Signal when SPIR-V legalization needs loop unroll

tools/clang/lib/Frontend/Rewrite/RewriteObjC.cpp

@@ -483,7 +483,7 @@ namespace {
         result =  Context->getObjCIdType();
       FunctionProtoType::ExtProtoInfo fpi;
       fpi.Variadic = variadic;
-      return Context->getFunctionType(result, args, fpi);
+      return Context->getFunctionType(result, args, fpi, {});
     }
 
     // Helper function: create a CStyleCastExpr with trivial type source info.

tools/clang/lib/SPIRV/SpirvEmitter.cpp

@@ -593,6 +593,8 @@ SpirvEmitter::SpirvEmitter(CompilerInstance &ci)
       constEvaluator(astContext, spvBuilder), entryFunction(nullptr),
       curFunction(nullptr), curThis(nullptr), seenPushConstantAt(),
       isSpecConstantMode(false), needsLegalization(false),
+      needsLegalizationLoopUnroll(false),
+      needsLegalizationSsaRewrite(false),
       beforeHlslLegalization(false), mainSourceFile(nullptr) {
 
   // Get ShaderModel from command line hlsl profile option.
@@ -954,6 +956,9 @@ void SpirvEmitter::HandleTranslationUnit(ASTContext &context) {
       declIdMapper.requiresFlatteningCompositeResources() ||
       !dsetbindingsToCombineImageSampler.empty() ||
       spirvOptions.signaturePacking;
+  needsLegalizationSsaRewrite =
+      needsLegalizationSsaRewrite ||
+      !dsetbindingsToCombineImageSampler.empty();
 
   // Run legalization passes
   if (spirvOptions.codeGenHighLevel) {
@@ -5823,8 +5828,11 @@ SpirvInstruction *SpirvEmitter::createImageSample(
     SpirvInstruction *minLod, SpirvInstruction *residencyCodeId,
     SourceLocation loc, SourceRange range) {
 
-  if (varOffset)
+  if (varOffset) {
     needsLegalization = true;
+    needsLegalizationLoopUnroll = true;
+    needsLegalizationSsaRewrite = true;
+  }
 
   // SampleDref* instructions in SPIR-V always return a scalar.
   // They also have the correct type in HLSL.
@@ -8045,7 +8053,7 @@ SpirvInstruction *SpirvEmitter::createVectorSplat(const Expr *scalarExpr,
   // Should find a more meaningful one.
   if (auto *constVal = dyn_cast<SpirvConstant>(scalarVal)) {
     llvm::SmallVector<SpirvConstant *, 4> elements(size_t(size), constVal);
-    const bool isSpecConst = constVal->getopcode() == spv::Op::OpSpecConstant;
+    const bool isSpecConst = constVal->isSpecConstant();
     auto *value =
         spvBuilder.getConstantComposite(vecType, elements, isSpecConst);
     if (!value)
@@ -16665,7 +16673,15 @@ bool SpirvEmitter::spirvToolsLegalize(std::vector<uint32_t> *mod,
     optimizer.RegisterPass(
         spvtools::CreateInterfaceVariableScalarReplacementPass());
   }
-  optimizer.RegisterLegalizationPasses(spirvOptions.preserveInterface);
+  auto legalizationSsaRewriteMode = spvtools::SSARewriteMode::None;
+  if (needsLegalizationLoopUnroll) {
+    legalizationSsaRewriteMode = spvtools::SSARewriteMode::All;
+  } else if (needsLegalizationSsaRewrite) {
+    legalizationSsaRewriteMode = spvtools::SSARewriteMode::OpaqueOnly;
+  }
+  optimizer.RegisterLegalizationPasses(
+      spirvOptions.preserveInterface, needsLegalizationLoopUnroll,
+      legalizationSsaRewriteMode);
   // Add flattening of resources if needed.
   if (spirvOptions.flattenResourceArrays) {
     optimizer.RegisterPass(

tools/clang/lib/SPIRV/SpirvEmitter.h

@@ -1578,6 +1578,8 @@ class SpirvEmitter : public ASTConsumer {
   ///
   /// Note: legalization specific code
   bool needsLegalization;
+  bool needsLegalizationLoopUnroll;
+  bool needsLegalizationSsaRewrite;
 
   /// Whether the translated SPIR-V binary passes --before-hlsl-legalization
   /// option to spirv-val because of illegal function parameter scope.

tools/clang/test/CodeGenSPIRV/cast.flat-conversion.matrix.hlsl

@@ -16,20 +16,18 @@ RWStructuredBuffer<T> t_output;
 // COL: OpMemberDecorate %S 0 RowMajor
 // ROW: OpMemberDecorate %S 0 ColMajor
 
-// The DXIL generated for the two cases seem to produce the same results,
-// and this matches that behaviour.
-// CHECK: [[array_const:%[0-9]+]] = OpConstantComposite %_arr_float_uint_6 %float_0 %float_1 %float_2 %float_3 %float_4 %float_5
-// CHECK: [[t:%[0-9]+]] = OpConstantComposite %T [[array_const]]
-
 // The DXIL that is generates different order for the values depending on
 // whether the matrix is column or row major. However, for SPIR-V, the value
 // stored in both cases is the same because the decoration, which is checked
 // above, is what determines the layout in memory for the value.
 
