We want to start reworking and expanding this code, but it'll be a lot
easier to do once we disentangle it from the rest of the stuff in nir.c.
Unfortunately, there are a few unavoidable dependencies in nir.c on
methods we'd rather not expose publicly, since if not used in very
specific situations they can cause Bad Things (tm) to happen. Namely, we
need to do some magical control flow munging when adding/removing jumps.
In the future, we may disallow adding/removing jumps in
nir_instr_insert_*() and nir_instr_remove(), and use separate functions
that are part of the control flow modification code, but for now we
expose them and put them in a separate, private header.
Signed-off-by: Connor Abbott <connor.w.abbott@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
cleanup_cf_node() is part of the control flow modification code, which
we're going to split into its own file, but remove_defs_uses() is an
internal function used by nir_instr_remove(). Break the dependency by
making cleanup_cf_node() use nir_instr_remove() instead, which simply
calls remove_defs_uses() and then removes the instruction from the list.
nir_instr_remove() does do extra things for jumps, though, so we avoid
calling it on jumps which matches the previous behavior (this will be
fixed later in the series).
Signed-off-by: Connor Abbott <connor.w.abbott@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
It was being used to initialize function impls and loops, even though
it's really a control flow modification helper. It's pretty trivial, so
just inline it to avoid the dependency.
Signed-off-by: Connor Abbott <connor.w.abbott@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
We should be checking almost everything now.
Signed-off-by: Connor Abbott <connor.w.abbott@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
It's simply the first nir_cf_node in the nir_function_impl::body list,
which is easy enough to access - we don't to store a pointer to it
explicitly. Removing it means we don't need to maintain the pointer
when, say, splitting the start block when modifying control flow.
Thanks to Connor Abbott for suggesting this.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
v2, review from Francisco Jerez:
- make the destination variable as large as what the nir instrinsic
defines (4) instead of the size of the return variable of glsl. This
is still safe for the already existing code because all the intrinsics
affected returned the same amount of components as expected by glsl IR.
In the case of image_size, it is not possible to do so because the
returned number of component depends on the image type and this case
is not well handled by nir.
v3:
- Style fix
Signed-off-by: Martin Peres <martin.peres@linux.intel.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
The positive and negative value of a float can only
be equal to each other if it is -0.0f and 0.0f.
This is safe for Nan and Inf, as -Nan != Nan, and -Inf != Inf
This gives no changes in my shader-db
Signed-off-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
-NaN != NaN, and -Inf != Inf, so this should be safe.
Found while working on my VRP pass.
Shader-db results on my IVB:
total instructions in shared programs: 1698267 -> 1698067 (-0.01%)
instructions in affected programs: 15785 -> 15585 (-1.27%)
helped: 36
HURT: 0
GAINED: 0
LOST: 0
Some shaders was found to have the following pattern in NIR:
vec1 ssa_26 = fneg ssa_21
vec1 ssa_27 = fne ssa_21, ssa_26
Make that:
vec1 ssa_27 = fne ssa_21, 0.0f
This is found in Dota2 and Brutal Legend.
One shader is cut by 8%, from 323 -> 296 instructons in SIMD8
Signed-off-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
NIR instruction count results on i965:
total instructions in shared programs: 1261954 -> 1261937 (-0.00%)
instructions in affected programs: 455 -> 438 (-3.74%)
One in yofrankie, two in tropics. Apparently i965 had also optimized all
of these out anyway.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
There are so many flags in textures, that the CSE pass would have a hard
time referencing the correct set when figuring out if two texture ops are
the same. By zeroing, we can avoid that fragility.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Avoids regressions in vc4 when trying to do our blending in NIR.
v2: Add the other unpack ops I meant to when writing the original commit
message.
Reviewed-by: Matt Turner <mattst88@gmail.com>
We may find a cause to do more undef optimization in the future, but for
now this fixes up things after if flattening. vc4 was handling this
internally most of the time, but a GLB2.7 shader that did a conditional
discard and assign gl_FragColor in the else was still emitting some extra
code.
total instructions in shared programs: 100809 -> 100795 (-0.01%)
instructions in affected programs: 37 -> 23 (-37.84%)
v2: Use nir_instr_rewrite_src() to update def/use on src[0] (by Thomas
Helland).
v3: Make sure to flag metadata dirties, and copy the swizzle and abs/neg
over to src[0], too (by anholt).
Reviewed-by: Thomas Helland <thomashelland90@gmail.com> (v2)
Tested-by: Thomas Helland <thomashelland90@gmail.com> (v2)
This is useful to increase the CSE opportunities for a scalar backend. It
avoids regressions when dropping vc4's custom CSE implementation.
v2: Cleanups by Matt (decl in the for loop, and unreachable()).
