Subzero assumes that stores and loads to/from stack memory can take an
immediate offset. This isn't true for sub-vector loads and stores. We
can just use regular full vector loads and stores instead.

Bug b/37496321

Change-Id: Iaffd60f8077a4c9d602665421f90a7288ab04050
Reviewed-on: https://swiftshader-review.googlesource.com/13008Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

Test: New test passes
Bug: b/37496078
Change-Id: Ieff1b080bdde3f9c86a0531566c8b08d7cdf24fe
Reviewed-on: https://swiftshader-review.googlesource.com/13009Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Casey Dahlin <sadmac@google.com>

Use VDUP for replicating a single element.
Use VZIP for interleaving vectors.
Use VMOV Dd, Dm for rearranging quadword vectors.

Bug b/67106219

Change-Id: I0de1457454c1db6d467bf870288b7af7cb59ac09
Reviewed-on: https://chromium-review.googlesource.com/695004Reviewed-by: Jim Stichnoth <stichnot@chromium.org>
Reviewed-on: https://swiftshader-review.googlesource.com/12968Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

When the first source register is different from the destination
register, but equal to the second source, we need to first narrow
the second source to prevent overwriting data we still need. Similar
when the destination is equal to the first source.

Bug b/37496082

Change-Id: I908a8e125a77ec4bf1eb5eab9e48c6112ee4ca13
Reviewed-on: https://chromium-review.googlesource.com/696032Reviewed-by: Jim Stichnoth <stichnot@chromium.org>
Reviewed-on: https://swiftshader-review.googlesource.com/12949Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

OpenGLES 3.0 allows for more combinations of format/internalformat
for floating point types. These types were already supported, just
not using the combinations added here.

Bug chromium:763384

Change-Id: I146548c2920799c7ea0d5d537d556ba562708147
Reviewed-on: https://swiftshader-review.googlesource.com/12928Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Test: New tests pass
Bug: b/37496856
Change-Id: Ifae1f802d61a7d51f6e8f5c135e3c8c67be30125
Reviewed-on: https://swiftshader-review.googlesource.com/12950Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Casey Dahlin <sadmac@google.com>

We weren't clamping the value when less than zero.

Bug b/37496082

Change-Id: I3170d14a622f8571f0142dd4ce80f96617c7e0aa
Reviewed-on: https://swiftshader-review.googlesource.com/12948Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Casey Dahlin <sadmac@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

When SSE4.1 isn't available, we were infinitely calling PackUnsigned()
recursively.

Bug b/37496082

Change-Id: Ie1fa8150d64740776c5b80baf07d997dde9d51be
Reviewed-on: https://swiftshader-review.googlesource.com/12929Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Casey Dahlin <sadmac@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Test: New tests pass
Bug: b/37496082
Change-Id: I4a0aae14318f03b5ce59f12f7542db13270f52ab
Reviewed-on: https://swiftshader-review.googlesource.com/12848Tested-by: Casey Dahlin <sadmac@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Integer division already had this modification, to avoid
division by 0, but integer modulo hadn't been modified
to match the other fix, which is now done here.

Bug chromium:763435

Change-Id: I40f4ed9f1c4b5889ddf47abc1bb7c70c21cbd7b8
Reviewed-on: https://swiftshader-review.googlesource.com/12908Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Alexis Hétu <sugoi@google.com>

MultiplyAddPairs is implemented using VMULL+VPADD.
MultiplyHighSigned/Unsigned is implemented using VMULL+VSHRN.
SubVectorLoad/Store is implemented using VLDR/VLD1/VSTR/VST1.
VectorPackSigned/Unsigned is implemented using two VQMOVN.

Bug b/37496078
Bug b/37496856
Bug b/37496321
Bug b/37496082

Change-Id: I141fd901d53da24ce780f503dc7ad17b94fc6ba8
Reviewed-on: https://chromium-review.googlesource.com/693049Reviewed-by: Jim Stichnoth <stichnot@chromium.org>
Reviewed-on: https://swiftshader-review.googlesource.com/12709Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

Test: New test passes
Bug: b/67106219
Change-Id: Ib21f73d312d39cf2b3aa15a11d2a207885a58978
Reviewed-on: https://swiftshader-review.googlesource.com/12828Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Casey Dahlin <sadmac@google.com>

If imageSize is 0, inputPitch may be 0, which would cause a division
by 0. In any case, if imageSize is 0, there's no work to perform, so
we just skip loadCompressedData() entirely in that case.

Bug:765094

Change-Id: Iedc6516ef6d025d24a8827597045cb3b83599e4a
Reviewed-on: https://swiftshader-review.googlesource.com/12648Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

WindowSurface::checkForResize() contained an assert which got
triggered in some tests. In order to avoid this issue, this cl
gracefully handles an invalid window handle.

