[SPIR-V] Add SPIR-V structurizer

This commit adds an initial SPIR-V structurizer.
It leverages the previously merged passes, and the convergence region
analysis to determine the correct merge and continue blocks for SPIR-V.

The first part does a branch cleanup (simplifying switches, and
legalizing them), then merge instructions are added to cycles,
convergent and later divergent blocks.
Then comes the important part: splitting critical edges, and making sure
the divergent construct boundaries don't cross.

- we split blocks with multiple headers into 2 blocks.
- we split blocks that are a merge blocks for 2 or more constructs:
  SPIR-V spec disallow a merge block to be shared by 2 loop/switch/condition
  construct.
- we split merge & continue blocks: SPIR-V spec disallow a basic block to be
  both a continue block, and a merge block.
- we remove superfluous headers: when a header doesn't bring more info than
  the parent on the divergence state, it must be removed.

This PR leverages the merged SPIR-V simulator for testing, as long as
spirv-val. For now, most DXC structurization tests are passing. The
unsupported ones are either caused by unsupported features like switches
on boolean types, or switches in region exits, because the MergeExit
pass doesn't support those yet (there is a FIXME).

This PR is quite large, and the addition not trivial, so I tried to keep
it simple. E.G: as soon as the CFG changes, I recompute the dominator trees
and other structures instead of updating them.

Author: Keenuts

clang/test/CodeGenHLSL/convergence/cf.for.plain.hlsl

@@ -0,0 +1,44 @@
+// RUN: %clang_cc1 -std=hlsl2021 -finclude-default-header -x hlsl -triple \
+// RUN:   spirv-pc-vulkan-library %s -emit-llvm -disable-llvm-passes -o - | FileCheck %s
+
+int process() {
+// CHECK: entry:
+// CHECK:   %[[#entry_token:]] = call token @llvm.experimental.convergence.entry()
+  int val = 0;
+
+// CHECK: for.cond:
+// CHECK-NEXT:   %[[#]] = call token @llvm.experimental.convergence.loop() [ "convergencectrl"(token %[[#entry_token]]) ]
+// CHECK: br i1 {{.*}}, label %for.body, label %for.end
+  for (int i = 0; i < 10; ++i) {
+
+// CHECK: for.body:
+// CHECK:   br label %for.inc
+    val = i;
+
+// CHECK: for.inc:
+// CHECK:   br label %for.cond
+  }
+
+// CHECK: for.end:
+// CHECK:   br label %for.cond1
+
+  // Infinite loop
+  for ( ; ; ) {
+// CHECK: for.cond1:
+// CHECK-NEXT:   %[[#]] = call token @llvm.experimental.convergence.loop() [ "convergencectrl"(token %[[#entry_token]]) ]
+// CHECK:   br label %for.cond1
+    val = 0;
+  }
+
+// CHECK-NEXT: }
+// This loop in unreachable. Not generated.
+  // Null body
+  for (int j = 0; j < 10; ++j)
+  ;
+  return val;
+}
+
+[numthreads(1, 1, 1)]
+void main() {
+  process();
+}

llvm/include/llvm/IR/IntrinsicsSPIRV.td

@@ -31,6 +31,8 @@ let TargetPrefix = "spv" in {
   def int_spv_bitcast : Intrinsic<[llvm_any_ty], [llvm_any_ty]>;
   def int_spv_ptrcast : Intrinsic<[llvm_any_ty], [llvm_any_ty, llvm_metadata_ty, llvm_i32_ty], [ImmArg<ArgIndex<2>>]>;
   def int_spv_switch : Intrinsic<[], [llvm_any_ty, llvm_vararg_ty]>;
+  def int_spv_loop_merge : Intrinsic<[], [llvm_vararg_ty]>;
+  def int_spv_selection_merge : Intrinsic<[], [llvm_vararg_ty]>;
   def int_spv_cmpxchg : Intrinsic<[llvm_i32_ty], [llvm_any_ty, llvm_vararg_ty]>;
   def int_spv_unreachable : Intrinsic<[], []>;
   def int_spv_alloca : Intrinsic<[llvm_any_ty], []>;

llvm/lib/Target/SPIRV/Analysis/SPIRVConvergenceRegionAnalysis.cpp

@@ -203,7 +203,8 @@ class ConvergenceRegionAnalyzer {
 
 private:
   bool isBackEdge(const BasicBlock *From, const BasicBlock *To) const {
-    assert(From != To && "From == To. This is awkward.");
+    if (From == To)
+      return true;
 
     // We only handle loop in the simplified form. This means:
     // - a single back-edge, a single latch.
@@ -230,6 +231,7 @@ class ConvergenceRegionAnalyzer {
     auto *Terminator = From->getTerminator();
     for (unsigned i = 0; i < Terminator->getNumSuccessors(); ++i) {
       auto *To = Terminator->getSuccessor(i);
+      // Ignore back edges.
       if (isBackEdge(From, To))
         continue;
 
@@ -276,7 +278,6 @@ class ConvergenceRegionAnalyzer {
     while (ToProcess.size() != 0) {
       auto *L = ToProcess.front();
       ToProcess.pop();
-      assert(L->isLoopSimplifyForm());
 
       auto CT = getConvergenceToken(L->getHeader());
       SmallPtrSet<BasicBlock *, 8> RegionBlocks(L->block_begin(),

llvm/lib/Target/SPIRV/Analysis/SPIRVConvergenceRegionAnalysis.h

@@ -130,6 +130,9 @@ class ConvergenceRegionInfo {
   }
 
   const ConvergenceRegion *getTopLevelRegion() const { return TopLevelRegion; }
+  ConvergenceRegion *getWritableTopLevelRegion() const {
+    return TopLevelRegion;
+  }
 };
 
