C that can
> be safely negated.
This regresses VAX code produced by the cmpelim-eq-notsi.c test case (and
its similar counterparts) with the `vax-netbsdelf' target.
Previously this assembly:
.text
.align 1
.globl eq_notsi
.type eq_notsi, @function
eq_notsi:
.word 0 # 35 [c=0] procedure_entry_mask
subl2 $4,%sp # 46 [c=32] *addsi3
mcoml 4(%ap),%r0 # 32 [c=16] *one_cmplsi2_ccz
jeql .L1 # 34 [c=26] *branch_ccz
addl2 $2,%r0 # 31 [c=32] *addsi3
.L1:
ret # 40 [c=0] return
.size eq_notsi, .-eq_notsi
was produced. Now this:
.text
.align 1
.globl eq_notsi
.type eq_notsi, @function
eq_notsi:
.word 0 # 36 [c=0] procedure_entry_mask
subl2 $4,%sp # 48 [c=32] *addsi3
movl 4(%ap),%r0 # 33 [c=16] *movsi_2
cmpl %r0,$-1 # 34 [c=8] *cmpsi_ccz/1
jeql .L3 # 35 [c=26] *branch_ccz
subl3 %r0,$1,%r0 # 32 [c=32] *subsi3/1
ret # 27 [c=0] return
.L3:
clrl %r0 # 31 [c=2] *movsi_2
ret # 41 [c=0] return
.size eq_notsi, .-eq_notsi
is, which is clearly worse, both in terms of performance and size.
The key here is that the cost of constant 0, here used with a comparison
operation eliminated after MCOML in the former assembly sequence, is lower
(as per `vax_rtx_costs') in the VAX ISA than the cost of constant -1, used
with CMPL in the latter sequence. Not only constant 0 is an implied
operand with some machine instructions saving cycles and space otherwise
used for an explicitly encoded operand, but if used with a comparison
operation it can usually be eliminated, so it should be preferred over all
other constants.
With the example you gave with the PR I can see progression with f3, f4,
f7, f8, regression with f1, f2, and no change in operation cost with f5,
f6.
Shouldn't a transformation like this respect target-specific expression
costs somehow then? Depending on the individual case one form or the
other might be cheaper, and somehow we assume here both are equivalent in
terms of performance and/or code size (as applicable for the optimisation
mode chosen).
Maciej
