Updated spirv-tools.

This commit is contained in:
Бранимир Караџић 2019-01-18 17:39:04 -08:00
parent 7d149d9105
commit 07191eb38b
30 changed files with 1579 additions and 123 deletions

View File

@ -80,6 +80,7 @@ SPVTOOLS_OPT_SRC_FILES := \
source/opt/cfg.cpp \
source/opt/cfg_cleanup_pass.cpp \
source/opt/ccp_pass.cpp \
source/opt/code_sink.cpp \
source/opt/combine_access_chains.cpp \
source/opt/common_uniform_elim_pass.cpp \
source/opt/compact_ids_pass.cpp \

View File

@ -460,6 +460,8 @@ static_library("spvtools_opt") {
"source/opt/cfg.h",
"source/opt/cfg_cleanup_pass.cpp",
"source/opt/cfg_cleanup_pass.h",
"source/opt/code_sink.cpp",
"source/opt/code_sink.h",
"source/opt/combine_access_chains.cpp",
"source/opt/combine_access_chains.h",
"source/opt/common_uniform_elim_pass.cpp",

View File

@ -1 +1 @@
"v2019.2-dev", "SPIRV-Tools v2019.2-dev 3c8b3c81ef7f378717ef88916c06cb221eea890b"
"v2019.2-dev", "SPIRV-Tools v2019.2-dev 7d149d91055744e5a9bc6368de7b3bb605d0e1a9"

View File

@ -722,6 +722,10 @@ Optimizer::PassToken CreateInstBindlessCheckPass(uint32_t desc_set,
// conform to that model's requirements.
Optimizer::PassToken CreateUpgradeMemoryModelPass();
// Create a pass to do code sinking. Code sinking is a transformation
// where an instruction is moved into a more deeply nested construct.
Optimizer::PassToken CreateCodeSinkingPass();
} // namespace spvtools
#endif // INCLUDE_SPIRV_TOOLS_OPTIMIZER_HPP_

View File

@ -19,6 +19,7 @@ set(SPIRV_TOOLS_OPT_SOURCES
ccp_pass.h
cfg_cleanup_pass.h
cfg.h
code_sink.h
combine_access_chains.h
common_uniform_elim_pass.h
compact_ids_pass.h
@ -111,6 +112,7 @@ set(SPIRV_TOOLS_OPT_SOURCES
ccp_pass.cpp
cfg_cleanup_pass.cpp
cfg.cpp
code_sink.cpp
combine_access_chains.cpp
common_uniform_elim_pass.cpp
compact_ids_pass.cpp

View File

@ -0,0 +1,316 @@
// Copyright (c) 2019 Google LLC
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "code_sink.h"
#include <set>
#include <vector>
#include "source/opt/instruction.h"
#include "source/opt/ir_builder.h"
#include "source/opt/ir_context.h"
#include "source/util/bit_vector.h"
namespace spvtools {
namespace opt {
Pass::Status CodeSinkingPass::Process() {
bool modified = false;
for (Function& function : *get_module()) {
cfg()->ForEachBlockInPostOrder(function.entry().get(),
[&modified, this](BasicBlock* bb) {
if (SinkInstructionsInBB(bb)) {
modified = true;
}
});
}
return modified ? Status::SuccessWithChange : Status::SuccessWithoutChange;
}
bool CodeSinkingPass::SinkInstructionsInBB(BasicBlock* bb) {
bool modified = false;
for (auto inst = bb->rbegin(); inst != bb->rend(); ++inst) {
if (SinkInstruction(&*inst)) {
inst = bb->rbegin();
modified = true;
}
}
return modified;
}
bool CodeSinkingPass::SinkInstruction(Instruction* inst) {
if (inst->opcode() != SpvOpLoad && inst->opcode() != SpvOpAccessChain) {
return false;
}
if (ReferencesMutableMemory(inst)) {
return false;
}
if (BasicBlock* target_bb = FindNewBasicBlockFor(inst)) {
Instruction* pos = &*target_bb->begin();
while (pos->opcode() == SpvOpPhi) {
pos = pos->NextNode();
}
inst->InsertBefore(pos);
context()->set_instr_block(inst, target_bb);
return true;
}
return false;
}
BasicBlock* CodeSinkingPass::FindNewBasicBlockFor(Instruction* inst) {
assert(inst->result_id() != 0 && "Instruction should have a result.");
BasicBlock* original_bb = context()->get_instr_block(inst);
BasicBlock* bb = original_bb;
std::unordered_set<uint32_t> bbs_with_uses;
get_def_use_mgr()->ForEachUse(
inst, [&bbs_with_uses, this](Instruction* use, uint32_t idx) {
if (use->opcode() != SpvOpPhi) {
bbs_with_uses.insert(context()->get_instr_block(use)->id());
} else {
bbs_with_uses.insert(use->GetSingleWordOperand(idx + 1));
}
});
while (true) {
// If |inst| is used in |bb|, then |inst| cannot be moved any further.
if (bbs_with_uses.count(bb->id())) {
break;
}
// If |bb| has one successor (succ_bb), and |bb| is the only predecessor
// of succ_bb, then |inst| can be moved to succ_bb. If succ_bb, has move
// then one predecessor, then moving |inst| into succ_bb could cause it to
// be executed more often, so the search has to stop.
if (bb->terminator()->opcode() == SpvOpBranch) {
uint32_t succ_bb_id = bb->terminator()->GetSingleWordInOperand(0);
if (cfg()->preds(succ_bb_id).size() == 1) {
bb = context()->get_instr_block(succ_bb_id);
continue;
} else {
break;
}
}
// The remaining checks need to know the merge node. If there is no merge
// instruction or an OpLoopMerge, then it is a break or continue. We could
// figure it out, but not worth doing it now.
Instruction* merge_inst = bb->GetMergeInst();
if (merge_inst == nullptr || merge_inst->opcode() != SpvOpSelectionMerge) {
break;
}
// Check all of the successors of |bb| it see which lead to a use of |inst|
// before reaching the merge node.
bool used_in_multiple_blocks = false;
uint32_t bb_used_in = 0;
bb->ForEachSuccessorLabel([this, bb, &bb_used_in, &used_in_multiple_blocks,
&bbs_with_uses](uint32_t* succ_bb_id) {
if (IntersectsPath(*succ_bb_id, bb->MergeBlockIdIfAny(), bbs_with_uses)) {
if (bb_used_in == 0) {
bb_used_in = *succ_bb_id;
} else {
used_in_multiple_blocks = true;
}
}
});
// If more than one successor, which is not the merge block, uses |inst|
// then we have to leave |inst| in bb because there is none of the
// successors dominate all uses of |inst|.
if (used_in_multiple_blocks) {
break;
}
if (bb_used_in == 0) {
// If |inst| is not used before reaching the merge node, then we can move
// |inst| to the merge node.
bb = context()->get_instr_block(bb->MergeBlockIdIfAny());
} else {
// If the only successor that leads to a used of |inst| has more than 1
// predecessor, then moving |inst| could cause it to be executed more
// often, so we cannot move it.
if (cfg()->preds(bb_used_in).size() != 1) {
break;
}
// If |inst| is used after the merge block, then |bb_used_in| does not
// dominate all of the uses. So we cannot move |inst| any further.
if (IntersectsPath(bb->MergeBlockIdIfAny(), original_bb->id(),
bbs_with_uses)) {
break;
}
// Otherwise, |bb_used_in| dominates all uses, so move |inst| into that
// block.
bb = context()->get_instr_block(bb_used_in);
}
continue;
}
return (bb != original_bb ? bb : nullptr);
}
bool CodeSinkingPass::ReferencesMutableMemory(Instruction* inst) {
if (!inst->IsLoad()) {
return false;
}
Instruction* base_ptr = inst->GetBaseAddress();
if (base_ptr->opcode() != SpvOpVariable) {
return true;
}
if (base_ptr->IsReadOnlyVariable()) {
return false;
}
if (HasUniformMemorySync()) {
return true;
}
if (base_ptr->GetSingleWordInOperand(0) != SpvStorageClassUniform) {
return true;
}
return HasPossibleStore(base_ptr);
}
bool CodeSinkingPass::HasUniformMemorySync() {
if (checked_for_uniform_sync_) {
return has_uniform_sync_;
}
bool has_sync = false;
get_module()->ForEachInst([this, &has_sync](Instruction* inst) {
switch (inst->opcode()) {
case SpvOpMemoryBarrier: {
uint32_t mem_semantics_id = inst->GetSingleWordInOperand(1);
if (IsSyncOnUniform(mem_semantics_id)) {
has_sync = true;
}
break;
}
case SpvOpControlBarrier:
case SpvOpAtomicLoad:
case SpvOpAtomicStore:
case SpvOpAtomicExchange:
case SpvOpAtomicIIncrement:
case SpvOpAtomicIDecrement:
case SpvOpAtomicIAdd:
case SpvOpAtomicISub:
case SpvOpAtomicSMin:
case SpvOpAtomicUMin:
case SpvOpAtomicSMax:
case SpvOpAtomicUMax:
case SpvOpAtomicAnd:
case SpvOpAtomicOr:
case SpvOpAtomicXor:
case SpvOpAtomicFlagTestAndSet:
case SpvOpAtomicFlagClear: {
uint32_t mem_semantics_id = inst->GetSingleWordInOperand(2);
if (IsSyncOnUniform(mem_semantics_id)) {
has_sync = true;
}
break;
}
case SpvOpAtomicCompareExchange:
case SpvOpAtomicCompareExchangeWeak:
if (IsSyncOnUniform(inst->GetSingleWordInOperand(2)) ||
IsSyncOnUniform(inst->GetSingleWordInOperand(3))) {
has_sync = true;
}
break;
default:
break;
}
});
has_uniform_sync_ = has_sync;
return has_sync;
}
bool CodeSinkingPass::IsSyncOnUniform(uint32_t mem_semantics_id) const {
const analysis::Constant* mem_semantics_const =
context()->get_constant_mgr()->FindDeclaredConstant(mem_semantics_id);
assert(mem_semantics_const != nullptr &&
"Expecting memory semantics id to be a constant.");
assert(mem_semantics_const->AsIntConstant() &&
"Memory semantics should be an integer.");
uint32_t mem_semantics_int = mem_semantics_const->GetU32();
// If it does not affect uniform memory, then it is does not apply to uniform
// memory.
if ((mem_semantics_int & SpvMemorySemanticsUniformMemoryMask) == 0) {
return false;
}
// Check if there is an acquire or release. If so not, this it does not add
// any memory constraints.
return (mem_semantics_int & (SpvMemorySemanticsAcquireMask |
SpvMemorySemanticsAcquireReleaseMask |
SpvMemorySemanticsReleaseMask)) != 0;
}
bool CodeSinkingPass::HasPossibleStore(Instruction* var_inst) {
assert(var_inst->opcode() == SpvOpVariable ||
var_inst->opcode() == SpvOpAccessChain ||
var_inst->opcode() == SpvOpPtrAccessChain);
return get_def_use_mgr()->WhileEachUser(var_inst, [this](Instruction* use) {
switch (use->opcode()) {
case SpvOpStore:
return true;
case SpvOpAccessChain:
case SpvOpPtrAccessChain:
return HasPossibleStore(use);
default:
return false;
}
});
}
bool CodeSinkingPass::IntersectsPath(uint32_t start, uint32_t end,
const std::unordered_set<uint32_t>& set) {
std::vector<uint32_t> worklist;
worklist.push_back(start);
std::unordered_set<uint32_t> already_done;
already_done.insert(start);
while (!worklist.empty()) {
BasicBlock* bb = context()->get_instr_block(worklist.back());
worklist.pop_back();
if (bb->id() == end) {
continue;
}
if (set.count(bb->id())) {
return true;
}
bb->ForEachSuccessorLabel([&already_done, &worklist](uint32_t* succ_bb_id) {
if (already_done.insert(*succ_bb_id).second) {
worklist.push_back(*succ_bb_id);
}
});
}
return false;
}
// namespace opt
} // namespace opt
} // namespace spvtools

