[CIR] Add if statement support

clang/include/clang/CIR/Dialect/IR/CIRDialect.h

@@ -60,6 +60,10 @@ class SameFirstOperandAndResultType
 using BuilderCallbackRef =
     llvm::function_ref<void(mlir::OpBuilder &, mlir::Location)>;
 
+namespace cir {
+void buildTerminatedBody(mlir::OpBuilder &builder, mlir::Location loc);
+} // namespace cir
+
 // TableGen'erated files for MLIR dialects require that a macro be defined when
 // they are included.  GET_OP_CLASSES tells the file to define the classes for
 // the operations of that dialect.

clang/include/clang/CIR/Dialect/IR/CIROps.td

@@ -472,8 +472,8 @@ def StoreOp : CIR_Op<"store", [
 // ReturnOp
 //===----------------------------------------------------------------------===//
 
-def ReturnOp : CIR_Op<"return", [ParentOneOf<["FuncOp", "ScopeOp", "DoWhileOp",
-                                              "WhileOp", "ForOp"]>,
+def ReturnOp : CIR_Op<"return", [ParentOneOf<["FuncOp", "ScopeOp", "IfOp",
+                                              "DoWhileOp", "WhileOp", "ForOp"]>,
                                  Terminator]> {
   let summary = "Return from function";
   let description = [{
@@ -510,6 +510,58 @@ def ReturnOp : CIR_Op<"return", [ParentOneOf<["FuncOp", "ScopeOp", "DoWhileOp",
   let hasVerifier = 1;
 }
 
+//===----------------------------------------------------------------------===//
+// IfOp
+//===----------------------------------------------------------------------===//
+
+def IfOp : CIR_Op<"if",
+     [DeclareOpInterfaceMethods<RegionBranchOpInterface>,
+     RecursivelySpeculatable, AutomaticAllocationScope, NoRegionArguments]>{
+
+  let summary = "the if-then-else operation";
+  let description = [{
+    The `cir.if` operation represents an if-then-else construct for
+    conditionally executing two regions of code. The operand is a `cir.bool`
+    type.
+
+    Examples:
+
+    ```mlir
+    cir.if %cond  {
+      ...
+    } else {
+      ...
+    }
+
+    cir.if %cond  {
+      ...
+    }
+
+    cir.if %cond  {
+      ...
+      cir.br ^a
+    ^a:
+      cir.yield
+    }
+    ```
+
+    `cir.if` defines no values and the 'else' can be omitted. The if/else
+    regions must be terminated. If the region has only one block, the terminator
+    can be left out, and `cir.yield` terminator will be inserted implictly.
+    Otherwise, the region must be explicitly terminated.
+  }];
+  let arguments = (ins CIR_BoolType:$condition);
+  let regions = (region AnyRegion:$thenRegion, AnyRegion:$elseRegion);
+  let hasCustomAssemblyFormat=1;
+  let hasVerifier=1;
+  let skipDefaultBuilders=1;
+  let builders = [
+    OpBuilder<(ins "mlir::Value":$cond, "bool":$withElseRegion,
+      CArg<"BuilderCallbackRef", "buildTerminatedBody">:$thenBuilder,
+      CArg<"BuilderCallbackRef", "nullptr">:$elseBuilder)>
+  ];
+}
+
 //===----------------------------------------------------------------------===//
 // ConditionOp
 //===----------------------------------------------------------------------===//
@@ -560,8 +612,8 @@ def ConditionOp : CIR_Op<"condition", [
 //===----------------------------------------------------------------------===//
 
 def YieldOp : CIR_Op<"yield", [ReturnLike, Terminator,
-                               ParentOneOf<["ScopeOp", "WhileOp", "ForOp",
-                                            "DoWhileOp"]>]> {
+                               ParentOneOf<["IfOp", "ScopeOp", "WhileOp",
+                                            "ForOp", "DoWhileOp"]>]> {
   let summary = "Represents the default branching behaviour of a region";
   let description = [{
     The `cir.yield` operation terminates regions on different CIR operations,

clang/include/clang/CIR/MissingFeatures.h

@@ -80,6 +80,7 @@ struct MissingFeatures {
 
   // Clang early optimizations or things defered to LLVM lowering.
   static bool mayHaveIntegerOverflow() { return false; }
+  static bool shouldReverseUnaryCondOnBoolExpr() { return false; }
 
