Commits · 68cfc78ee7f74d69901ba0dd54b7cba27b0b60d9 · Chen Yisong / swiftshader

04 Jul, 2019 7 commits

Reactor: Rework optimization flags. · 68cfc78e

authored Jun 29, 2019

Replace the old 'runOptimizations' flag on Nucleus::acquireRoutine() and Nucleus::acquireCoroutine() with a new OptimizationLevel enumerator.
runOptimizations was always true, and there is already the rr::optimization extern to control the optimization passes performed.

All Function and Coroutines are now passed the new vk::ReactorOptimizationLevel. This will be changed to a non-Default value in another CL.

Bug: b/135609394
Change-Id: I1154da05d413b18a471a3818fbb03f356a3d0e6c
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33482Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Ben Clayton <bclayton@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>

68cfc78e

Refactor LLVMReactor globals into 3 classes. · 6f8e565b

authored Jun 29, 2019

Split the scattered LLVM global variables into 3 new classes:
• JITGlobals is a singleton that holds all the immutable machine specific
  information for the host device.
• JITBuilder holds all the LLVM state for building routines.
• JITRoutine is a rr::Routine that holds a LLVM JIT session, compiler and
  object layer as each routine may require different target machine
  settings and no Reactor routine directly links against another.

This reduce all the global varibles to down to one - 'jit' a JITBuilder.
In theory, this can be marked as TLS to remove the global reactor mutex lock.

These change also make it possible to specify per-function JIT optimization
settings (implemented in a follow-up change).

Change-Id: If4c90de444830f398eb79081fd349f313a8d3bd7
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33481Tested-by: Ben Clayton <bclayton@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

6f8e565b

Regres: Bump MaxProcMemory from 2GB to 3GB · 18699c63

authored Jul 03, 2019

dEQP-VK.synchronization.internally_synchronized_objects.pipeline_cache_compute

When the wind is blowing in the right direction, this test can exceed the per-test 2GB sandbox memory limit, causing a flaky out-of-memory crash.

Bug: b/133127573
Change-Id: Ie403d91738c05258639be077aa31ec36d92d109e
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33688Tested-by: Ben Clayton <bclayton@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>

18699c63

Regres: Post significant changes to test times · 9d6e6f61

authored Jul 01, 2019

Change-Id: I3833a203ae7635845ea3b8ceb5e323565a92be18
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33487
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Ben Clayton <bclayton@google.com>

9d6e6f61

Regres: Treat ABORT()s as failures · 4dafba90

authored Jul 01, 2019

Bug: b/132353920
Change-Id: Ic7950ea29be6fd0dc7712813c10cdd2dfab73b90
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33548Reviewed-by: Chris Forbes <chrisforbes@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Ben Clayton <bclayton@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>

4dafba90

Regres: Add flag to prioritize a single change · 8c126fa1

authored May 20, 2019

Change-Id: Ieaec8e9dd2017803d862ce8bcdc6b00582b7d45b
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33486
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Ben Clayton <bclayton@google.com>

8c126fa1

Regres: Add flag to run the daily test run on launch · 1a3f167a

authored May 20, 2019

Change-Id: I6247e0b5eef4639527ea15e7c5af115e5954a342
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33485
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Ben Clayton <bclayton@google.com>

1a3f167a

03 Jul, 2019 2 commits

Context cleanup · 72d81048

authored Jun 10, 2019

- Removed twoSidedStencil, which was always true
  when stencil is enabled
- Removed unused enums and functions
- Removed custom new/delete operators
- Used VkFrontFace type for frontFace

Bug b/132280877

Change-Id: I440d5e910e50de0ac8049d3a56848fa8d2de3493
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33531
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Tested-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

72d81048

Removed the AtomicInt class · 463fab94

authored Jun 12, 2019

Deleted Thread.hpp and replaced AtomicInt with std::atomic<int>.
Should be noop.