View File

@ -0,0 +1,107 @@
// Copyright (c) 2019 Google LLC
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#ifndef SOURCE_OPT_CODE_SINK_H_
#define SOURCE_OPT_CODE_SINK_H_
#include <unordered_map>
#include "source/opt/ir_context.h"
#include "source/opt/module.h"
#include "source/opt/pass.h"
namespace spvtools {
namespace opt {
// This pass does code sinking for OpAccessChain and OpLoad on variables in
// uniform storage or in read only memory. Code sinking is a transformation
// where an instruction is moved into a more deeply nested construct.
//
// The goal is to move these instructions as close as possible to their uses
// without having to execute them more often or to replicate the instruction.
// Moving the instruction in this way can lead to shorter live ranges, which can
// lead to less register pressure. It can also cause instructions to be
// executed less often because they could be moved into one path of a selection
// construct.
//
// This optimization can cause register pressure to rise if the operands of the
// instructions go dead after the instructions being moved. That is why we only
// move certain OpLoad and OpAccessChain instructions. They generally have
// constants, loop induction variables, and global pointers as operands. The
// operands are live for a longer time in most cases.
class CodeSinkingPass : public Pass {
public:
const char* name() const override { return "code-sink"; }
Status Process() override;
// Return the mask of preserved Analyses.
IRContext::Analysis GetPreservedAnalyses() override {
return IRContext::kAnalysisDefUse |
IRContext::kAnalysisInstrToBlockMapping |
IRContext::kAnalysisCombinators | IRContext::kAnalysisCFG |
IRContext::kAnalysisDominatorAnalysis |
IRContext::kAnalysisLoopAnalysis | IRContext::kAnalysisNameMap |
IRContext::kAnalysisConstants | IRContext::kAnalysisTypes;
}
private:
// Sinks the instructions in |bb| as much as possible. Returns true if
// something changes.
bool SinkInstructionsInBB(BasicBlock* bb);
// Tries the sink |inst| as much as possible. Returns true if the instruction
// is moved.
bool SinkInstruction(Instruction* inst);
// Returns the basic block in which to move |inst| to move is as close as
// possible to the uses of |inst| without increasing the number of times
// |inst| will be executed. Return |nullptr| if there is no need to move
// |inst|.
BasicBlock* FindNewBasicBlockFor(Instruction* inst);
// Return true if |inst| reference memory and it is possible that the data in
// the memory changes at some point.
bool ReferencesMutableMemory(Instruction* inst);
// Returns true if the module contains an instruction that has a memory
// semantics id as an operand, and the memory semantics enforces a
// synchronization of uniform memory. See section 3.25 of the SPIR-V
// specification.
bool HasUniformMemorySync();
// Returns true if there may be a store to the variable |var_inst|.
bool HasPossibleStore(Instruction* var_inst);
// Returns true if one of the basic blocks in |set| exists on a path from the
// basic block |start| to |end|.
bool IntersectsPath(uint32_t start, uint32_t end,
const std::unordered_set<uint32_t>& set);
// Returns true if |mem_semantics_id| is the id of a constant that, when
// interpreted as a memory semantics mask enforces synchronization of uniform
// memory. See section 3.25 of the SPIR-V specification.
bool IsSyncOnUniform(uint32_t mem_semantics_id) const;
// True if a check has for uniform storage has taken place.
bool checked_for_uniform_sync_;
// Cache of whether or not the module has a memory sync on uniform storage.
// only valid if |check_for_uniform_sync_| is true.
bool has_uniform_sync_;
};
} // namespace opt
} // namespace spvtools
#endif // SOURCE_OPT_CODE_SINK_H_

View File

@ -45,8 +45,13 @@ namespace {
uint32_t InstructionFolder::UnaryOperate(SpvOp opcode, uint32_t operand) const {
switch (opcode) {
// Arthimetics
case SpvOp::SpvOpSNegate:
return -static_cast<int32_t>(operand);
case SpvOp::SpvOpSNegate: {
int32_t s_operand = static_cast<int32_t>(operand);
if (s_operand == std::numeric_limits<int32_t>::min()) {
return s_operand;
}
return -s_operand;
}
case SpvOp::SpvOpNot:
return ~operand;
case SpvOp::SpvOpLogicalNot:

View File

@ -127,6 +127,9 @@ void IRContext::InvalidateAnalyses(IRContext::Analysis analyses_to_invalidate) {
constant_mgr_.reset(nullptr);
}
if (analyses_to_invalidate & kAnalysisTypes) {
// The ConstantManager contains Type pointers. If the TypeManager goes
// away, the ConstantManager has to go away.
constant_mgr_.reset(nullptr);
type_mgr_.reset(nullptr);
}

View File

@ -937,22 +937,23 @@ void LoopDescriptor::PostModificationCleanup() {
for (Loop* loop : loops_to_remove_) {
loops_.erase(std::find(loops_.begin(), loops_.end(), loop));
delete loop;
}
for (auto& pair : loops_to_add_) {
Loop* parent = pair.first;
Loop* loop = pair.second;
std::unique_ptr<Loop> loop = std::move(pair.second);
if (parent) {
loop->SetParent(nullptr);
parent->AddNestedLoop(loop);
parent->AddNestedLoop(loop.get());
for (uint32_t block_id : loop->GetBlocks()) {
parent->AddBasicBlock(block_id);
}
}
loops_.emplace_back(loop);
loops_.emplace_back(loop.release());
}
loops_to_add_.clear();

View File

@ -499,8 +499,8 @@ class LoopDescriptor {
// Mark the loop |loop_to_add| as needing to be added when the user calls
// PostModificationCleanup. |parent| may be null.
inline void AddLoop(Loop* loop_to_add, Loop* parent) {
loops_to_add_.emplace_back(std::make_pair(parent, loop_to_add));
inline void AddLoop(std::unique_ptr<Loop>&& loop_to_add, Loop* parent) {
loops_to_add_.emplace_back(std::make_pair(parent, std::move(loop_to_add)));
}
// Checks all loops in |this| and will create pre-headers for all loops
@ -537,7 +537,9 @@ class LoopDescriptor {
// TODO(dneto): This should be a vector of unique_ptr. But VisualStudio 2013
// is unable to compile it.
using LoopContainerType = std::vector<Loop*>;
using LoopsToAddContainerType = std::vector<std::pair<Loop*, Loop*>>;
using LoopsToAddContainerType =
std::vector<std::pair<Loop*, std::unique_ptr<Loop>>>;
// Creates loop descriptors for the function |f|.
void PopulateList(IRContext* context, const Function* f);

View File

@ -394,12 +394,12 @@ void LoopUnrollerUtilsImpl::PartiallyUnrollResidualFactor(Loop* loop,
// This is a naked new due to the VS2013 requirement of not having unique
// pointers in vectors, as it will be inserted into a vector with
// loop_descriptor.AddLoop.
Loop* new_loop = new Loop(*loop);
std::unique_ptr<Loop> new_loop = MakeUnique<Loop>(*loop);
// Clear the basic blocks of the new loop.
new_loop->ClearBlocks();
DuplicateLoop(loop, new_loop);
DuplicateLoop(loop, new_loop.get());
// Add the blocks to the function.
AddBlocksToFunction(loop->GetMergeBlock());
@ -414,10 +414,10 @@ void LoopUnrollerUtilsImpl::PartiallyUnrollResidualFactor(Loop* loop,
loop_induction_variable_ = state_.new_phi;
// Unroll the new loop by the factor with the usual -1 to account for the
// existing block iteration.
Unroll(new_loop, factor);
Unroll(new_loop.get(), factor);
LinkLastPhisToStart(new_loop);
AddBlocksToLoop(new_loop);
LinkLastPhisToStart(new_loop.get());
AddBlocksToLoop(new_loop.get());
// Add the new merge block to the back of the list of blocks to be added. It
// needs to be the last block added to maintain dominator order in the binary.
@ -507,7 +507,7 @@ void LoopUnrollerUtilsImpl::PartiallyUnrollResidualFactor(Loop* loop,
LoopDescriptor& loop_descriptor = *context_->GetLoopDescriptor(&function_);
loop_descriptor.AddLoop(new_loop, loop->GetParent());
loop_descriptor.AddLoop(std::move(new_loop), loop->GetParent());
RemoveDeadInstructions();
}

View File

@ -21,6 +21,7 @@
#include <vector>
#include <source/spirv_optimizer_options.h>
#include "code_sink.h"
#include "source/opt/build_module.h"
#include "source/opt/log.h"
#include "source/opt/pass_manager.h"
@ -181,7 +182,8 @@ Optimizer& Optimizer::RegisterPerformancePasses() {
.RegisterPass(CreateRedundancyEliminationPass())
.RegisterPass(CreateDeadBranchElimPass())
.RegisterPass(CreateBlockMergePass())
.RegisterPass(CreateSimplificationPass());
.RegisterPass(CreateSimplificationPass())
.RegisterPass(CreateCodeSinkingPass());
// Currently exposing driver bugs resulting in crashes (#946)
// .RegisterPass(CreateCommonUniformElimPass())
}
@ -439,6 +441,8 @@ bool Optimizer::RegisterPassFromFlag(const std::string& flag) {
}
} else if (pass_name == "ccp") {
RegisterPass(CreateCCPPass());
} else if (pass_name == "code-sink") {
RegisterPass(CreateCodeSinkingPass());
} else if (pass_name == "O") {
RegisterPerformancePasses();
} else if (pass_name == "Os") {
@ -790,4 +794,9 @@ Optimizer::PassToken CreateInstBindlessCheckPass(uint32_t desc_set,
MakeUnique<opt::InstBindlessCheckPass>(desc_set, shader_id));
}
Optimizer::PassToken CreateCodeSinkingPass() {
return MakeUnique<Optimizer::PassToken::Impl>(
MakeUnique<opt::CodeSinkingPass>());
}
} // namespace spvtools

View File

@ -21,6 +21,7 @@
#include "source/opt/block_merge_pass.h"
#include "source/opt/ccp_pass.h"
#include "source/opt/cfg_cleanup_pass.h"
#include "source/opt/code_sink.h"
#include "source/opt/combine_access_chains.h"
#include "source/opt/common_uniform_elim_pass.h"
#include "source/opt/compact_ids_pass.h"

View File

@ -638,8 +638,8 @@ bool ScalarReplacementPass::CheckUses(const Instruction* inst,
switch (user->opcode()) {
case SpvOpAccessChain:
case SpvOpInBoundsAccessChain:
if (index == 2u) {
uint32_t id = user->GetSingleWordOperand(3u);
if (index == 2u && user->NumInOperands() > 1) {
uint32_t id = user->GetSingleWordInOperand(1u);
const Instruction* opInst = get_def_use_mgr()->GetDef(id);
if (!IsCompileTimeConstantInst(opInst->opcode())) {
ok = false;
@ -783,16 +783,6 @@ ScalarReplacementPass::GetUsedComponents(Instruction* inst) {
return false;
}
}
case SpvOpCopyObject: {
// Follow the copy to see which components are used.
auto t = GetUsedComponents(use);
if (!t) {
result.reset(nullptr);
return false;
}
result->insert(t->begin(), t->end());
return true;
}
default:
// We do not know what is happening. Have to assume the worst.
result.reset(nullptr);

