Propagate OpLine to all applied instructions in spirv-opt by jaebaek · Pull Request #3951 · KhronosGroup/SPIRV-Tools

jaebaek · 2020-10-20T22:08:50Z

Based on the OpLine spec, an OpLine instruction must be applied to
the instructions physically following it up to the first occurrence
of the next end of block, the next OpLine instruction, or the next
OpNoLine instruction.

OpLine %file 0 0
OpNoLine
OpLine %file 1 1
OpStore %foo %int_1
%value = OpLoad %int %foo
OpLine %file 2 2

For the above code, the current spirv-opt keeps three line
instructions OpLine %file 0 0, OpNoLine, and OpLine %file 1 1
in std::vector<Instruction> dbg_line_insts_ of Instruction class
for OpStore %foo %int_1. It does not put any line instruction to
std::vector<Instruction> dbg_line_insts_ of
%value = OpLoad %int %foo even though OpLine %file 1 1 must be
applied to %value = OpLoad %int %foo based on the spec.

This results in the missing line information for
%value = OpLoad %int %foo while each spirv-opt pass optimizes the
code. We have to put OpLine %file 1 1 to
std::vector<Instruction> dbg_line_insts_ of
both %value = OpLoad %int %foo and OpStore %foo %int_1.

KhronosGroup/glslang#2440 is a related PR that stop using
PropagateLineInfoPass and RedundantLineInfoElimPass from glslang.
When the code in this PR applied, the glslang tests will pass.

Based on the OpLine spec, an OpLine instruction must be applied to the instructions physically following it up to the first occurrence of the next end of block, the next OpLine instruction, or the next OpNoLine instruction. The current implementation of the debug info preservation in each spirv-opt pass assumes that each instruction has OpLine instructions applied to it. Therefore, we have to assign the OpLine instruction that physically appears before an instruction to the instruction. The existing code does not assign the OpLine instruction to the instruction physically following it if it is not the next instruction of the OpLine instruction. This commit correctly assigns the OpLine instruction not only to the next instruction of the OpLine instruction but also instructions physically following it (up to the occurrence of one instruction that ends the effectiveness of the OpLine).

dnovillo

Nit/question: this is not actually a re-factoring, right? This is fixing the bug described in the commit message?

jaebaek · 2020-10-22T14:42:26Z

Ah yes. Let me update the title.

dnovillo

If I'm understanding this change properly, before we were adding OpLine markers to every contiguous instruction in a block and now, we simply emit it to the first instruction in the block?

To prevent optimizations from losing this association, each instruction keeps its own line number info in dbg_line_info_.

Is my understanding correct?

dnovillo · 2020-10-22T20:44:31Z

source/opt/ir_loader.cpp

+    if (dbg_line_info_.back().opcode() == SpvOpNoLine)
+      dbg_line_info_.clear();
+    else
+      dbg_line_info_.erase(dbg_line_info_.begin(), dbg_line_info_.end() - 1);


I'm not understanding why we're removing the first N - 1 elements of dbg_line_info_ here.

I wanted to keep the last line instruction because it must be applied to the next instruction.
After seeing your comment, I realized that I must keep the last line instruction using a separate variable, not dbg_line_info_ because processing the new OpLine will accumulate the new one in dbg_line_info_. It would be a bug. I updated the code to fix it.

jaebaek · 2020-10-23T17:51:16Z

before we were adding OpLine markers to every contiguous instruction in a block and now, we simply emit it to the first instruction in the block?

Before: we simply emit OpLine to the first instruction that is next to the OpLine.
Now: we add the last OpLine marker to every contiguous instruction. Only the last OpLine must be applied to following instructions based on the spec.

To prevent optimizations from losing this association, each instruction keeps its own line number info in dbg_line_info_.

Yes. That's the main point of this PR.

kuhar

I don't have much context here, but the patch seems fine to me.

source/opt/ir_loader.cpp

kuhar · 2020-10-23T17:54:15Z

test/opt/ir_loader_test.cpp

 namespace opt {
 namespace {

+const uint32_t kOpLineOperandLineIndex = 1;


nit: contexpr?

Done. It's weird that constexpr looks better than const for similar constants but spirv-tools does not use it. We intentionally do not use it for spirv-tools.

