Vector-aligned operands are handled by temporarily allocating
a vector-SSA value for the duration of the instruction.
On completion of the register assignment, the individual
operands are assigned to the reserved register space and,
if necessary, parallelcopies are emitted.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34359>
The old version worked by precoloring both the operand and definition to
whatever register the operand was at the time. This didn't allow moving
the operand/definition after precoloring them, or two tied operands with
the same temporary.
The new version works by temporarily making the operands late kill and
creating a copy if the temporary is live-through or used as another tied
operand. Then we can simply later make the operands early kill again and
assign the definitions to the operand's register.
This way, we can move the operand to make space and the new location will
not intersect with any other definition and won't cause the operand and
definition registers to mismatch.
fossil-db (gfx1201):
Totals from 2253 (2.84% of 79377) affected shaders:
Instrs: 1634747 -> 1630799 (-0.24%); split: -0.27%, +0.03%
CodeSize: 8688148 -> 8672348 (-0.18%); split: -0.20%, +0.02%
VGPRs: 106500 -> 106512 (+0.01%)
Latency: 11385480 -> 11382965 (-0.02%); split: -0.04%, +0.01%
InvThroughput: 1754430 -> 1754326 (-0.01%); split: -0.01%, +0.00%
SClause: 38954 -> 38964 (+0.03%); split: -0.01%, +0.04%
Copies: 110772 -> 110800 (+0.03%); split: -0.02%, +0.04%
Branches: 29093 -> 29092 (-0.00%)
VALU: 902011 -> 902008 (-0.00%)
SALU: 260175 -> 260203 (+0.01%); split: -0.01%, +0.02%
fossil-db (navi31):
Totals from 2 (0.00% of 79377) affected shaders:
Latency: 1766 -> 1765 (-0.06%)
InvThroughput: 3219 -> 3215 (-0.12%)
fossil-db (navi21):
Totals from 14 (0.02% of 79377) affected shaders:
(no affected stats)
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34700>
We have to handle tied definitions after optimize_encoding, but a rewrite
of that will need the register file from before clearing the killed
operands.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34700>
This probably doesn't fix anything in practice. I don't think this path is
ever taken for SGPRs except for pseudo-scalar transcendental instructions,
and those don't have any precolored operands/definitions.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34343>
Aligning these can create situations where register allocation is
impossible. Only pseudo-instructions can use these, and they don't require
any alignment.
I'm not sure if these temporaries actually happen in practice. This
probably only affects the get_reg()'s compact_relocate_vars fallback path,
which doesn't usually happen for SGPRs.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34343>
D16 MIMG or pseudo-scalar transcendental instructions might give lower
limits to the maximum register available for their definitions, so just
try to place them earlier.
This is also part of fixing compact_relocate_vars with aligned NPOT
def/killed-op space (the second part is the later commit which changes
get_stride()).
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34343>
The register file tests here should be done after update_renames().
Normally, get_reg() wouldn't have to move anything to make space for a 1-3
byte definition. This was encountered with skip_optimistic_path=true and a
get_reg_impl() bug (fixed in a later commit) which caused suboptimal
register assignment.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34679>
info.bounds might be smaller than the bounds available for the moved
variables.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Fixes: 626aa7b648 ("aco: workaround GFX9 hardware bug for D16 image instructions")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34158>
This lets us remove some special casing from update_renames and make it
simpler. Doing this in update_renames was also fragile, since the
parallelcopies vector is not final when update_renames is called, so it
might not have been safe to do so in the end.
No fossil-db changes.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33444>
ttmp sgprs are readonly outside of trap handlers, so the instructions were
probably skipped. RA should also never create additional exec writes.
Fixes: e06773281b ("aco/ra: Optimize some SOP2 instructions with literal to SOPK.")
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32545>
Also ensure that 'needs_scratch_reg' is always true if SCC might be overwritten.
Few changes, because some p_split_vector get SCC as scratch reg assigned,
and thus, can inhibit some postRA optimizations.
Totals from 3 (0.00% of 79395) affected shaders: (Navi31)
Instrs: 10501 -> 10500 (-0.01%); split: -0.02%, +0.01%
CodeSize: 51580 -> 51520 (-0.12%); split: -0.12%, +0.01%
Latency: 84166 -> 84174 (+0.01%)
InvThroughput: 13109 -> 13111 (+0.02%)
SALU: 859 -> 860 (+0.12%)
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32217>