View File

@ -66,7 +66,8 @@ void VectorDCE::FindLiveComponents(Function* function,
switch (current_inst->opcode()) {
case SpvOpCompositeExtract:
MarkExtractUseAsLive(current_inst, live_components, &work_list);
MarkExtractUseAsLive(current_inst, current_item.components,
live_components, &work_list);
break;
case SpvOpCompositeInsert:
MarkInsertUsesAsLive(current_item, live_components, &work_list);
@ -92,6 +93,7 @@ void VectorDCE::FindLiveComponents(Function* function,
}
void VectorDCE::MarkExtractUseAsLive(const Instruction* current_inst,
const utils::BitVector& live_elements,
LiveComponentMap* live_components,
std::vector<WorkListItem>* work_list) {
analysis::DefUseManager* def_use_mgr = context()->get_def_use_mgr();
@ -102,7 +104,11 @@ void VectorDCE::MarkExtractUseAsLive(const Instruction* current_inst,
if (HasVectorOrScalarResult(operand_inst)) {
WorkListItem new_item;
new_item.instruction = operand_inst;
new_item.components.Set(current_inst->GetSingleWordInOperand(1));
if (current_inst->NumInOperands() < 2) {
new_item.components = live_elements;
} else {
new_item.components.Set(current_inst->GetSingleWordInOperand(1));
}
AddItemToWorkListIfNeeded(new_item, live_components, work_list);
}
}

View File

@ -129,6 +129,7 @@ class VectorDCE : public MemPass {
// live. If anything becomes live they are added to |work_list| and
// |live_components| is updated accordingly.
void MarkExtractUseAsLive(const Instruction* current_inst,
const utils::BitVector& live_elements,
LiveComponentMap* live_components,
std::vector<WorkListItem>* work_list);

View File

@ -276,6 +276,16 @@ uint32_t GetMakeVisibleScope(const Instruction* inst, uint32_t mask) {
return scope_id;
}
bool DoesStructContainRTA(const ValidationState_t& _, const Instruction* inst) {
for (size_t member_index = 1; member_index < inst->operands().size();
++member_index) {
const auto member_id = inst->GetOperandAs<uint32_t>(member_index);
const auto member_type = _.FindDef(member_id);
if (member_type->opcode() == SpvOpTypeRuntimeArray) return true;
}
return false;
}
spv_result_t CheckMemoryAccess(ValidationState_t& _, const Instruction* inst,
uint32_t index) {
SpvStorageClass dst_sc, src_sc;
@ -283,9 +293,9 @@ spv_result_t CheckMemoryAccess(ValidationState_t& _, const Instruction* inst,
if (inst->operands().size() <= index) {
if (src_sc == SpvStorageClassPhysicalStorageBufferEXT ||
dst_sc == SpvStorageClassPhysicalStorageBufferEXT) {
return _.diag(SPV_ERROR_INVALID_ID, inst) << "Memory accesses with "
"PhysicalStorageBufferEXT "
"must use Aligned.";
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "Memory accesses with PhysicalStorageBufferEXT must use "
"Aligned.";
}
return SPV_SUCCESS;
}
@ -318,7 +328,7 @@ spv_result_t CheckMemoryAccess(ValidationState_t& _, const Instruction* inst,
if (!(mask & SpvMemoryAccessNonPrivatePointerKHRMask)) {
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "NonPrivatePointerKHR must be specified if "
"MakePointerVisibleKHR is specified.";
<< "MakePointerVisibleKHR is specified.";
}
// Check the associated scope for MakeVisibleKHR.
@ -335,8 +345,8 @@ spv_result_t CheckMemoryAccess(ValidationState_t& _, const Instruction* inst,
dst_sc != SpvStorageClassPhysicalStorageBufferEXT) {
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "NonPrivatePointerKHR requires a pointer in Uniform, "
"Workgroup, CrossWorkgroup, Generic, Image or StorageBuffer "
"storage classes.";
<< "Workgroup, CrossWorkgroup, Generic, Image or StorageBuffer "
<< "storage classes.";
}
if (src_sc != SpvStorageClassMax && src_sc != SpvStorageClassUniform &&
src_sc != SpvStorageClassWorkgroup &&
@ -346,17 +356,17 @@ spv_result_t CheckMemoryAccess(ValidationState_t& _, const Instruction* inst,
src_sc != SpvStorageClassPhysicalStorageBufferEXT) {
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "NonPrivatePointerKHR requires a pointer in Uniform, "
"Workgroup, CrossWorkgroup, Generic, Image or StorageBuffer "
"storage classes.";
<< "Workgroup, CrossWorkgroup, Generic, Image or StorageBuffer "
<< "storage classes.";
}
}
if (!(mask & SpvMemoryAccessAlignedMask)) {
if (src_sc == SpvStorageClassPhysicalStorageBufferEXT ||
dst_sc == SpvStorageClassPhysicalStorageBufferEXT) {
return _.diag(SPV_ERROR_INVALID_ID, inst) << "Memory accesses with "
"PhysicalStorageBufferEXT "
"must use Aligned.";
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "Memory accesses with PhysicalStorageBufferEXT must use "
"Aligned.";
}
}
@ -442,15 +452,15 @@ spv_result_t ValidateVariable(ValidationState_t& _, const Instruction* inst) {
// VariablePointersStorageBuffer is implied by VariablePointers.
if (pointee->opcode() == SpvOpTypePointer) {
if (!_.HasCapability(SpvCapabilityVariablePointersStorageBuffer)) {
return _.diag(SPV_ERROR_INVALID_ID, inst) << "In Logical addressing, "
"variables may not "
"allocate a pointer type";
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "In Logical addressing, variables may not allocate a pointer "
<< "type";
} else if (storage_class != SpvStorageClassFunction &&
storage_class != SpvStorageClassPrivate) {
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "In Logical addressing with variable pointers, variables "
"that allocate pointers must be in Function or Private "
"storage classes";
<< "that allocate pointers must be in Function or Private "
<< "storage classes";
}
}
}
@ -493,10 +503,9 @@ spv_result_t ValidateVariable(ValidationState_t& _, const Instruction* inst) {
<< "' has illegal type.\n"
<< "From Vulkan spec, section 14.5.2:\n"
<< "Variables identified with the Uniform storage class are "
"used "
<< "to access transparent buffer backed resources. Such "
"variables "
<< "must be typed as OpTypeStruct, or an array of this type";
<< "used to access transparent buffer backed resources. Such "
<< "variables must be typed as OpTypeStruct, or an array of "
<< "this type";
}
}
}
@ -564,13 +573,13 @@ spv_result_t ValidateVariable(ValidationState_t& _, const Instruction* inst) {
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "OpVariable " << inst->id()
<< ": expected AliasedPointerEXT or RestrictPointerEXT for "
"PhysicalStorageBufferEXT pointer.";
<< "PhysicalStorageBufferEXT pointer.";
}
if (foundAliased && foundRestrict) {
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "OpVariable " << inst->id()
<< ": can't specify both AliasedPointerEXT and "
"RestrictPointerEXT for PhysicalStorageBufferEXT pointer.";
<< "RestrictPointerEXT for PhysicalStorageBufferEXT pointer.";
}
}
}
@ -588,9 +597,9 @@ spv_result_t ValidateVariable(ValidationState_t& _, const Instruction* inst) {
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "OpVariable, <id> '" << _.getIdName(inst->id())
<< "', is attempting to create memory for an illegal type, "
"OpTypeRuntimeArray.\nFor Vulkan OpTypeRuntimeArray can only "
"appear as the final member of an OpTypeStruct, thus cannot "
"be instantiated via OpVariable";
<< "OpTypeRuntimeArray.\nFor Vulkan OpTypeRuntimeArray can only "
<< "appear as the final member of an OpTypeStruct, thus cannot "
<< "be instantiated via OpVariable";
} else {
// A bare variable OpTypeRuntimeArray is allowed in this context, but
// still need to check the storage class.
@ -599,8 +608,8 @@ spv_result_t ValidateVariable(ValidationState_t& _, const Instruction* inst) {
storage_class != SpvStorageClassUniformConstant) {
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "For Vulkan with RuntimeDescriptorArrayEXT, a variable "
"containing OpTypeRuntimeArray must have storage class of "
"StorageBuffer, Uniform, or UniformConstant.";
<< "containing OpTypeRuntimeArray must have storage class of "
<< "StorageBuffer, Uniform, or UniformConstant.";
}
}
}
@ -610,39 +619,64 @@ spv_result_t ValidateVariable(ValidationState_t& _, const Instruction* inst) {
// with Block, or it must be in the Uniform storage class and be decorated
// as BufferBlock.
if (value_type && value_type->opcode() == SpvOpTypeStruct) {
bool contains_RTA = false;
for (size_t member_type_index = 1;
member_type_index < value_type->operands().size();
++member_type_index) {
const auto member_type_id =
value_type->GetOperandAs<uint32_t>(member_type_index);
const auto member_type = _.FindDef(member_type_id);
if (member_type->opcode() == SpvOpTypeRuntimeArray) {
contains_RTA = true;
break;
}
}
if (contains_RTA) {
if (DoesStructContainRTA(_, value_type)) {
if (storage_class == SpvStorageClassStorageBuffer) {
if (!_.HasDecoration(value_id, SpvDecorationBlock)) {
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "For Vulkan, an OpTypeStruct variable containing an "
"OpTypeRuntimeArray must be decorated with Block if it "
"has storage class StorageBuffer.";
<< "OpTypeRuntimeArray must be decorated with Block if it "
<< "has storage class StorageBuffer.";
}
} else if (storage_class == SpvStorageClassUniform) {
if (!_.HasDecoration(value_id, SpvDecorationBufferBlock)) {
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "For Vulkan, an OpTypeStruct variable containing an "
"OpTypeRuntimeArray must be decorated with BufferBlock "
"if it has storage class Uniform.";
<< "OpTypeRuntimeArray must be decorated with BufferBlock "
<< "if it has storage class Uniform.";
}
} else {
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "For Vulkan, OpTypeStruct variables containing "
"OpTypeRuntimeArray must have storage class of "
"StorageBuffer or Uniform.";
<< "OpTypeRuntimeArray must have storage class of "
<< "StorageBuffer or Uniform.";
}
}
}
}
// WebGPU specific validation rules for OpTypeRuntimeArray
if (spvIsWebGPUEnv(_.context()->target_env)) {
const auto type_index = 2;
const auto value_id = result_type->GetOperandAs<uint32_t>(type_index);
auto value_type = _.FindDef(value_id);
// OpTypeRuntimeArray should only ever be in an OpTypeStruct,
// so should never appear as a bare variable.
if (value_type && value_type->opcode() == SpvOpTypeRuntimeArray) {
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "OpVariable, <id> '" << _.getIdName(inst->id())
<< "', is attempting to create memory for an illegal type, "
<< "OpTypeRuntimeArray.\nFor WebGPU OpTypeRuntimeArray can only "
<< "appear as the final member of an OpTypeStruct, thus cannot "
<< "be instantiated via OpVariable";
}
// If an OpStruct has an OpTypeRuntimeArray somewhere within it, then it
// must have the storage class StorageBuffer and be decorated
// with Block.
if (value_type && value_type->opcode() == SpvOpTypeStruct) {
if (DoesStructContainRTA(_, value_type)) {
if (storage_class == SpvStorageClassStorageBuffer) {
if (!_.HasDecoration(value_id, SpvDecorationBlock)) {
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "For WebGPU, an OpTypeStruct variable containing an "
<< "OpTypeRuntimeArray must be decorated with Block if it "
<< "has storage class StorageBuffer.";
}
} else {
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "For WebGPU, OpTypeStruct variables containing "
<< "OpTypeRuntimeArray must have storage class of "
<< "StorageBuffer";
}
}
}