For a long time we supported VisualStudio 2013 which did not support constexpr.

``` OpLine %file 5 9 OpStore %foo %bar OpLine %file 5 9 // <-- we can skip emitting it OpReturn ``` We can skip the same OpLine instruction if it appeared before.

jaebaek · 2020-10-28T01:32:59Z

@greg-lunarg Hi Greg, I want to know your opinion about this PR.

The current spirv-opt has PropagateLineInfoPass and RedundantLineInfoElimPass. I realize that they must be just a part of the default IR loader and IR emission because the debug information preservation of the current spirv-opt without PropagateLineInfoPass and RedundantLineInfoElimPass lost some line information, which is critical for some simple shaders.

The high-level idea of this PR is the same with PropagateLineInfoPass and RedundantLineInfoElimPass but I slightly changed the logic to fix some bugs e.g., OpLine emission between OpLoopMerge and OpBranch.

If you agree that it would be better to merge them as a part of the default IR loader/emission, I want to remove PropagateLineInfoPass and RedundantLineInfoElimPass. Since I found them in glslang source code, I will prepare a PR to remove them from glslang as well. What do you think?

jaebaek · 2020-10-28T14:15:28Z

KhronosGroup/glslang#2440 is a related PR that stop using PropagateLineInfoPass and RedundantLineInfoElimPass from glslang.
When the code in this PR applied, the glslang tests pass.

greg-lunarg · 2020-10-28T18:55:08Z

The only concern I have with doing this by default is the additional compile time and space for someone who doesn't care if OpLine info is preserved. The philosophy of spirv-opt has been to keep default optimization to a minimum and let each user decide what they want/need.

Do you have any statistics on the time and size costs on a significantly sized SPIR-V file?

greg-lunarg · 2020-10-28T20:13:14Z

Update to my last comment: as this change is required to prevent loss of information, which is undesirable if not unacceptable, I think this makes sense as the default. I still think it would be prudent to check compile size and time costs, just in case it is out of line with with what would be an acceptable cost.

jaebaek · 2020-10-28T20:27:37Z

Thank you for the feedback @greg-lunarg . I will check the compiler time and the output SPIR-V file size difference using SaschaWillems shaders and get back to you.

jaebaek · 2020-10-28T22:03:19Z

I tested it using SaschaWillems shaders. I compared the spirv-opt built with this PR and the upstream one.
The file "opline.txt" is the result from the spirv-opt built with this PR.
The file "upstream.txt" is the one of the upstream.
"run.txt" is the shell script I ran.

In average:

time increases %6.66
output SPIR-V size increases %1.65

In worst case:

time increases 4.3 times
output SPIR-V size increases %6.63

I simply got (user time for the one with this PR) / (user time for the upstream) and the same for the size except the divisor 0.

Note that the time measurement is inaccurate (because of the fluctuation?).
Sometimes the one built with this PR has a better performance (354 among 776).

In my opinion, the performance degradation looks acceptable.

opline.txt
upstream.txt
run.txt

greg-lunarg · 2020-10-28T22:39:39Z

Agree. I think the compile costs are ok. Thanks!

kuhar · 2020-10-28T22:40:16Z

@jaebaek SW shaders are fairly small, aren't they? Would it be more representative to also run it on some real-world shaders from a large game?

kuhar · 2020-10-28T22:42:21Z

Since you are comparing very close binaries (spirv-tools before and after your patch), it may also make more sense to measure the number of executed instructions instead of cpu time.

jaebaek · 2020-10-29T16:21:56Z

OK. I tested it with shaders of a real world game. I randomly selected 496 shaders.
I ran the previous SW experiment on Linux but for this experiment I run it on Windows with Visual Studio build.
My script executes DXC with/without this PR and it enables the legalization for all the shaders.
