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>

Test: foo
Change-Id: I27a139266d34d26f5d3468e3993a502f0596ae7d
Reviewed-on: https://swiftshader-review.googlesource.com/8650Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Greg Hartman <ghartman@google.com>

After a pthread_join() call the handle becomes invalid or could be
recycled by another thread, leading to undefined behavior when
attempting to join it again. Also, on Windows the handle has to be
closed to free it.

Bug b/34883464

Change-Id: Ib20d0539b6b46e331c6378b3a9f0c2a334d34892
Reviewed-on: https://swiftshader-review.googlesource.com/8612Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Nicolas Capens <capn@google.com>

Modified the BUILD.gn files in order to be able to easily switch
between Subzero and LLVM back ends. Note that Subzero
compilation should already work on Linux, as soon as the Wheezy
to Jessie update lands. For now, only Windows will build Subzero.

Also removed linux hack from libEGL build file.

Change-Id: Ic584b604496c4b4f745b05d72a752f056495055b
Reviewed-on: https://swiftshader-review.googlesource.com/8630Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Linux handled the extra / properly, but Windows did not,
so it had to be removed.

Change-Id: Ie42d8777aef83c0024bc9b3e5c631ece7e87e236
Reviewed-on: https://swiftshader-review.googlesource.com/8629Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

When second arch is translated we don't have 32-bit platform and
64-bit platform.  Instead we have two 32-bit ones, just one is
translated.

Test: Build system refactoring CL. Existing unit tests still pass.

BUG=31422117
BUG=27526885

Change-Id: Ia5fdd2f485dd5e4e8bc6799f0691cd03c2bff1a0
Reviewed-on: https://swiftshader-review.googlesource.com/8609Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Nicolas Capens <capn@google.com>
(cherry picked from commit f24b1b445274b0dbb4ffc860847a26d0b33d3a21)
Reviewed-on: https://swiftshader-review.googlesource.com/8611

The 3.0 camera HAL makes extensive use of
HAL_PIXEL_FORMAT_IMPLEMENTATION_DEFINED, which Android defines as the
format that is best for the hardware. The fake camera implementation
decided that that was RGBX8888, so it seem like that is what we should
use.

Test: untested
Change-Id: Ia02ebbefd015f9ca099b4393daf5d464f38fe8fc
Reviewed-on: https://swiftshader-review.googlesource.com/8588Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Nicolas Capens <capn@google.com>

Bug b/34799114

Change-Id: Id42d3b784946291d8754774a7916a186a4694ba1
Reviewed-on: https://swiftshader-review.googlesource.com/8591Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Change-Id: I82f8215ce4366e0795ce249b4d8f6c8e391af96c
Reviewed-on: https://swiftshader-review.googlesource.com/8568Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Fixed MacOS build after the recent Linux fix caused an issue on
MacOS. Removed a few flags and changed visibility to fix it.

Change-Id: Ic9d08d8ec0ccf8a2002f2787c3ec79861bb59139
Reviewed-on: https://swiftshader-review.googlesource.com/8589Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Change-Id: I4136781005cbd0551461adecdd94d8e3cc8688d0
Reviewed-on: https://swiftshader-review.googlesource.com/8570Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

operator[] on a pointer was using the vector register size (16 bytes)
instead of the size of the (emulated) vector type.

Change-Id: Ice8677b41a3af81176fe9495fd02ede9f8b0482f
Reviewed-on: https://swiftshader-review.googlesource.com/8569Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Stores are dead if the address is from an alloca and all uses of the
address are stores, not just when it's a single store.

Change-Id: I5b65a64e21b9e398922e5440c35f0d318c03e911
Reviewed-on: https://swiftshader-review.googlesource.com/8574Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

If a store address has no definition, it must be a pointer argument.
Unlike stores to alloca's, these are not dead.

Change-Id: I66e50b14ffea1cfc61fc756b5545ab8ae28b4bf8
Reviewed-on: https://swiftshader-review.googlesource.com/8573Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

In order to add Subzero as a dependency in Chromium, Chromium's
DEPS file will be modified to use SwiftShader's DEPS file in
order to pull SwiftShader's desired revision of Subzero. This
will ensure that SwiftShader and Subzero are always in sync.

Change-Id: I0b24845c0ddbf974baac4ef6cc3ff6ab397d5f69
Reviewed-on: https://swiftshader-review.googlesource.com/8529Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Although SwiftShader was building properly on the build bots, it
wasn't exporting its symbols properly. I compared the flags used in
the BUILD.gn files and in the CMakeLists.txt file and attempted to
get both build systems to use the same flags and options.