Change-Id: I6bace88370980e94676d4af9de0d88e561175835
Reviewed-on: https://swiftshader-review.googlesource.com/12808Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Bug b/37495545

Test: New tests pass
Change-Id: Icc81729d8c65aadccdd685490218d298e671b638
Reviewed-on: https://swiftshader-review.googlesource.com/12768Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Casey Dahlin <sadmac@google.com>

Bug b/37495545

Change-Id: I1d025f894bb7cf08dcaafd950605781633596ab3
Reviewed-on: https://chromium-review.googlesource.com/689098Reviewed-by: Jim Stichnoth <stichnot@chromium.org>
Reviewed-on: https://swiftshader-review.googlesource.com/12748Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

This also fixes the encoding of right shifts with unsigned elements.

Bug b/37496338

Change-Id: I9a1dc91359daea5f4391a137b7f9e03bd941146b
Reviewed-on: https://chromium-review.googlesource.com/688057Reviewed-by: Jim Stichnoth <stichnot@chromium.org>
Reviewed-on: https://swiftshader-review.googlesource.com/12668Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

Change-Id: I24a2ac536c24732538cc5c7f2aad4e7763989f9e
Reviewed-on: https://chromium-review.googlesource.com/568673Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Jim Stichnoth <stichnot@chromium.org>
Reviewed-on: https://swiftshader-review.googlesource.com/12690Tested-by: Nicolas Capens <nicolascapens@google.com>

R=bradnelson@chromium.org

Bug: 665593
Change-Id: I1b9a5fd31c5e426bd193a32a2feb291eb17af5a6
Reviewed-on: https://chromium-review.googlesource.com/587993Reviewed-by: Brad Nelson <bradnelson@chromium.org>
Reviewed-on: https://swiftshader-review.googlesource.com/12689Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

The gold linker thinks that causes an ODR violation.

BUG=chromium:449754
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/2896133003 .

Patch from Lei Zhang <thestig@chromium.org>.

Change-Id: I7f39b1f6e638ad3462b9b462ebdb6348473b9fc5
Reviewed-on: https://swiftshader-review.googlesource.com/12688Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

x86 vector packing instructions always treat the input as having signed
integer components, but can perform signed or unsigned saturation on
the output. In Reactor the Pack() intrinsic has overloads which
differentiate between them based on the signedness of the input, but
this is confusing.

Also simplify emulation of saturating add/subtract.

Bug b/37496082

Change-Id: I0625fff429ffb40f42baf9600c7760d9858b5d89
Reviewed-on: https://swiftshader-review.googlesource.com/12548Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

Google Test by default disables breaking into the debugger to be able
to run to completion without interaction. Enable it for debug builds
within Visual Studio.

Change-Id: If36e3d74ab1ec98eefa221761e02a33bd8117151
Reviewed-on: https://swiftshader-review.googlesource.com/12588Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

Call make before launching a debug session. Also move from the 'build'
subdirectory to the 'debug' subdirectory to allow for a separate
release build or other configuration.

Change-Id: I6db00a5bb0a707ebbf9c160897852526a8bb73e1
Reviewed-on: https://swiftshader-review.googlesource.com/12528Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

We don't start timing until after the first frame finished rendering
and is presented, so the frame count has to be 0 at that point.

Change-Id: Ic242bb5625c6c50694e7625008565ee421859624
Reviewed-on: https://swiftshader-review.googlesource.com/12488Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

The problem was at line 827, where "(qHead - qSize) % 32" could give
unexpected results if qSize > gHead (which wasn't a problem with
unsigned values). Also removed other % operations which could have
cause some issues.

Change-Id: Ia443e05ce1add3879720e90f7dbac771e712d2ab
Reviewed-on: https://swiftshader-review.googlesource.com/12568Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Alexis Hétu <sugoi@google.com>

The copy constructor of std::atomic is not allowed, so initialization
through operator= is not allowed. Explicitly using the AtomicInt
constructor to fix this issue.

Change-Id: If77763e89e0d958d8a4b5a83d5d7bd9e3ba08ff6
Reviewed-on: https://swiftshader-review.googlesource.com/12508Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

MSAN detected uninitialized memory which caused many test failures.
This can be easily fixed by zeroing out the memory before calling
XCreateImage.

Change-Id: I913c818ef38446055af7fd6aa885028e5f3f8bfb
Reviewed-on: https://swiftshader-review.googlesource.com/12388Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

A few more variables were causing possible data races,
so they were changed to AtomicInt variables.

