Reapply "Introduce -fexperimental-loop-fusion to clang and flang (#158844)

madhur13490 · web-flow · commit a134b0621798 · 2025-09-16T16:43:37.000+05:30
This PR is a reapplication of #142686
diff --git a/clang/include/clang/Basic/CodeGenOptions.def b/clang/include/clang/Basic/CodeGenOptions.def
@@ -322,6 +322,7 @@ CODEGENOPT(TimeTrace         , 1, 0, Benign) ///< Set when -ftime-trace is enabl
 VALUE_CODEGENOPT(TimeTraceGranularity, 32, 500, Benign) ///< Minimum time granularity (in microseconds),
                                                         ///< traced by time profiler
 CODEGENOPT(InterchangeLoops  , 1, 0, Benign) ///< Run loop-interchange.
+CODEGENOPT(FuseLoops         , 1, 0, Benign) ///< Run loop-fusion.
 CODEGENOPT(UnrollLoops       , 1, 0, Benign) ///< Control whether loops are unrolled.
 CODEGENOPT(RerollLoops       , 1, 0, Benign) ///< Control whether loops are rerolled.
 CODEGENOPT(NoUseJumpTables   , 1, 0, Benign) ///< Set when -fno-jump-tables is enabled.
diff --git a/clang/include/clang/Driver/Options.td b/clang/include/clang/Driver/Options.td
@@ -4304,6 +4304,10 @@ def floop_interchange : Flag<["-"], "floop-interchange">, Group<f_Group>,
   HelpText<"Enable the loop interchange pass">, Visibility<[ClangOption, CC1Option, FlangOption, FC1Option]>;
 def fno_loop_interchange: Flag<["-"], "fno-loop-interchange">, Group<f_Group>,
   HelpText<"Disable the loop interchange pass">, Visibility<[ClangOption, CC1Option, FlangOption, FC1Option]>;
+defm experimental_loop_fusion
+    : OptInCC1FFlag<"experimental-loop-fusion", "Enable", "Disable",
+                    "Enable the loop fusion pass",
+                    [ClangOption, FlangOption, FC1Option]>;
 def funroll_loops : Flag<["-"], "funroll-loops">, Group<f_Group>,
   HelpText<"Turn on loop unroller">, Visibility<[ClangOption, CC1Option, FlangOption, FC1Option]>;
 def fno_unroll_loops : Flag<["-"], "fno-unroll-loops">, Group<f_Group>,
diff --git a/clang/lib/CodeGen/BackendUtil.cpp b/clang/lib/CodeGen/BackendUtil.cpp
@@ -896,6 +896,7 @@ void EmitAssemblyHelper::RunOptimizationPipeline(
   PipelineTuningOptions PTO;
   PTO.LoopUnrolling = CodeGenOpts.UnrollLoops;
   PTO.LoopInterchange = CodeGenOpts.InterchangeLoops;
+  PTO.LoopFusion = CodeGenOpts.FuseLoops;
   // For historical reasons, loop interleaving is set to mirror setting for loop
   // unrolling.
   PTO.LoopInterleaving = CodeGenOpts.UnrollLoops;
@@ -1331,6 +1332,7 @@ runThinLTOBackend(CompilerInstance &CI, ModuleSummaryIndex *CombinedIndex,
   Conf.SampleProfile = std::move(SampleProfile);
   Conf.PTO.LoopUnrolling = CGOpts.UnrollLoops;
   Conf.PTO.LoopInterchange = CGOpts.InterchangeLoops;
+  Conf.PTO.LoopFusion = CGOpts.FuseLoops;
   // For historical reasons, loop interleaving is set to mirror setting for loop
   // unrolling.
   Conf.PTO.LoopInterleaving = CGOpts.UnrollLoops;
diff --git a/clang/lib/Driver/ToolChains/Clang.cpp b/clang/lib/Driver/ToolChains/Clang.cpp
@@ -6875,6 +6875,8 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
                   options::OPT_fno_unroll_loops);
   Args.AddLastArg(CmdArgs, options::OPT_floop_interchange,
                   options::OPT_fno_loop_interchange);
+  Args.addOptInFlag(CmdArgs, options::OPT_fexperimental_loop_fusion,
+                    options::OPT_fno_experimental_loop_fusion);
 