Change-Id: Idc73d4ad5b45311881e88720bf611e25110dcc70
Reviewed-on: https://swiftshader-review.googlesource.com/8576Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Constant operands of commutative operations preferably go on the right
hand side to avoid requiring an extra register for two operand
instructions. Also, Subzero assumes constants in pointer arithmetic are
on the right hand side to consider optimizing it into an addressing
mode.

Change-Id: Ife5a471903d5f4bef0c19b6c908d75715f06bfec
Reviewed-on: https://swiftshader-review.googlesource.com/8548Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Nicolas Capens <capn@google.com>

Change-Id: I40922c89056cacd00f9d728dfe1f3f12824c81a1
Reviewed-on: https://swiftshader-review.googlesource.com/3253Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

BUG: 31072273
Change-Id: I037505ad3ab1ba80aecab4e24ec8d1932df2dcf7
Reviewed-on: https://swiftshader-review.googlesource.com/7030Reviewed-by: Lingfeng Yang <lfy@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Lingfeng Yang <lfy@google.com>

MAP_ANON has been deprecated in favor of MAP_ANONYMOUS, but Mac OS only
defines MAP_ANON.

Bug chromium:630728

Change-Id: I7345a5f9227acbc0caa29d8103855ab531f16499
Reviewed-on: https://swiftshader-review.googlesource.com/8531Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Nicolas Capens <capn@google.com>

Fixed:
- type comparison mismatch
- redefined macros
- unused functions
- line endings

Change-Id: I965ca4b887aee968e6cf95feb36c3cbf90b85580
Reviewed-on: https://swiftshader-review.googlesource.com/8528Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Change-Id: I7979eb2557a6103ad10b7be3c274fadf12d27ae7
Reviewed-on: https://swiftshader-review.googlesource.com/8508Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Bug swiftshader:20

Change-Id: I67818bfe10cb29211559fb2ee047f6bec6ce46d4
Reviewed-on: https://swiftshader-review.googlesource.com/8451Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Bug swiftshader:20

Change-Id: I20c2ab7cb4c00c365520ff8b8500f7594127498b
Reviewed-on: https://swiftshader-review.googlesource.com/8468Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

When we have an uninitialized variable that is declared outside of a
loop and is read and written to inside the loop, we should not eliminate
the load and replace it with an undefined value, because in the next
iteration we want the value from the previous iteration.

Change-Id: Ic996c796307f62fc835079af6c58adf1eb288259
Reviewed-on: https://swiftshader-review.googlesource.com/8488Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Bug swiftshader:20

Change-Id: I81ad267d450713ffe2a5a84e1d7f7f140b515c85
Reviewed-on: https://swiftshader-review.googlesource.com/8454Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Nicolas Capens <capn@google.com>

Provide default constructors for each type, and initialize the swizzle
parent in a less bug prone manner.

Change-Id: Ia7f406a66274c7b1737cbe82db733e2e3ab65175
Reviewed-on: https://swiftshader-review.googlesource.com/8453Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Nicolas Capens <capn@google.com>

Change-Id: Ic8785bd56ac7303b0502f0303f9270a6999a1df9
Reviewed-on: https://swiftshader-review.googlesource.com/8452Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Nicolas Capens <capn@google.com>

Subzero silently allows inserting a 32-bit integer into a vector with
16-bit elements. This generates a pinsrd instruction, which requires
SSE4.1, even when specifying to only support SSE2. This change emits
a pinsrw instruction instead, which is SSE2.

Bug swiftshader:20

Change-Id: I37bd0ebb6d9cfdd2cdef4788e89e9672386c685e
Reviewed-on: https://swiftshader-review.googlesource.com/8450Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>
Tested-by: Nicolas Capens <capn@google.com>

Subzero returns vNi1 type results for vector comparisons, which we were
sign-extending to v4i32, even for smaller fields. We don't actually
ever have to sign-extend these results.

Change-Id: Ifdd30edd498d66fc4c557804035794a659eacd87
Reviewed-on: https://swiftshader-review.googlesource.com/8449Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>

Subzero uses a helper function to convert uint to float. It's faster to
just emit a sequence of operations to perform the cast manually.
This implementation converts the lower 31 bits as a signed integer, and
adds 0x80000000 as a floating-point value when the upper bit is set.
This approach does not produce the correct rounding in rare cases, but
should still be adequate. For consistency, we're also using this method
with the LLVM back-end.

Change-Id: Ic5d7b73cd2a9e154056365cdbe9af0962bdbe1cb
Reviewed-on: https://swiftshader-review.googlesource.com/8312Tested-by: Nicolas Capens <capn@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <capn@google.com>