Note: AtomicInt's operator++(int) and operator--(int) were
      Post-increment operators implemented as Pre-increment
      operators, so expressions <AtomicInt object>++ have
      been replaced with ++<std::atomic object> (same for
      decrement operators).

Bug b/129403963

Change-Id: I907b8c91581fed4b76647327f316b87369c6c82f
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33668Tested-by: Alexis Hétu <sugoi@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

463fab94

02 Jul, 2019 4 commits

Don't emit bounds checks when robustness is disabled · a29aa771

authored Jun 26, 2019

The Vulkan device's robustBufferAccess feature determines whether
accesses to buffers must be bounds-checked against the range of the
buffer descriptor.

The spec states that "Some features, such as robustBufferAccess, may
incur a run-time performance cost. Application writers should carefully
consider the implications of enabling all supported features."
Hence in many cases applications will leave the feature disabled to
ensure maximum performance.

Bug: b/131224163
Change-Id: I69e150d5043f0118db8a4b2751b24b5d87bccd75
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33389
Presubmit-Ready: Nicolas Capens <nicolascapens@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Chris Forbes <chrisforbes@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>

a29aa771

Fix treating Image storage class as non-interleaved · 8f5bdcf0

authored Jun 28, 2019

This was causing the dEQP-VK.memory_model.message_passing.core11.u32.coherent.fence_fence.atomicwrite.workgroup.*
tests to read out of bounds when robust buffer access is disabled.

Bug: b/136561217
Change-Id: I4891f5cd531e99b3dd87b4b84ae545ba47355875
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33528Reviewed-by: Chris Forbes <chrisforbes@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

8f5bdcf0

SpirvShader: Optimize SIMD sequential, fully-in-bounds loads & stores · af1eedb3

authored Jul 02, 2019

For sequential, fully-in-bound vectors:
* Loads can safely be a regular vector load. We mask just to keep behavior consistent with rr::MaskedLoad and rr::Gather.
* Stores of non-atomics can be implemented as a read-modify-write.

These optimizations have drastic performance improvements on architectures where there are no masked-read and masked-write instructions.

Bug: b/135609394
Change-Id: I552cc38f4aeae73f8db079a0a11da6a8db857710
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33628Tested-by: Ben Clayton <bclayton@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

af1eedb3

Regres: Update test lists @ · 142f4581

authored Jul 02, 2019

Change-Id: I5e007143e62f5fea83d0294332a690ef0d58a52e
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33608
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Tested-by: Ben Clayton <bclayton@google.com>

142f4581

28 Jun, 2019 9 commits

PipelineCache implementation · 52edb176

authored Jun 26, 2019

The PipelineCache class now contains members for caching programs and
shaders, which allows significant speedups when using it to prevent
recompiling the same shaders over and over again.

For now, each PipelineCache object contains its own cache and the
cache is not shared between PipelineCache objects. It remains to be
seen if SwiftShader would benefit from always caching shaders and
programs in a global cache.

Notes:
- Merging 2 PipelineCache objects was also implemented
- Added a few "const" where appropriate

IMPORTANT NOTE: This cl DOES NOT allow pipeline caches to be saved
                and loaded through the pipeline cache data.

Bug b/123588002

Change-Id: I95b183033c03e114d69d4432e5831e26be477033
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33428Tested-by: Alexis Hétu <sugoi@google.com>
Presubmit-Ready: Alexis Hétu <sugoi@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

52edb176

Reactor: Assert that array indexing is in bounds · 83dd452c

authored Jun 27, 2019

For compile time indices, we can do some basic sanity checking.

Change-Id: I9ae4ea5bbf106c4653cbf9bbe6776a5a188b4289
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33470Tested-by: Ben Clayton <bclayton@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>

83dd452c

CMakeLists: Use ccache if found · 4901ffd5

authored Jun 27, 2019

Dramatically speeds up builds when switching build flags.

