Early returns were erroneous and were causing memory leaks

Change-Id: Ifd6e37d79b6abd3f7d8d8899c951f1a0b4a66fb0
Reviewed-on: https://swiftshader-review.googlesource.com/19588Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

The use of std::unordered_map was the main source of slowdowns in
the optimizer code, so it was re-written without any maps. In order
to do so, the information originally carried by the maps was moved
to user-defined information stored within Subzero classes. The
optimizer now manages the memory used to store this information.

Bug swiftshader:69
Bug b/67872293

Change-Id: I2757169f0d3d467766317af6e00e149b4317fb9c
Reviewed-on: https://swiftshader-review.googlesource.com/19508Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Previously we were only attempting to obtain a handle to an already
loaded sibling library residing in the same directory. If it wasn't
already pre-loaded, we only loaded the library by name without any
directory prefix. This could lead to not finding the correct library.

Also, it's fine if the directory prefix is empty, so there's no need for
duplicate loops.

Bug b/110407478

Change-Id: Ia55b426e58b9f4039064bba9253dae80e4fdf5cf
Reviewed-on: https://swiftshader-review.googlesource.com/19549Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

Added MSAN, ASAN, TSAN, and UBSAN build options. Disabled use of
--no-undefined if a Santizer is used, since by design they leave some
symbols unresolved until run time. UBSAN required blacklisting of our
old copy of LLVM.

Bug swiftshader:108

Change-Id: Iab94b6815bc188c2a7f07d5b7a461234fb1035f3
Reviewed-on: https://swiftshader-review.googlesource.com/19548Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

BUG=chromium:854817

Change-Id: Icd81848fbbf3691eca46c9582b3c053ba1b00353
Reviewed-on: https://swiftshader-review.googlesource.com/19528Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Adrienne Walker <enne@chromium.org>

ChromeOS runs tests on some bots which run on Linux, but do not support X11.
In order to allow SwiftShader to successfully build on these bots, all X11
dependencies are #ifdefed behind the USE_X11 flag, which is already supplied
by the Chromium build system.

Change-Id: I6b914b1e662d9fbf101eb7caea7ac59e43cc7b56
Reviewed-on: https://swiftshader-review.googlesource.com/19488Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Bug b/110185979
Bug b/72330874

Change-Id: I593e975dca0798d027c8c5476e384eeca0fd1f7f
Reviewed-on: https://swiftshader-review.googlesource.com/19428Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

Whenever a cube map was being used, these asserts would fire as soon
as any border computation would happen. Fixed the asserts to allow
cube maps to work properly.

Change-Id: Iedc2661e63db37e5b4e77e08f97ce044e9f88837
Reviewed-on: https://swiftshader-review.googlesource.com/19468Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Alexis Hétu <sugoi@google.com>

When clear operations fall back to the slow path (i.e. neither fastClear
nor blitReactor is used), we were copying a rectangle the size of the
destination image. It should only sample within the 1x1 source pixel
instead.

Bug chromium:852641, chromium:851707

Change-Id: I9f247483f6167f92be8308b8470c021f5641b657
Reviewed-on: https://swiftshader-review.googlesource.com/19448Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

After [1], a manual dependency on exe_and_shlib_deps is no longer necessary
since it's automatically added.  This CL removes all remaining manual references
to exe_and_shlib_deps.

[1] https://chromium.googlesource.com/chromium/src.git/+/d7ed1f0a9c28c932fddc834ca5de44f28266c7f5

BUG=chromium:845700
R=capn

Change-Id: I20a88d716ed4cddd55047529f87267b5383366bc
Reviewed-on: https://swiftshader-review.googlesource.com/19348Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Tom Anderson <thomasanderson@google.com>

Optimizer::optimizeStoresInSingleBasicBlock() was taking an extremely long time due to
looping through all the instructions of a single basic block at every iteration. By
creating a much smaller list of only load/store operations the first time a single basic
block is encountered and reusing that list when the same block is encountered again,
this function now runs ~10X faster.

In my test case:
out/Default/cc_unittests --gtest_filter=ImageBackgroundFilter.BackgroundFilterRotated_GL
The sw::optimize function went from taking almost 16s to roughly 1.5s.

Bug b/67872293

Change-Id: I7e2e2aa8dc1bf2663988fb59b58d72d9ee986e36
Reviewed-on: https://swiftshader-review.googlesource.com/19408Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

SwiftShader should be able to build on Fuchsia, but it would
hit a collision error if added as a dependency in Fuchsia,
so the same logic currently used on MacOS could also be used
to avoid the collision on Fuchsia.