I measure the DXC execution time, which includes the execution time of spirv-opt.

Since the upstream DXC debug info generation (without spirv-opt) has segfaults for many shaders (I will work on it but it will be another big project to make the DXC generate all the debug info for the real game correctly .. the current implementation is prototype and our current goal is to enable debugging SW shaders), I tested it with/without -Zi option that emits OpLine without OpenCL.DebugInfo.100 instructions.

Note that -Zi instruction embeds the source code into the output SPIR-V, which increases the file size by ~10x.
DXC compilation of the real world game with -Zi fails for 4 shaders among those 496 shaders, which is not because of OpLine but the failures are caused by the long source code emission (they are validation errors for OpSourceContinued but I did not dig into them).

Compile time without OpLine emission
- The one with this PR is 0.335% slower than the upstream
- Definitely the file size is the same between them
Compile time with OpLine emission and preservation
- The one with this PR is 9.6% faster than the upstream. I believe this result just came from the flakyness.
- The file size increases 0.352% with this PR.
The OpLine emission and preservation do not degrade the compile time performance that much --> just 1.06% slower with this PR

Conclusion

The performance difference for the release version of the compilation i.e., without OpLine is very small (0.335%). Therefore, it is completely acceptable for the normal use case.
For the line information, 0.352% file size increment is very small as well.

I believe it is fine to add this PR in the view of the performance.

kuhar

Thanks a lot for running the experiments, the performance numbers on real-world shaders look much better than on SW. This is excellent data.

dneto0 · 2020-11-24T19:39:38Z

source/opt/module.h

+  // Pushes the binary segments for this instruction into the back of *|binary|
+  // including all OpLine and OpNoLine even if we can skip emitting some line
+  // instructions. If |skip_nop| is true and this is a OpNop, do nothing.
+  void ToBinaryWithAllOpLines(std::vector<uint32_t>* binary,


Looks like this is dead. I'll remove it.

Roll third_party/glslang/ 517f39e..142cb87 (35 commits) KhronosGroup/glslang@517f39e...142cb87 $ git log 517f39e..142cb87 --date=short --no-merges --format='%ad %ae %s' 2020-10-23 rex.xu HLSL: Add support for printf(). 2020-10-21 bclayton Fix GN build and presubmits 2020-10-20 john SPV: Update to the latest SPIR-V header, includes variable-rate shading 2020-07-02 laddoc Add GL_EXT_fragment_shading_rate 2020-10-20 bclayton Kokoro: Add configurations for GN presubmit 2020-10-19 bclayton Fix uninitialized use of TIntermediate::resource (KhronosGroup#2424) 2020-10-16 bclayton Add GN build instructions to README.md 2020-10-16 bclayton Add basic GN configurations 2020-10-12 hwguy.siplus SPIR-V: Remove SpvTools.h include from disassemble.cpp (KhronosGroup#2417) 2020-10-12 rverschelde Remove executable bits from code/data files (KhronosGroup#2420) 2020-10-07 dneto Add test case for read-only storage texture passed to helper function (KhronosGroup#2414) 2020-10-07 8729214+jonahryandavis Disable -Wno-conversion on MSVC compiler (KhronosGroup#2410) 2020-10-05 cepheus Revert "Add new SpirvToolsDisassemble API interface + Improve Doc on existing API interface (KhronosGroup#2408)" 2020-10-05 dev Add new SpirvToolsDisassemble API interface + Improve Doc on existing API interface (KhronosGroup#2408) 2020-09-27 cepheus Revert "Add more flexible SpirvToolsDisassemble interface to allow specifying spv_target_env for disassembly output. (KhronosGroup#2406)" 2020-09-27 dev Add more flexible SpirvToolsDisassemble interface to allow specifying spv_target_env for disassembly output. (KhronosGroup#2406) 2020-09-26 cstout [spirv-remap] Fix undefined behavior in hashing (KhronosGroup#2403) 2020-09-26 cstout [Wconversion] Suppress glslang issue (KhronosGroup#2404) 2020-09-24 greg Update spirv-tools and spirv-headers known goods (KhronosGroup#2401) 2020-09-18 rex.xu SPIRV: Add more utility functions to build some opcodes (KhronosGroup#2398) 2020-09-15 laddoc Preprocessor related issue fix (KhronosGroup#2378) 2020-09-14 rex.xu SPIRV: Add disassembly support for multiple literal strings (KhronosGroup#2397) 2020-09-14 laddoc Fix scope definition in ES 100. (KhronosGroup#2379) 2020-09-14 john Fix KhronosGroup#2385: guard against constant_id on non-const. 