View File

@ -37,9 +37,9 @@ spv_result_t ValidateExecutionScope(ValidationState_t& _,
if (!is_const_int32) {
if (_.HasCapability(SpvCapabilityShader)) {
return _.diag(SPV_ERROR_INVALID_DATA, inst) << "Scope ids must be "
"OpConstant when Shader "
"capability is present";
return _.diag(SPV_ERROR_INVALID_DATA, inst)
<< "Scope ids must be OpConstant when Shader capability is "
<< "present";
}
return SPV_SUCCESS;
}
@ -54,7 +54,7 @@ spv_result_t ValidateExecutionScope(ValidationState_t& _,
return _.diag(SPV_ERROR_INVALID_DATA, inst)
<< spvOpcodeString(opcode)
<< ": in Vulkan environment Execution scope is limited to "
"Subgroup";
<< "Subgroup";
}
}
@ -86,7 +86,7 @@ spv_result_t ValidateExecutionScope(ValidationState_t& _,
return _.diag(SPV_ERROR_INVALID_DATA, inst)
<< spvOpcodeString(opcode)
<< ": in Vulkan environment Execution Scope is limited to "
"Workgroup and Subgroup";
<< "Workgroup and Subgroup";
}
}
@ -97,7 +97,7 @@ spv_result_t ValidateExecutionScope(ValidationState_t& _,
return _.diag(SPV_ERROR_INVALID_DATA, inst)
<< spvOpcodeString(opcode)
<< ": in WebGPU environment Execution Scope is limited to "
"Workgroup and Subgroup";
<< "Workgroup and Subgroup";
}
}
@ -131,9 +131,9 @@ spv_result_t ValidateMemoryScope(ValidationState_t& _, const Instruction* inst,
if (!is_const_int32) {
if (_.HasCapability(SpvCapabilityShader)) {
return _.diag(SPV_ERROR_INVALID_DATA, inst) << "Scope ids must be "
"OpConstant when Shader "
"capability is present";
return _.diag(SPV_ERROR_INVALID_DATA, inst)
<< "Scope ids must be OpConstant when Shader capability is "
<< "present";
}
return SPV_SUCCESS;
}
@ -145,7 +145,7 @@ spv_result_t ValidateMemoryScope(ValidationState_t& _, const Instruction* inst,
return _.diag(SPV_ERROR_INVALID_DATA, inst)
<< spvOpcodeString(opcode)
<< ": Memory Scope QueueFamilyKHR requires capability "
"VulkanMemoryModelKHR";
<< "VulkanMemoryModelKHR";
}
}
@ -154,7 +154,7 @@ spv_result_t ValidateMemoryScope(ValidationState_t& _, const Instruction* inst,
!_.HasCapability(SpvCapabilityVulkanMemoryModelDeviceScopeKHR)) {
return _.diag(SPV_ERROR_INVALID_DATA, inst)
<< "Use of device scope with VulkanKHR memory model requires the "
"VulkanMemoryModelDeviceScopeKHR capability";
<< "VulkanMemoryModelDeviceScopeKHR capability";
}
// Vulkan Specific rules
@ -171,8 +171,7 @@ spv_result_t ValidateMemoryScope(ValidationState_t& _, const Instruction* inst,
return _.diag(SPV_ERROR_INVALID_DATA, inst)
<< spvOpcodeString(opcode)
<< ": in Vulkan 1.0 environment Memory Scope is limited to "
"Device, "
"Workgroup and Invocation";
<< "Device, Workgroup and Invocation";
}
// Vulkan 1.1 specifc rules
if (_.context()->target_env == SPV_ENV_VULKAN_1_1 &&
@ -181,8 +180,18 @@ spv_result_t ValidateMemoryScope(ValidationState_t& _, const Instruction* inst,
return _.diag(SPV_ERROR_INVALID_DATA, inst)
<< spvOpcodeString(opcode)
<< ": in Vulkan 1.1 environment Memory Scope is limited to "
"Device, "
"Workgroup and Invocation";
<< "Device, Workgroup and Invocation";
}
}
// WebGPU specific rules
if (spvIsWebGPUEnv(_.context()->target_env)) {
if (value != SpvScopeWorkgroup && value != SpvScopeSubgroup &&
value != SpvScopeQueueFamilyKHR) {
return _.diag(SPV_ERROR_INVALID_DATA, inst)
<< spvOpcodeString(opcode)
<< ": in WebGPU environment Memory Scope is limited to "
<< "Workgroup, Subgroup and QueuFamilyKHR";
}
}

View File

@ -107,11 +107,14 @@ spv_result_t ValidateTypeArray(ValidationState_t& _, const Instruction* inst) {
<< "' is a void type.";
}
if (spvIsVulkanEnv(_.context()->target_env) &&
if ((spvIsVulkanEnv(_.context()->target_env) ||
spvIsWebGPUEnv(_.context()->target_env)) &&
element_type->opcode() == SpvOpTypeRuntimeArray) {
const char* env_text =
spvIsVulkanEnv(_.context()->target_env) ? "Vulkan" : "WebGPU";
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "OpTypeArray Element Type <id> '" << _.getIdName(element_type_id)
<< "' is not valid in Vulkan environment.";
<< "' is not valid in " << env_text << " environment.";
}
const auto length_index = 2;
@ -169,12 +172,15 @@ spv_result_t ValidateTypeRuntimeArray(ValidationState_t& _,
<< _.getIdName(element_id) << "' is a void type.";
}
if (spvIsVulkanEnv(_.context()->target_env) &&
if ((spvIsVulkanEnv(_.context()->target_env) ||
spvIsWebGPUEnv(_.context()->target_env)) &&
element_type->opcode() == SpvOpTypeRuntimeArray) {
const char* env_text =
spvIsVulkanEnv(_.context()->target_env) ? "Vulkan" : "WebGPU";
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "OpTypeRuntimeArray Element Type <id> '"
<< _.getIdName(element_id)
<< "' is not valid in Vulkan environment.";
<< _.getIdName(element_id) << "' is not valid in " << env_text
<< " environment.";
}
return SPV_SUCCESS;
@ -200,11 +206,11 @@ spv_result_t ValidateTypeStruct(ValidationState_t& _, const Instruction* inst) {
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "Structure <id> " << _.getIdName(member_type_id)
<< " contains members with BuiltIn decoration. Therefore this "
"structure may not be contained as a member of another "
"structure "
"type. Structure <id> "
<< _.getIdName(struct_id) << " contains structure <id> "
<< _.getIdName(member_type_id) << ".";
<< "structure may not be contained as a member of another "
<< "structure "
<< "type. Structure <id> " << _.getIdName(struct_id)
<< " contains structure <id> " << _.getIdName(member_type_id)
<< ".";
}
if (_.IsForwardPointer(member_type_id)) {
// If we're dealing with a forward pointer:
@ -223,14 +229,17 @@ spv_result_t ValidateTypeStruct(ValidationState_t& _, const Instruction* inst) {
}
}
if (spvIsVulkanEnv(_.context()->target_env) &&
if ((spvIsVulkanEnv(_.context()->target_env) ||
spvIsWebGPUEnv(_.context()->target_env)) &&
member_type->opcode() == SpvOpTypeRuntimeArray) {
const bool is_last_member =
member_type_index == inst->operands().size() - 1;
if (!is_last_member) {
const char* env_text =
spvIsVulkanEnv(_.context()->target_env) ? "Vulkan" : "WebGPU";
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "In Vulkan, OpTypeRuntimeArray must only be used for the "
"last member of an OpTypeStruct";
<< "In " << env_text << ", OpTypeRuntimeArray must only be used "
<< "for the last member of an OpTypeStruct";
}
}
}
@ -247,9 +256,9 @@ spv_result_t ValidateTypeStruct(ValidationState_t& _, const Instruction* inst) {
if (num_builtin_members > 0 && num_builtin_members != num_struct_members) {
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "When BuiltIn decoration is applied to a structure-type member, "
"all members of that structure type must also be decorated with "
"BuiltIn (No allowed mixing of built-in variables and "
"non-built-in variables within a single structure). Structure id "
<< "all members of that structure type must also be decorated with "
<< "BuiltIn (No allowed mixing of built-in variables and "
<< "non-built-in variables within a single structure). Structure id "
<< struct_id << " does not meet this requirement.";
}
if (num_builtin_members > 0) {
@ -332,7 +341,7 @@ spv_result_t ValidateTypeForwardPointer(ValidationState_t& _,
pointer_type_inst->GetOperandAs<uint32_t>(1)) {
return _.diag(SPV_ERROR_INVALID_ID, inst)
<< "Storage class in OpTypeForwardPointer does not match the "
"pointer definition.";
<< "pointer definition.";
}
return SPV_SUCCESS;

View File

@ -21,6 +21,7 @@ add_spvtools_unittest(TARGET opt
block_merge_test.cpp
ccp_test.cpp
cfg_cleanup_test.cpp
code_sink_test.cpp
combine_access_chains_test.cpp
common_uniform_elim_test.cpp
compact_ids_test.cpp