Change-Id: I9ac216dfc2180165a82bd5ae81c843ee4aea7ce6
Reviewed-on: https://swiftshader-review.googlesource.com/19328Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Sergey Ulanov <sergeyu@chromium.org>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

The parameters of the matrix constructor form an aggregate node that
was passed directly to the constructor of the structure, instead of
creating a one-element aggregate containing the matrix.

https://github.com/mc-imperial/shader-compiler-bugs/issues/74

Bug swiftshader:56

Change-Id: Iff9a2d8dc60d79a0dde28f2aad76407028486ec8
Reviewed-on: https://swiftshader-review.googlesource.com/19308Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

It appears that not only render targets, but also samplers that use
IOSurfaces require SwiftShader to perform synchronized draws,
otherwise it seems like using a shared IOSurface's CPU memory is
unsafe and can cause crashes in Layout Tests.

Bug chromium:846693

Change-Id: I0ce24700d34c657ac2447ceb2f6f837bfa3a9a58
Reviewed-on: https://swiftshader-review.googlesource.com/19288Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Change-Id: If953be0cce3cd11b833d9646e10ede0b31732729
Reviewed-on: https://swiftshader-review.googlesource.com/19248Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

Change-Id: I0984f43de3846428c30232192c7ba83342b1debd
Reviewed-on: https://swiftshader-review.googlesource.com/19228Reviewed-by: Krzysztof Kosiński <krzysio@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

Bug chromium:849217

Change-Id: I69d2269c117500d21e5d0860aed0acea212948ba
Reviewed-on: https://swiftshader-review.googlesource.com/19268Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

Added missing function declaration.

Change-Id: I52ddb666d70733187242906d298fb986f413756a
Reviewed-on: https://swiftshader-review.googlesource.com/19208Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Alexis Hétu <sugoi@google.com>

Because of the way SwiftShader is integrated into Chromium on MacOS,
a synchronization issue exists when Chromium attempts to use an
IOSurface before SwiftShader is done rendering into it. In order to
solve this, all draw calls that end up rendering into an IOSurface
must be synchronized within SwiftShader and can't yield to Chromium
until rendering is complete.

Bug chromium:847094

Change-Id: If2dba4fa998e7437ec414d3b4aff9e8b19ecc128
Reviewed-on: https://swiftshader-review.googlesource.com/19188Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Fixes "error: conditional expression is ambiguous; 'sw::Float' can be
converted to 'RValue<sw::Float>' and vice versa".

Change-Id: I2bd21036803dc54ce212df61a7520b7276b94b2a
Reviewed-on: https://swiftshader-review.googlesource.com/19168Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

The numeric imprecision can accumulate quickly and cause visible
sampling errors.

Bug chromium:848238

Change-Id: Ie41c10a0462e5b5e5c3ed5f7329ece00dab3f7f9
Reviewed-on: https://swiftshader-review.googlesource.com/19148Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

On Windows, GLAPI defaults to __declspec(dllimport), while GLAPICALL is
empty, leading to declaration mismatches. Either should be empty for
the unit tests since we link statically, for all platforms.

Change-Id: I2dde0ca35108678fd9767895e0120ee4b0d7030f
Reviewed-on: https://swiftshader-review.googlesource.com/19128Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

GLAPI is the desktop OpenGL equivalent of OpenGL ES's GL_APICALL macro.
They need to be consistent to prevent CLang from producing errors when
compiling with -Werror,-Winconsistent-dllimport.

Change-Id: Iceb73f42f52628a31aba7ce18fe60221529f890c
Reviewed-on: https://swiftshader-review.googlesource.com/19108Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

https://github.com/KhronosGroup/EGL-Registry/pull/43 made eglplatform.h
include X11 headers on Mac OS, which aren't pre-installed and breaks
the Chrome build.

https://github.com/KhronosGroup/EGL-Registry/issues/52

Change-Id: I73e3c11244a07e269684323ce76cc9700fe9c1aa
Reviewed-on: https://swiftshader-review.googlesource.com/19088Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

Fixing https://github.com/KhronosGroup/OpenGL-Registry/pull/149 and
https://github.com/KhronosGroup/OpenGL-Registry/pull/183 enabled
including desktop OpenGL headers without conflicts, so we no longer
have to define enums for ARB extensions ourselves.

Change-Id: I4f2e2d3b79a6f5cacf53f2a8ace78b6d124cb5d1
Reviewed-on: https://swiftshader-review.googlesource.com/19029Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

Change-Id: If56e711f15236397ac0f8b15f4ca182a7b1e7c49
Reviewed-on: https://swiftshader-review.googlesource.com/19028Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

This allows compilers targeting the MS ABI to select the
correct inheritance model for the member function pointer field
DrawCall::setupPrimitives using the complete type of Renderer. It
will allow us to enable the new Clang flag -fcomplete-member-pointers
globally.

