[flang] Add -f[no-]associative-math and -mreassociate

[tblah]

Summary

Only add the option processing and store the result. No attributes are added to FIR yet.

Clang only forwards -mreassociate

if (AssociativeMath && !SignedZeros && !TrappingMath)

Flang doesn't have -f[no-]trapping-math, so this part of the condition has been omitted. !TrappingMath is the default.

[kiranchandramohan]

LGTM.

[awarzynski]

LGTM, thanks!

[clementval]

Wouldn't it be good to have a RFC for all these options and what they will do in Flang instead of just adding them all? Or did I miss the RFC?

[awarzynski]

In D137329#3909082, @clementval wrote:

Wouldn't it be good to have a RFC for all these options and what they will do in Flang instead of just adding them all? Or did I miss the RFC?

+1

[kiranchandramohan]

In D137329#3909943, @awarzynski wrote:

In D137329#3909082, @clementval wrote:

Wouldn't it be good to have a RFC for all these options and what they will do in Flang instead of just adding them all? Or did I miss the RFC?

+1

Fast Math attributes supported in LLVM IR are documented in https://llvm.org/docs/LangRef.html#fast-math-flags. This set of patches (details below) provides a way to set or unset each of these attributes.

While making policy decisions about these flags, like for ffp-contract, we have created an RFC (https://discourse.llvm.org/t/rfc-ffp-contract-default-value/66301). When adding umbrella flags like -ffast-math/-Ofast/-ffp-model/-funsafe-math-optimizations, we will submit RFCs.

Patch : Flag Name : LLVM IR Attribute
https://reviews.llvm.org/D137325 : f[no-]honor-nans : nnan
https://reviews.llvm.org/D137072 : f[no-]honor-infinities : ninf
https://reviews.llvm.org/D137328 : f[no-]signed-zeros : nsz
https://reviews.llvm.org/D137330 : f[no-]reciprocal-math : arcp
https://reviews.llvm.org/D136080 : ffp-contract option : contract
https://reviews.llvm.org/D137326 : f[no-]approx-func : afn
https://reviews.llvm.org/D137329 : f[no-]associative-math: reassoc

[awarzynski]

In D137329#3910249, @kiranchandramohan wrote:

In D137329#3909943, @awarzynski wrote:

In D137329#3909082, @clementval wrote:

Wouldn't it be good to have a RFC for all these options and what they will do in Flang instead of just adding them all? Or did I miss the RFC?

+1

Fast Math attributes supported in LLVM IR are documented in https://llvm.org/docs/LangRef.html#fast-math-flags. This set of patches (details below) provides a way to set or unset each of these attributes.

While making policy decisions about these flags, like for ffp-contract, we have created an RFC (https://discourse.llvm.org/t/rfc-ffp-contract-default-value/66301). When adding umbrella flags like -ffast-math/-Ofast/-ffp-model/-funsafe-math-optimizations, we will submit RFCs.

Patch : Flag Name : LLVM IR Attribute
https://reviews.llvm.org/D137325 : f[no-]honor-nans : nnan
https://reviews.llvm.org/D137072 : f[no-]honor-infinities : ninf
https://reviews.llvm.org/D137328 : f[no-]signed-zeros : nsz
https://reviews.llvm.org/D137330 : f[no-]reciprocal-math : arcp
https://reviews.llvm.org/D136080 : ffp-contract option : contract
https://reviews.llvm.org/D137326 : f[no-]approx-func : afn
https://reviews.llvm.org/D137329 : f[no-]associative-math: reassoc

Thanks for this summary Kiran!

clang/include/clang/Driver/Options.td

def new_struct_path_tbaa : Flag<["-"], "new-struct-path-tbaa">,
   HelpText<"Enable enhanced struct-path aware Type Based Alias Analysis">;
 def mdebug_pass : Separate<["-"], "mdebug-pass">,
   HelpText<"Enable additional debug output">,
   MarshallingInfoString<CodeGenOpts<"DebugPass">>;
 def mframe_pointer_EQ : Joined<["-"], "mframe-pointer=">,
   HelpText<"Specify which frame pointers to retain.">, Values<"all,non-leaf,none">,
   NormalizedValuesScope<"CodeGenOptions::FramePointerKind">, NormalizedValues<["All", "NonLeaf", "None"]>,
   MarshallingInfoEnum<CodeGenOpts<"FramePointer">, "None">;
-def mreassociate : Flag<["-"], "mreassociate">,
-  HelpText<"Allow reassociation transformations for floating-point instructions">,
-  MarshallingInfoFlag<LangOpts<"AllowFPReassoc">>, ImpliedByAnyOf<[funsafe_math_optimizations.KeyPath]>;
 def mabi_EQ_ieeelongdouble : Flag<["-"], "mabi=ieeelongdouble">,
   HelpText<"Use IEEE 754 quadruple-precision for long double">,
   MarshallingInfoFlag<LangOpts<"PPCIEEELongDouble">>;
 def mfloat_abi : Separate<["-"], "mfloat-abi">,
   HelpText<"The float ABI to use">,
   MarshallingInfoString<CodeGenOpts<"FloatABI">>;
 def mtp : Separate<["-"], "mtp">,
   HelpText<"Mode for reading thread pointer">;
def main_file_name : Separate<["-"], "main-file-name">,
   MarshallingInfoString<CodeGenOpts<"MainFileName">>;
 def split_dwarf_output : Separate<["-"], "split-dwarf-output">,
   HelpText<"File name to use for split dwarf debug info output">,
   Flags<[CC1Option, CC1AsOption, NoDriverOption]>,
   MarshallingInfoString<CodeGenOpts<"SplitDwarfOutput">>;
 