 } // namespace SPIRV

llvm/lib/Target/SPIRV/CMakeLists.txt

@@ -31,6 +31,7 @@ add_llvm_target(SPIRVCodeGen
   SPIRVMCInstLower.cpp
   SPIRVMetadata.cpp
   SPIRVModuleAnalysis.cpp
+  SPIRVStructurizer.cpp
   SPIRVPreLegalizer.cpp
   SPIRVPostLegalizer.cpp
   SPIRVPrepareFunctions.cpp

llvm/lib/Target/SPIRV/SPIRV.h

@@ -20,6 +20,7 @@ class InstructionSelector;
 class RegisterBankInfo;
 
 ModulePass *createSPIRVPrepareFunctionsPass(const SPIRVTargetMachine &TM);
+FunctionPass *createSPIRVStructurizerPass();
 FunctionPass *createSPIRVMergeRegionExitTargetsPass();
 FunctionPass *createSPIRVStripConvergenceIntrinsicsPass();
 FunctionPass *createSPIRVRegularizerPass();

llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp

@@ -2296,6 +2296,19 @@ bool SPIRVInstructionSelector::selectIntrinsic(Register ResVReg,
     }
     return MIB.constrainAllUses(TII, TRI, RBI);
   }
+  case Intrinsic::spv_loop_merge:
+  case Intrinsic::spv_selection_merge: {
+    const auto Opcode = IID == Intrinsic::spv_selection_merge
+                            ? SPIRV::OpSelectionMerge
+                            : SPIRV::OpLoopMerge;
+    auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode));
+    for (unsigned i = 1; i < I.getNumExplicitOperands(); ++i) {
+      assert(I.getOperand(i).isMBB());
+      MIB.addMBB(I.getOperand(i).getMBB());
+    }
+    MIB.addImm(SPIRV::SelectionControl::None);
+    return MIB.constrainAllUses(TII, TRI, RBI);
+  }
   case Intrinsic::spv_cmpxchg:
     return selectAtomicCmpXchg(ResVReg, ResType, I);
   case Intrinsic::spv_unreachable:

llvm/lib/Target/SPIRV/SPIRVMergeRegionExitTargets.cpp

@@ -133,7 +133,7 @@ class SPIRVMergeRegionExitTargets : public FunctionPass {
   // Run the pass on the given convergence region, ignoring the sub-regions.
   // Returns true if the CFG changed, false otherwise.
   bool runOnConvergenceRegionNoRecurse(LoopInfo &LI,
-                                       const SPIRV::ConvergenceRegion *CR) {
+                                       SPIRV::ConvergenceRegion *CR) {
     // Gather all the exit targets for this region.
     SmallPtrSet<BasicBlock *, 4> ExitTargets;
     for (BasicBlock *Exit : CR->Exits) {
@@ -198,14 +198,19 @@ class SPIRVMergeRegionExitTargets : public FunctionPass {
     for (auto Exit : CR->Exits)
       replaceBranchTargets(Exit, ExitTargets, NewExitTarget);
 
+    CR = CR->Parent;
+    while (CR) {
+      CR->Blocks.insert(NewExitTarget);
+      CR = CR->Parent;
+    }
+
     return true;
   }
 
   /// Run the pass on the given convergence region and sub-regions (DFS).
   /// Returns true if a region/sub-region was modified, false otherwise.
   /// This returns as soon as one region/sub-region has been modified.
-  bool runOnConvergenceRegion(LoopInfo &LI,
-                              const SPIRV::ConvergenceRegion *CR) {
+  bool runOnConvergenceRegion(LoopInfo &LI, SPIRV::ConvergenceRegion *CR) {
     for (auto *Child : CR->Children)
       if (runOnConvergenceRegion(LI, Child))
         return true;
@@ -235,20 +240,17 @@ class SPIRVMergeRegionExitTargets : public FunctionPass {
 
   virtual bool runOnFunction(Function &F) override {
     LoopInfo &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
-    const auto *TopLevelRegion =
+    auto *TopLevelRegion =
         getAnalysis<SPIRVConvergenceRegionAnalysisWrapperPass>()
             .getRegionInfo()
-            .getTopLevelRegion();
+            .getWritableTopLevelRegion();
 
     // FIXME: very inefficient method: each time a region is modified, we bubble
     // back up, and recompute the whole convergence region tree. Once the
     // algorithm is completed and test coverage good enough, rewrite this pass
     // to be efficient instead of simple.
     bool modified = false;
     while (runOnConvergenceRegion(LI, TopLevelRegion)) {
-      TopLevelRegion = getAnalysis<SPIRVConvergenceRegionAnalysisWrapperPass>()
-                           .getRegionInfo()
-                           .getTopLevelRegion();
       modified = true;
     }
 
@@ -262,6 +264,8 @@ class SPIRVMergeRegionExitTargets : public FunctionPass {
     AU.addRequired<DominatorTreeWrapperPass>();
     AU.addRequired<LoopInfoWrapperPass>();
     AU.addRequired<SPIRVConvergenceRegionAnalysisWrapperPass>();
+
+    AU.addPreserved<SPIRVConvergenceRegionAnalysisWrapperPass>();
     FunctionPass::getAnalysisUsage(AU);
   }
 };