Non-Functional: Add postprocess as more robust way to add capabilities

SPIRV/CMakeLists.txt

@@ -3,6 +3,7 @@ set(SOURCES
     InReadableOrder.cpp
     Logger.cpp
     SpvBuilder.cpp
+    SpvPostProcess.cpp
     doc.cpp
     disassemble.cpp)
 

SPIRV/GlslangToSpv.cpp

@@ -1191,6 +1191,7 @@ TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion, const gl
 // Finish creating SPV, after the traversal is complete.
 void TGlslangToSpvTraverser::finishSpv()
 {
+    // Finish the entry point function
     if (! entryPointTerminated) {
         builder.setBuildPoint(shaderEntry->getLastBlock());
         builder.leaveFunction();
@@ -1200,7 +1201,9 @@ void TGlslangToSpvTraverser::finishSpv()
     for (auto it = iOSet.cbegin(); it != iOSet.cend(); ++it)
         entryPoint->addIdOperand(*it);
 
-    builder.eliminateDeadDecorations();
+    // Add capabilities, extensions, remove unneeded decorations, etc., 
+    // based on the resulting SPIR-V.
+    builder.postProcess();
 }
 
 // Write the SPV into 'out'.
@@ -4627,27 +4630,21 @@ spv::Id TGlslangToSpvTraverser::createUnaryOperation(glslang::TOperator op, OpDe
         unaryOp = spv::OpFwidth;
         break;
     case glslang::EOpDPdxFine:
-        builder.addCapability(spv::CapabilityDerivativeControl);
         unaryOp = spv::OpDPdxFine;
         break;
     case glslang::EOpDPdyFine:
-        builder.addCapability(spv::CapabilityDerivativeControl);
         unaryOp = spv::OpDPdyFine;
         break;
     case glslang::EOpFwidthFine:
-        builder.addCapability(spv::CapabilityDerivativeControl);
         unaryOp = spv::OpFwidthFine;
         break;
     case glslang::EOpDPdxCoarse:
-        builder.addCapability(spv::CapabilityDerivativeControl);
         unaryOp = spv::OpDPdxCoarse;
         break;
     case glslang::EOpDPdyCoarse:
-        builder.addCapability(spv::CapabilityDerivativeControl);
         unaryOp = spv::OpDPdyCoarse;
         break;
     case glslang::EOpFwidthCoarse:
-        builder.addCapability(spv::CapabilityDerivativeControl);
         unaryOp = spv::OpFwidthCoarse;
         break;
     case glslang::EOpInterpolateAtCentroid:
@@ -4655,7 +4652,6 @@ spv::Id TGlslangToSpvTraverser::createUnaryOperation(glslang::TOperator op, OpDe
         if (typeProxy == glslang::EbtFloat16)
             builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float);
 #endif
-        builder.addCapability(spv::CapabilityInterpolationFunction);
         libCall = spv::GLSLstd450InterpolateAtCentroid;
         break;
     case glslang::EOpAny:
@@ -4791,8 +4787,6 @@ spv::Id TGlslangToSpvTraverser::createUnaryOperation(glslang::TOperator op, OpDe
 #endif
 #ifdef NV_EXTENSIONS
     case glslang::EOpSubgroupPartition:
-        builder.addExtension(spv::E_SPV_NV_shader_subgroup_partitioned);
-        builder.addCapability(spv::CapabilityGroupNonUniformPartitionedNV);
         unaryOp = spv::OpGroupNonUniformPartitionNV;
         break;
 #endif
@@ -6087,7 +6081,6 @@ spv::Id TGlslangToSpvTraverser::createMiscOperation(glslang::TOperator op, spv::
         if (typeProxy == glslang::EbtFloat16)
             builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float);
 #endif
-        builder.addCapability(spv::CapabilityInterpolationFunction);
         libCall = spv::GLSLstd450InterpolateAtSample;
         break;
     case glslang::EOpInterpolateAtOffset:
@@ -6095,7 +6088,6 @@ spv::Id TGlslangToSpvTraverser::createMiscOperation(glslang::TOperator op, spv::
         if (typeProxy == glslang::EbtFloat16)
             builder.addExtension(spv::E_SPV_AMD_gpu_shader_half_float);
 #endif
-        builder.addCapability(spv::CapabilityInterpolationFunction);
         libCall = spv::GLSLstd450InterpolateAtOffset;
         break;
     case glslang::EOpAddCarry:

SPIRV/SpvBuilder.cpp

@@ -1774,9 +1774,6 @@ Id Builder::createTextureCall(Decoration precision, Id resultType, bool sparse,
 // Comments in header
 Id Builder::createTextureQueryCall(Op opCode, const TextureParameters& parameters, bool isUnsignedResult)
 {
-    // All these need a capability
-    addCapability(CapabilityImageQuery);
-
     // Figure out the result type
     Id resultType = 0;
     switch (opCode) {
@@ -2483,42 +2480,6 @@ Id Builder::accessChainGetInferredType()
     return type;
 }
 
-// comment in header
-void Builder::eliminateDeadDecorations() {
-    std::unordered_set<const Block*> reachable_blocks;
-    std::unordered_set<Id> unreachable_definitions;
-    // Collect IDs defined in unreachable blocks. For each function, label the
-    // reachable blocks first. Then for each unreachable block, collect the
-    // result IDs of the instructions in it.
-    for (std::vector<Function*>::const_iterator fi = module.getFunctions().cbegin();
-        fi != module.getFunctions().cend(); fi++) {
-        Function* f = *fi;
-        Block* entry = f->getEntryBlock();
-        inReadableOrder(entry, [&reachable_blocks](const Block* b) {
-            reachable_blocks.insert(b);
-        });
-        for (std::vector<Block*>::const_iterator bi = f->getBlocks().cbegin();
-            bi != f->getBlocks().cend(); bi++) {
-            Block* b = *bi;
-            if (!reachable_blocks.count(b)) {
-                for (std::vector<std::unique_ptr<Instruction> >::const_iterator
-                         ii = b->getInstructions().cbegin();
-                    ii != b->getInstructions().cend(); ii++) {
-                    Instruction* i = ii->get();
-                    unreachable_definitions.insert(i->getResultId());
-                }
-            }
-        }
-    }
-    decorations.erase(std::remove_if(decorations.begin(), decorations.end(),
-        [&unreachable_definitions](std::unique_ptr<Instruction>& I) -> bool {
-            Instruction* inst = I.get();
-            Id decoration_id = inst->getIdOperand(0);
-            return unreachable_definitions.count(decoration_id) != 0;
-        }),
-        decorations.end());
-}
-
 void Builder::dump(std::vector<unsigned int>& out) const
 {
     // Header, before first instructions:

SPIRV/SpvBuilder.h

@@ -563,9 +563,15 @@ public:
     // based on the type of the base and the chain of dereferences.
     Id accessChainGetInferredType();
 
-    // Remove OpDecorate instructions whose operands are defined in unreachable
-    // blocks.
-    void eliminateDeadDecorations();
+    // Add capabilities, extensions, remove unneeded decorations, etc., 
+    // based on the resulting SPIR-V.
+    void postProcess();
+
+    // Hook to visit each instruction in a block in a function
+    void postProcess(Instruction& inst);
+    // Hook to visit each instruction in a reachable block in a function.
+    void postProcessReachable(Instruction& inst);
+
     void dump(std::vector<unsigned int>&) const;
 
     void createBranch(Block* block);