   Args.AddLastArg(CmdArgs, options::OPT_fstrict_flex_arrays_EQ);
 
diff --git a/clang/lib/Driver/ToolChains/Flang.cpp b/clang/lib/Driver/ToolChains/Flang.cpp
@@ -156,6 +156,9 @@ void Flang::addCodegenOptions(const ArgList &Args,
       !stackArrays->getOption().matches(options::OPT_fno_stack_arrays))
     CmdArgs.push_back("-fstack-arrays");
 
+  Args.addOptInFlag(CmdArgs, options::OPT_fexperimental_loop_fusion,
+                    options::OPT_fno_experimental_loop_fusion);
+
   handleInterchangeLoopsArgs(Args, CmdArgs);
   handleVectorizeLoopsArgs(Args, CmdArgs);
   handleVectorizeSLPArgs(Args, CmdArgs);
diff --git a/clang/lib/Frontend/CompilerInvocation.cpp b/clang/lib/Frontend/CompilerInvocation.cpp
@@ -1680,6 +1680,9 @@ void CompilerInvocationBase::GenerateCodeGenArgs(const CodeGenOptions &Opts,
   else
     GenerateArg(Consumer, OPT_fno_loop_interchange);
 
+  if (Opts.FuseLoops)
+    GenerateArg(Consumer, OPT_fexperimental_loop_fusion);
+
   if (!Opts.BinutilsVersion.empty())
     GenerateArg(Consumer, OPT_fbinutils_version_EQ, Opts.BinutilsVersion);
 
@@ -2001,6 +2004,8 @@ bool CompilerInvocation::ParseCodeGenArgs(CodeGenOptions &Opts, ArgList &Args,
                    (Opts.OptimizationLevel > 1));
   Opts.InterchangeLoops =
       Args.hasFlag(OPT_floop_interchange, OPT_fno_loop_interchange, false);
+  Opts.FuseLoops = Args.hasFlag(OPT_fexperimental_loop_fusion,
+                                OPT_fno_experimental_loop_fusion, false);
   Opts.BinutilsVersion =
       std::string(Args.getLastArgValue(OPT_fbinutils_version_EQ));
 
diff --git a/clang/test/Driver/clang_f_opts.c b/clang/test/Driver/clang_f_opts.c
@@ -52,6 +52,15 @@
 // CHECK-INTERCHANGE-LOOPS: "-floop-interchange"
 // CHECK-NO-INTERCHANGE-LOOPS: "-fno-loop-interchange"
 
+// RUN: %clang -### -S -fexperimental-loop-fusion %s 2>&1 | FileCheck -check-prefix=CHECK-FUSE-LOOPS %s
+// CHECK-FUSE-LOOPS: "-fexperimental-loop-fusion"
+//
+// RUN: %clang -c -fexperimental-loop-fusion -mllvm -print-pipeline-passes -O3 %s 2>&1 | FileCheck --check-prefixes=LOOP-FUSION-ON %s
+// RUN: %clang -c -mllvm -print-pipeline-passes -O3 %s 2>&1 | FileCheck --check-prefixes=LOOP-FUSION-OFF %s
+
+// LOOP-FUSION-ON: loop-fusion
+// LOOP-FUSION-OFF-NOT: loop-fusion
+
 // RUN: %clang -### -S -fprofile-sample-accurate %s 2>&1 | FileCheck -check-prefix=CHECK-PROFILE-SAMPLE-ACCURATE %s
 // CHECK-PROFILE-SAMPLE-ACCURATE: "-fprofile-sample-accurate"
 
diff --git a/flang/docs/ReleaseNotes.md b/flang/docs/ReleaseNotes.md
@@ -35,6 +35,8 @@ page](https://llvm.org/releases/).
 