View File

@ -0,0 +1,533 @@
// Copyright (c) 2019 Google LLC
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include <string>
#include "gmock/gmock.h"
#include "test/opt/assembly_builder.h"
#include "test/opt/pass_fixture.h"
#include "test/opt/pass_utils.h"
namespace spvtools {
namespace opt {
namespace {
using CodeSinkTest = PassTest<::testing::Test>;
TEST_F(CodeSinkTest, MoveToNextBlock) {
const std::string text = R"(
;CHECK: OpFunction
;CHECK: OpLabel
;CHECK: OpLabel
;CHECK: [[ac:%\w+]] = OpAccessChain
;CHECK: [[ld:%\w+]] = OpLoad %uint [[ac]]
;CHECK: OpCopyObject %uint [[ld]]
OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %1 "main"
%void = OpTypeVoid
%uint = OpTypeInt 32 0
%uint_0 = OpConstant %uint 0
%uint_4 = OpConstant %uint 4
%_arr_uint_uint_4 = OpTypeArray %uint %uint_4
%_ptr_Uniform_uint = OpTypePointer Uniform %uint
%_ptr_Uniform__arr_uint_uint_4 = OpTypePointer Uniform %_arr_uint_uint_4
%9 = OpVariable %_ptr_Uniform__arr_uint_uint_4 Uniform
%10 = OpTypeFunction %void
%1 = OpFunction %void None %10
%11 = OpLabel
%12 = OpAccessChain %_ptr_Uniform_uint %9 %uint_0
%13 = OpLoad %uint %12
OpBranch %14
%14 = OpLabel
%15 = OpCopyObject %uint %13
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<CodeSinkingPass>(text, true);
}
TEST_F(CodeSinkTest, MovePastSelection) {
const std::string text = R"(
;CHECK: OpFunction
;CHECK: OpLabel
;CHECK: OpSelectionMerge [[merge_bb:%\w+]]
;CHECK: [[merge_bb]] = OpLabel
;CHECK: [[ac:%\w+]] = OpAccessChain
;CHECK: [[ld:%\w+]] = OpLoad %uint [[ac]]
;CHECK: OpCopyObject %uint [[ld]]
OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %1 "main"
%void = OpTypeVoid
%bool = OpTypeBool
%true = OpConstantTrue %bool
%uint = OpTypeInt 32 0
%uint_0 = OpConstant %uint 0
%uint_4 = OpConstant %uint 4
%_arr_uint_uint_4 = OpTypeArray %uint %uint_4
%_ptr_Uniform_uint = OpTypePointer Uniform %uint
%_ptr_Uniform__arr_uint_uint_4 = OpTypePointer Uniform %_arr_uint_uint_4
%11 = OpVariable %_ptr_Uniform__arr_uint_uint_4 Uniform
%12 = OpTypeFunction %void
%1 = OpFunction %void None %12
%13 = OpLabel
%14 = OpAccessChain %_ptr_Uniform_uint %11 %uint_0
%15 = OpLoad %uint %14
OpSelectionMerge %16 None
OpBranchConditional %true %17 %16
%17 = OpLabel
OpBranch %16
%16 = OpLabel
%18 = OpCopyObject %uint %15
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<CodeSinkingPass>(text, true);
}
TEST_F(CodeSinkTest, MoveIntoSelection) {
const std::string text = R"(
;CHECK: OpFunction
;CHECK: OpLabel
;CHECK: OpSelectionMerge [[merge_bb:%\w+]]
;CHECK-NEXT: OpBranchConditional %true [[bb:%\w+]] [[merge_bb]]
;CHECK: [[bb]] = OpLabel
;CHECK-NEXT: [[ac:%\w+]] = OpAccessChain
;CHECK-NEXT: [[ld:%\w+]] = OpLoad %uint [[ac]]
;CHECK-NEXT: OpCopyObject %uint [[ld]]
OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %1 "main"
%void = OpTypeVoid
%bool = OpTypeBool
%true = OpConstantTrue %bool
%uint = OpTypeInt 32 0
%uint_0 = OpConstant %uint 0
%uint_4 = OpConstant %uint 4
%_arr_uint_uint_4 = OpTypeArray %uint %uint_4
%_ptr_Uniform_uint = OpTypePointer Uniform %uint
%_ptr_Uniform__arr_uint_uint_4 = OpTypePointer Uniform %_arr_uint_uint_4
%11 = OpVariable %_ptr_Uniform__arr_uint_uint_4 Uniform
%12 = OpTypeFunction %void
%1 = OpFunction %void None %12
%13 = OpLabel
%14 = OpAccessChain %_ptr_Uniform_uint %11 %uint_0
%15 = OpLoad %uint %14
OpSelectionMerge %16 None
OpBranchConditional %true %17 %16
%17 = OpLabel
%18 = OpCopyObject %uint %15
OpBranch %16
%16 = OpLabel
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<CodeSinkingPass>(text, true);
}
TEST_F(CodeSinkTest, LeaveBeforeSelection) {
const std::string text = R"(
OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %1 "main"
%void = OpTypeVoid
%bool = OpTypeBool
%true = OpConstantTrue %bool
%uint = OpTypeInt 32 0
%uint_0 = OpConstant %uint 0
%uint_4 = OpConstant %uint 4
%_arr_uint_uint_4 = OpTypeArray %uint %uint_4
%_ptr_Uniform_uint = OpTypePointer Uniform %uint
%_ptr_Uniform__arr_uint_uint_4 = OpTypePointer Uniform %_arr_uint_uint_4
%11 = OpVariable %_ptr_Uniform__arr_uint_uint_4 Uniform
%12 = OpTypeFunction %void
%1 = OpFunction %void None %12
%13 = OpLabel
%14 = OpAccessChain %_ptr_Uniform_uint %11 %uint_0
%15 = OpLoad %uint %14
OpSelectionMerge %16 None
OpBranchConditional %true %17 %20
%20 = OpLabel
OpBranch %16
%17 = OpLabel
%18 = OpCopyObject %uint %15
OpBranch %16
%16 = OpLabel
%19 = OpCopyObject %uint %15
OpReturn
OpFunctionEnd
)";
auto result = SinglePassRunAndDisassemble<CodeSinkingPass>(
text, /* skip_nop = */ true, /* do_validation = */ true);
EXPECT_EQ(Pass::Status::SuccessWithoutChange, std::get<1>(result));
}
TEST_F(CodeSinkTest, LeaveAloneUseInSameBlock) {
const std::string text = R"(
OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %1 "main"
%void = OpTypeVoid
%bool = OpTypeBool
%true = OpConstantTrue %bool
%uint = OpTypeInt 32 0
%uint_0 = OpConstant %uint 0
%uint_4 = OpConstant %uint 4
%_arr_uint_uint_4 = OpTypeArray %uint %uint_4
%_ptr_Uniform_uint = OpTypePointer Uniform %uint
%_ptr_Uniform__arr_uint_uint_4 = OpTypePointer Uniform %_arr_uint_uint_4
%11 = OpVariable %_ptr_Uniform__arr_uint_uint_4 Uniform
%12 = OpTypeFunction %void
%1 = OpFunction %void None %12
%13 = OpLabel
%14 = OpAccessChain %_ptr_Uniform_uint %11 %uint_0
%15 = OpLoad %uint %14
%cond = OpIEqual %bool %15 %uint_0
OpSelectionMerge %16 None
OpBranchConditional %cond %17 %16
%17 = OpLabel
OpBranch %16
%16 = OpLabel
%19 = OpCopyObject %uint %15
OpReturn
OpFunctionEnd
)";
auto result = SinglePassRunAndDisassemble<CodeSinkingPass>(
text, /* skip_nop = */ true, /* do_validation = */ true);
EXPECT_EQ(Pass::Status::SuccessWithoutChange, std::get<1>(result));
}
TEST_F(CodeSinkTest, DontMoveIntoLoop) {
const std::string text = R"(
OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %1 "main"
%void = OpTypeVoid
%bool = OpTypeBool
%true = OpConstantTrue %bool
%uint = OpTypeInt 32 0
%uint_0 = OpConstant %uint 0
%uint_4 = OpConstant %uint 4
%_arr_uint_uint_4 = OpTypeArray %uint %uint_4
%_ptr_Uniform_uint = OpTypePointer Uniform %uint
%_ptr_Uniform__arr_uint_uint_4 = OpTypePointer Uniform %_arr_uint_uint_4
%11 = OpVariable %_ptr_Uniform__arr_uint_uint_4 Uniform
%12 = OpTypeFunction %void
%1 = OpFunction %void None %12
%13 = OpLabel
%14 = OpAccessChain %_ptr_Uniform_uint %11 %uint_0
%15 = OpLoad %uint %14
OpBranch %17
%17 = OpLabel
OpLoopMerge %merge %cont None
OpBranch %cont
%cont = OpLabel
%cond = OpIEqual %bool %15 %uint_0
OpBranchConditional %cond %merge %17
%merge = OpLabel
OpReturn
OpFunctionEnd
)";
auto result = SinglePassRunAndDisassemble<CodeSinkingPass>(
text, /* skip_nop = */ true, /* do_validation = */ true);
EXPECT_EQ(Pass::Status::SuccessWithoutChange, std::get<1>(result));
}
TEST_F(CodeSinkTest, DontMoveIntoLoop2) {
const std::string text = R"(
OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %1 "main"
%void = OpTypeVoid
%bool = OpTypeBool
%true = OpConstantTrue %bool
%uint = OpTypeInt 32 0
%uint_0 = OpConstant %uint 0
%uint_4 = OpConstant %uint 4
%_arr_uint_uint_4 = OpTypeArray %uint %uint_4
%_ptr_Uniform_uint = OpTypePointer Uniform %uint
%_ptr_Uniform__arr_uint_uint_4 = OpTypePointer Uniform %_arr_uint_uint_4
%11 = OpVariable %_ptr_Uniform__arr_uint_uint_4 Uniform
%12 = OpTypeFunction %void
%1 = OpFunction %void None %12
%13 = OpLabel
%14 = OpAccessChain %_ptr_Uniform_uint %11 %uint_0
%15 = OpLoad %uint %14
OpSelectionMerge %16 None
OpBranchConditional %true %17 %16
%17 = OpLabel
OpLoopMerge %merge %cont None
OpBranch %cont
%cont = OpLabel
%cond = OpIEqual %bool %15 %uint_0
OpBranchConditional %cond %merge %17
%merge = OpLabel
OpBranch %16
%16 = OpLabel
OpReturn
OpFunctionEnd
)";
auto result = SinglePassRunAndDisassemble<CodeSinkingPass>(
text, /* skip_nop = */ true, /* do_validation = */ true);
EXPECT_EQ(Pass::Status::SuccessWithoutChange, std::get<1>(result));
}
TEST_F(CodeSinkTest, DontMoveSelectionUsedInBothSides) {
const std::string text = R"(
OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %1 "main"
%void = OpTypeVoid
%bool = OpTypeBool
%true = OpConstantTrue %bool
%uint = OpTypeInt 32 0
%uint_0 = OpConstant %uint 0
%uint_4 = OpConstant %uint 4
%_arr_uint_uint_4 = OpTypeArray %uint %uint_4
%_ptr_Uniform_uint = OpTypePointer Uniform %uint
%_ptr_Uniform__arr_uint_uint_4 = OpTypePointer Uniform %_arr_uint_uint_4
%11 = OpVariable %_ptr_Uniform__arr_uint_uint_4 Uniform
%12 = OpTypeFunction %void
%1 = OpFunction %void None %12
%13 = OpLabel
%14 = OpAccessChain %_ptr_Uniform_uint %11 %uint_0
%15 = OpLoad %uint %14
OpSelectionMerge %16 None
OpBranchConditional %true %17 %20
%20 = OpLabel
%19 = OpCopyObject %uint %15
OpBranch %16
%17 = OpLabel
%18 = OpCopyObject %uint %15
OpBranch %16
%16 = OpLabel
OpReturn
OpFunctionEnd
)";
auto result = SinglePassRunAndDisassemble<CodeSinkingPass>(
text, /* skip_nop = */ true, /* do_validation = */ true);
EXPECT_EQ(Pass::Status::SuccessWithoutChange, std::get<1>(result));
}
TEST_F(CodeSinkTest, DontMoveBecauseOfStore) {
const std::string text = R"(
OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %1 "main"
%void = OpTypeVoid
%bool = OpTypeBool
%true = OpConstantTrue %bool
%uint = OpTypeInt 32 0
%uint_0 = OpConstant %uint 0
%uint_4 = OpConstant %uint 4
%_arr_uint_uint_4 = OpTypeArray %uint %uint_4
%_ptr_Uniform_uint = OpTypePointer Uniform %uint
%_ptr_Uniform__arr_uint_uint_4 = OpTypePointer Uniform %_arr_uint_uint_4
%11 = OpVariable %_ptr_Uniform__arr_uint_uint_4 Uniform
%12 = OpTypeFunction %void
%1 = OpFunction %void None %12
%13 = OpLabel
%14 = OpAccessChain %_ptr_Uniform_uint %11 %uint_0
%15 = OpLoad %uint %14
OpStore %14 %15
OpSelectionMerge %16 None
OpBranchConditional %true %17 %20
%20 = OpLabel
OpBranch %16
%17 = OpLabel
%18 = OpCopyObject %uint %15
OpBranch %16
%16 = OpLabel
OpReturn
OpFunctionEnd
)";
auto result = SinglePassRunAndDisassemble<CodeSinkingPass>(
text, /* skip_nop = */ true, /* do_validation = */ true);
EXPECT_EQ(Pass::Status::SuccessWithoutChange, std::get<1>(result));
}
TEST_F(CodeSinkTest, MoveReadOnlyLoadWithSync) {
const std::string text = R"(
OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %1 "main"
%void = OpTypeVoid
%bool = OpTypeBool
%true = OpConstantTrue %bool
%uint = OpTypeInt 32 0
%uint_0 = OpConstant %uint 0
%uint_4 = OpConstant %uint 4
%mem_semantics = OpConstant %uint 0x42 ; Uniform memeory arquire
%_arr_uint_uint_4 = OpTypeArray %uint %uint_4
%_ptr_Uniform_uint = OpTypePointer Uniform %uint
%_ptr_Uniform__arr_uint_uint_4 = OpTypePointer Uniform %_arr_uint_uint_4
%11 = OpVariable %_ptr_Uniform__arr_uint_uint_4 Uniform
%12 = OpTypeFunction %void
%1 = OpFunction %void None %12
%13 = OpLabel
%14 = OpAccessChain %_ptr_Uniform_uint %11 %uint_0
%15 = OpLoad %uint %14
OpMemoryBarrier %uint_4 %mem_semantics
OpSelectionMerge %16 None
OpBranchConditional %true %17 %20
%20 = OpLabel
OpBranch %16
%17 = OpLabel
%18 = OpCopyObject %uint %15
OpBranch %16
%16 = OpLabel
OpReturn
OpFunctionEnd
)";
auto result = SinglePassRunAndDisassemble<CodeSinkingPass>(
text, /* skip_nop = */ true, /* do_validation = */ true);
EXPECT_EQ(Pass::Status::SuccessWithChange, std::get<1>(result));
}
TEST_F(CodeSinkTest, DontMoveBecauseOfSync) {
const std::string text = R"(
OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %1 "main"
OpDecorate %_arr_uint_uint_4 BufferBlock
OpMemberDecorate %_arr_uint_uint_4 0 Offset 0
%void = OpTypeVoid
%bool = OpTypeBool
%true = OpConstantTrue %bool
%uint = OpTypeInt 32 0
%uint_0 = OpConstant %uint 0
%uint_4 = OpConstant %uint 4
%mem_semantics = OpConstant %uint 0x42 ; Uniform memeory arquire
%_arr_uint_uint_4 = OpTypeStruct %uint
%_ptr_Uniform_uint = OpTypePointer Uniform %uint
%_ptr_Uniform__arr_uint_uint_4 = OpTypePointer Uniform %_arr_uint_uint_4
%11 = OpVariable %_ptr_Uniform__arr_uint_uint_4 Uniform
%12 = OpTypeFunction %void
%1 = OpFunction %void None %12
%13 = OpLabel
%14 = OpAccessChain %_ptr_Uniform_uint %11 %uint_0
%15 = OpLoad %uint %14
OpMemoryBarrier %uint_4 %mem_semantics
OpSelectionMerge %16 None
OpBranchConditional %true %17 %20
%20 = OpLabel
OpBranch %16
%17 = OpLabel
%18 = OpCopyObject %uint %15
OpBranch %16
%16 = OpLabel
OpReturn
OpFunctionEnd
)";
auto result = SinglePassRunAndDisassemble<CodeSinkingPass>(
text, /* skip_nop = */ true, /* do_validation = */ true);
EXPECT_EQ(Pass::Status::SuccessWithoutChange, std::get<1>(result));
}
TEST_F(CodeSinkTest, DontMoveBecauseOfAtomicWithSync) {
const std::string text = R"(
OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %1 "main"
OpDecorate %_arr_uint_uint_4 BufferBlock
OpMemberDecorate %_arr_uint_uint_4 0 Offset 0
%void = OpTypeVoid
%bool = OpTypeBool
%true = OpConstantTrue %bool
%uint = OpTypeInt 32 0
%uint_0 = OpConstant %uint 0
%uint_4 = OpConstant %uint 4
%mem_semantics = OpConstant %uint 0x42 ; Uniform memeory arquire
%_arr_uint_uint_4 = OpTypeStruct %uint
%_ptr_Uniform_uint = OpTypePointer Uniform %uint
%_ptr_Uniform__arr_uint_uint_4 = OpTypePointer Uniform %_arr_uint_uint_4
%11 = OpVariable %_ptr_Uniform__arr_uint_uint_4 Uniform
%12 = OpTypeFunction %void
%1 = OpFunction %void None %12
%13 = OpLabel
%14 = OpAccessChain %_ptr_Uniform_uint %11 %uint_0
%15 = OpLoad %uint %14
%al = OpAtomicLoad %uint %14 %uint_4 %mem_semantics
OpSelectionMerge %16 None
OpBranchConditional %true %17 %20
%20 = OpLabel
OpBranch %16
%17 = OpLabel
%18 = OpCopyObject %uint %15
OpBranch %16
%16 = OpLabel
OpReturn
OpFunctionEnd
)";
auto result = SinglePassRunAndDisassemble<CodeSinkingPass>(
text, /* skip_nop = */ true, /* do_validation = */ true);
EXPECT_EQ(Pass::Status::SuccessWithoutChange, std::get<1>(result));
}
TEST_F(CodeSinkTest, MoveWithAtomicWithoutSync) {
const std::string text = R"(
OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint GLCompute %1 "main"
OpDecorate %_arr_uint_uint_4 BufferBlock
OpMemberDecorate %_arr_uint_uint_4 0 Offset 0
%void = OpTypeVoid
%bool = OpTypeBool
%true = OpConstantTrue %bool
%uint = OpTypeInt 32 0
%uint_0 = OpConstant %uint 0
%uint_4 = OpConstant %uint 4
%_arr_uint_uint_4 = OpTypeStruct %uint
%_ptr_Uniform_uint = OpTypePointer Uniform %uint
%_ptr_Uniform__arr_uint_uint_4 = OpTypePointer Uniform %_arr_uint_uint_4
%11 = OpVariable %_ptr_Uniform__arr_uint_uint_4 Uniform
%12 = OpTypeFunction %void
%1 = OpFunction %void None %12
%13 = OpLabel
%14 = OpAccessChain %_ptr_Uniform_uint %11 %uint_0
%15 = OpLoad %uint %14
%al = OpAtomicLoad %uint %14 %uint_4 %uint_0
OpSelectionMerge %16 None
OpBranchConditional %true %17 %20
%20 = OpLabel
OpBranch %16
%17 = OpLabel
%18 = OpCopyObject %uint %15
OpBranch %16
%16 = OpLabel
OpReturn
OpFunctionEnd
)";
auto result = SinglePassRunAndDisassemble<CodeSinkingPass>(
text, /* skip_nop = */ true, /* do_validation = */ true);
EXPECT_EQ(Pass::Status::SuccessWithChange, std::get<1>(result));
}
} // namespace
} // namespace opt
} // namespace spvtools