Change-Id: Icf233482528e086fc58f6919232ab2e3f0784689
Reviewed-on: https://swiftshader-review.googlesource.com/12368Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

TSAN detected many data race errors in the SwiftShader Renderer
class. x86 has a strong memory ordering model which guarantees
that changes are observed in the same order by other threads.
However, C++ does not provide such guarantees unless specified
using atomic operations. In order to fix these, a new AtomicInt
class was added which is a basically a wrapper class for
std::atomic<int> and which only exposes the portion of the API
required by SwiftShader.

Since std::atomic isn't available on older versions of Android,
a fallback class was implemented without using std::atomic, which
is closer to the previous implementation. Both classes appear to
work properly after performing a few dEQP tests. Both also perform
similarly.

A few minor changes were made in order to attempt to reduce the use
of atomic integer operations when possible.

Change-Id: Ife6d3a2b6113346f8f8163b692e79c2a0e03b22f
Reviewed-on: https://swiftshader-review.googlesource.com/12308Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

In order to catch errors faster, this change enables asserts on
Chromium trybots.

Change-Id: Ifdc4c9e0ec7d54aa7f6f333d992054c4e0062efe
Reviewed-on: https://swiftshader-review.googlesource.com/12328Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

This FPS counter is disabled by default.

Change-Id: Ida04352dece25a212cb678c9ceca4c31d654f373
Reviewed-on: https://swiftshader-review.googlesource.com/12128Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

The spec defines that the initial state for a newly bound renderbuffer
has a format of RGBA4. However, when querying the number of red/green/
blue/alpha/depth/stencil bits, we should report the actual number of
bits allocated (0), and not assert.

Bug chromium:765912
Bug b/25351344

Change-Id: Icc3eb59e76bcbd6f0f2f16a2a75ace706ae17251
Reviewed-on: https://swiftshader-review.googlesource.com/12148Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

OpenGL requires depth values to be clamped to the [0, 1] range. Due to
frustum clipping already limiting the range, this can only happen when
non-zero polygon offset parameters are active.

Also fix the 'minimum resolvable difference' for the 32-bit floating-
point internal depth format that we use.

Bug swiftshader:82

Change-Id: Ic9ebcac182a2bc81ab51d79cfe0bb451d340bd1e
Reviewed-on: https://swiftshader-review.googlesource.com/12108Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

This does not fix the OGLSimpleCube sample. See swiftshader:81.

Change-Id: I4bb5755d8b63f234965c3d12f774706968f2d34e
Reviewed-on: https://swiftshader-review.googlesource.com/12088Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

Some applications may not call eglReleaseThread() before terminating a
thread, which can cause a minor memory leak. pthread_key_create()
supports specifying a destructor callback to implicitly free the
thread-local storage on thread termination. On Windows we use the
DLL_THREAD_DETACH signal to free it explicitly. This change also
simplifies TLS management by removing the ability to set a new value.

Bug swiftshader:80

Change-Id: I249ca519c1c9ab722e570516d0931cb17ea081bb
Reviewed-on: https://swiftshader-review.googlesource.com/12068Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

The stencil test compares the unsigned masked reference value against
the unsigned masked stencil value. But because x86 vector instructions
only support signed comparison, we have to add 0x80 to both sides.
However, this offset was incorrectly added before the masking.

Bug b/64683344

Change-Id: I49748e25d8d99e1c2b3c87d3dbe74a9dba75fb1c
Reviewed-on: https://swiftshader-review.googlesource.com/12028Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

eglReleaseThread() was leaking some memory because even though it
deletes the 'current' thread state, it was inadvertently getting
re-allocated when recording the success error state.

Change-Id: I92efb1b34e62a21998b4fc19aaeba6473b06548b
Reviewed-on: https://swiftshader-review.googlesource.com/11968Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

Emulated vectors have fewer elements than their underlying type, e.g.
v4i16 is stored in the lower half of a v8i16 vector, but for constant
vector construction we need to provide values for all the elements of
the underlying vector.

Bug b/65124545

Change-Id: Iaf942c4d517b1d4d464894ec501f7d339ae86438
Reviewed-on: https://swiftshader-review.googlesource.com/11908Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

Change-Id: Ia4feb6791c6da8dc4beef86aab7eb4fa06af461b
Reviewed-on: https://swiftshader-review.googlesource.com/11869Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

When BOARD_VNDK_VERSION is set, a vendor lib cannot link against a
platform lib and is not provided with global include path. So, some
internal libraries such as swiftshader_top_[release|debug] were also
marked as vendor module and dependency to some header libs
(libhardware_headers, libnativebase_headers, etc.) were added
explicitly.

Bug: 64704866
Test: lunch sdk_gphone_x86-user, BOARD_VNDK_VERSION=current m -j
libEGL_swiftshader libGLESv1_CM_swiftshader libGLESv2_swiftshader
successful

Change-Id: I37851c9876f7f7c247bbc5d6e18946cb5bf72fa2
Reviewed-on: https://swiftshader-review.googlesource.com/11868Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>