 let Flags = [CC1Option, FC1Option, NoDriverOption] in {
 
+def mreassociate : Flag<["-"], "mreassociate">,
+  HelpText<"Allow reassociation transformations for floating-point instructions">,
+  MarshallingInfoFlag<LangOpts<"AllowFPReassoc">>, ImpliedByAnyOf<[funsafe_math_optimizations.KeyPath]>;
 def menable_no_nans : Flag<["-"], "menable-no-nans">,
   HelpText<"Allow optimization to assume there are no NaNs.">,
   MarshallingInfoFlag<LangOpts<"NoHonorNaNs">>, ImpliedByAnyOf<[ffinite_math_only.KeyPath]>;
 def menable_no_infinities : Flag<["-"], "menable-no-infs">,
   HelpText<"Allow optimization to assume there are no infinities.">,
   MarshallingInfoFlag<LangOpts<"NoHonorInfs">>, ImpliedByAnyOf<[ffinite_math_only.KeyPath]>;
 
 def pic_level : Separate<["-"], "pic-level">,

clang/lib/Driver/ToolChains/Flang.cpp

 
 static void addFloatingPointOptions(const Driver &D, const ArgList &Args,
                                     ArgStringList &CmdArgs) {
   StringRef FPContract;
   bool HonorINFs = true;
   bool HonorNaNs = true;
   bool ApproxFunc = false;
   bool SignedZeros = true;
+  bool AssociativeMath = false;
 
   if (const Arg *A = Args.getLastArg(options::OPT_ffp_contract)) {
     const StringRef Val = A->getValue();
     if (Val == "fast" || Val == "off") {
       FPContract = Val;
     } else if (Val == "on") {
       // Warn instead of error because users might have makefiles written for
       // gfortran (which accepts -ffp-contract=on)
case options::OPT_fno_approx_func:
       ApproxFunc = false;
       break;
     case options::OPT_fsigned_zeros:
       SignedZeros = true;
       break;
     case options::OPT_fno_signed_zeros:
       SignedZeros = false;
       break;
+    case options::OPT_fassociative_math:
+      AssociativeMath = true;
+      break;
+    case options::OPT_fno_associative_math:
+      AssociativeMath = false;
+      break;
     }
 
     // If we handled this option claim it
     A->claim();
   }
 
   if (!FPContract.empty())
     CmdArgs.push_back(Args.MakeArgString("-ffp-contract=" + FPContract));
 
   if (!HonorINFs)
     CmdArgs.push_back("-menable-no-infs");
 
   if (!HonorNaNs)
     CmdArgs.push_back("-menable-no-nans");
 
   if (ApproxFunc)
     CmdArgs.push_back("-fapprox-func");
 
   if (!SignedZeros)
     CmdArgs.push_back("-fno-signed-zeros");
+
+  if (AssociativeMath && !SignedZeros)
+    CmdArgs.push_back("-mreassociate");
 }
 
 void Flang::ConstructJob(Compilation &C, const JobAction &JA,
                          const InputInfo &Output, const InputInfoList &Inputs,
                          const ArgList &Args, const char *LinkingOutput) const {
   const auto &TC = getToolChain();
   const llvm::Triple &Triple = TC.getEffectiveTriple();
   const std::string &TripleStr = Triple.getTriple();

flang/include/flang/Frontend/LangOptions.def

 /// Permit floating point optimization without regard to infinities
 LANGOPT(NoHonorInfs, 1, false)
 /// Permit floating point optimization without regard to NaN
 LANGOPT(NoHonorNaNs, 1, false)
 /// Allow math functions to be replaced with an approximately equivalent calculation
 LANGOPT(ApproxFunc, 1, false)
 /// Allow optimizations that ignore the sign of floating point zeros
 LANGOPT(NoSignedZeros, 1, false)
+/// Allow reassociation transformations for floating-point instructions
+LANGOPT(AssociativeMath, 1, false)
 