 ## New Compiler Flags
 
+* -fexperimental-loop-fusion is now recognized by flang.
+
 ## Windows Support
 
 ## Fortran Language Changes in Flang
diff --git a/flang/include/flang/Frontend/CodeGenOptions.def b/flang/include/flang/Frontend/CodeGenOptions.def
@@ -43,6 +43,7 @@ CODEGENOPT(StackArrays, 1, 0) ///< -fstack-arrays (enable the stack-arrays pass)
 CODEGENOPT(VectorizeLoop, 1, 0) ///< Enable loop vectorization.
 CODEGENOPT(VectorizeSLP, 1, 0) ///< Enable SLP vectorization.
 CODEGENOPT(InterchangeLoops, 1, 0) ///< Enable loop interchange.
+CODEGENOPT(FuseLoops, 1, 0) ///< Enable loop fusion.
 CODEGENOPT(LoopVersioning, 1, 0) ///< Enable loop versioning.
 CODEGENOPT(UnrollLoops, 1, 0) ///< Enable loop unrolling
 CODEGENOPT(AliasAnalysis, 1, 0) ///< Enable alias analysis pass
diff --git a/flang/lib/Frontend/CompilerInvocation.cpp b/flang/lib/Frontend/CompilerInvocation.cpp
@@ -280,6 +280,9 @@ static void parseCodeGenArgs(Fortran::frontend::CodeGenOptions &opts,
   if (args.getLastArg(clang::driver::options::OPT_floop_interchange))
     opts.InterchangeLoops = 1;
 
+  if (args.getLastArg(clang::driver::options::OPT_fexperimental_loop_fusion))
+    opts.FuseLoops = 1;
+
   if (args.getLastArg(clang::driver::options::OPT_vectorize_loops))
     opts.VectorizeLoop = 1;
 
diff --git a/flang/lib/Frontend/FrontendActions.cpp b/flang/lib/Frontend/FrontendActions.cpp
@@ -958,6 +958,7 @@ void CodeGenAction::runOptimizationPipeline(llvm::raw_pwrite_stream &os) {
     si.getTimePasses().setOutStream(ci.getTimingStreamLLVM());
   pto.LoopUnrolling = opts.UnrollLoops;
   pto.LoopInterchange = opts.InterchangeLoops;
+  pto.LoopFusion = opts.FuseLoops;
   pto.LoopInterleaving = opts.UnrollLoops;
   pto.LoopVectorization = opts.VectorizeLoop;
   pto.SLPVectorization = opts.VectorizeSLP;
diff --git a/flang/test/Driver/loop-fuse.f90 b/flang/test/Driver/loop-fuse.f90
@@ -0,0 +1,17 @@
+! RUN: %flang -### -S -fexperimental-loop-fusion %s 2>&1 | FileCheck -check-prefix=CHECK-LOOP-FUSE %s
+! RUN: %flang -### -S -fno-experimental-loop-fusion %s 2>&1 | FileCheck -check-prefix=CHECK-NO-LOOP-FUSE %s
+! RUN: %flang -### -S -O0 %s 2>&1 | FileCheck -check-prefix=CHECK-NO-LOOP-FUSE %s
+! RUN: %flang -### -S -O1 %s 2>&1 | FileCheck -check-prefix=CHECK-NO-LOOP-FUSE %s
+! RUN: %flang -### -S -O2 %s 2>&1 | FileCheck -check-prefix=CHECK-NO-LOOP-FUSE %s
+! RUN: %flang -### -S -O3 %s 2>&1 | FileCheck -check-prefix=CHECK-NO-LOOP-FUSE %s
+! RUN: %flang -### -S -Os %s 2>&1 | FileCheck -check-prefix=CHECK-NO-LOOP-FUSE %s
+! RUN: %flang -### -S -Oz %s 2>&1 | FileCheck -check-prefix=CHECK-NO-LOOP-FUSE %s
+! CHECK-LOOP-FUSE: "-fexperimental-loop-fusion"
+! CHECK-NO-LOOP-FUSE-NOT: "-fexperimental-loop-fusion"
+! RUN: %flang_fc1 -emit-llvm -O2 -fexperimental-loop-fusion -mllvm -print-pipeline-passes -o /dev/null %s 2>&1 | FileCheck -check-prefix=CHECK-LOOP-FUSE-PASS %s
+! RUN: %flang_fc1 -emit-llvm -O2 -fno-experimental-loop-fusion -mllvm -print-pipeline-passes -o /dev/null %s 2>&1 | FileCheck -check-prefix=CHECK-NO-LOOP-FUSE-PASS %s
+! CHECK-LOOP-FUSE-PASS: loop-fusion
+! CHECK-NO-LOOP-FUSE-PASS-NOT: loop-fusion
+
+program test
+end program
diff --git a/llvm/include/llvm/Passes/PassBuilder.h b/llvm/include/llvm/Passes/PassBuilder.h
@@ -65,6 +65,9 @@ class PipelineTuningOptions {
   /// false.
   bool LoopInterchange;
 
