Every nir_ssa_def is part of a chain of uses, implemented with doubly linked
lists. That means each requires 2 * 64-bit = 16 bytes per def, which is
memory intensive. Together they require 32 bytes per def. Not cool.
To cut that memory use in half, we can combine the two linked lists into a
single use list that contains both regular instruction uses and if-uses. To do
this, we augment the nir_src with a boolean "is_if", and reimplement the
abstract if-uses operations on top of that list. That boolean should fit into
the padding already in nir_src so should not actually affect memory use, and in
the future we sneak it into the bottom bit of a pointer.
However, this creates a new inefficiency: now iterating over regular uses
separate from if-uses is (nominally) more expensive. It turns out virtually
every caller of nir_foreach_if_use(_safe) also calls nir_foreach_use(_safe)
immediately before, so we rewrite most of the callers to instead call a new
single `nir_foreach_use_including_if(_safe)` which predicates the logic based on
`src->is_if`. This should mitigate the performance difference.
There's a bit of churn, but this is largely a mechanical set of changes.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Faith Ekstrand <faith.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22343>
On some architectures, Boolean values used to control conditional
branches or condtional selection must be propagated into a flag. This
generally means that a stored Boolean value must be compared with zero.
Rather than force the generation of extra compares with zero, re-emit
the original comparison instruction. This can save register pressure by
not needing to store the Boolean value.
There are several possible ares for future improvement to this pass:
1. Be more conservative. If both sources to the comparison instruction
are non-constants, it may be better for register pressure to emit the
extra compare. The current shader-db results on Intel GPUs (next
commit) lead me to believe that this is not currently a problem.
2. Be less conservative. Currently the pass requires that all users of
the comparison match the pattern. The idea is that after the pass is
complete, no instruction will use the resulting Boolean value. The only
uses will be of the flag value. It may be beneficial to relax this
requirement in some cases.
3. Be less conservative. Also try to rematerialize comparisons used for
discard_if intrinsics. After changing the way the Intel compiler
generates cod e for discard_if (see MR!935), I tried implementing this
already. The changes were pretty small. Instructions were helped in 19
shaders, but, overall, cycles were hurt. A commit "nir: Rematerialize
comparisons for nir_intrinsic_discard_if too" is on my fd.o cgit.
4. Copy the preceeding ALU instruction. If the comparison is a
comparison with zero, and it is the only user of a particular ALU
instruction (e.g., (a+b) != 0.0), it may be a further improvment to also
copy the preceeding ALU instruction. On Intel GPUs, this may enable
cmod propagation to make additional progress.
v2: Use much simpler method to get the prev_block for an if-statement.
Suggested by Tim.
Reviewed-by: Matt Turner <mattst88@gmail.com>
