Bug swiftshader:38

Change-Id: Ice466d96e4d71d546e5770551cc4188f8364edd0
Reviewed-on: https://swiftshader-review.googlesource.com/9029Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Bug swiftshader:38

Change-Id: I74c642b3c7346541d042acde78aa2e0213beb5f6
Reviewed-on: https://swiftshader-review.googlesource.com/9028Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

This CL fixes building the 'all' target in a Chromium checkout with the following args.gn:
is_debug = false
is_clang = false
use_sysroot = false
proprietary_codecs = true
ffmpeg_branding = "Chrome"
is_component_build = true
enable_nacl = false
use_goma = true

Bug chromium:697528
Change-Id: Ie725988e8a1cb9ee672eb1e8e20d718ac91004fa

Change-Id: Ie725988e8a1cb9ee672eb1e8e20d718ac91004fa
Reviewed-on: https://swiftshader-review.googlesource.com/8888Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Tom Anderson <thomasanderson@google.com>

String literals are const in C++11, and should only be assigned to
const pointers.

Bug swiftshader:33

Change-Id: Iefe0a1c06d58dffb69d7fbe199a2f0aa8bf3bcb1
Reviewed-on: https://swiftshader-review.googlesource.com/8990Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

When building with VC++ 2017 RC and its default header files the
intrin.h header file is less aggressively included by Windows headers.
This causes build breaks on Chrome x64 builds with VC++ 2017 RC when
X86MCTargetDesc.cpp uses __cpuid without first including its header.

https://github.com/google/swiftshader/pull/2

Change-Id: I664d7b509e0517b7621d41312ccdd0d3de60de84
Reviewed-on: https://swiftshader-review.googlesource.com/8951Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Nicolas Capens <capn@google.com>

WaitForDebugger function was undefined.

Change-Id: I578e99638bddc40c56817cfc66ed5787f1cdc6e7
Reviewed-on: https://swiftshader-review.googlesource.com/8953Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Alexis Hétu <sugoi@google.com>

The debugger wait dialog hinders tests on debug builds
on chromium bots, so this change sets the debugger wait
dialog to only be allowed only on Visual Studio builds
within the SwiftShader solution.

Change-Id: I0b126f9b0d0451d0fa8941fcb3f9097100488f45
Reviewed-on: https://swiftshader-review.googlesource.com/8952Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Alexis Hétu <sugoi@google.com>

Bug b/36232368

Change-Id: I0a62629db6737134771861710ee0b18826807008
Reviewed-on: https://swiftshader-review.googlesource.com/8949Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Greg Hartman <ghartman@google.com>

Rather than specifying defines using explicit commands in "cflags"
sections, all defines were moved to generic "defines" sections.

Change-Id: Ica6ee2a2447aae4f8fc6decc3a4f6b40af7ce9fd
Reviewed-on: https://swiftshader-review.googlesource.com/8948Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

Rendering on the main thread can cause segmentation faults on Windows
due to uncommitted stack memory, since Subzero does not currently call
__chkstk. In practice this would have mainly been an issue on VMs with
a single virtual CPU.

Bug swiftshader:25

Change-Id: Ic3be7e5a41ef09b7e056d3c3df2983c225101fe4
Reviewed-on: https://swiftshader-review.googlesource.com/8934Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

The __chkstk function can be at a greater than 4 GiB offset from our
generated function. Fix derived from https://reviews.llvm.org/D7267

Change-Id: Ife87dcd42541676828c4a0ca77dcded6649ce278
Reviewed-on: https://swiftshader-review.googlesource.com/8932Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Nicolas Capens <capn@google.com>

https://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html#Clobbers specifies
that the clobber list should not overlap with the input or output.

Change-Id: I77fc2bfc6209ed325355193758f00fc171c8aebf
Reviewed-on: https://swiftshader-review.googlesource.com/8930Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Nicolas Capens <capn@google.com>

Subzero doesn't actually use any ABI breaking LLVM functionality, so
we compile a minimized version of LLVMSupport which does not include
the symbol used to detect a mismatch in the ABI-breaking check at link-
time. Hence we need to disable this check.

Change-Id: I04c37deb359b81c21954ad92a6665d44fe2312a5
Reviewed-on: https://swiftshader-review.googlesource.com/8929Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

GCC warns against bitfields not fitting the entire range of an enum
with a fixed underlying type of unsigned int, which gets promoted to
an error with -Werror and cannot be suppressed. However, GCC already
defaults to using unsigned int as the underlying type of an unscoped
enum without a fixed underlying type. So we can just omit it.

Bug chromium:697528

Change-Id: I8db8fef09e3397d4ec5ebe9f7c3c5e4ae9c0edff
Reviewed-on: https://swiftshader-review.googlesource.com/8928Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Nicolas Capens <capn@google.com>

to be able to override chromium_git.