Change-Id: I661d18458db69f6663381c252b06b3ed96f3a5c6
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33472Tested-by: Ben Clayton <bclayton@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>

4901ffd5

Cleanup unittests a bit · 2e5d5b46

authored Jun 21, 2019

Made the following changes:
 * Changed our multi-line strings to use C++11 raw string literals
 * Added argument-less macros for checking if GL or EGL had an error.
 * Made checkCompileFails return the error log, in case a test wants to
   ensure a specific failure is reported.
 * Made some more tests use the new MakeShader function.

Bug chromium:977665

Change-Id: I5a410d88449d1349806215bd21bb8c2b6d7acc56
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33268
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Tested-by: Sean Risser <srisser@google.com>

2e5d5b46

Pipeline/ShaderCore: Reimplement halfToFloatBits to not rely on denorm float multiply · 249b092d

authored Jun 28, 2019

This function behaved differently on different architectures. Reimplement using mostly integer arithmatic.

Bug: b/136095012
Change-Id: I5df87e9400f2952386ebdf7b7b0866b0197145a8
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33480Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Presubmit-Ready: Ben Clayton <bclayton@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Tested-by: Ben Clayton <bclayton@google.com>

249b092d

Wire up remaining 32b __sync_* functions required for arm32 · fd4c96de

authored Jun 20, 2019

Bug: b/135691587
Change-Id: I796342057829f416bc6a2dfdb0c03c59f033127f
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33228Tested-by: Chris Forbes <chrisforbes@google.com>
Presubmit-Ready: Chris Forbes <chrisforbes@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>

fd4c96de

Regres: Update test lists @ · 74561df3

authored Jun 28, 2019

Change-Id: Id3546c1c53dc3aa7313cbd9d6353310fa7b52bf2
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33477
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Tested-by: Ben Clayton <bclayton@google.com>

74561df3

Fix extra semicolon after function def in FrameBufferOSX.mm · 40495290

authored Jun 27, 2019

Change-Id: I4a1b1a981096fbda616101b7d47bf8f82481c64b
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33509
Presubmit-Ready: Lingfeng Yang <lfy@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Lingfeng Yang <lfy@google.com>

40495290

Allow 3d texture layers for framebuffer blit · fb9c05cc

authored Jun 27, 2019

b/135141616

Change-Id: Ia3c804c0bef7a000a54359d24dcd8bacc063a121
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33508Tested-by: Lingfeng Yang <lfy@google.com>
Kokoro-Presubmit: Lingfeng Yang <lfy@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

fb9c05cc

27 Jun, 2019 4 commits

Add support for R8_SRGB textures · 762741be

authored Jun 26, 2019

This texture format is needed to run the PowerVR Glass demo

Bug: b/136178423
Tests: dEQP-VK.*r8_srgb*

Change-Id: Ifff800e25db69daf099ce9c748a6c16de6d93db1
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33449Reviewed-by: Alexis Hétu <sugoi@google.com>
Tested-by: Hernan Liatis <hliatis@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>

762741be

Selectively suppress class-memaccess warning/error · 6d874f05

authored Jun 26, 2019

To still produce the class-memaccess diagnostic on unintended cases,
suppress it only for our Memset<T> case and LLVM.

Bug: b/135744933
Change-Id: Id3b421a3b831600097d198ac793f582467d0a520
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33429
Presubmit-Ready: Nicolas Capens <nicolascapens@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>

6d874f05

Simplify Vulkan ICD loading for testing · 522d5121

authored Jun 21, 2019

The STANDALONE preprocessor macro is only for controlling the loading of
a delay loaded DLL, on Windows.

Change-Id: Ife765f27c7e03c542d7e81d90c09c97e4e2cbe5d
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33248
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>

522d5121

SpirvShader: Optimize stores with static equal offsets · cd98c739

authored Jun 22, 2019

This is heavily used in dEQP-VK.ssbo.*. Avoiding generating the scatter
is profitable on all non-AVX512-capable targets;
ScalarizeMaskedMemIntrin is incredibly slow.