   // Misc
   static bool cxxABI() { return false; }
@@ -110,6 +111,8 @@ struct MissingFeatures {
   static bool lvalueBaseInfo() { return false; }
   static bool alignCXXRecordDecl() { return false; }
   static bool setNonGC() { return false; }
+  static bool incrementProfileCounter() { return false; }
+  static bool insertBuiltinUnpredictable() { return false; }
 
   // Missing types
   static bool dataMemberType() { return false; }

clang/lib/CIR/CodeGen/CIRGenExpr.cpp

@@ -18,6 +18,7 @@
 #include "clang/AST/CharUnits.h"
 #include "clang/AST/Decl.h"
 #include "clang/AST/Expr.h"
+#include "clang/AST/ExprCXX.h"
 #include "clang/CIR/Dialect/IR/CIRDialect.h"
 #include "clang/CIR/MissingFeatures.h"
 
@@ -474,6 +475,90 @@ void CIRGenFunction::emitIgnoredExpr(const Expr *e) {
   emitLValue(e);
 }
 
+/// Emit an `if` on a boolean condition, filling `then` and `else` into
+/// appropriated regions.
+mlir::LogicalResult CIRGenFunction::emitIfOnBoolExpr(const Expr *cond,
+                                                     const Stmt *thenS,
+                                                     const Stmt *elseS) {
+  mlir::Location thenLoc = getLoc(thenS->getSourceRange());
+  std::optional<mlir::Location> elseLoc;
+  if (elseS)
+    elseLoc = getLoc(elseS->getSourceRange());
+
+  mlir::LogicalResult resThen = mlir::success(), resElse = mlir::success();
+  emitIfOnBoolExpr(
+      cond, /*thenBuilder=*/
+      [&](mlir::OpBuilder &, mlir::Location) {
+        LexicalScope lexScope{*this, thenLoc, builder.getInsertionBlock()};
+        resThen = emitStmt(thenS, /*useCurrentScope=*/true);
+      },
+      thenLoc,
+      /*elseBuilder=*/
+      [&](mlir::OpBuilder &, mlir::Location) {
+        assert(elseLoc && "Invalid location for elseS.");
+        LexicalScope lexScope{*this, *elseLoc, builder.getInsertionBlock()};
+        resElse = emitStmt(elseS, /*useCurrentScope=*/true);
+      },
+      elseLoc);
+
+  return mlir::LogicalResult::success(resThen.succeeded() &&
+                                      resElse.succeeded());
+}
+
+/// Emit an `if` on a boolean condition, filling `then` and `else` into
+/// appropriated regions.
+cir::IfOp CIRGenFunction::emitIfOnBoolExpr(
+    const clang::Expr *cond, BuilderCallbackRef thenBuilder,
+    mlir::Location thenLoc, BuilderCallbackRef elseBuilder,
+    std::optional<mlir::Location> elseLoc) {
+  // Attempt to be as accurate as possible with IfOp location, generate
+  // one fused location that has either 2 or 4 total locations, depending
+  // on else's availability.
+  SmallVector<mlir::Location, 2> ifLocs{thenLoc};
+  if (elseLoc)
+    ifLocs.push_back(*elseLoc);
+  mlir::Location loc = mlir::FusedLoc::get(&getMLIRContext(), ifLocs);
+
+  // Emit the code with the fully general case.
+  mlir::Value condV = emitOpOnBoolExpr(loc, cond);
+  return builder.create<cir::IfOp>(loc, condV, elseLoc.has_value(),
+                                   /*thenBuilder=*/thenBuilder,
+                                   /*elseBuilder=*/elseBuilder);
+}
+
+/// TODO(cir): see EmitBranchOnBoolExpr for extra ideas).
+mlir::Value CIRGenFunction::emitOpOnBoolExpr(mlir::Location loc,
+                                             const Expr *cond) {
+  assert(!cir::MissingFeatures::pgoUse());
+  assert(!cir::MissingFeatures::generateDebugInfo());
+  cond = cond->IgnoreParens();
+
+  // In LLVM the condition is reversed here for efficient codegen.
+  // This should be done in CIR prior to LLVM lowering, if we do now
+  // we can make CIR based diagnostics misleading.
+  //  cir.ternary(!x, t, f) -> cir.ternary(x, f, t)
+  assert(!cir::MissingFeatures::shouldReverseUnaryCondOnBoolExpr());
+
+  if (isa<ConditionalOperator>(cond)) {
+    cgm.errorNYI(cond->getExprLoc(), "Ternary NYI");
+    assert(!cir::MissingFeatures::ternaryOp());
+    return createDummyValue(loc, cond->getType());
+  }
+
+  if (isa<CXXThrowExpr>(cond)) {
+    cgm.errorNYI("NYI");
+    return createDummyValue(loc, cond->getType());
+  }
+
+  // If the branch has a condition wrapped by __builtin_unpredictable,
+  // create metadata that specifies that the branch is unpredictable.
+  // Don't bother if not optimizing because that metadata would not be used.
+  assert(!cir::MissingFeatures::insertBuiltinUnpredictable());
+
+  // Emit the code with the fully general case.
+  return evaluateExprAsBool(cond);
+}
+
 mlir::Value CIRGenFunction::emitAlloca(StringRef name, mlir::Type ty,
                                        mlir::Location loc, CharUnits alignment,
                                        bool insertIntoFnEntryBlock,