Change-Id: I11419a81fb492affcfa92ba69661cc90ad88ae98
Reviewed-on: https://swiftshader-review.googlesource.com/8868Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Alexis Hétu <sugoi@google.com>

Added the missing part of the FenceSync object
implementation based on the assumption that all
fence operations are always signaled before
calling any sync related wait functions.

Passes all related dEQP tests.

Change-Id: I2fd2ac160dfd1250f55240fc3f696f00579d12d4
Reviewed-on: https://swiftshader-review.googlesource.com/8855Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Dequeuing and queuing the graphics buffer takes care of ownership, and
a previous patch ensured the buffer is unlocked before we queue it so
we don't need to hold a reference for longer.

Bug b/34981145
Bug b/35443034

Change-Id: I6cbaeff08f5d930dde8bdd1522052ac33a6ae9ac
Reviewed-on: https://swiftshader-review.googlesource.com/8849Reviewed-by: Greg Hartman <ghartman@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Greg Hartman <ghartman@google.com>

The buffer changes ownership after queuing it, so unlock it first.
The decRef() gets removed in a subsequent patch.

Bug b/34981145

Change-Id: I4710699af46a7cbde12e97bc444ad35747509b07
Reviewed-on: https://swiftshader-review.googlesource.com/8848Reviewed-by: Greg Hartman <ghartman@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Greg Hartman <ghartman@google.com>

cancelBuffer() returns a graphics buffer to the buffer queue but
doesn't pass it on to the consumer side. There's currently no reason
for us to call it when the EGL surface gets destroyed or resized, since
we only briefly dequeue the buffer to blit the already rendered result
and immediately queue it again. Even if the queuing fails, we shouldn't
try to cancel the buffer because that's essentially the same as a
queueing but without valid content. It would fail with the same error,
or worse, transfer ownership twice and mess up the reference counting.

Bug b/34981145

Change-Id: Ia37a711db5ebd61bcde14cf4bc45097d4e2d33e8
Reviewed-on: https://swiftshader-review.googlesource.com/8814Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Greg Hartman <ghartman@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Greg Hartman <ghartman@google.com>

Many new OpenGL ES 3 specific entries in Context::getIntegerv
were not hidden when an OpenGL ES 2 context is created. The
switch statement was separated in 2 statements in order to
fix this.

Change-Id: Ib218b72e854f5857958cfa14ecdfef5ae03ee92b
Reviewed-on: https://swiftshader-review.googlesource.com/8851Tested-by: Alexis Hétu <sugoi@google.com>
Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Change-Id: I20b385c316f24b703da1d3071e393b4cde450173
Reviewed-on: https://swiftshader-review.googlesource.com/8811Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Silence a few warnings within the LLVM code used by Subzero
which ended up being treated as errors when compiling
Chromium for x86.

Change-Id: If4d3c13ce0b01185d59a2cd064b2bad537639b94
Reviewed-on: https://swiftshader-review.googlesource.com/8813Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Change-Id: I21ec9c20a594fd322d17f2bcd1c58b298d4c0ce9
Reviewed-on: https://swiftshader-review.googlesource.com/8812Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Once Subzero is pulled by Chromium (see
https://codereview.chromium.org/2695813002/),
we'll enable Subzero compilation on Windows.

Change-Id: I3464dbd98224b3ac602aef3c4a27596366d275b3
Reviewed-on: https://swiftshader-review.googlesource.com/8772Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

A few remaining cases in Context.cpp are actually
implemented, so we can remove the UNIMPLEMENTED()
calls in these cases: Primitive restart, uniform
blocks, fragment inputs and vertex outputs are all
implemented. Also, the implementation can legally
support 0 program binary formats.

Change-Id: I74b51b511c66170e284ac5d82f573453332872cc
Reviewed-on: https://swiftshader-review.googlesource.com/8788Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Change-Id: I6077e9f439dc20217e9e7fb6c6f98f969e7d0b51
Reviewed-on: https://swiftshader-review.googlesource.com/8790Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Nicolas Capens <capn@google.com>

Make min(x, y) equal to (x < y ? x : y) and max(x, y) equal to
(x > y ? x : y), including the behavior for signed zeros and NaNs.
This also enables optimizing them into min and max SSE2 instructions
on x86.

Bug swiftshader:19

Change-Id: I047b90e9da9f3c72657ab7c619bc91b92a700a45
Reviewed-on: https://swiftshader-review.googlesource.com/8771Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Nicolas Capens <capn@google.com>

Bug chromium:691427

The official build uses the LLVM gold linker, which seems
to have an issue with some visibility settings. Restoring
these settings for now to solve the issue.