View File

@ -538,7 +538,15 @@ INSTANTIATE_TEST_CASE_P(TestCase, IntegerInstructionFoldingTest,
"%2 = OpShiftLeftLogical %int %int_2 %uint_32\n" +
"OpReturn\n" +
"OpFunctionEnd",
2, 0)
2, 0),
// Test case 29: fold -INT_MIN
InstructionFoldingCase<uint32_t>(
Header() + "%main = OpFunction %void None %void_func\n" +
"%main_lab = OpLabel\n" +
"%2 = OpSNegate %int %int_min\n" +
"OpReturn\n" +
"OpFunctionEnd",
2, std::numeric_limits<int32_t>::min())
));
// clang-format on

View File

@ -107,8 +107,7 @@ class DuplicateInstPass : public Pass {
public:
const char* name() const override { return "DuplicateInst"; }
Status Process() override {
auto inst =
MakeUnique<Instruction>(*(--context()->debug1_end())->Clone(context()));
auto inst = MakeUnique<Instruction>(*(--context()->debug1_end()));
context()->AddDebug1Inst(std::move(inst));
return Status::SuccessWithChange;
}
@ -121,21 +120,21 @@ TEST_F(PassManagerTest, Run) {
AddPass<AppendOpNopPass>();
AddPass<AppendOpNopPass>();
RunAndCheck(text.c_str(), (text + "OpNop\nOpNop\n").c_str());
RunAndCheck(text, text + "OpNop\nOpNop\n");
RenewPassManger();
AddPass<AppendOpNopPass>();
AddPass<DuplicateInstPass>();
RunAndCheck(text.c_str(), (text + "OpNop\nOpNop\n").c_str());
RunAndCheck(text, text + "OpNop\nOpNop\n");
RenewPassManger();
AddPass<DuplicateInstPass>();
AddPass<AppendOpNopPass>();
RunAndCheck(text.c_str(), (text + "OpSource ESSL 310\nOpNop\n").c_str());
RunAndCheck(text, text + "OpSource ESSL 310\nOpNop\n");
RenewPassManger();
AddPass<AppendMultipleOpNopPass>(3);
RunAndCheck(text.c_str(), (text + "OpNop\nOpNop\nOpNop\n").c_str());
RunAndCheck(text, text + "OpNop\nOpNop\nOpNop\n");
}
// A pass that appends an OpTypeVoid instruction that uses a given id.

View File

@ -1593,6 +1593,33 @@ TEST_F(ScalarReplacementTest, AmbigousPointer) {
SinglePassRunAndMatch<ScalarReplacementPass>(text, true);
}
// Test that scalar replacement does not crash when there is an OpAccessChain
// with no index. If we choose to handle this case in the future, then the
// result can change.
TEST_F(ScalarReplacementTest, TestAccessChainWithNoIndexes) {
const std::string text = R"(
OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %1 "main"
OpExecutionMode %1 OriginLowerLeft
%void = OpTypeVoid
%3 = OpTypeFunction %void
%float = OpTypeFloat 32
%_struct_5 = OpTypeStruct %float
%_ptr_Function__struct_5 = OpTypePointer Function %_struct_5
%1 = OpFunction %void None %3
%7 = OpLabel
%8 = OpVariable %_ptr_Function__struct_5 Function
%9 = OpAccessChain %_ptr_Function__struct_5 %8
OpReturn
OpFunctionEnd
)";
auto result =
SinglePassRunAndDisassemble<ScalarReplacementPass>(text, true, false);
EXPECT_EQ(Pass::Status::SuccessWithoutChange, std::get<1>(result));
}
} // namespace
} // namespace opt
} // namespace spvtools

