https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83253
--- Comment #12 from Bill Schmidt ---
By contrast, on powerpc64le, we see:
- mult_by_coeff_cost (3, E_SImode, true) returns a cost of 8
- mult_by_coeff_cost (4, E_SImode, true) returns a cost of 4
These are the sort of costs one would expect,
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83253
--- Comment #11 from Bill Schmidt ---
OK, for the i386 case, this simply comes down to the following.
- mult_by_coeff_cost (3, E_SImode, true) returns a cost of 4
- mult_by_coeff_cost (4, E_SImode, true) returns a cost of 8
Garbage in, garbag
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83253
--- Comment #10 from Bill Schmidt ---
I see that I was looking at the wrong leg here. This is a CAND_ADD, not a
CAND_MULT, and I'm getting strange cost results on that path. The proposed
change is still appropriate in my view, but not relevant
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83253
--- Comment #9 from Bill Schmidt ---
I was able to build an i386 cross, and this wasn't sufficient to solve the
problem. I see:
Processing dependency tree rooted at 1.
Inserting initializer: slsr_10 = scale_7(D) * 3;
Increment vector:
0 in
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83253
--- Comment #8 from Bill Schmidt ---
Revised patch:
Index: gcc/gimple-ssa-strength-reduction.c
===
--- gcc/gimple-ssa-strength-reduction.c (rev
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83253
--- Comment #7 from Bill Schmidt ---
Thanks, Jakub, good catch on the HWI overflow. I'll make those changes and
re-test. FWIW, the patch as proposed above passes regstrap and will be good
enough for testing whether this indeed is the right appr
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83253
--- Comment #6 from Jakub Jelinek ---
(In reply to Bill Schmidt from comment #5)
> Here's an untested patch (bootstraps successfully but regtest is still
> ongoing):
>
> Index: gcc/gimple-ssa-strength-reduction.c
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83253
--- Comment #5 from Bill Schmidt ---
Here's an untested patch (bootstraps successfully but regtest is still
ongoing):
Index: gcc/gimple-ssa-strength-reduction.c
==
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83253
--- Comment #4 from Bill Schmidt ---
I think the issue may be in this code:
/* For any other increment, if this is a multiply candidate, we
must introduce a temporary T and initialize it with
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83253
--- Comment #3 from Bill Schmidt ---
Apologies for not responding sooner -- I've been out of pocket for the last
week. I'll have a look when I get unburied (probably next week).
Bill
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83253
--- Comment #2 from Jakub Jelinek ---
Perhaps slsr should then if it is considering
Y = B + (i' * S)
X = B + (i * S)
to
Y = B + (i' * S)
X = Y + (i - i') * S
and if i and i' are INTEGER_CSTs call choose_mult_variant on both
i and (i - i') and see
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83253
Richard Biener changed:
What|Removed |Added
Keywords||missed-optimization
Status|
12 matches
Mail list logo