2020-09-12 shuizhuyuanluo Try to find python interpreter from host first 2020-09-11 ShabbyX Allow subpassLoad for ANGLE 2020-09-09 greg Add texture sample to nonuniform test 2020-09-08 greg Add buffer store to nonuniform tests 2020-09-03 bas SPV: Add NonUniform decoration for constructors. 2020-09-03 bas SPV: Add NonUniform decoration for OpImages created during lowering. 2020-08-24 bas SPV: Add NonUniform decorations for stores. 2020-09-08 tobias.hector Added missing copyright amendment 2020-09-07 rex.xu SPIRV: Fix some disassembly issues 2020-09-03 tobias.hector Error when initializing rayQuery with assignment 2020-09-02 rex.xu Parser: Fix wrong names of extension macros Created with: roll-dep third_party/glslang Roll third_party/googletest/ df6b75949..282877317 (41 commits) google/googletest@df6b759...2828773 $ git log df6b75949..282877317 --date=short --no-merges --format='%ad %ae %s' 2020-10-27 absl-team Googletest export 2020-10-26 absl-team Googletest export 2020-10-20 sonzogniarthur Fix typo "definedin in" => "defined in" 2020-10-15 absl-team Googletest export 2020-10-15 absl-team Googletest export 2020-10-15 dmauro Googletest export 2020-10-14 absl-team Googletest export 2020-10-14 dmauro Googletest export 2020-10-14 dmauro Googletest export 2020-10-14 absl-team Googletest export 2020-10-14 dmauro Googletest export 2020-10-14 absl-team Googletest export 2020-10-13 dmauro Googletest export 2020-10-13 dmauro Googletest export 2020-10-13 absl-team Googletest export 2020-10-13 absl-team Googletest export 2020-10-09 ofats Googletest export 2020-10-09 absl-team Googletest export 2020-10-08 absl-team Googletest export 2020-10-12 peternewman Fix a typo 2020-10-07 manavrion Improve FilePath::Normalize method 2020-10-07 pravin1992 Issue 2135: Change template args in NiceMock, NaggyMock and StrictMock from A1, A2, ... to TArg1, TArg2,... to avoid clash with legacy header files 2020-09-29 absl-team Googletest export 2020-10-01 63450189+ranodeepbanerjee A slight Gramatical change. 2020-09-29 dmauro Googletest export 2020-09-29 absl-team Googletest export 2020-09-25 absl-team Googletest export 2020-09-27 56075233+keshavgbpecdelhi Update cook_book.md 2020-09-23 absl-team Googletest export 2020-09-23 absl-team Googletest export 2020-09-21 absl-team Googletest export 2020-09-24 thomas.barbier Fix warning maybe-uninitialized 2020-09-18 absl-team Googletest export 2020-09-17 absl-team Googletest export 2020-09-18 63900998+JethroSama Update README.md, added missing 'a' 2020-09-08 absl-team Googletest export 2020-09-02 dmauro Googletest export 2020-09-01 absl-team Googletest export 2020-09-01 absl-team Googletest export 2020-08-25 27jf Add timestamp to in old method mock macro guide 2020-05-05 igor.n.nazarenko Detect proto messages based on presense of DebugString. Created with: roll-dep third_party/googletest Roll third_party/re2/ ca11026a0..166dbbeb3 (25 commits) google/re2@ca11026...166dbbe $ git log ca11026a0..166dbbeb3 --date=short --no-merges --format='%ad %ae %s' 2020-10-26 junyer Fix symbol visibility and add test coverage. 2020-10-13 junyer Get the conditional right this time. Sigh. 2020-10-13 junyer Don't support ParseFrom() on MSVC. It can cause ICEs. 2020-10-09 junyer Ensure that RE2::Arg works even with overloaded ParseFrom(). 2020-10-08 junyer Refactor the RE2::Arg templates for readability. 2020-10-07 junyer Rename namespace internal to namespace re2_internal. 2020-10-07 junyer Address `-Wunused-parameter' warnings. 2020-10-07 junyer Add missing #include. Mea culpa. 