gallium/util: return FALSE for intersection if there's empty rectangles

It isn't really obvious if intersection test should take into account empty
rectangles or if the caller should do it. But it looks like most callers
actually verified one of the rects but not the other, but since correctly
returning an empty rect that other rect could actually be empty leading to
more bugs. Hence just verify both rects for emptyness in the intersection
test itself which makes the code easier in the caller (though it will be
slower if the caller knows the rectangles are non-empty).

Reviewed-by: Zack Rusin <zackr@vmware.com>

tgsi: add some more helper functions

This patch adds some more helper functions such as
   . tgsi_transform_temps_decl
   . tgsi_transform_output_decl
   . tgsi_transform_dst_reg
   . tgsi_transform_src_reg

Reviewed-by: Brian Paul <brianp@vmware.com>

tgsi: added tgsi_is_shadow_target() helper

tgsi: add negate parameter to tgsi_transform_kill_inst()

Reviewed-by: Charmaine Lee <charmainel@vmware.com>

util: added ffsll() function

v2: fix errant _GNU_SOURCE test, per Matt Turner.

Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>

util: added util_set_index_buffer()

Like util_set_vertex_buffers_count(), this basically just copies a
pipe_index_buffer object, taking care of refcounting.

mesa: Move gl_vert_attrib from mtypes.h to shader_enums.h

It is a shader enum after all...

Acked-by: Brian Paul <brianp@vmware.com>

glapi: Inline x86_64_current_tls().

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

r600g: Simplify out a couple of unnecessary branches

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>

radeonsi: use an indirect buffer for init_config

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: add IB2 indirect buffer support for pm4 states

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

winsys/radeon: add a flag telling how gfx IBs should be padded

This is always false on amdgpu (set by calloc).

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

winsys/amdgpu: remove IB padding for SI

SI is unsupported by amdgpu

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: remove unused macro si_pm4_set_state

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: remove si_pm4_cleanup

All remaining pm4 state are created and destroyed by state trackers.

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: rework uploading border colors

The border colors are uploaded only once when the state is created.

This brings truly immutable sampler descriptors, because they don't have
to be updated every time a sampler state is re-bound.

It also moves the TA_BC_BASE_ADDR registers to init_config, removing one
more state. The catch is there is now a limit: only 4096 border colors can
be used by one context. I don't think that will be a problem.

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: use all built-in border colors

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: inline si_cmd_context_control

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: remove unused si_pm4_state code

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: reorder si_context variables

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: don't send IB dword usage to si_need_cs_space

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: don't set number of IB dwords for states

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: don't count IB space for states, just use an upper bound

Since we don't put any resource descriptors in IBs, the space used by draw
calls is quite small.

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: convert SPI state to an atom

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

gallium/radeon: rename r600_context_bo_reloc -> radeon_add_to_buffer_list

this name should be easy to understand without other knowledge

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

gallium/radeon: rename write_*_reg functions

e.g. radeon_set_context_reg is nicer and looks consistent next to
radeon_emit().

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: rename and precalculate polygon offset states

one less calloc and state construction while drawing

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: convert CB_TARGET_MASK setup to an atom

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: don't set VGT_VTX_CNT_EN twice in init_config

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: convert stencil ref state into an atom

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: convert blend color state into an atom

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: convert sample mask state into an atom

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: convert clip state into an atom

Reducing calloc overhead.

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: avoid redundant CB and DB register updates

The main idea is to avoid setting CB_COLORi_INFO = 0 for i>0 repeatedly
when those colorbuffers aren't used. This is mainly for glamor.

Same for DB. Z_INFO and STENCIL_INFO need to be cleared only once.

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: don't rebind GSVS ring buffers every draw call using GS

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: don't clear the tessellation factor ring buffer

Leftover from the bring-up.

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: remove the tf_ring state, add the registers to init_config

One less state to worry about.

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: remove the gs_rings state, add the registers to init_config

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: use a bitmask for tracking dirty atoms

This mainly removes the cache misses when checking the dirty flags.
Not much else though.

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: initialize atom IDs for external atoms

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: call si_init_atom for remaining radeonsi atoms

I need to initialize more atom IDs.

This adds 4 more si_init_atom calls, which simplifies the code.
(si_init_atom needs a different context type of the emit functions though)

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: initialize atom IDs

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: define the state atom array separately

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: optimize viewport states

same as scissors

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: optimize scissor states

- convert 16 states to 1 atom
- only emit 1 scissor if VIEWPORT_INDEX isn't written
- use only one packet when emitting consecutive scissors

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: add SI_MAX_ATTRIBS

PIPE_MAX_ATTRIBS is 32, but we currently only support 16.