Reduces runtime on dEQP-VK.ssbo.layout.random.all_shared_buffer.5 from
24s to 14s on my threadripper (on top of stack of other optimizations).

Bug: b/135609394
Change-Id: I2d6840522a5bd30b4fd532b9c7e2a4712879caa9
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33289Tested-by: Chris Forbes <chrisforbes@google.com>
Presubmit-Ready: Chris Forbes <chrisforbes@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>

cd98c739

26 Jun, 2019 5 commits

Replace broken prime in Y'CbCr comment · 61f2a46a

authored Jun 26, 2019

Change-Id: I79303a65a2f578164d36917b503a1aed188cd2a2
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33448Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Chris Forbes <chrisforbes@google.com>

61f2a46a

LLVMReactor: Perform atomic load / stores of floats as ints. · 2fa1dad2

authored Jun 26, 2019

LLVM claims to support atomic loads and stores of integers, pointers and floating point types, but certain backends cannot deal with floats.

Mimic what Clang does - bitcast to/from an equal width integer.

Bug: b/136037244
Change-Id: I40167279de96f586414f4fb406bf100432a39abf
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33408
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Chris Forbes <chrisforbes@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Ben Clayton <bclayton@google.com>

2fa1dad2

Make the ICD manifest use a relative path to the library · a0b3a026

authored Jun 24, 2019

Change-Id: I61aff0500bfeafd00a0a2b4ba8df2cb8c1f1e282
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33217
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Paul Thomson <paulthomson@google.com>

a0b3a026

Don't treat class-memaccess warning as error · 84b037e0

authored Jun 25, 2019

Cache keys must be guaranteed to have no uninitialized bits so that they
can be compared with memcmpy() and hash values can be computed from
their raw memory representation. We do this by calling memset(this, ...)
in the Memset<T> class used as the first base class. This is safe since
no other constructor has run yet (similar to doing placement new in
zeroed out memory). GCC 8.2's class-memset warning should thus not be
treated as an error.

Bug: b/135744933
Bug: b/134932616
Change-Id: I378b663281b1737359330c29c8c73934e7dd20f7
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33368Tested-by: Nicolas Capens <nicolascapens@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Alexis Hétu <sugoi@google.com>

84b037e0

Remove incorrect and unused line to avoid confusion · 1f456938

authored Jun 21, 2019

Change-Id: I7c595a97f61cfef54ca2b0778b7eb906a002fade
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33252
Presubmit-Ready: Alexis Hétu <sugoi@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Alexis Hétu <sugoi@google.com>

1f456938

25 Jun, 2019 2 commits

Implement Reactor breakpoints · 9770a466

authored Jun 25, 2019

For LLVM, uses the debugtrap intrinsic which is more debugging specific
than the trap intrinsic. The Subzero implementationuses a Trap
intrinsic. Both translate to 'int 3' on x86.

Bug: b/135997638
Change-Id: I6eca71e74752c4646baeade9a39b338e17053336
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33369Reviewed-by: Ben Clayton <bclayton@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>

9770a466

Use correct alignment for input attachment loads · a25f5438

authored Jun 24, 2019

Previously SIMD::Load would assume that all pointers are 4-byte aligned.
However, texel pointers generated by SpirvShader::GetTexelAddress are
not aligned for small formats.

Add an alignment parameter to SIMD::Load and use it in EmitImageRead.

Bug: b/135954761
Change-Id: I6a420049e98f42a68960d557dee933fee9487af3
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33328Tested-by: Chris Forbes <chrisforbes@google.com>
Presubmit-Ready: Chris Forbes <chrisforbes@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>

a25f5438

24 Jun, 2019 7 commits

Refactor remaining cases of memset(this, ...) · e899321b

authored Jun 21, 2019

Use Memset<T> as the first base class of cache key types to ensure they
get initialized before any other base classes or members get
constructed.