2020-10-07 junyer Rework RE2::Arg with templates instead of macros. 2020-10-06 junyer Write `typename' in templates rather than `class'. 2020-09-27 junyer Fix some indentation. 2020-09-27 junyer Set BAZELISK_GITHUB_TOKEN. 2020-09-25 junyer Point to the official Python wrapper. 2020-09-25 junyer Disable fail-fast in GitHub Actions. 2020-09-25 junyer Fix a template that will break with GCC 11.x. 2020-09-22 junyer Try to use the Clang packages instead. 2020-09-22 junyer Try to make the Clang containers work. 2020-09-22 junyer Configure a build matrix for Clang using containers. 2020-09-22 junyer Address `-Wclass-memaccess' warnings from GCC 10.x. 2020-09-22 junyer Configure a build matrix for GCC using containers. 2020-09-22 junyer Migrate from Kokoro to GitHub Actions for Bazel. 2020-09-22 junyer Migrate from Kokoro to GitHub Actions for CMake. 2020-09-22 junyer Remove Travis CI configuration. 2020-09-22 junyer Initial GitHub Actions CI configuration. 2020-09-22 junyer Improve the comments for RE2::FullMatch() et al. Created with: roll-dep third_party/re2 Roll third_party/spirv-headers/ 3fdabd0..7845730 (7 commits) KhronosGroup/SPIRV-Headers@3fdabd0...7845730 $ git log 3fdabd0..7845730 --date=short --no-merges --format='%ad %ae %s' 2020-10-23 john Bump revision to 4, for SPIR-V 1.5. 2020-10-19 TobyHector Add SPV_EXT_shader_image_int64 (KhronosGroup#170) 2020-10-19 TobyHector Added SPV_KHR_fragment_shading_rate (KhronosGroup#172) 2020-10-12 hwguy.siplus Register the Xenia emulator as a generator (KhronosGroup#171) 2020-09-27 atyuwen Register the Messiah SPIR-V CodeGen (KhronosGroup#169) 2020-09-10 syoussefi Register the ANGLE compiler (KhronosGroup#168) 2020-09-08 cepheus Rebuild of latest headers, which slightly moves OpTerminateInvocation Created with: roll-dep third_party/spirv-headers Roll third_party/spirv-tools/ 8a0ebd4..f7da527 (138 commits) KhronosGroup/SPIRV-Tools@8a0ebd4...f7da527 $ git log 8a0ebd4..f7da527 --date=short --no-merges --format='%ad %ae %s' 2020-10-30 jaebaek Temporarily add EmptyPass to prevent glslang from failing (KhronosGroup#4004) 2020-10-30 Junda.Liu spirv-opt: Add support to prevent functions from being inlined if they have DontInline flag (KhronosGroup#3858) 2020-10-29 jaebaek Propagate OpLine to all applied instructions in spirv-opt (KhronosGroup#3951) 2020-10-29 bclayton CMake: Add SPIRV_TOOLS_BUILD_STATIC flag (KhronosGroup#3910) 2020-10-29 dneto Avoid copying a ref in a loop (KhronosGroup#4000) 2020-10-28 justsid spirv-val: Allow the ViewportIndex and Layer built-ins on SPIR-V 1.5 (KhronosGroup#3986) 2020-10-28 dnovillo Simplify logic to decide whether CCP modified the IR (KhronosGroup#3997) 2020-10-27 jaebaek Add DebugValue for DebugDecl invisible to value assignment (KhronosGroup#3973) 2020-10-26 greg Fix bounds check instrumentation to handle 16-bit values (KhronosGroup#3983) 2020-10-23 andreperezmaselco.developer spirv-fuzz: Add expand vector reduction transformation (KhronosGroup#3869) 2020-10-23 vasniktel spirv-fuzz: Don't replace irrelevant indices in OpAccessChain (KhronosGroup#3988) 2020-10-23 vasniktel spirv-fuzz: Add FuzzerPassAddCompositeExtract (KhronosGroup#3904) 2020-10-22 afdx spirv-fuzz: Fix mismatch with shrinker step limit (KhronosGroup#3985) 2020-10-22 afdx spirv-fuzz: Fix off-by-one error in replayer (KhronosGroup#3982) 2020-10-22 afdx spirv-fuzz: Get order right for OpSelect arguments (KhronosGroup#3974) 2020-10-22 afdx spirv-fuzz: Do not add synonym-creating loops in dead blocks (KhronosGroup#3975) 2020-10-22 afdx spirv-fuzz: Skip OpTypeSampledImage when propagating up (KhronosGroup#3976) 2020-10-22 afdx spirv-fuzz: Pass OpUndef in function call if needed (KhronosGroup#3978) 2020-10-22 afdx spirv-fuzz: Fix off-by-one in TransformationCompositeConstruct (KhronosGroup#3979) 2020-10-22 afdx spirv-fuzz: Tolerate absent ids in data synonym fact management (KhronosGroup#3966) 2020-10-21 afdx spirv-fuzz: Fix to id availability (KhronosGroup#3971) 2020-10-21 afdx