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: fix memory usage checking for big IBs

Cc: 11.0 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: set all 16 viewport Z bounds for GL 4.1

Cc: 11.0 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

radeonsi: fix a Unigine Heaven hang when drirc is missing

Cc: 10.6 11.0 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

winsys/amdgpu: use small IBs for better performance on VI

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

gallium/util: add u_bit_scan_consecutive_range

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>

i965: Prevent coordinate overflow in intel_emit_linear_blit

Fixes regression from
commit 8c17d53823c77ac1c56b0548e4e54f69a33285f1
Author: Kenneth Graunke <kenneth@whitecape.org>
Date:   Wed Apr 15 03:04:33 2015 -0700

    i965: Make intel_emit_linear_blit handle Gen8+ alignment restrictions.

which adjusted the coordinates to be relative to the nearest cacheline.
However, this then offsets the coordinates by up to 63 and this may then
cause them to overflow the BLT limits. For the well aligned large
transfer case, we can use 32bpp pixels and so reduce the coordinates by
4 (versus the current 8bpp pixels). We also have to be more careful
doing the last line just in case it may exceed the coordinate limit.

Reported-and-tested-by: kaillasse91@hotmail.fr
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=90734
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Kenneth Graunke <kenneth@whitecape.org>
Cc: Ian Romanick <ian.d.romanick@intel.com>
Cc: Anuj Phogat <anuj.phogat@gmail.com>
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>

i965/nir: enable the dead control flow optimization

total instructions in shared programs: 7541551 -> 7541381 (-0.00%)
instructions in affected programs:     3054 -> 2884 (-5.57%)
helped:                                29

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

nir/dead_cf: add support for removing useless loops

v2: fix detecting if the loop has any phi nodes after it.
v2: use nir_foreach_ssa_def() instead of nir_foreach_dest() when
    checking for values live after the loop to catch const_load
    instructions.
v2: fix handling return instructions
v2: add some documentation to loop_is_dead()

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>

nir: add a helper for iterating over blocks in a cf node

We were already doing this internally for iterating over a function
implementation, so just expose it directly.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

nir: add nir_block_get_following_loop() helper

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

nir/dead_cf: delete code that's unreachable due to jumps

v2: use nir_cf_node_remove_after().
v2: use foreach_list_typed() instead of hardcoding a list walk.
v3: update to new control flow modification helpers.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

nir: add an optimization for removing dead control flow

v2: use nir_cf_node_remove_after() instead of our own broken thing.
v3: use the new control flow modification helpers.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

r600g: fix calculation for gpr allocation

I've been chasing a geom shader hang on rv635 since I wrote
r600 geom code, and finally I hacked some values from fglrx
in and I could run texelfetch without failures.

This is totally my fault as well, maths fail 101.

This makes geom shaders on r600 not fail heavily.

Cc: "10.6" "11.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

mesa: Limit Framebuffer Parameter OpenGL ES 3.1 usage

According to OpenGL ES 3.1 specification, section 9.2.1 for
glFramebufferParameter and section 9.2.3 for glGetFramebufferParameteriv:

"An INVALID_ENUM error is generated if pname is not FRAMEBUFFER_DEFAULT_WIDTH,
FRAMEBUFFER_DEFAULT_HEIGHT, FRAMEBUFFER_DEFAULT_SAMPLES, or
FRAMEBUFFER_DEFAULT_FIXED_SAMPLE_LOCATIONS."

Therefore exclude OpenGL ES 3.1 from using the GL_FRAMEBUFFER_DEFAULT_LAYERS
parameter.

Signed-off-by: Marta Lofstedt <marta.lofstedt@intel.com>
Reviewed-by: Kevin Rogovin <kevin.rogovin at intel.com>

mesa: Expose GL_ARB_framebuffer_no_attachments to GLES 3.1

V2: Conform to new standard for exposing enums for OpenGL ES 3.1.

Signed-off-by: Marta Lofstedt <marta.lofstedt@intel.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>

nir/builder: Use nir_after_instr to advance the cursor

This *should* ensure that the cursor gets properly advanced in all cases.
We had a problem before where, if the cursor was created using
nir_after_cf_node on a non-block cf_node, that would call nir_before_block
on the block following the cf node.  Instructions would then get inserted
in backwards order at the top of the block which is not at all what you
would expect from nir_after_cf_node.  By just resetting to after_instr, we
avoid all these problems.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965: advertise ASTC support for Skylake

v2: remove OES ASTC extension reference.

Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
Signed-off-by: Nanley Chery <nanley.g.chery@intel.com>

mesa/glformats: recognize ASTC formats as color formats

ASTC formats contain RGBA components.

Reviewed-by: Chad Versace <chad.versace@intel.com>
Signed-off-by: Nanley Chery <nanley.g.chery@intel.com>

mesa/texformat: use format conversion function in _mesa_choose_tex_format

This function's cases for non-generic compressed formats duplicate
the GL to MESA translation in _mesa_glenum_to_compressed_format().
This patch replaces the switch cases with a call to the translation
function. This change teaches this function about ASTC, thus enabling
ASTC for glTex*Storage*() calls.

Reviewed-by: Chad Versace <chad.versace@intel.com>
Signed-off-by: Nanley Chery <nanley.g.chery@intel.com>

mesa/texcompress: correct mapping of S3TC formats in conversion function

MESA_FORMAT_RGBA_DXT5 should actually be reserved for GL_RGBA[4]_DXT5_S3TC.
Also, Gallium and other dri drivers (radeon and nouveau) follow this mapping
scheme.

Reviewed-by: Chad Versace <chad.versace@intel.com>
Signed-off-by: Nanley Chery <nanley.g.chery@intel.com>

r600/sb: update last_cf for finalize if.

As Glenn did for finalize_loop we need to update_cf when we
add a POP at the end of a shader.

I think this fixes one of the earlier shader going off end
of memory problems we've stopped.

Reviewed-by: Glenn Kennard <glenn.kennard@gmail.com>
Cc: "10.6" "11.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>

i965/fs: Use greater-equal cmod to implement maximum.

The docs specifically call out SEL with .l and .ge as the
implementations of MIN and MAX respectively. Among other things,
SEL with these conditional mods are commutative.

See commit 3b7f683f.

Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>

i965/chv|skl: Apply sampler bypass w/a

Certain compressed formats require this setting. The docs don't go into much
detail as to why it's needed exactly.

This patch introduces no piglit regressions on gen9 (bsw is untested). Note that
the SKL "regressions" are fixed tests, and the egl_khr_gl_colorspace tests are
WTF. The patch also fixes nothing I can find.
http://otc-mesa-ci.jf.intel.com/job/Leeroy/127820/

v2:
Reworded commit message (Matt); Added piglit results link.
Restructured condition (Matt)
Moved check out to function (Nanley). I left the setting of the bit in the
  surface state open coded because it seems to go better with the existing code.

v3:
Use and inline function only in gen8_emit_texture_surface_state() (Matt).

Cc: Matt Turner <mattst88@gmail.com>
Cc: Nanley Chery <nanleychery@gmail.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Matt Turner <mattst88@gmail.com>

st/mesa: move to renumbering registers in a group

This can be done with a single pass for the instruction base,
and takes renumber_registers out of its spot on the profile.

Acked-by: Marek Olšák <marek.olsak@amd.com
Signed-off-by: Dave Airlie <airlied@redhat.com>

st/mesa: reduce time spent in calculating temp read/writes

The glsl->tgsi convertor does some temporary register reduction
however in profiling shader-db this shows up quite highly,

so optimise things to reduce the number of loops through
all the instructions we do. This drops merge_registers
from 4-5% on the profile to 1%. I think this can be reduced
further by possibly optimising the renumber pass.

Acked-by: Marek Olšák <marek.olsak@amd.com
Signed-off-by: Dave Airlie <airlied@redhat.com>

st/mesa: cache tgsi opcode info in the instruction

Instead of looking this up lots, lets just cache it in the instruction
translation up front. I just noticed this function what high in a profile
of shader-db on radeonsi.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

r600: move prim convert from geom shader to function.

This should avoid C++ fail including this header.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

glsl: remove specical case subroutine type counting

Unlike samplers we can get the correct value for subroutines from
component_slots()

Reviewed-by: Dave Airlie <airlied@redhat.com>

r600g: Use TGSI parse results instead of manually exfiltrating

This makes better use of the work that the TGSI API has done for
us.

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>

r600g: Set geometry properties in r600_create_shader_state()

The selector is shared by all shader variants, so the
individual shaders shouldn't change it. Use tgsi_shader_scan()
results to set geometry properties within a
r600_create_shader_state() call and treat said propertices in
the selector as read-only within r600_shader_from_tgsi().