-// CHECK: [[row0:%[0-9]+]] = OpConstantComposite %v3float %float_0 %float_1 %float_2
-// CHECK: [[row1:%[0-9]+]] = OpConstantComposite %v3float %float_3 %float_4 %float_5
-// CHECK: [[mat:%[0-9]+]] = OpConstantComposite %mat2v3float %33 %34
-// CHECK: [[s:%[0-9]+]] = OpConstantComposite %S %35
+// CHECK: [[mat:%[0-9]+]] = OpLoad %mat2v3float
+// CHECK: [[e00:%[0-9]+]] = OpCompositeExtract %float [[mat]] 0 0
+// CHECK: [[e01:%[0-9]+]] = OpCompositeExtract %float [[mat]] 0 1
+// CHECK: [[e02:%[0-9]+]] = OpCompositeExtract %float [[mat]] 0 2
+// CHECK: [[e10:%[0-9]+]] = OpCompositeExtract %float [[mat]] 1 0
+// CHECK: [[e11:%[0-9]+]] = OpCompositeExtract %float [[mat]] 1 1
+// CHECK: [[e12:%[0-9]+]] = OpCompositeExtract %float [[mat]] 1 2
 
 void main() {
   S s;
@@ -40,13 +38,16 @@ void main() {
       s.a[i][j] = i*3+j;
     }
   }
-// CHECK: [[ac:%[0-9]+]] = OpAccessChain %_ptr_Uniform_T %t_output %int_0 %uint_0
-// CHECK: OpStore [[ac]] [[t]]
+// CHECK: [[tptr:%[0-9]+]] = OpAccessChain %_ptr_Uniform_T %t_output %int_0 %uint_0
+// CHECK: [[tarr:%[0-9]+]] = OpCompositeConstruct %_arr_float_uint_6 [[e00]] [[e01]] [[e02]] [[e10]] [[e11]] [[e12]]
+// CHECK: [[tval:%[0-9]+]] = OpCompositeConstruct %T [[tarr]]
+// CHECK: OpStore [[tptr]] [[tval]]
   T t = (T)(s);
   t_output[0] = t;
 
-// CHECK: [[ac:%[0-9]+]] = OpAccessChain %_ptr_Uniform_S %s_output %int_0 %uint_0
-// CHECK: OpStore [[ac]] [[s]]
+// CHECK: [[sptr:%[0-9]+]] = OpAccessChain %_ptr_Uniform_S %s_output %int_0 %uint_0
+// CHECK: [[sval:%[0-9]+]] = OpCompositeConstruct %S [[mat]]
+// CHECK: OpStore [[sptr]] [[sval]]
   s = (S)t;
   s_output[0] = s;
 }

tools/clang/test/CodeGenSPIRV/legal-examples/15-loop-var-unroll-ok.hlsl

@@ -1,14 +1,12 @@
 // RUN: %dxc -T cs_6_0 -E main -O3 -Vd %s -spirv | FileCheck %s
 
-// CHECK:      [[ptr:%[0-9]+]] = OpAccessChain %_ptr_Uniform_S %gSBuffer2
-// CHECK-NEXT: [[val:%[0-9]+]] = OpLoad %S [[ptr]]
-// CHECK-NEXT: [[ptr:%[0-9]+]] = OpAccessChain %_ptr_Uniform_S %gRWSBuffer
-// CHECK-NEXT:                OpStore [[ptr]] [[val]]
-
-// CHECK:      [[ptr:%[0-9]+]] = OpAccessChain %_ptr_Uniform_S %gSBuffer2
-// CHECK-NEXT: [[val:%[0-9]+]] = OpLoad %S [[ptr]]
-// CHECK-NEXT: [[ptr:%[0-9]+]] = OpAccessChain %_ptr_Uniform_S %gRWSBuffer
-// CHECK-NEXT:                OpStore [[ptr]] [[val]]
+// CHECK: OpLoopMerge {{%[0-9]+}} {{%[0-9]+}} Unroll
+// CHECK: [[which:%[0-9]+]] = OpSelect %_ptr_Uniform_type_StructuredBuffer_S {{%[0-9]+}} %gSBuffer1 %gSBuffer2
+// CHECK: [[idx:%[0-9]+]] = OpBitcast %uint {{%[0-9]+}}
+// CHECK: [[src:%[0-9]+]] = OpAccessChain %_ptr_Uniform_S [[which]] %int_0 [[idx]]
+// CHECK: [[val:%[0-9]+]] = OpLoad %S [[src]]
+// CHECK: [[dst:%[0-9]+]] = OpAccessChain %_ptr_Uniform_S %gRWSBuffer %int_0 [[idx]]
+// CHECK: OpStore [[dst]] [[val]]
 
