https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70287
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |compile-time-hog
Status|UNCONFIRMED |NEW
Last reconfirmed| |2016-03-18
Summary|Slow compilation time |RTL LSM very slow with very
| |many loads/stores in a BB
Ever confirmed|0 |1
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Works fine with GCC 6 btw. Note that compile-time hog testcases where you
need to disable dead code removal are ... a bit fishy ;)
Note that tree optimizers optimize this to a single basic-block (everything
unrolled) but w/o all the scalar cleanups 90000 stmts remain in that
basic-block. With -ffloat-store another 180000 stores/loads are emitted
at RTL expansion time and GCSE SM barfs on this (well, it could specially
detect that in a single-basic-block function there is nothing to do ... ;))
But as said, I think the testcase is quite artificial though it has a point
in that it looks like LSM is awfully slow for basic-blocks with very many
loads/stores (well, not that surprising).
For GCC 6 I can't find enough options to disable, I always end up with just
<bb 2>:
var0_110 = -212204486;
ivtmp_111 = 9;
var0_116 = var0_110 + 1;
var0_121 = var0_116 + 1;
var0_126 = var0_121 + 1;
var0_133 = var0_126 + 1;
var0_137 = var0_133 + 1;
var0_142 = var0_137 + 1;
var0_149 = var0_142 + 1;
var0_154 = var0_149 + 1;
ivtmp_155 = ivtmp_111 + 4294967288;
var0_37 = var0_154 + 1;
ivtmp_18 = ivtmp_155 + 4294967295;
return 0;
before RTL expansion (tried to disable forwprop, DOM and VRP).
So, kind of "confirmed", but the artificial testcase should be rewritten
to not require all those disabled opts but directly have a very large
single BB with many loads/stores (or better at least some LSM opportunities)
that are _not_ dead.
