There seems to be a hardware bug that sometimes causes a GPU hang when
an alias...sam sequence crosses an instruction cache line boundary. This
commit adds a workaround pass that inserts padding nops to ensure no
such sequence cross a cache line. Until an alternative solution is
found, this is the best we can do.
While the number of nops we have to insert is fixed at this point, we
can try to minimize the number of nops executed at runtime by replacing
nops encoded in instructions by standalone nops. That is, if this pass
has to insert one nop, it will try to make one of the following
replacements:
- (rptN)nop -> (rptN-1)nop; nop
- (nopN)foo -> (nopN-1)foo; nop
It does so by keeping track of "insert points". Each insert point keeps
track of the instruction and the maximum number of nops that can be
inserted there without pushing any subsequent alias sequences over the
next cache line. Whenever we need to insert nops, we first try it at the
encountered insert points and only if that doesn't work, we insert them
right before the first alias. The pass makes sure the insert points are
only visited a bounded number of times in total to keep the whole pass
O(n).
Totals:
Instrs: 48207402 -> 48278230 (+0.15%)
CodeSize: 101907026 -> 102294524 (+0.38%)
NOPs: 8386320 -> 8457148 (+0.84%)
(ss)-stall: 4013046 -> 4012931 (-0.00%)
(sy)-stall: 16741190 -> 16741033 (-0.00%)
Preamble Instrs: 11506988 -> 11520671 (+0.12%)
Last helper: 11686328 -> 11701615 (+0.13%)
Cat0: 9241457 -> 9312285 (+0.77%)
Totals from 25237 (15.32% of 164705) affected shaders:
Instrs: 22172360 -> 22243188 (+0.32%)
CodeSize: 44372164 -> 44759662 (+0.87%)
NOPs: 4201698 -> 4272526 (+1.69%)
(ss)-stall: 1982473 -> 1982358 (-0.01%)
(sy)-stall: 7379552 -> 7379395 (-0.00%)
Preamble Instrs: 4552074 -> 4565757 (+0.30%)
Last helper: 6260280 -> 6275567 (+0.24%)
Cat0: 4616677 -> 4687505 (+1.53%)
Signed-off-by: Job Noorman <jnoorman@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36639>
Job Noorman
a1278095d3