 struct S {
   float4 f;

tools/clang/test/CodeGenSPIRV/max_id.hlsl

@@ -8,7 +8,7 @@
 // CHECK-30: fatal error: failed to optimize SPIR-V: ID overflow. Try running compact-ids.
 
 // With a larger limit, the test case can compile successfully.
-// CHECK-400: Bound: 204
+// CHECK-400: Bound:
 
 
 RWStructuredBuffer<int> data;
@@ -20,4 +20,4 @@ void main(uint3 id : SV_DispatchThreadID)
   for( int i = 0; i < 64; i++ ) {
     data[i] = i;
   }
-}
+}

tools/clang/test/CodeGenSPIRV/type.rwstructured-buffer.array.counter.indirect.hlsl

@@ -20,11 +20,13 @@ void func(RWStructuredBuffer<uint> local) {
 
 float4 main(PSInput input) : SV_TARGET
 {
-// CHECK: [[ac2:%[0-9]+]] = OpAccessChain %_ptr_Uniform_int %counter_var_g_rwbuffer {{%[0-9]+}} %uint_0
+// CHECK: [[counter_struct:%[0-9]+]] = OpAccessChain %_ptr_Uniform_type_ACSBuffer_counter %counter_var_g_rwbuffer {{%[0-9]+}}
+// CHECK: [[ac2:%[0-9]+]] = OpAccessChain %_ptr_Uniform_int [[counter_struct]] %uint_0
 // CHECK: OpAtomicIAdd %int [[ac2]] %uint_1 %uint_0 %int_1
     func(g_rwbuffer[input.idx]);
 
-// CHECK: [[ac2_0:%[0-9]+]] = OpAccessChain %_ptr_Uniform_uint %g_rwbuffer {{%[0-9]+}} %int_0 %uint_0
+// CHECK: [[rwbuffer_struct:%[0-9]+]] = OpAccessChain %_ptr_Uniform_type_RWStructuredBuffer_uint %g_rwbuffer {{%[0-9]+}}
+// CHECK: [[ac2_0:%[0-9]+]] = OpAccessChain %_ptr_Uniform_uint [[rwbuffer_struct]] %int_0 %uint_0
 // CHECK: OpLoad %uint [[ac2_0]]
     return g_rwbuffer[input.idx][0];
 }

tools/clang/test/CodeGenSPIRV/type.rwstructured-buffer.array.counter.indirect2.hlsl

@@ -18,11 +18,13 @@ float4 main(PSInput input) : SV_TARGET
 {
     RWStructuredBuffer<uint> l_rwbuffer[5] = g_rwbuffer;
 
-// CHECK: [[ac2:%[0-9]+]] = OpAccessChain %_ptr_Uniform_int %counter_var_g_rwbuffer %int_0 %uint_0
+// CHECK: [[counter_struct:%[0-9]+]] = OpAccessChain %_ptr_Uniform_type_ACSBuffer_counter %counter_var_g_rwbuffer %int_0
+// CHECK: [[ac2:%[0-9]+]] = OpAccessChain %_ptr_Uniform_int [[counter_struct]] %uint_0
 // CHECK: OpAtomicIAdd %int [[ac2]] %uint_1 %uint_0 %int_1
     l_rwbuffer[0].IncrementCounter();
 
-// CHECK: [[ac2_0:%[0-9]+]] = OpAccessChain %_ptr_Uniform_uint %g_rwbuffer {{%[0-9]+}} %int_0 %uint_0
+// CHECK: [[rwbuffer_struct:%[0-9]+]] = OpAccessChain %_ptr_Uniform_type_RWStructuredBuffer_uint %g_rwbuffer {{%[0-9]+}}
+// CHECK: [[ac2_0:%[0-9]+]] = OpAccessChain %_ptr_Uniform_uint [[rwbuffer_struct]] %int_0 %uint_0
 // CHECK: OpLoad %uint [[ac2_0]]
     return l_rwbuffer[input.idx][0];
 }

tools/clang/test/CodeGenSPIRV/vk.binding.global-struct-of-resources.optimized.hlsl