Bug: b/134932616
Change-Id: I8f28252696d6e017db11da068180d2425cdf1d57
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33249Tested-by: Nicolas Capens <nicolascapens@google.com>
Presubmit-Ready: Nicolas Capens <nicolascapens@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>

e899321b

Fix initialization of LRU cache keys · 302a9723

authored Jun 21, 2019

In particular this fixes Blitter::Options state having uninitialized
bits after the bitfield, and Blitter::State having uninitialized padding
bytes after Options so 'sourceFormat' is 32-bit aligned.

Defines a Memset<T> class to be used as the first base class of cache
key types, which makes it explicit that their underlying memory will be
fully initialized before any member constructors are run.

Also adds is_memcmparable<T> for checking if memcmp() can be used to
implement operator==() for cache keys. It's equivalent to
std::is_trivially_copyable except it provides a fallback for STL
implementations that don't support it.

Bug: b/134932616
Change-Id: I6569e78b380e67aee02d5bfd39d80d210bd225fd
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/32929
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Tested-by: Nicolas Capens <nicolascapens@google.com>
Reviewed-by: Chris Forbes <chrisforbes@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>

302a9723

LLVMReactor: Bind double variants of trig / exp functions · b5f1a13d

authored Jun 24, 2019

These are used instead of the *f float forms on x86.
We should probably investigate why, but having LLVM attempting, failing and crashing to resolve functions under any circumstances is Not Good.

Bug: b/126873455
Change-Id: Ic31866ebea7deba2dc29a9a4fb54a457b13e622a
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33214Tested-by: Ben Clayton <bclayton@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

b5f1a13d

LLVMReactor: Remove "stack-probe-size" fudge, bind _chkstk extern · 2f58df3f

authored Jun 23, 2019

Fixes a number of dEQP tests that use over 4kB's of allocas on Win32.

Bug: b/135866359
Change-Id: I0e50eea6a70d284c8d653fb070f4ae982c02914c
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33213Tested-by: Ben Clayton <bclayton@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

2f58df3f

Reactor: Various build fixes for MSVC 2019 · 68a2901f

authored Jun 23, 2019

In MSVC 2019 __LINE__ is treated as a long, and will now error (C2397) if the literal is narrowed to the initializer type. Lots of people moaning that there's no way to disable this or treat it as a warning.

Another attempt at working around variadic macro bugs in MSVC. Previously, none of the RR_WATCH or RR_LOG macros would compile if used. They now work.

Change-Id: I05312df1d88167d44f2582eb6eb45a0085d09434
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33212Tested-by: Ben Clayton <bclayton@google.com>
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

68a2901f

LLVMReactor: Align the coroutine stack frame. · 40a885e0

authored Jun 23, 2019

This was being allocated with no particular alignment, causing spurious crashes when vectors were loaded or stored in the stack frame.

Bug: b/135691587
Change-Id: I0240fdeb08df17dbbc156e508013376e15c514b4
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33211
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>
Tested-by: Ben Clayton <bclayton@google.com>

40a885e0

Put LLVM IR verification behind a default-disabled flag · 5375f478

authored Jun 24, 2019

Verification can take upwards of 10% of a dEQP test's execution time.

Verification is enabled for Kokoro builds and unit tests as silently regressing would be a Bad Thing.

Verification is also enabled for debug (!NDEBUG) builds.

Bug: b/135609394
Change-Id: I6c354c40f524e1e8f3e08f24129b7e008163b4bd
Reviewed-on: https://swiftshader-review.googlesource.com/c/SwiftShader/+/33215
Kokoro-Presubmit: kokoro <noreply+kokoro@google.com>
Tested-by: Ben Clayton <bclayton@google.com>
Reviewed-by: Nicolas Capens <nicolascapens@google.com>

5375f478