Bug: chromium:847724
Change-Id: Ied5859ec2f5d38b3ccf51608527506caec53f470
Reviewed-on: https://swiftshader-review.googlesource.com/19068Tested-by: Peter Collingbourne <pcc@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

delete[] should be called on memory allocated with new[].

Change-Id: Ic21e270a7aa0e6e15132cbc3c454ea4f1e0222ea
Reviewed-on: https://swiftshader-review.googlesource.com/19048Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

Loops were no longer getting unrolled because we analyzed the entire
loop, not just the body, for changes to the index variable. Also, the
logic was inverted. Also, be conservative about loops with return
statements not being unrollable.

Bug chromium:845103

Change-Id: I5957e17f7b985ae90b10053216b6945f3f64338e
Reviewed-on: https://swiftshader-review.googlesource.com/19008Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

Previously we only copied the GLSL shader compiler's info log on
failure. This hid any warnings or debug info from the application.

Bug chromium:845103

Change-Id: Ia1877a405db2017d327dfc68037596fbda1579fa
Reviewed-on: https://swiftshader-review.googlesource.com/19009Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

Previously only dynamic indexing of uniforms was supported.

Since this is essentially a gather operation within the register file,
it is slow. We optimize the common case of using the loop index
variable as relative address, where the index value would be the same
for all shader invocations running in lock-step across SIMD lanes.

Bug chromium:845103
Bug skia:7846

Change-Id: Idb36b512dd560d740ac9088691b633ff3a1561c1
Reviewed-on: https://swiftshader-review.googlesource.com/18968Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

This allows to address the registers with a vector of indices.

Also rename 'dynamic' register files to 'indirect addressable', to
disambiguate from 'dynamic indexing' at the shader level. Indexing with
a uniform does not require gather/scatter operations, but does require
indirect addressing.

Bug chromium:845103
Bug skia:7846

Change-Id: I3c42be33def66328688f2900c61c80246bf1e584
Reviewed-on: https://swiftshader-review.googlesource.com/18989Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

Loops can only be unrolled if their loop index variable is not being
modified in the loop body.

Also check that the increment step of the loop operates on the initial
index variable.

Also remove some UNIMPLEMENTED's that were benign.

Bug chromium:845103
Bug chromium:843867
Bug skia:7846

Change-Id: Ib2b39f2d58763f0299ce7f6f75a8a75e6bdc7963
Reviewed-on: https://swiftshader-review.googlesource.com/18988Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

Change-Id: Icdeb96a3044ddf215a59ab832703bb447d85c3ab
Reviewed-on: https://swiftshader-review.googlesource.com/18948Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

Change-Id: I3cd5b2514149c5bf44dfab483eda415082da7793
Reviewed-on: https://swiftshader-review.googlesource.com/18928Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

This usage is not explicitly specified in EXT_texture_filter_anisotropic,
but in practice all notable ES 3.0 implementations treat maximum anisotropy
as sampler state. Furthermore, it's already accepted as a valid sampler
parameter by ValidateSamplerObjectParameter in Context.cpp.

This avoids a crash when calling
glSamplerParameteri(sampler, GL_TEXTURE_MAX_ANISOTROPY_EXT, 1.f);

Bug: b/79951966
Change-Id: I4013f621781fc74a0fbfbf14ab326871fbb87b7c
Reviewed-on: https://swiftshader-review.googlesource.com/18908Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Krzysztof Kosiński <krzysio@google.com>

We assume the pixels of an image can be addressed with a signed 32-bit
offset, including any padding. For a 3D image it's possible to exceed
this without exceeding the per-dimension limits. Lowering the per-
dimension limit so the allocation is always less than 2 GiB makes them
unreasonably small, so instead we must check the total size.

Use 1 GiB as the soft limit in OpenGL.

Bug chromium:835299

Change-Id: I9c5184002c1710e3923b549f8c21e7f6a516e1c7
Reviewed-on: https://swiftshader-review.googlesource.com/18869Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

This eliminates the duplication between pitchB() and size(), and
ensures that the latter is the full allocation size.

Bug chromium:835299

Change-Id: Icf555ad497fb3b92fd00e9a3e6ced6810b2d310d
Reviewed-on: https://swiftshader-review.googlesource.com/18789Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

OpenGL has separate implementation-defined texture size limits, for
3D textures and array textures. For now just give them the same value.

Bug chromium:835299

Change-Id: Ifaf537511f016e21992388f56598d5ec12a393b8
Reviewed-on: https://swiftshader-review.googlesource.com/18788Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

https://gitlab.khronos.org/opengl/API/issues/72 was resolved to treat
these cases as a no-op and not generate an error.

Bug chromium:825545

Change-Id: Ic4cfbc728156bb88a6dc70486ae57c3e12ea43ae
Reviewed-on: https://swiftshader-review.googlesource.com/18870Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>