The current implementation only moves the joinAt when splitting after
the given instruction, not before it. So if you have a BB with
foo
instr
bar
joinat
and thus with joinAt set, we end up first splitting before instr, at
which point the instr's bb is updated to the new bb. Since that bb
doesn't have a joinAt set (despite containing one), when splitting after
the instr, there is nothing to copy over. Since the joinat will be in
the "split" bb irrespective of whether we're splitting before or after
the instruction, move it over in either case.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=91124
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "10.5 10.6" <mesa-stable@lists.freedesktop.org>
This adds support for ARB_gpu_shader_fp64 and ARB_vertex_attrib_64bit to
llvmpipe.
Two things that don't mix well are SoA and doubles, see
emit_fetch_double, and emit_store_double_chan in this.
I've also had to split emit_data.chan, to add src_chan,
which can be different for doubles.
It handles indirect double fetches from temps, inputs, constants
and immediates. It doesn't handle double stores to indirects,
however it appears the mesa/st doesn't currently emit these,
it always does UARL/MOV combos, which will work fine.
tested with piglit, no regressions, all the fp64 tests seem to pass.
v2:
switch to using shuffles for fetch/store (Roland)
assert on indirect double stores - mesa/st never emits these (it uses MOV)
fix indirect temp/input/constant/immediates (Roland)
typos/formatting fixes (Roland)
v2.1:
cleanup some long lines, emit_store_double_chan cleanups.
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
As of now, the width field is no longer used for anything. The width field
"seemed like a good idea at the time" but is actually entirely redundant
with the instruction's execution size. Initially, it gave us the ability
to easily set the instructions execution size based entirely on register
widths. With the builder, we can easiliy set the sizes explicitly and the
width field doesn't have as much purpose. At this point, it's just
redundant information that can get out of sync so it really needs to go.
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Acked-by: Francisco Jerez <currojerez@riseup.net>
There are a variety of places where we use dst.width / 8 to compute the
size of a single logical channel. Instead, we should be using exec_size.
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Acked-by: Francisco Jerez <currojerez@riseup.net>
Now that all of the non-explicit constructors are gone, we don't need to
guess anymore.
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Acked-by: Francisco Jerez <currojerez@riseup.net>
Previously we used dst.width but the two *should* be the same.
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Acked-by: Francisco Jerez <currojerez@riseup.net>
Previously, we were just depending on register widths to ensure that
various things were exec_size of 1 etc. Now, we do so explicitly using the
builder.
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Acked-by: Francisco Jerez <currojerez@riseup.net>
Shortly, offset() will depend on the builder so we need it moved to some
place where it has access to that.
Reviewed-by: Iago Toral Quiroga <itoral@igali.com>
Acked-by: Francisco Jerez <currojerez@riseup.net>
Soon we will start using the builder to explicitly set all the execution
sizes. We could make a 32-wide builder, but the builder asserts that we
never grow it which is usually a reasonable assumption. Since this one
instruction is a bit of an odd-ball, we just set the exec_size explicitly.
v2: Explicitly new the fs_inst instead of using the builder and setting
exec_size after the fact.
v3: Set force_writemask_all with the builder instead of directly. The
builder over-writes it if we set it manually. Also, if we don't have
force_writemask_all in the builder it will assert-fail on SIMD32.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Acked-by: Francisco Jerez <currojerez@riseup.net>
Previously, fs_inst::regs_read() fell back to depending on the register
width for the second source. This isn't really correct since it isn't a
SIMD8 value at all, but a SIMD4x2 value. This commit changes it to
explicitly be always one register.
v2: Use mlen for determining the number of registers read
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Acked-by: Francisco Jerez <currojerez@riseup.net>
Previously, we were allocating the payload with different sizes per gen and
then figuring out the mlen in the generator based on gen. This meant,
among other things, that the higher level passes knew nothing about it.
Acked-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This makes things a little simpler, more efficient, and quite a bit more
readable.
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Reviewed-by: Topi Pohjolainen <topi.pohjolainen@intel.com>
Before, we would lazily emit a MOV whenever we encountered a use of a
constant. Now that we have a dedicated file for SSA values, we can
instead only emit the MOV's once, which is more consistent and prevents
us from relying on CSE to re-combine the constants when they aren't
absorbed into the instruction.
total instructions in shared programs: 6078991 -> 6073118 (-0.10%)
instructions in affected programs: 402221 -> 396348 (-1.46%)
helped: 1527
HURT: 0
GAINED: 8
LOST: 2
v2: split this out from the previous commit (Jason)
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Before, we would use registers, but set a magical "parent_instr" field
to indicate that it was actually purely an SSA value (i.e., it wasn't
involved in any phi nodes). Instead, just use SSA values directly, which
lets us get rid of the hack and reduces memory usage since we're not
allocating a nir_register for every value. It also makes our handling of
load_const more consistent compared to the other instructions.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
We already don't convert constants out of SSA, and in our backend we'd
like to have only one way of saying something is still in SSA.
The one tricky part about this is that we may now leave some undef
instructions around if they aren't part of a phi-web, so we have to be
more careful about deleting them.
v2: rename and flip meaning of flag (Jason)
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
0 is not used as a valid drawable id, as such there is no point in
attempting to query its geometry. Just bail out early and provide the
more meaningful EGL_BAD_NATIVE_WINDOW to the user.
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Raise EGL_BAD_NATIVE_WINDOW instead of crashing.
v2: s/Rise/Raise/ (spotted by Michel)
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Free the memory for dri2_surf in the unlikely case that one provides
NULL for native_window. Also set the relevant EGL_ERROR to provide
feedback to the user.
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Previously we were unconditionally doing ttn_get_src() even for
instructions with no src's. Which created a lot of unnecessary
load_const instructions. These were mostly harmless since NIR opt
passes would strip them back out. But for an ENDIF following a
BRK, it would result in load_const instructions created after the
NIR break instruction. Which nir_validate dislikes.
But we can actually just dtrt by using NumSrcRegs instead.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
It isn't quite yet practical to enable TGSI_ANY_INOUT_DECL_RANGE shader
cap yet, at least not in drivers that need lower_to_scalar pass (which
right now is all of the ttn users), since the register arrays do not get
converted to SSA, which angers nir_lower_alu_to_scalar.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
It is silly to traverse back to find first instruction that writes part
of a larger "virtual" register many times per instruction (plus per use
as a src to later instructions). Cache this information so we only
figure it out once.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
The fanin source could be grouped, for example with shaders like:
VERT
DCL IN[0]
DCL IN[1]
DCL OUT[0], POSITION
DCL OUT[1], GENERIC[9]
DCL SAMP[0]
DCL SVIEW[0], 2D, FLOAT
DCL TEMP[0], LOCAL
0: MOV TEMP[0].xy, IN[1].xyyy
1: MOV TEMP[0].w, IN[1].wwww
2: TXF TEMP[0], TEMP[0], SAMP[0], 2D
3: MOV OUT[1], TEMP[0]
4: MOV OUT[0], IN[0]
5: END
The second arg to the isaml is IN[1].w, so we need to look at the fanin
source to get the correct offset.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Split out most of dump_info() from ir3_cmdline compiler into a function
that can be used both by cmdline compiler and also for the disasm debug
option. This way, for FD_MESA_DEBUG=disasm we also get to see intput/
output registers, etc.
Signed-off-by: Rob Clark <robclark@freedesktop.org>
Some piglit tests, like arb_fragment_program-sparse-samplers, result in
having a null samp#0 but valid samp#1.
TODO: a3xx probably needs similar fix
Signed-off-by: Rob Clark <robclark@freedesktop.org>