SPIRV/SpvPostProcess.cpp

+//
+// Copyright (C) 2016-2018 Google, Inc.
+//
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions
+// are met:
+//
+//    Redistributions of source code must retain the above copyright
+//    notice, this list of conditions and the following disclaimer.
+//
+//    Redistributions in binary form must reproduce the above
+//    copyright notice, this list of conditions and the following
+//    disclaimer in the documentation and/or other materials provided
+//    with the distribution.
+//
+//    Neither the name of 3Dlabs Inc. Ltd. nor the names of its
+//    contributors may be used to endorse or promote products derived
+//    from this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
+// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
+// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+// POSSIBILITY OF SUCH DAMAGE.
+
+//
+// Post-processing for SPIR-V IR, in internal form, not standard binary form.
+//
+
+#include <cassert>
+#include <cstdlib>
+
+#include <unordered_set>
+#include <algorithm>
+
+#include "SpvBuilder.h"
+
+#include "spirv.hpp"
+#include "GlslangToSpv.h"
+#include "SpvBuilder.h"
+namespace spv {
+    #include "GLSL.std.450.h"
+    #include "GLSL.ext.KHR.h"
+    #include "GLSL.ext.EXT.h"
+#ifdef AMD_EXTENSIONS
+    #include "GLSL.ext.AMD.h"
+#endif
+#ifdef NV_EXTENSIONS
+    #include "GLSL.ext.NV.h"
+#endif
+}
+
+namespace spv {
+
+// Called for each instruction in a block.
+void Builder::postProcess(Instruction& inst)
+{
+    // Add capabilities based simply on the opcode.
+    switch (inst.getOpCode()) {
+    case OpExtInst:
+        switch (inst.getIdOperand(1)) {
+        case GLSLstd450InterpolateAtCentroid:
+        case GLSLstd450InterpolateAtSample:
+        case GLSLstd450InterpolateAtOffset:
+            addCapability(CapabilityInterpolationFunction);
+            break;
+        default:
+            break;
+        }
+        break;
+    case OpDPdxFine:
+    case OpDPdyFine:
+    case OpFwidthFine:
+    case OpDPdxCoarse:
+    case OpDPdyCoarse:
+    case OpFwidthCoarse:
+        addCapability(CapabilityDerivativeControl);
+        break;
+
+    case OpImageQueryLod:
+    case OpImageQuerySize:
+    case OpImageQuerySizeLod:
+    case OpImageQuerySamples:
+    case OpImageQueryLevels:
+        addCapability(CapabilityImageQuery);
+        break;
+
+#ifdef NV_EXTENSIONS
+    case OpGroupNonUniformPartitionNV:
+        addExtension(E_SPV_NV_shader_subgroup_partitioned);
+        addCapability(CapabilityGroupNonUniformPartitionedNV);
+        break;
+#endif
+
+    default:
+        break;
+    }
+}
+
+// Called for each instruction in a reachable block.
+void Builder::postProcessReachable(Instruction& inst)
+{
+    // did have code here, but questionable to do so without deleting the instructions
+}
+
+// comment in header
+void Builder::postProcess()
+{
+    std::unordered_set<const Block*> reachableBlocks;
+    std::unordered_set<Id> unreachableDefinitions;
+    // Collect IDs defined in unreachable blocks. For each function, label the
+    // reachable blocks first. Then for each unreachable block, collect the
+    // result IDs of the instructions in it.
+    for (auto fi = module.getFunctions().cbegin(); fi != module.getFunctions().cend(); fi++) {
+        Function* f = *fi;
+        Block* entry = f->getEntryBlock();
+        inReadableOrder(entry, [&reachableBlocks](const Block* b) { reachableBlocks.insert(b); });
+        for (auto bi = f->getBlocks().cbegin(); bi != f->getBlocks().cend(); bi++) {
+            Block* b = *bi;
+            if (reachableBlocks.count(b) == 0) {
+                for (auto ii = b->getInstructions().cbegin(); ii != b->getInstructions().cend(); ii++)
+                    unreachableDefinitions.insert(ii->get()->getResultId());
+            }
+        }
+    }
+
+    // Remove unneeded decorations, for unreachable instructions
+    decorations.erase(std::remove_if(decorations.begin(), decorations.end(),
+        [&unreachableDefinitions](std::unique_ptr<Instruction>& I) -> bool {
+            Id decoration_id = I.get()->getIdOperand(0);
+            return unreachableDefinitions.count(decoration_id) != 0;
+        }),
+        decorations.end());
+
+    // Add per-instruction capabilities, extensions, etc.,
+
+    // process all reachable instructions...
+    for (auto bi = reachableBlocks.cbegin(); bi != reachableBlocks.cend(); ++bi) {
+        const Block* block = *bi;
+        const auto function = [this](const std::unique_ptr<Instruction>& inst) { postProcessReachable(*inst.get()); };
+        std::for_each(block->getInstructions().begin(), block->getInstructions().end(), function);
+    }
+
+    // process all block-contained instructions
+    for (auto fi = module.getFunctions().cbegin(); fi != module.getFunctions().cend(); fi++) {
+        Function* f = *fi;
+        for (auto bi = f->getBlocks().cbegin(); bi != f->getBlocks().cend(); bi++) {
+            Block* b = *bi;
+            for (auto ii = b->getInstructions().cbegin(); ii != b->getInstructions().cend(); ii++)
+                postProcess(*ii->get());
+        }
+    }
+}
+
+}; // end spv namespace