clang/lib/CIR/CodeGen/CIRGenFunction.cpp

@@ -135,6 +135,71 @@ mlir::Location CIRGenFunction::getLoc(mlir::Location lhs, mlir::Location rhs) {
   return mlir::FusedLoc::get(locs, metadata, &getMLIRContext());
 }
 
+bool CIRGenFunction::containsLabel(const Stmt *s, bool ignoreCaseStmts) {
+  // Null statement, not a label!
+  if (!s)
+    return false;
+
+  // If this is a label, we have to emit the code, consider something like:
+  // if (0) {  ...  foo:  bar(); }  goto foo;
+  //
+  // TODO: If anyone cared, we could track __label__'s, since we know that you
+  // can't jump to one from outside their declared region.
+  if (isa<LabelStmt>(s))
+    return true;
+
+  // If this is a case/default statement, and we haven't seen a switch, we
+  // have to emit the code.
+  if (isa<SwitchCase>(s) && !ignoreCaseStmts)
+    return true;
+
+  // If this is a switch statement, we want to ignore case statements when we
+  // recursively process the sub-statements of the switch. If we haven't
+  // encountered a switch statement, we treat case statements like labels, but
+  // if we are processing a switch statement, case statements are expected.
+  if (isa<SwitchStmt>(s))
+    ignoreCaseStmts = true;
+
+  // Scan subexpressions for verboten labels.
+  return std::any_of(s->child_begin(), s->child_end(),
+                     [=](const Stmt *subStmt) {
+                       return containsLabel(subStmt, ignoreCaseStmts);
+                     });
+}
+
+/// If the specified expression does not fold to a constant, or if it does but
+/// contains a label, return false.  If it constant folds return true and set
+/// the boolean result in Result.
+bool CIRGenFunction::constantFoldsToBool(const Expr *cond, bool &resultBool,
+                                         bool allowLabels) {
+  llvm::APSInt resultInt;
+  if (!constantFoldsToSimpleInteger(cond, resultInt, allowLabels))
+    return false;
+
+  resultBool = resultInt.getBoolValue();
+  return true;
+}
+
+/// If the specified expression does not fold to a constant, or if it does
+/// fold but contains a label, return false. If it constant folds, return
+/// true and set the folded value.
+bool CIRGenFunction::constantFoldsToSimpleInteger(const Expr *cond,
+                                                  llvm::APSInt &resultInt,
+                                                  bool allowLabels) {
+  // FIXME: Rename and handle conversion of other evaluatable things
+  // to bool.
+  Expr::EvalResult result;
+  if (!cond->EvaluateAsInt(result, getContext()))
+    return false; // Not foldable, not integer or not fully evaluatable.
+
+  llvm::APSInt intValue = result.Val.getInt();
+  if (!allowLabels && containsLabel(cond))
+    return false; // Contains a label.
+
+  resultInt = intValue;
+  return true;
+}
+
 void CIRGenFunction::emitAndUpdateRetAlloca(QualType type, mlir::Location loc,
                                             CharUnits alignment) {
   if (!type->isVoidType()) {