spirv-fuzz: Fix operand types (KhronosGroup#3962) 2020-10-21 8729214+jonahryandavis Update SPIRV-Headers revision in DEPS file (KhronosGroup#3961) 2020-10-21 afdx spirv-fuzz: Don't flatten conditional if condition is irrelevant (KhronosGroup#3944) 2020-10-21 afdx spirv-fuzz: Do not produce OpPhis of type OpTypeSampledImage (KhronosGroup#3964) 2020-10-21 afdx spirv-fuzz: Restrict fuzzer pass to reachable blocks (KhronosGroup#3970) 2020-10-21 afdx spirv-fuzz: Handle more types when extending OpPhi instructions (KhronosGroup#3969) 2020-10-21 afdx spirv-fuzz: Skip early terminator wrappers when merging returns (KhronosGroup#3968) 2020-10-21 afdx spirv-fuzz: Avoid irrelevant constants in synonym-creating loops (KhronosGroup#3967) 2020-10-21 afdx spirv-fuzz: Skip dead blocks in FuzzerPassAddOpPhiSynonyms (KhronosGroup#3965) 2020-10-21 afdx spirv-fuzz: Avoid the type manager when looking for struct types (KhronosGroup#3963) 2020-10-20 afdx spirv-fuzz: Fix to TransformationDuplicateRegionWithSelection (KhronosGroup#3941) 2020-10-20 afdx spirv-fuzz: Skip OpFunction when replacing irrelevant ids (KhronosGroup#3932) 2020-10-20 afdx spirv-fuzz: Use component-wise selectors when flattening conditional branches (KhronosGroup#3921) 2020-10-20 TobyHector Add SPV_EXT_shader_image_int64 (KhronosGroup#3852) 2020-10-20 TobyHector Support SPV_KHR_fragment_shading_rate (KhronosGroup#3943) 2020-10-19 afdx spirv-val: Fix validation of OpPhi instructions (KhronosGroup#3919) 2020-10-19 afdx spirv-fuzz: Avoid void struct member when outlining functions (KhronosGroup#3936) 2020-10-19 afdx spirv-fuzz: Do not allow Block-decorated structs when adding parameters (KhronosGroup#3931) 2020-10-19 afdx spirv-fuzz: Fix to operand id type (KhronosGroup#3937) 2020-10-19 afdx spirv-fuzz: Handle dead blocks in TransformationEquationInstruction (KhronosGroup#3933) 2020-10-19 afdx spirv-fuzz: Do not allow sampled image load when flattening conditionals (KhronosGroup#3930) 2020-10-19 afdx spirv-fuzz: Take care of OpPhi instructions when inlining (KhronosGroup#3939) 2020-10-16 afdx spirv-fuzz: Fix to TransformationInlineFunction (KhronosGroup#3913) 2020-10-16 afdx spirv-fuzz: Wrap early terminators before merging returns (KhronosGroup#3925) 2020-10-16 jaebaek Add DebugValue for function param regardless of scope (KhronosGroup#3923) 2020-10-16 afdx Temporary fix to make GoogleTest compile. (KhronosGroup#3922) 2020-10-15 afdx spirv-fuzz: Lower probability of adding bit instruction synonyms (KhronosGroup#3917) 2020-10-15 afdx spirv-fuzz: Fix handling of OpPhi in FlattenConditionalBranch (KhronosGroup#3916) 2020-10-13 afdx spirv-fuzz: Avoid creating blocks without parents (KhronosGroup#3908) (...) 2020-09-24 ehsannas Start SPIRV-Tools v2020.6 2020-09-24 ehsannas Finalize SPIRV-Tools v2020.5 2020-09-24 ehsannas Update CHANGES 2020-09-24 vasniktel spirv-fuzz: Support dead blocks in TransformationAddSynonym (KhronosGroup#3832) 2020-09-24 vasniktel spirv-fuzz: Move IRContext parameter into constructor (KhronosGroup#3837) 2020-09-24 Simran-B Add missing backticks around <result-id> (KhronosGroup#3840) 2020-09-23 rharrison Validate SPIRV Version number when parsing binary header (KhronosGroup#3834) 2020-09-23 stefanomil spirv-fuzz: Create synonym of int constant using a loop (KhronosGroup#3790) 2020-09-22 58573781+richard-lunarg Fix compiler error on macOS with XCode12 (KhronosGroup#3836) 2020-09-22 vasniktel spirv-fuzz: Handle OpPhis in TransformationInlineFunction (KhronosGroup#3833) 2020-09-22 stevenperron Update CHANGES 2020-09-22 afdx spirv-fuzz: Refactor fuzzer, replayer and shrinker (KhronosGroup#3818) 2020-09-18 afdx spirv-fuzz: Add pass recommendations (KhronosGroup#3757) 2020-09-18 stefanomil spirv-fuzz: Consider all ids from dead blocks irrelevant (KhronosGroup#3795) 2020-09-18 afdx Fix header guard macros (KhronosGroup#3811) 2020-09-18 antonikarp spirv-fuzz: Fix TransformationDuplicateRegionWithSelection (KhronosGroup#3815) 2020-09-17 46493288+sfricke-samsung spirv-val: Add