[flang] Add nsw flag to do-variable increment with a new option

This patch adds nsw flag to the increment of do-variables when
a new option is enabled.
NOTE 11.10 in the Fortran 2018 standard says they never overflow.

See also the discussion in 74709 and
discourse post.

Author: yus3710-fj

clang/include/clang/Driver/Options.td

@@ -6530,6 +6530,10 @@ def flang_deprecated_no_hlfir : Flag<["-"], "flang-deprecated-no-hlfir">,
   Flags<[HelpHidden]>, Visibility<[FlangOption, FC1Option]>,
   HelpText<"Do not use HLFIR lowering (deprecated)">;
 
+def flang_experimental_integer_overflow : Flag<["-"], "flang-experimental-integer-overflow">,
+  Flags<[HelpHidden]>, Visibility<[FlangOption, FC1Option]>,
+  HelpText<"Add nsw flag to internal operations such as do-variable increment (experimental)">;
+
 //===----------------------------------------------------------------------===//
 // FLangOption + CoreOption + NoXarchOption
 //===----------------------------------------------------------------------===//

clang/lib/Driver/ToolChains/Flang.cpp

@@ -139,6 +139,7 @@ void Flang::addCodegenOptions(const ArgList &Args,
 
   Args.addAllArgs(CmdArgs, {options::OPT_flang_experimental_hlfir,
                             options::OPT_flang_deprecated_no_hlfir,
+                            options::OPT_flang_experimental_integer_overflow,
                             options::OPT_fno_ppc_native_vec_elem_order,
                             options::OPT_fppc_native_vec_elem_order});
 }

flang/include/flang/Lower/LoweringOptions.def

@@ -34,5 +34,9 @@ ENUM_LOWERINGOPT(NoPPCNativeVecElemOrder, unsigned, 1, 0)
 /// On by default.
 ENUM_LOWERINGOPT(Underscoring, unsigned, 1, 1)
 
+/// If true, add nsw flags to arithmetic operations for integer.
+/// Off by default.
+ENUM_LOWERINGOPT(NoSignedWrap, unsigned, 1, 0)
+
 #undef LOWERINGOPT
 #undef ENUM_LOWERINGOPT

flang/include/flang/Optimizer/Transforms/Passes.h

@@ -54,6 +54,7 @@ namespace fir {
 std::unique_ptr<mlir::Pass> createAffineDemotionPass();
 std::unique_ptr<mlir::Pass>
 createArrayValueCopyPass(fir::ArrayValueCopyOptions options = {});
+std::unique_ptr<mlir::Pass> createCFGConversionPassWithNSW();
 std::unique_ptr<mlir::Pass> createExternalNameConversionPass();
 std::unique_ptr<mlir::Pass>
 createExternalNameConversionPass(bool appendUnderscore);
@@ -89,7 +90,8 @@ createFunctionAttrPass(FunctionAttrTypes &functionAttr, bool noInfsFPMath,
                        bool noSignedZerosFPMath, bool unsafeFPMath);
 
 void populateCfgConversionRewrites(mlir::RewritePatternSet &patterns,
-                                   bool forceLoopToExecuteOnce = false);
+                                   bool forceLoopToExecuteOnce = false,
+                                   bool setNSW = false);
 
 // declarative passes
 #define GEN_PASS_REGISTRATION

flang/include/flang/Optimizer/Transforms/Passes.td

@@ -151,7 +151,10 @@ def CFGConversion : Pass<"cfg-conversion"> {
   let options = [
     Option<"forceLoopToExecuteOnce", "always-execute-loop-body", "bool",
            /*default=*/"false",
-           "force the body of a loop to execute at least once">
+           "force the body of a loop to execute at least once">,
+    Option<"setNSW", "set-nsw", "bool",
+           /*default=*/"false",
+           "set nsw on loop variable increment">
   ];
 }
 

flang/include/flang/Tools/CLOptions.inc

@@ -148,9 +148,14 @@ static void addCanonicalizerPassWithoutRegionSimplification(
   pm.addPass(mlir::createCanonicalizerPass(config));
 }
 
-inline void addCfgConversionPass(mlir::PassManager &pm) {
-  addNestedPassToAllTopLevelOperationsConditionally(
-      pm, disableCfgConversion, fir::createCFGConversion);
+inline void addCfgConversionPass(
+    mlir::PassManager &pm, const MLIRToLLVMPassPipelineConfig &config) {
+  if (config.NoSignedWrap)
+    addNestedPassToAllTopLevelOperationsConditionally(
+        pm, disableCfgConversion, fir::createCFGConversionPassWithNSW);
+  else
+    addNestedPassToAllTopLevelOperationsConditionally(
+        pm, disableCfgConversion, fir::createCFGConversion);
 }
 
 inline void addAVC(
@@ -290,7 +295,7 @@ inline void createDefaultFIROptimizerPassPipeline(
     pm.addPass(fir::createAliasTagsPass());
 