clang/lib/CIR/CodeGen/CIRGenFunction.h

@@ -24,6 +24,7 @@
 #include "clang/AST/ASTContext.h"
 #include "clang/AST/CharUnits.h"
 #include "clang/AST/Decl.h"
+#include "clang/AST/Stmt.h"
 #include "clang/AST/Type.h"
 #include "clang/CIR/Dialect/IR/CIRDialect.h"
 #include "clang/CIR/MissingFeatures.h"
@@ -166,6 +167,20 @@ class CIRGenFunction : public CIRGenTypeCache {
   /// that it requires no code to be generated.
   bool isTrivialInitializer(const Expr *init);
 
+  /// If the specified expression does not fold to a constant, or if it does but
+  /// contains a label, return false.  If it constant folds return true and set
+  /// the boolean result in Result.
+  bool constantFoldsToBool(const clang::Expr *cond, bool &resultBool,
+                           bool allowLabels = false);
+  bool constantFoldsToSimpleInteger(const clang::Expr *cond,
+                                    llvm::APSInt &resultInt,
+                                    bool allowLabels = false);
+
+  /// Return true if the statement contains a label in it.  If
+  /// this statement is not executed normally, it not containing a label means
+  /// that we can just remove the code.
+  bool containsLabel(const clang::Stmt *s, bool ignoreCaseStmts = false);
+
   struct AutoVarEmission {
     const clang::VarDecl *Variable;
     /// The address of the alloca for languages with explicit address space
@@ -459,6 +474,25 @@ class CIRGenFunction : public CIRGenTypeCache {
   mlir::LogicalResult emitDeclStmt(const clang::DeclStmt &s);
   LValue emitDeclRefLValue(const clang::DeclRefExpr *e);
 
+  /// Emit an `if` on a boolean condition to the specified blocks.
+  /// FIXME: Based on the condition, this might try to simplify the codegen of
+  /// the conditional based on the branch.
+  /// In the future, we may apply code generation simplifications here,
+  /// similar to those used in classic LLVM  codegen
+  /// See `EmitBranchOnBoolExpr` for inspiration.
+  mlir::LogicalResult emitIfOnBoolExpr(const clang::Expr *cond,
+                                       const clang::Stmt *thenS,
+                                       const clang::Stmt *elseS);
+  cir::IfOp emitIfOnBoolExpr(const clang::Expr *cond,
+                             BuilderCallbackRef thenBuilder,
+                             mlir::Location thenLoc,
+                             BuilderCallbackRef elseBuilder,
+                             std::optional<mlir::Location> elseLoc = {});
+
+  mlir::Value emitOpOnBoolExpr(mlir::Location loc, const clang::Expr *cond);
+
+  mlir::LogicalResult emitIfStmt(const clang::IfStmt &s);
+
   /// Emit code to compute the specified expression,
   /// ignoring the result.
   void emitIgnoredExpr(const clang::Expr *e);