+  /// Tuning option to enable/disable loop fusion. Its default value is false.
+  bool LoopFusion;
+
   /// Tuning option to forget all SCEV loops in LoopUnroll. Its default value
   /// is that of the flag: `-forget-scev-loop-unroll`.
   bool ForgetAllSCEVInLoopUnroll;
diff --git a/llvm/lib/Passes/PassBuilderPipelines.cpp b/llvm/lib/Passes/PassBuilderPipelines.cpp
@@ -104,6 +104,7 @@
 #include "llvm/Transforms/Scalar/LoopDeletion.h"
 #include "llvm/Transforms/Scalar/LoopDistribute.h"
 #include "llvm/Transforms/Scalar/LoopFlatten.h"
+#include "llvm/Transforms/Scalar/LoopFuse.h"
 #include "llvm/Transforms/Scalar/LoopIdiomRecognize.h"
 #include "llvm/Transforms/Scalar/LoopInstSimplify.h"
 #include "llvm/Transforms/Scalar/LoopInterchange.h"
@@ -313,6 +314,7 @@ PipelineTuningOptions::PipelineTuningOptions() {
   SLPVectorization = false;
   LoopUnrolling = true;
   LoopInterchange = EnableLoopInterchange;
+  LoopFusion = false;
   ForgetAllSCEVInLoopUnroll = ForgetSCEVInLoopUnroll;
   LicmMssaOptCap = SetLicmMssaOptCap;
   LicmMssaNoAccForPromotionCap = SetLicmMssaNoAccForPromotionCap;
@@ -1551,6 +1553,11 @@ PassBuilder::buildModuleOptimizationPipeline(OptimizationLevel Level,
   OptimizePM.addPass(createFunctionToLoopPassAdaptor(
       std::move(LPM), /*UseMemorySSA=*/false, /*UseBlockFrequencyInfo=*/false));
 
+  // FIXME: This may not be the right place in the pipeline.
+  // We need to have the data to support the right place.
+  if (PTO.LoopFusion)
+    OptimizePM.addPass(LoopFusePass());
+
   // Distribute loops to allow partial vectorization.  I.e. isolate dependences
   // into separate loop that would otherwise inhibit vectorization.  This is
   // currently only performed for loops marked with the metadata
@@ -2355,4 +2362,4 @@ AAManager PassBuilder::buildDefaultAAPipeline() {
 bool PassBuilder::isInstrumentedPGOUse() const {
   return (PGOOpt && PGOOpt->Action == PGOOptions::IRUse) ||
          !UseCtxProfile.empty();
-}
+}
diff --git a/llvm/tools/opt/NewPMDriver.cpp b/llvm/tools/opt/NewPMDriver.cpp
@@ -60,6 +60,9 @@ cl::opt<bool> VerifyEachDebugInfoPreserve(
     cl::desc("Start each pass with collecting and end it with checking of "
              "debug info preservation."));
 
+static cl::opt<bool> EnableLoopFusion("enable-loopfusion", cl::init(false),
+                                      cl::Hidden,
+                                      cl::desc("Enable the LoopFuse Pass"));
 cl::opt<std::string>
     VerifyDIPreserveExport("verify-di-preserve-export",
                    cl::desc("Export debug info preservation failures into "
@@ -446,6 +449,7 @@ bool llvm::runPassPipeline(
   // option has been enabled.
   PTO.LoopUnrolling = !DisableLoopUnrolling;
   PTO.UnifiedLTO = UnifiedLTO;
+  PTO.LoopFusion = EnableLoopFusion;
   PassBuilder PB(TM, PTO, P, &PIC);
   registerEPCallbacks(PB);
 