DeviceIndex (KhronosGroup#3812) 2020-09-16 rharrison Fix missed modification flagging (KhronosGroup#3814) 2020-09-16 andreperezmaselco.developer spirv-fuzz: Use an irrelevant id for the unused components (KhronosGroup#3810) 2020-09-16 stefanomil spirv-fuzz: Improvements to random number generation (KhronosGroup#3809) 2020-09-16 greg Add buffer oob check to bindless instrumentation (KhronosGroup#3800) 2020-09-16 vasniktel spirv-fuzz: Remove CanFindOrCreateZeroConstant (KhronosGroup#3807) 2020-09-15 andreperezmaselco.developer spirv-fuzz: Add bit instruction synonym transformation (KhronosGroup#3775) 2020-09-16 vasniktel spirv-fuzz: Skip unreachable blocks (KhronosGroup#3729) 2020-09-15 afdx Fix build errors (KhronosGroup#3804) 2020-09-15 vasniktel spirv-fuzz: Handle invalid ids in fact manager (KhronosGroup#3742) 2020-09-15 vasniktel spirv-fuzz: Support memory instructions MoveInstructionDown (KhronosGroup#3700) 2020-09-15 stefanomil spirv-fuzz: Pass submanagers to other submanagers when necessary (KhronosGroup#3796) 2020-09-15 stefanomil spirv-fuzz: Transformation to flatten conditional branch (KhronosGroup#3667) 2020-09-14 46493288+sfricke-samsung spirv-val: Add BaseInstance, BaseVertex, DrawIndex, and ViewIndex (KhronosGroup#3782) 2020-09-14 dnovillo Properly mark IR changed if instruction folder creates more than one constant. (KhronosGroup#3799) 2020-09-11 afdx Add missing file to BUILD.gn (KhronosGroup#3798) 2020-09-11 antonikarp spirv-fuzz: Add TransformationDuplicateRegionWithSelection (KhronosGroup#3773) 2020-09-11 afdx spirv-reduce: Support reducing a specific function (KhronosGroup#3774) 2020-09-10 afdx spirv-reduce: Refactoring (KhronosGroup#3793) 2020-09-10 afdx Favour 'integrity' over 'coherence' as a replacement for 'sanity'. (KhronosGroup#3619) 2020-09-10 antonikarp spirv-fuzz: Fix header guards in transformations/fuzzer passes (KhronosGroup#3784) 2020-09-10 paulthomson spirv-fuzz: Add SPIRV_FUZZ_PROTOC_COMMAND (KhronosGroup#3789) 2020-09-10 paulthomson Add missing include (KhronosGroup#3788) 2020-09-09 paulthomson Improve spirv-fuzz CMake code (KhronosGroup#3781) 2020-09-08 stevenperron Allow SPV_KHR_8bit_storage extension. (KhronosGroup#3780) 2020-09-08 stefanomil spirv-opt: Add function to compute nesting depth of a block (KhronosGroup#3771) 2020-09-03 stefanomil spirv-fuzz: Transformation to convert OpSelect to conditional branch (KhronosGroup#3681) 2020-09-02 46493288+sfricke-samsung spirv-val: Add Vulkan VUID labels to BuiltIn (KhronosGroup#3756) 2020-09-02 vasniktel spirv-fuzz: Add support for BuiltIn decoration (KhronosGroup#3736) 2020-09-02 stefanomil spirv-fuzz: Fix GetIdEquivalenceClasses (KhronosGroup#3767) 2020-09-02 stefanomil spirv-fuzz: Replace id in OpPhi coming from a dead predecessor (KhronosGroup#3744) 2020-09-01 stefanomil spirv-fuzz: Transformation to replace the use of an irrelevant id (KhronosGroup#3697) 2020-09-01 vasniktel spirv-fuzz: TransformationMutatePointer (KhronosGroup#3737) 2020-09-01 stefanomil spirv-fuzz: Compute interprocedural loop nesting depth of blocks (KhronosGroup#3753) Created with: roll-dep third_party/spirv-tools

jaebaek added the component:optimization label Oct 20, 2020

jaebaek requested review from ben-clayton, dnovillo and kuhar October 20, 2020 22:08

jaebaek self-assigned this Oct 20, 2020

jaebaek requested a review from dneto0 October 21, 2020 15:49

dnovillo reviewed Oct 22, 2020

View reviewed changes

jaebaek changed the title ~~Refactor handling OpLine in spirv-opt~~ Distribute OpLine to all applied instructions Oct 22, 2020

jaebaek changed the title ~~Distribute OpLine to all applied instructions~~ Distribute OpLine to all applied instructions in spirv-opt Oct 22, 2020

dnovillo reviewed Oct 22, 2020

View reviewed changes

Correctly keep the last effective line instruction

31a9020

kuhar approved these changes Oct 23, 2020

View reviewed changes

jaebaek added 4 commits October 26, 2020 13:51

Skip emitting dead OpLine and OpNoLine instructions

a6184cf