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>

r600g: Move geometry properties state from shader to selector

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>

r600g: Remove dead assigment to 'gs_input_prim' in shader state

Note that 'geometry shader properties' should be carried in the
selector state over the shader state in any case.

Signed-off-by: Edward O'Callaghan <eocallaghan@alterapraxis.com>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>

radeonsi: don't use the emit qt keyword in si_init_atom

It confuses my editor.

radeonsi: remove no-op 32-bit masking

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

gallium/radeon: fix the ADDRESS_HI mask for EVENT_WRITE CIK packets

Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

winsys/radeon: handle non-zero finite timeout when waiting for buffers

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

freedreno/a3xx: implement half-z clipping

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>

freedreno/a3xx: add basic clip plane support

The hardware is capable of dealing with GL1-style user clip planes.
No clip vertex, no clip distances. Fixes a number of ucp tests, as well
as neverball.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Cc: "11.0" <mesa-stable@lists.freedesktop.org>

nvc0: change prefix of MP performance counters to HW_SM

According to NVIDIA, local performance counters (MP) are prefixed
with SM, while global performance counters (PCOUNTER) are called PM.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

nvc0: sort performance counter queries by name

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

nvc0: make names of performance counter queries consistent

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

nvc0: use enumerations for driver queries

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

nvc0: remove commented out code related to PCOUNTER queries

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

r600: port si_conv_prim_to_gs_out from radeonsi

This code was broken by the tess merge, and I totally missed it
until now. I'm not sure this fixes anything but it stops the assert.

Cc: "11.0" <mesa-stable@lists.freedesktop.org>
Reviewed-by: Glenn Kennard <glenn.kennard@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

r600g: use PRIi64 for some compute debug printfs

Otherwise this will crash on 32-bit, and it gets rid of
warnings building on 32-bit.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

gallium/util: fix debug_get_flags_option on 32-bit

On 32-bit we need to use PRIu64 flags for printfs,
otherwise this segfaults in R600_DEBUG=help otherwise.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Cc: "11.0" <mesa-stable@lists.freedesktop.org>
Signed-off-by: Dave Airlie <airlied@redhat.com>

glsl: provide the option of using BFE for unpack builting lowering

This greatly improves generated code, especially for the snorm variants,
since it is able to get rid of the lshift/rshift for sext, as well as
replacing each shift + mask with a single op.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Matt Turner <mattst88@gmail.com>

glsl: use bitfield_insert instead of and + shift + or for packing

It is fairly tricky to detect the proper conditions for using bitfield
insert, but easy to just use it up front. This removes a lot of
instructions on nvc0 when invoking the packing builtins.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Matt Turner <mattst88@gmail.com>

i965/fs: Remove fs_visitor::try_replace_with_sel().

No shader-db changes on g4x, snb, hsw, or bdw.

Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965/fs: Replace awful variable names.

   start_to      -> dst_start
   end_to        -> dst_end
   start_from    -> src_start
   end_from      -> src_end
   var_to        -> dst_var
   var_from      -> src_var
   reg_to        -> dst_reg
   reg_to_offset -> dst_reg_offset
   reg_from      -> src_reg

Not sure how these made sense to me before.

Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965/fs: Skip blocks in register coalescing interference check.

No need to walk through instructions in blocks we know don't contain our
registers' live ranges.

Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965/fs: Improve register coalescing interference check.

I always thought that the is_control_flow() -> return false check was a
bad hack, and some previous attempts to remove it have failed and have
been reverted.

The previous two patches fix some problems that caused register
coalescing to not notice some interference between registers, which the
is_control_flow() check apparently works around.

With that fixed, we can calculate interference more accurately.

total instructions in shared programs: 6261319 -> 6257917 (-0.05%)
instructions in affected programs:     346282 -> 342880 (-0.98%)
helped:                                1552

Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965/fs: Use overwrites_reg() instead of dst.equals().

equals() returns false for registers with different types, using it
isn't appropriate to determine whether an is overwriting a register.

Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965: Only consider fixed_hw_reg in equals() if file is HW_REG/IMM.

Noticed when debugging things that lead to the next patch.

On G45 (and presumably ILK) this helps register coalescing:

total instructions in shared programs: 4077373 -> 4077340 (-0.00%)
instructions in affected programs:     43751 -> 43718 (-0.08%)
helped:                                52
HURT:                                  2

Reviewed-by: Jason Ekstrand <jason.ekstrand@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>