   // convert control flow to CFG form
-  fir::addCfgConversionPass(pm);
+  fir::addCfgConversionPass(pm, pc);
   pm.addPass(mlir::createConvertSCFToCFPass());
 
   pm.addPass(mlir::createCanonicalizerPass(config));

flang/include/flang/Tools/CrossToolHelpers.h

@@ -122,6 +122,7 @@ struct MLIRToLLVMPassPipelineConfig : public FlangEPCallBacks {
   bool NoSignedZerosFPMath =
       false; ///< Set no-signed-zeros-fp-math attribute for functions.
   bool UnsafeFPMath = false; ///< Set unsafe-fp-math attribute for functions.
+  bool NoSignedWrap = false; ///< Add nsw flag to numeric operations.
 };
 
 struct OffloadModuleOpts {

flang/lib/Frontend/CompilerInvocation.cpp

@@ -1198,6 +1198,12 @@ bool CompilerInvocation::createFromArgs(
     invoc.loweringOpts.setNoPPCNativeVecElemOrder(true);
   }
 
+  // -flang-experimental-integer-overflow
+  if (args.hasArg(
+          clang::driver::options::OPT_flang_experimental_integer_overflow)) {
+    invoc.loweringOpts.setNoSignedWrap(true);
+  }
+
   // Preserve all the remark options requested, i.e. -Rpass, -Rpass-missed or
   // -Rpass-analysis. This will be used later when processing and outputting the
   // remarks generated by LLVM in ExecuteCompilerInvocation.cpp.

flang/lib/Frontend/FrontendActions.cpp

@@ -809,6 +809,9 @@ void CodeGenAction::generateLLVMIR() {
     config.VScaleMax = vsr->second;
   }
 
+  if (ci.getInvocation().getLoweringOpts().getNoSignedWrap())
+    config.NoSignedWrap = true;
+
   // Create the pass pipeline
   fir::createMLIRToLLVMPassPipeline(pm, config, getCurrentFile());
   (void)mlir::applyPassManagerCLOptions(pm);

flang/lib/Lower/Bridge.cpp

@@ -1996,6 +1996,11 @@ class FirConverter : public Fortran::lower::AbstractConverter {
   void genFIRIncrementLoopEnd(IncrementLoopNestInfo &incrementLoopNestInfo) {
     assert(!incrementLoopNestInfo.empty() && "empty loop nest");
     mlir::Location loc = toLocation();
+    mlir::arith::IntegerOverflowFlags flags{};
+    if (getLoweringOptions().getNoSignedWrap())
+      flags = bitEnumSet(flags, mlir::arith::IntegerOverflowFlags::nsw);
+    auto iofAttr = mlir::arith::IntegerOverflowFlagsAttr::get(
+        builder->getContext(), flags);
     for (auto it = incrementLoopNestInfo.rbegin(),
               rend = incrementLoopNestInfo.rend();
          it != rend; ++it) {
@@ -2010,15 +2015,16 @@ class FirConverter : public Fortran::lower::AbstractConverter {
         builder->setInsertionPointToEnd(info.doLoop.getBody());
         llvm::SmallVector<mlir::Value, 2> results;
         results.push_back(builder->create<mlir::arith::AddIOp>(
-            loc, info.doLoop.getInductionVar(), info.doLoop.getStep()));
+            loc, info.doLoop.getInductionVar(), info.doLoop.getStep(),
+            iofAttr));
         // Step loopVariable to help optimizations such as vectorization.
         // Induction variable elimination will clean up as necessary.
         mlir::Value step = builder->createConvert(
             loc, info.getLoopVariableType(), info.doLoop.getStep());
         mlir::Value loopVar =
             builder->create<fir::LoadOp>(loc, info.loopVariable);
         results.push_back(
-            builder->create<mlir::arith::AddIOp>(loc, loopVar, step));
+            builder->create<mlir::arith::AddIOp>(loc, loopVar, step, iofAttr));
         builder->create<fir::ResultOp>(loc, results);
         builder->setInsertionPointAfter(info.doLoop);
         // The loop control variable may be used after the loop.
@@ -2043,7 +2049,7 @@ class FirConverter : public Fortran::lower::AbstractConverter {
       if (info.hasRealControl)
         value = builder->create<mlir::arith::AddFOp>(loc, value, step);
       else
-        value = builder->create<mlir::arith::AddIOp>(loc, value, step);
+        value = builder->create<mlir::arith::AddIOp>(loc, value, step, iofAttr);
       builder->create<fir::StoreOp>(loc, value, info.loopVariable);
 
       genBranch(info.headerBlock);