``` OpLine %file 5 9 OpStore %foo %bar OpLine %file 5 9 // <-- we can skip emitting it OpReturn ``` We can skip the same OpLine instruction if it appeared before.

Support emit all line instructions for debugging purpose

4e93eb8

Code review

7f13269

Update unit test for process-lines-pass

52a627f

jaebaek force-pushed the opline_refactoring branch from ad7eb3b to 52a627f Compare October 27, 2020 20:09

jaebaek changed the title ~~Distribute OpLine to all applied instructions in spirv-opt~~ Propagate OpLine to all applied instructions in spirv-opt Oct 28, 2020

Remove PropagateLineInfoPass and RedundantLineInfoElimPass

3c8e197

jaebaek mentioned this pull request Oct 28, 2020

Do not use PropagateLineInfoPass and RedundantLineInfoElimPass KhronosGroup/glslang#2440

Merged

kuhar approved these changes Oct 29, 2020

View reviewed changes

jaebaek merged commit 56d0f50 into KhronosGroup:master Oct 29, 2020

jaebaek deleted the opline_refactoring branch October 30, 2020 14:47

dneto0 mentioned this pull request Nov 23, 2020

spirv-opt module builder injects extra OpLine instructions that don't exist on the input #4029

Closed

dneto0 reviewed Nov 24, 2020

View reviewed changes

Conversation

jaebaek commented Oct 20, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

dnovillo left a comment

Choose a reason for hiding this comment

Uh oh!

jaebaek commented Oct 22, 2020

Uh oh!

dnovillo left a comment

Choose a reason for hiding this comment

Uh oh!

dnovillo Oct 22, 2020

Choose a reason for hiding this comment

Uh oh!

jaebaek Oct 23, 2020

Choose a reason for hiding this comment

Uh oh!

jaebaek commented Oct 23, 2020

Uh oh!

kuhar left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

kuhar Oct 23, 2020

Choose a reason for hiding this comment

Uh oh!

jaebaek Oct 26, 2020

Choose a reason for hiding this comment

Uh oh!

dneto0 Nov 23, 2020

Choose a reason for hiding this comment

Uh oh!

jaebaek commented Oct 28, 2020

Uh oh!

jaebaek commented Oct 28, 2020

Uh oh!

greg-lunarg commented Oct 28, 2020

Uh oh!

greg-lunarg commented Oct 28, 2020

Uh oh!

jaebaek commented Oct 28, 2020

Uh oh!

jaebaek commented Oct 28, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

greg-lunarg commented Oct 28, 2020

Uh oh!

kuhar commented Oct 28, 2020

Uh oh!

kuhar commented Oct 28, 2020

Uh oh!

jaebaek commented Oct 29, 2020

Conclusion

Uh oh!

kuhar left a comment

Choose a reason for hiding this comment

Uh oh!

dneto0 Nov 24, 2020

Choose a reason for hiding this comment

Uh oh!

dneto0 Nov 24, 2020

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

jaebaek commented Oct 20, 2020 •

edited

Loading

jaebaek commented Oct 28, 2020 •

edited

Loading