Reviewed-by: Matt Turner <mattst88@gmail.com>
Avoid copying an overwritten swizzle, use the original values.
Example:
Former swizzle[] = xyzw
src->swizzle[] = zyxx
The expected output swizzle = zyxx but if we reuse swizzle in the loop,
then output swizzle would be zyzz.
Signed-off-by: Samuel Iglesias Gonsalvez <siglesias@igalia.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
The current implementation operates in scalar mode only, so add a vec4
mode where types are padded to vec4 sizes.
This will be useful in the i965 driver for its vec4 nir backend
(and possbly other drivers that have vec4-based shaders).
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
The only values allowed are 0 and 1, and the value is checked before
assigning.
This is a copy of 8eeca7a56c that seems to have been made to the glsl
ir type after it was copied for use in nir but before nir landed.
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
This just adds some missing pieces to nir/i965,
it is lightly tested on my Haswell.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This type will be used to store the name of subroutine types
as in subroutine void myfunc(void);
will store myfunc into a subroutine type.
This is required to the parser can identify a subroutine
type in a uniform decleration as a valid type, and also for
looking up the type later.
Also add contains_subroutine method.
v2: handle subroutine to int comparisons, needed
for lowering pass.
v3: do subroutine to int with it's own IR
operation to avoid hacking on asserts (Kayden)
v3.1: fix warnings in this patch, fix nir,
fix tgsi
v3.2: fixup tests
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>
Signed-off-by: Dave Airlie <airlied@redhat.com>
tests: fix warnings
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>
We already recognize min(max(a, 0.0), 1.0) as a saturate, but neglected
this variant (which is also handled by the GLSL IR pass).
shader-db results on Broadwell:
total instructions in shared programs: 7363046 -> 7362788 (-0.00%)
instructions in affected programs: 11928 -> 11670 (-2.16%)
helped: 64
HURT: 0
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
These are basically just moves, so they should be safe as well.
When disabling i965's GLSL IR level scalarizer (channel expressions)
pass, I started seeing NIR code like this:
if ssa_21 {
block block_1:
/* preds: block_0 */
vec4 ssa_120 = vec4 ssa_82, ssa_83, ssa_84, ssa_30
/* succs: block_3 */
} else {
block block_2:
/* preds: block_0 */
/* succs: block_3 */
}
block block_3:
/* preds: block_1 block_2 */
vec4 ssa_33 = phi block_1: ssa_120, block_2: ssa_2
Previously, the GLSL IR scalarizer pass would break the vec4 into a
series of fmovs, which were allowed by the peephole pass. But with
the vec4 operation, they were not. We want to keep getting selects.
Normal i965 on Broadwell:
instructions in affected programs: 200 -> 176 (-12.00%)
helped: 4
With brw_fs_channel_expressions() disabled:
instructions in affected programs: 1832 -> 1646 (-10.15%)
helped: 30
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
lower_phis_to_scalar() pass recurses the instruction dependence graph to
determine if all the sources of a given instruction are scalarizable.
To prevent cycles, it temporary marks the phi instruction before recursing in,
then updates the entry with the resulting value. However, it does not consider
that the entry value may have changed after a recursion pass, hence causing
a use-after-free situation and a crash.
This patch fixes this by reloading the entry corresponding to the 'phi'
after recursing and before updating its value.
The crash can be reproduced ~20% of times with the dEQP test:
dEQP-GLES3.functional.shaders.loops.while_constant_iterations.nested_sequence_fragment
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
When we compute the output swizzle we want to consider the number of
components in the add operation. So far we were using the writemask
of the multiplication for this instead, which is not correct.
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Some shaders in Civilization V and Beyond Earth do
pow(pow(x, 2.2), 0.454545)
which is converting to and from sRGB colorspace.
A more general rule that replaces pow(pow(a, b), c) with pow(a, b * c)
actually regresses two shaders in Sun Temple in which the result of the
inner pow is used twice, once by another pow and once by another
instruction. Also, since 2.2 * 0.454545 isn't exactly one, the more
general pattern would have still left us with a pow, and I'm 2.2 *
0.454545 percent sure that's not what they want.
instructions in affected programs: 934 -> 886 (-5.14%)
helped: 16
Previously, we used intrinsic->const_index[1] to represent "the number of
array elements to load" for load/store intrinsics. However, this set to 1
by every pass that ever creates a load/store intrinsic. Also, while it
might make some sense for registers, it makes no sense whatsoever in SSA.
On top of that, the i965 backend was the only backend to ever support it;
freedreno and vc4 just assert that it's always 1. Let's just delete it.
Signed-off-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Rob Clark <robclark@freedesktop.org>