Change-Id: I4f448c38bb7c964b2a53d19e064ecb23c119ab61
Reviewed-on: https://swiftshader-review.googlesource.com/8770Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Alexis Hétu <sugoi@google.com>

This change adds 2 platform specific files for Linux, which
are used to define the typeinfo() functions used in some
classes in order to satisfy the ubsan compiler on Linux.
Unfortunately, adding those definitions on all platforms
breaks the build on Mac. Those aren't necessary on Windows,
so the Linux specific files are enough to make the build
work properly on all platforms.

Change-Id: If731ab2e5df47731695412da04f9b2db576ea140
Reviewed-on: https://swiftshader-review.googlesource.com/8768Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

The Linux build fix had broken the Mac build, so the new
"typeinfo" virtual function were moved from the header files
to the source files to fix both builds simultaneously.

Change-Id: I37f5ef4bd5ec4da34a19531dcb85df6a98728aa1
Reviewed-on: https://swiftshader-review.googlesource.com/8748Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

To prevent Chromium from building the "swiftshader_subzero"
source set when building all, the source set is now located
within an "if (use_swiftshader_with_subzero)" statement.

Change-Id: I4363487103ec8cd6856b1afb0dd861146a3e56bd
Reviewed-on: https://swiftshader-review.googlesource.com/8728Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

The current way "use_subzero" was used seems to fail when
compiling with goma (I don't see why, honestly). Anyway,
most build configuration related booleans are declared in
a declare_args() section, so I moved the setting there,
which will hopefully solve the issue.

Change-Id: I063d9938d04bfdf8140878a740f3710309f8bf1e
Reviewed-on: https://swiftshader-review.googlesource.com/8714Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Bug b/34268167
Change-Id: Id2200a9c7d6fea522e3405829447e031e2da15b3
Reviewed-on: https://swiftshader-review.googlesource.com/8711Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Greg Hartman <ghartman@google.com>

Only the instructions following a return statement were being marked as
affected by it so predication would be applied. But in a loop the instructions
above the return statement are also affected by it.

Bug b/25220690

Change-Id: If2490a6e0b4e9cf8b6e28b33cbbbcec8b4ebfdaa
Reviewed-on: https://swiftshader-review.googlesource.com/5183Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Nicolas Capens <capn@google.com>

Cleaned up a few things before updating SwiftShader's version
in Chromium's DEPS file:
- Added Windows Subzero fix
- Removed commented out sysroot related lines
- Fixed formatting using git
- Temporarily disabled Subzero compilation for now

Change-Id: Ib6ae75c20869522675d28086ab170f01fd730f93
Reviewed-on: https://swiftshader-review.googlesource.com/8709Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Chromium UBSan builds require RTTI, for which GCC/Clang requires each
class' first non-inline virtual method (the "key method") to have a known
definition so that its address can be used as a unique type identifier:
https://gcc.gnu.org/onlinedocs/gcc/Vague-Linkage.html

libEGL and libGLESv2 each use objects who's full definition only exists
within one or the other. Methods for which the definition is unknown can
still be called if they are virtual, because the compiler/linker only needs
to know the vtable entry offset. But because of the GCC/Clang requirement
of having the first non-inline virtual method be fully defined, we need to
add dummy virtual methods and their definitions.

Bug swiftshader:31

Change-Id: Ib146cac811388086b29dbb099266c43795d6ed31
Reviewed-on: https://swiftshader-review.googlesource.com/8708Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Nicolas Capens <capn@google.com>

The BackoffLock spins idle for a while when waiting for a locked mutex
before yielding the core, thus wasting many CPU cycles. Modern pthread
implementations have low overhead mutexes which make the thread sleep
if the lock is already held, and efficiently resume them it becomes
available.

Change-Id: I26b64c86db620739671373fd0d82085744d34fa8
Reviewed-on: https://swiftshader-review.googlesource.com/8648Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Nicolas Capens <capn@google.com>

Since floating-point scalars are stored in vector registers, Subzero
allows us to just bitcast between them, eliminating a load and insert
before shuffling.

Change-Id: Ibccf242fd4cfc28604f35f420a04fd4ee6eabe52
Reviewed-on: https://swiftshader-review.googlesource.com/8575Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Array accesses with unsigned indices can be faster on x86-64 because
we can take advantage of implicit zero-extension of 32-bit integers to
64-bit during pointer arithmetic.

Change-Id: I17d531d9ad05c2d2994f007d5444b2a514a591b8
Reviewed-on: https://swiftshader-review.googlesource.com/8571Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Nicolas Capens <capn@google.com>

BUG: 30979264

Change-Id: I315bc6e6b64a012896542667d7db58f3516b29de
Reviewed-on: https://swiftshader-review.googlesource.com/8649Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Greg Hartman <ghartman@google.com>