View File

@ -1190,6 +1190,42 @@ TEST_F(VectorDCETest, InsertWithNoIndices) {
SinglePassRunAndMatch<VectorDCE>(text, true);
}
TEST_F(VectorDCETest, ExtractWithNoIndices) {
const std::string text = R"(
; CHECK: OpLoad %float
; CHECK: [[ld:%\w+]] = OpLoad %v4float
; CHECK: [[ex1:%\w+]] = OpCompositeExtract %v4float [[ld]]
; CHECK: [[ex2:%\w+]] = OpCompositeExtract %float [[ex1]] 1
; CHECK: OpStore {{%\w+}} [[ex2]]
OpCapability Shader
OpMemoryModel Logical GLSL450
OpEntryPoint Fragment %1 "PSMain" %2 %14 %3
OpExecutionMode %1 OriginUpperLeft
%void = OpTypeVoid
%5 = OpTypeFunction %void
%float = OpTypeFloat 32
%v4float = OpTypeVector %float 4
%_ptr_Input_float = OpTypePointer Input %float
%_ptr_Input_v4float = OpTypePointer Input %v4float
%_ptr_Output_float = OpTypePointer Output %float
%2 = OpVariable %_ptr_Input_v4float Input
%14 = OpVariable %_ptr_Input_float Input
%3 = OpVariable %_ptr_Output_float Output
%1 = OpFunction %void None %5
%10 = OpLabel
%13 = OpLoad %float %14
%11 = OpLoad %v4float %2
%12 = OpCompositeInsert %v4float %13 %11 0
%20 = OpCompositeExtract %v4float %12
%21 = OpCompositeExtract %float %20 1
OpStore %3 %21
OpReturn
OpFunctionEnd
)";
SinglePassRunAndMatch<VectorDCE>(text, true);
}
} // namespace
} // namespace opt
} // namespace spvtools

View File

@ -340,7 +340,7 @@ TEST_F(ValidateAtomics, AtomicLoadVulkanInt64) {
TEST_F(ValidateAtomics, AtomicLoadWebGPUShaderSuccess) {
const std::string body = R"(
%val1 = OpAtomicLoad %u32 %u32_var %device %relaxed
%val1 = OpAtomicLoad %u32 %u32_var %queuefamily %relaxed
%val2 = OpAtomicLoad %u32 %u32_var %workgroup %acquire
)";
@ -544,7 +544,7 @@ OpAtomicStore %u32_var %device %sequentially_consistent %u32_1
TEST_F(ValidateAtomics, AtomicStoreWebGPUSuccess) {
const std::string body = R"(
OpAtomicStore %u32_var %device %release %u32_1
OpAtomicStore %u32_var %queuefamily %release %u32_1
)";
CompileSuccessfully(GenerateWebGPUShaderCode(body), SPV_ENV_WEBGPU_0);
@ -553,7 +553,7 @@ OpAtomicStore %u32_var %device %release %u32_1
TEST_F(ValidateAtomics, AtomicStoreWebGPUSequentiallyConsistent) {
const std::string body = R"(
OpAtomicStore %u32_var %device %sequentially_consistent %u32_1
OpAtomicStore %u32_var %queuefamily %sequentially_consistent %u32_1
)";
CompileSuccessfully(GenerateWebGPUShaderCode(body), SPV_ENV_WEBGPU_0);
@ -564,7 +564,8 @@ OpAtomicStore %u32_var %device %sequentially_consistent %u32_1
"WebGPU spec disallows any bit masks in Memory Semantics that are "
"not Acquire, Release, AcquireRelease, UniformMemory, "
"WorkgroupMemory, ImageMemory, OutputMemoryKHR, MakeAvailableKHR, or "
"MakeVisibleKHR\n OpAtomicStore %29 %uint_1_0 %uint_16 %uint_1\n"));
"MakeVisibleKHR\n"
" OpAtomicStore %29 %uint_5 %uint_16 %uint_1\n"));
}
TEST_F(ValidateAtomics, AtomicStoreWrongPointerType) {
@ -1858,6 +1859,75 @@ OpExtension "SPV_KHR_vulkan_memory_model"
EXPECT_EQ(SPV_SUCCESS, ValidateInstructions(SPV_ENV_UNIVERSAL_1_3));
}
TEST_F(ValidateAtomics, WebGPUCrossDeviceMemoryScopeBad) {
const std::string body = R"(
%val1 = OpAtomicLoad %u32 %u32_var %cross_device %relaxed
)";
CompileSuccessfully(GenerateWebGPUShaderCode(body), SPV_ENV_WEBGPU_0);
EXPECT_EQ(SPV_ERROR_INVALID_DATA, ValidateInstructions(SPV_ENV_WEBGPU_0));
EXPECT_THAT(
getDiagnosticString(),
HasSubstr("AtomicLoad: in WebGPU environment Memory Scope is limited to "
"Workgroup, Subgroup and QueuFamilyKHR\n"
" %34 = OpAtomicLoad %uint %29 %uint_0_0 %uint_0_1\n"));
}
TEST_F(ValidateAtomics, WebGPUDeviceMemoryScopeBad) {
const std::string body = R"(
%val1 = OpAtomicLoad %u32 %u32_var %device %relaxed
)";
CompileSuccessfully(GenerateWebGPUShaderCode(body), SPV_ENV_WEBGPU_0);
EXPECT_EQ(SPV_ERROR_INVALID_DATA, ValidateInstructions(SPV_ENV_WEBGPU_0));
EXPECT_THAT(
getDiagnosticString(),
HasSubstr("AtomicLoad: in WebGPU environment Memory Scope is limited to "
"Workgroup, Subgroup and QueuFamilyKHR\n"
" %34 = OpAtomicLoad %uint %29 %uint_1_0 %uint_0_1\n"));
}
TEST_F(ValidateAtomics, WebGPUWorkgroupMemoryScopeGood) {
const std::string body = R"(
%val1 = OpAtomicLoad %u32 %u32_var %workgroup %relaxed
)";
CompileSuccessfully(GenerateWebGPUShaderCode(body), SPV_ENV_WEBGPU_0);
EXPECT_EQ(SPV_SUCCESS, ValidateInstructions(SPV_ENV_WEBGPU_0));
}
TEST_F(ValidateAtomics, WebGPUSubgroupMemoryScopeGood) {
const std::string body = R"(
%val1 = OpAtomicLoad %u32 %u32_var %subgroup %relaxed
)";
CompileSuccessfully(GenerateWebGPUShaderCode(body), SPV_ENV_WEBGPU_0);
EXPECT_EQ(SPV_SUCCESS, ValidateInstructions(SPV_ENV_WEBGPU_0));
}
TEST_F(ValidateAtomics, WebGPUInvocationMemoryScopeBad) {
const std::string body = R"(
%val1 = OpAtomicLoad %u32 %u32_var %invocation %relaxed
)";
CompileSuccessfully(GenerateWebGPUShaderCode(body), SPV_ENV_WEBGPU_0);
EXPECT_EQ(SPV_ERROR_INVALID_DATA, ValidateInstructions(SPV_ENV_WEBGPU_0));
EXPECT_THAT(
getDiagnosticString(),
HasSubstr("AtomicLoad: in WebGPU environment Memory Scope is limited to "
"Workgroup, Subgroup and QueuFamilyKHR\n"
" %34 = OpAtomicLoad %uint %29 %uint_4 %uint_0_1\n"));
}
TEST_F(ValidateAtomics, WebGPUQueueFamilyMemoryScopeGood) {
const std::string body = R"(
%val1 = OpAtomicLoad %u32 %u32_var %queuefamily %relaxed
)";
CompileSuccessfully(GenerateWebGPUShaderCode(body), SPV_ENV_WEBGPU_0);
EXPECT_EQ(SPV_SUCCESS, ValidateInstructions(SPV_ENV_WEBGPU_0));
}
} // namespace
} // namespace val
} // namespace spvtools

View File

@ -872,6 +872,68 @@ TEST_F(ValidateData, vulkan_RTA_not_at_end_of_struct) {
"OpTypeStruct %_runtimearr_uint %uint\n"));
}
TEST_F(ValidateData, webgpu_RTA_array_at_end_of_struct) {
std::string str = R"(
OpCapability Shader
OpCapability VulkanMemoryModelKHR
OpExtension "SPV_KHR_vulkan_memory_model"
OpMemoryModel Logical VulkanKHR
OpEntryPoint Fragment %func "func"
OpExecutionMode %func OriginUpperLeft
OpDecorate %array_t ArrayStride 4
OpMemberDecorate %struct_t 0 Offset 0
OpMemberDecorate %struct_t 1 Offset 4
OpDecorate %struct_t Block
%uint_t = OpTypeInt 32 0
%array_t = OpTypeRuntimeArray %uint_t
%struct_t = OpTypeStruct %uint_t %array_t
%struct_ptr = OpTypePointer StorageBuffer %struct_t
%2 = OpVariable %struct_ptr StorageBuffer
%void = OpTypeVoid
%func_t = OpTypeFunction %void
%func = OpFunction %void None %func_t
%1 = OpLabel
OpReturn
OpFunctionEnd
)";
CompileSuccessfully(str.c_str(), SPV_ENV_WEBGPU_0);
ASSERT_EQ(SPV_SUCCESS, ValidateInstructions(SPV_ENV_WEBGPU_0));
}
TEST_F(ValidateData, webgpu_RTA_not_at_end_of_struct) {
std::string str = R"(
OpCapability Shader
OpCapability VulkanMemoryModelKHR
OpExtension "SPV_KHR_vulkan_memory_model"
OpMemoryModel Logical VulkanKHR
OpEntryPoint Fragment %func "func"
OpExecutionMode %func OriginUpperLeft
OpDecorate %array_t ArrayStride 4
OpMemberDecorate %struct_t 0 Offset 0
OpMemberDecorate %struct_t 1 Offset 4
OpDecorate %struct_t Block
%uint_t = OpTypeInt 32 0
%array_t = OpTypeRuntimeArray %uint_t
%struct_t = OpTypeStruct %array_t %uint_t
%struct_ptr = OpTypePointer StorageBuffer %struct_t
%2 = OpVariable %struct_ptr StorageBuffer
%void = OpTypeVoid
%func_t = OpTypeFunction %void
%func = OpFunction %void None %func_t
%1 = OpLabel
OpReturn
OpFunctionEnd
)";
CompileSuccessfully(str.c_str(), SPV_ENV_WEBGPU_0);
ASSERT_EQ(SPV_ERROR_INVALID_ID, ValidateInstructions(SPV_ENV_WEBGPU_0));
EXPECT_THAT(getDiagnosticString(),
HasSubstr("In WebGPU, OpTypeRuntimeArray must only be used for "
"the last member of an OpTypeStruct\n %_struct_3 = "
"OpTypeStruct %_runtimearr_uint %uint\n"));
}
} // namespace
} // namespace val
} // namespace spvtools