@@ -3,20 +3,20 @@
 // Check the names
 //
 // CHECK: OpName %secondGlobal_t "secondGlobal.t"
-// CHECK: OpName %[[fg0:firstGlobal_[0-9]+__t]] "firstGlobal{{.*}}.t"
-// CHECK: OpName %[[fg1:firstGlobal_[0-9]+__t]] "firstGlobal{{.*}}.t"
-// CHECK: OpName %[[fg2:firstGlobal_[0-9]+__t]] "firstGlobal{{.*}}.t"
-// CHECK: OpName %[[fg3:firstGlobal_[0-9]+__t]] "firstGlobal{{.*}}.t"
+// CHECK: OpName %[[fg0:[0-9A-Za-z_]+]] "firstGlobal{{.*}}.t"
+// CHECK: OpName %[[fg1:[0-9A-Za-z_]+]] "firstGlobal{{.*}}.t"
+// CHECK: OpName %[[fg2:[0-9A-Za-z_]+]] "firstGlobal{{.*}}.t"
+// CHECK: OpName %[[fg3:[0-9A-Za-z_]+]] "firstGlobal{{.*}}.t"
 // CHECK: OpName %secondGlobal_tt_0__s "secondGlobal.tt[0].s"
 // CHECK: OpName %secondGlobal_tt_1__s "secondGlobal.tt[1].s"
-// CHECK: OpName %[[fgtt0_0:firstGlobal_[0-9]+__tt_0__s]] "firstGlobal{{.*}}.tt[0].s"
-// CHECK: OpName %[[fgtt0_1:firstGlobal_[0-9]+__tt_1__s]] "firstGlobal{{.*}}.tt[1].s"
-// CHECK: OpName %[[fgtt1_0:firstGlobal_[0-9]+__tt_0__s]] "firstGlobal{{.*}}.tt[0].s"
-// CHECK: OpName %[[fgtt1_1:firstGlobal_[0-9]+__tt_1__s]] "firstGlobal{{.*}}.tt[1].s"
-// CHECK: OpName %[[fgtt2_0:firstGlobal_[0-9]+__tt_0__s]] "firstGlobal{{.*}}.tt[0].s"
-// CHECK: OpName %[[fgtt2_1:firstGlobal_[0-9]+__tt_1__s]] "firstGlobal{{.*}}.tt[1].s"
-// CHECK: OpName %[[fgtt3_0:firstGlobal_[0-9]+__tt_0__s]] "firstGlobal{{.*}}.tt[0].s"
-// CHECK: OpName %[[fgtt3_1:firstGlobal_[0-9]+__tt_1__s]] "firstGlobal{{.*}}.tt[1].s"
+// CHECK: OpName %[[fgtt0_0:[0-9A-Za-z_]+]] "firstGlobal{{.*}}.tt[0].s"
+// CHECK: OpName %[[fgtt0_1:[0-9A-Za-z_]+]] "firstGlobal{{.*}}.tt[1].s"
+// CHECK: OpName %[[fgtt1_0:[0-9A-Za-z_]+]] "firstGlobal{{.*}}.tt[0].s"
+// CHECK: OpName %[[fgtt1_1:[0-9A-Za-z_]+]] "firstGlobal{{.*}}.tt[1].s"
+// CHECK: OpName %[[fgtt2_0:[0-9A-Za-z_]+]] "firstGlobal{{.*}}.tt[0].s"
+// CHECK: OpName %[[fgtt2_1:[0-9A-Za-z_]+]] "firstGlobal{{.*}}.tt[1].s"
+// CHECK: OpName %[[fgtt3_0:[0-9A-Za-z_]+]] "firstGlobal{{.*}}.tt[0].s"
+// CHECK: OpName %[[fgtt3_1:[0-9A-Za-z_]+]] "firstGlobal{{.*}}.tt[1].s"
 
 // Check flattening of bindings
 // Explanation: Only the resources that are used will have a binding assignment

tools/clang/test/CodeGenSPIRV/vk.spec-constant.reuse.hlsl

@@ -5,6 +5,6 @@
 
 [shader("pixel")]
 float4 main(float4 position : SV_Position) : SV_Target0 {
-// CHECK: OpConstantComposite %v4bool %y %y %y %y
+// CHECK: OpSpecConstantComposite %v4bool %y %y %y %y
     return y ? position : 1.0;
 }