View File

@ -1806,6 +1806,38 @@ OpFunctionEnd
"%_ptr_UniformConstant__runtimearr_2 UniformConstant\n"));
}
TEST_F(ValidateMemory, WebGPURTAOutsideOfStructBad) {
std::string spirv = R"(
OpCapability Shader
OpCapability VulkanMemoryModelKHR
OpExtension "SPV_KHR_vulkan_memory_model"
OpMemoryModel Logical VulkanKHR
OpEntryPoint Fragment %func "func"
OpExecutionMode %func OriginUpperLeft
%sampler_t = OpTypeSampler
%array_t = OpTypeRuntimeArray %sampler_t
%array_ptr = OpTypePointer UniformConstant %array_t
%2 = OpVariable %array_ptr UniformConstant
%void = OpTypeVoid
%func_t = OpTypeFunction %void
%func = OpFunction %void None %func_t
%1 = OpLabel
OpReturn
OpFunctionEnd
)";
CompileSuccessfully(spirv.c_str(), SPV_ENV_WEBGPU_0);
EXPECT_EQ(SPV_ERROR_INVALID_ID, ValidateInstructions(SPV_ENV_WEBGPU_0));
EXPECT_THAT(
getDiagnosticString(),
HasSubstr(
"OpVariable, <id> '5[%5]', is attempting to create memory for an "
"illegal type, OpTypeRuntimeArray.\nFor WebGPU OpTypeRuntimeArray "
"can only appear as the final member of an OpTypeStruct, thus cannot "
"be instantiated via OpVariable\n %5 = OpVariable "
"%_ptr_UniformConstant__runtimearr_2 UniformConstant\n"));
}
TEST_F(ValidateMemory, VulkanRTAOutsideOfStructWithRuntimeDescriptorArrayGood) {
std::string spirv = R"(
OpCapability Shader
@ -1890,6 +1922,34 @@ OpFunctionEnd
EXPECT_EQ(SPV_SUCCESS, ValidateInstructions(SPV_ENV_VULKAN_1_1));
}
TEST_F(ValidateMemory, WebGPURTAInsideStorageBufferStructGood) {
std::string spirv = R"(
OpCapability Shader
OpCapability VulkanMemoryModelKHR
OpExtension "SPV_KHR_vulkan_memory_model"
OpMemoryModel Logical VulkanKHR
OpEntryPoint Fragment %func "func"
OpExecutionMode %func OriginUpperLeft
OpDecorate %array_t ArrayStride 4
OpMemberDecorate %struct_t 0 Offset 0
OpDecorate %struct_t Block
%uint_t = OpTypeInt 32 0
%array_t = OpTypeRuntimeArray %uint_t
%struct_t = OpTypeStruct %array_t
%struct_ptr = OpTypePointer StorageBuffer %struct_t
%2 = OpVariable %struct_ptr StorageBuffer
%void = OpTypeVoid
%func_t = OpTypeFunction %void
%func = OpFunction %void None %func_t
%1 = OpLabel
OpReturn
OpFunctionEnd
)";
CompileSuccessfully(spirv.c_str(), SPV_ENV_WEBGPU_0);
EXPECT_EQ(SPV_SUCCESS, ValidateInstructions(SPV_ENV_WEBGPU_0));
}
TEST_F(ValidateMemory, VulkanRTAInsideWrongStorageClassStructBad) {
std::string spirv = R"(
OpCapability Shader
@ -1919,6 +1979,36 @@ OpFunctionEnd
"OpVariable %_ptr_Workgroup__struct_4 Workgroup\n"));
}
TEST_F(ValidateMemory, WebGPURTAInsideWrongStorageClassStructBad) {
std::string spirv = R"(
OpCapability Shader
OpCapability VulkanMemoryModelKHR
OpExtension "SPV_KHR_vulkan_memory_model"
OpMemoryModel Logical VulkanKHR
OpEntryPoint Fragment %func "func"
OpExecutionMode %func OriginUpperLeft
%uint_t = OpTypeInt 32 0
%array_t = OpTypeRuntimeArray %uint_t
%struct_t = OpTypeStruct %array_t
%struct_ptr = OpTypePointer Workgroup %struct_t
%2 = OpVariable %struct_ptr Workgroup
%void = OpTypeVoid
%func_t = OpTypeFunction %void
%func = OpFunction %void None %func_t
%1 = OpLabel
OpReturn
OpFunctionEnd
)";
CompileSuccessfully(spirv.c_str(), SPV_ENV_WEBGPU_0);
EXPECT_EQ(SPV_ERROR_INVALID_ID, ValidateInstructions(SPV_ENV_WEBGPU_0));
EXPECT_THAT(
getDiagnosticString(),
HasSubstr("For WebGPU, OpTypeStruct variables containing "
"OpTypeRuntimeArray must have storage class of StorageBuffer\n "
" %6 = OpVariable %_ptr_Workgroup__struct_4 Workgroup\n"));
}
TEST_F(ValidateMemory, VulkanRTAInsideStorageBufferStructWithoutBlockBad) {
std::string spirv = R"(
OpCapability Shader
@ -1947,6 +2037,36 @@ OpFunctionEnd
"%_ptr_StorageBuffer__struct_4 StorageBuffer\n"));
}
TEST_F(ValidateMemory, WebGPURTAInsideStorageBufferStructWithoutBlockBad) {
std::string spirv = R"(
OpCapability Shader
OpCapability VulkanMemoryModelKHR
OpExtension "SPV_KHR_vulkan_memory_model"
OpMemoryModel Logical VulkanKHR
OpEntryPoint Fragment %func "func"
OpExecutionMode %func OriginUpperLeft
%uint_t = OpTypeInt 32 0
%array_t = OpTypeRuntimeArray %uint_t
%struct_t = OpTypeStruct %array_t
%struct_ptr = OpTypePointer StorageBuffer %struct_t
%2 = OpVariable %struct_ptr StorageBuffer
%void = OpTypeVoid
%func_t = OpTypeFunction %void
%func = OpFunction %void None %func_t
%1 = OpLabel
OpReturn
OpFunctionEnd
)";
CompileSuccessfully(spirv.c_str(), SPV_ENV_WEBGPU_0);
EXPECT_EQ(SPV_ERROR_INVALID_ID, ValidateInstructions(SPV_ENV_WEBGPU_0));
EXPECT_THAT(getDiagnosticString(),
HasSubstr("For WebGPU, an OpTypeStruct variable containing an "
"OpTypeRuntimeArray must be decorated with Block if it "
"has storage class StorageBuffer.\n %6 = OpVariable "
"%_ptr_StorageBuffer__struct_4 StorageBuffer\n"));
}
TEST_F(ValidateMemory, VulkanRTAInsideUniformStructGood) {
std::string spirv = R"(
OpCapability Shader
@ -1973,6 +2093,39 @@ OpFunctionEnd
EXPECT_EQ(SPV_SUCCESS, ValidateInstructions(SPV_ENV_VULKAN_1_1));
}
TEST_F(ValidateMemory, WebGPURTAInsideUniformStructBad) {
std::string spirv = R"(
OpCapability Shader
OpCapability VulkanMemoryModelKHR
OpExtension "SPV_KHR_vulkan_memory_model"
OpMemoryModel Logical VulkanKHR
OpEntryPoint Fragment %func "func"
OpExecutionMode %func OriginUpperLeft
OpDecorate %array_t ArrayStride 4
OpMemberDecorate %struct_t 0 Offset 0
OpDecorate %struct_t BufferBlock
%uint_t = OpTypeInt 32 0
%array_t = OpTypeRuntimeArray %uint_t
%struct_t = OpTypeStruct %array_t
%struct_ptr = OpTypePointer Uniform %struct_t
%2 = OpVariable %struct_ptr Uniform
%void = OpTypeVoid
%func_t = OpTypeFunction %void
%func = OpFunction %void None %func_t
%1 = OpLabel
OpReturn
OpFunctionEnd
)";
CompileSuccessfully(spirv.c_str(), SPV_ENV_WEBGPU_0);
EXPECT_EQ(SPV_ERROR_INVALID_ID, ValidateInstructions(SPV_ENV_WEBGPU_0));
EXPECT_THAT(
getDiagnosticString(),
HasSubstr("For WebGPU, OpTypeStruct variables containing "
"OpTypeRuntimeArray must have storage class of StorageBuffer\n "
" %6 = OpVariable %_ptr_Uniform__struct_3 Uniform\n"));
}
TEST_F(ValidateMemory, VulkanRTAInsideUniformStructWithoutBufferBlockBad) {
std::string spirv = R"(
OpCapability Shader
@ -2030,6 +2183,37 @@ OpFunctionEnd
"OpTypeRuntimeArray %_runtimearr_2\n"));
}
TEST_F(ValidateMemory, WebGPURTAInsideRTABad) {
std::string spirv = R"(
OpCapability Shader
OpCapability VulkanMemoryModelKHR
OpExtension "SPV_KHR_vulkan_memory_model"
OpMemoryModel Logical VulkanKHR
OpEntryPoint Fragment %func "func"
OpExecutionMode %func OriginUpperLeft
%sampler_t = OpTypeSampler
%inner_array_t = OpTypeRuntimeArray %sampler_t
%array_t = OpTypeRuntimeArray %inner_array_t
%array_ptr = OpTypePointer UniformConstant %array_t
%2 = OpVariable %array_ptr UniformConstant
%void = OpTypeVoid
%func_t = OpTypeFunction %void
%func = OpFunction %void None %func_t
%1 = OpLabel
OpReturn
OpFunctionEnd
)";
CompileSuccessfully(spirv.c_str(), SPV_ENV_WEBGPU_0);
EXPECT_EQ(SPV_ERROR_INVALID_ID, ValidateInstructions(SPV_ENV_WEBGPU_0));
EXPECT_THAT(
getDiagnosticString(),
HasSubstr(
"OpTypeRuntimeArray Element Type <id> '3[%_runtimearr_2]' is not "
"valid in WebGPU environment.\n %_runtimearr__runtimearr_2 = "
"OpTypeRuntimeArray %_runtimearr_2\n"));
}
TEST_F(ValidateMemory, VulkanRTAInsideRTAWithRuntimeDescriptorArrayBad) {
std::string spirv = R"(
OpCapability RuntimeDescriptorArrayEXT
@ -2190,6 +2374,38 @@ OpFunctionEnd
"OpTypeArray %_runtimearr_4 %uint_1\n"));
}
TEST_F(ValidateMemory, WebGPURTAInsideArrayBad) {
std::string spirv = R"(
OpCapability Shader
OpCapability VulkanMemoryModelKHR
OpExtension "SPV_KHR_vulkan_memory_model"
OpMemoryModel Logical VulkanKHR
OpEntryPoint Fragment %func "func"
OpExecutionMode %func OriginUpperLeft
%uint_t = OpTypeInt 32 0
%dim = OpConstant %uint_t 1
%sampler_t = OpTypeSampler
%inner_array_t = OpTypeRuntimeArray %sampler_t
%array_t = OpTypeArray %inner_array_t %dim
%array_ptr = OpTypePointer UniformConstant %array_t
%2 = OpVariable %array_ptr UniformConstant
%void = OpTypeVoid
%func_t = OpTypeFunction %void
%func = OpFunction %void None %func_t
%1 = OpLabel
OpReturn
OpFunctionEnd
)";
CompileSuccessfully(spirv.c_str(), SPV_ENV_WEBGPU_0);
EXPECT_EQ(SPV_ERROR_INVALID_ID, ValidateInstructions(SPV_ENV_WEBGPU_0));
EXPECT_THAT(
getDiagnosticString(),
HasSubstr("OpTypeArray Element Type <id> '5[%_runtimearr_4]' is not "
"valid in WebGPU environment.\n %_arr__runtimearr_4_uint_1 = "
"OpTypeArray %_runtimearr_4 %uint_1\n"));
}
TEST_F(ValidateMemory, VulkanRTAInsideArrayWithRuntimeDescriptorArrayBad) {
std::string spirv = R"(
OpCapability RuntimeDescriptorArrayEXT

View File

@ -478,8 +478,11 @@ OptStatus ParseOconfigFlag(const char* prog_name, const char* opt_flag,
new_argv[i] = flags[i].c_str();
}
return ParseFlags(static_cast<int>(flags.size()), new_argv, optimizer,
in_file, out_file, validator_options, optimizer_options);
auto ret_val =
ParseFlags(static_cast<int>(flags.size()), new_argv, optimizer, in_file,
out_file, validator_options, optimizer_options);
delete[] new_argv;
return ret_val;
}
// Canonicalize the flag in |argv[argi]| of the form '--pass arg' into
@ -660,7 +663,6 @@ int main(int argc, const char** argv) {
spv_target_env target_env = kDefaultEnvironment;
spvtools::Optimizer optimizer(target_env);
optimizer.SetMessageConsumer(spvtools::utils::CLIMessageConsumer);