 #undef LANGOPT
 #undef ENUM_LANGOPT

flang/lib/Frontend/CompilerInvocation.cpp

static bool parseFloatingPointArgs(CompilerInvocation &invoc,
   }
 
   if (const llvm::opt::Arg *a =
           args.getLastArg(clang::driver::options::OPT_fno_signed_zeros)) {
     diags.Report(diagUnimplemented) << a->getOption().getName();
     opts.NoSignedZeros = true;
   }
 
+  if (const llvm::opt::Arg *a =
+          args.getLastArg(clang::driver::options::OPT_mreassociate)) {
+    diags.Report(diagUnimplemented) << a->getOption().getName();
+    opts.AssociativeMath = true;
+  }
+
   return true;
 }
 
 bool CompilerInvocation::createFromArgs(
     CompilerInvocation &res, llvm::ArrayRef<const char *> commandLineArgs,
     clang::DiagnosticsEngine &diags) {
 
   bool success = true;

flang/test/Driver/driver-help.f90

 ! HELP-FC1-NEXT: -I <dir>               Add directory to the end of the list of include search paths
 ! HELP-FC1-NEXT: -load <dsopath>        Load the named plugin (dynamic shared object)
 ! HELP-FC1-NEXT: -menable-no-infs       Allow optimization to assume there are no infinities.
 ! HELP-FC1-NEXT: -menable-no-nans       Allow optimization to assume there are no NaNs.
 ! HELP-FC1-NEXT: -mllvm <value>         Additional arguments to forward to LLVM's option processing
 ! HELP-FC1-NEXT: -mmlir <value>         Additional arguments to forward to MLIR's option processing
 ! HELP-FC1-NEXT: -module-dir <dir>      Put MODULE files in <dir>
 ! HELP-FC1-NEXT: -module-suffix <suffix> Use <suffix> as the suffix for module files (the default value is `.mod`)
+! HELP-FC1-NEXT: -mreassociate          Allow reassociation transformations for floating-point instructions
 ! HELP-FC1-NEXT: -mrelocation-model <value>
 ! HELP-FC1-NEXT:                        The relocation model to use
 ! HELP-FC1-NEXT: -nocpp                 Disable predefined and command line preprocessor macros
 ! HELP-FC1-NEXT: -o <file>              Write output to <file>
 ! HELP-FC1-NEXT: -pedantic              Warn on language extensions
 ! HELP-FC1-NEXT: -pic-is-pie             File is for a position independent executable
 ! HELP-FC1-NEXT: -pic-level <value>      Value for __PIC__
 ! HELP-FC1-NEXT: -plugin <name>         Use the named plugin action instead of the default action (use "help" to list available options)

flang/test/Driver/flang_fp_opts.f90

 ! Test for handling of floating point options within the frontend driver
 
 ! RUN: %flang_fc1 \
 ! RUN:      -ffp-contract=fast \
 ! RUN:      -menable-no-infs \
 ! RUN:      -menable-no-nans \
 ! RUN:      -fapprox-func \
 ! RUN:      -fno-signed-zeros \
+! RUN:      -mreassociate \
 ! RUN:      %s 2>&1 | FileCheck %s
 ! CHECK: ffp-contract= is not currently implemented
 ! CHECK: menable-no-infs is not currently implemented
 ! CHECK: menable-no-nans is not currently implemented
 ! CHECK: fapprox-func is not currently implemented
 ! CHECK: fno-signed-zeros is not currently implemented
+! CHECK: mreassociate is not currently implemented

flang/test/Driver/frontend-forwarding.f90

 ! Test that flang-new forwards Flang frontend
 ! options to flang-new -fc1 as expected.
 
 ! RUN: %flang -fsyntax-only -### %s -o %t 2>&1 \
 ! RUN:     -finput-charset=utf-8 \
 ! RUN:     -fdefault-double-8 \
 ! RUN:     -fdefault-integer-8 \
 ! RUN:     -fdefault-real-8 \
 ! RUN:     -flarge-sizes \
 ! RUN:     -fconvert=little-endian \
 ! RUN:     -ffp-contract=fast \
 ! RUN:     -fno-honor-infinities \
 ! RUN:     -fno-honor-nans \
 ! RUN:     -fapprox-func \
 ! RUN:     -fno-signed-zeros \
+! RUN:     -fassociative-math \
 ! RUN:     -mllvm -print-before-all\
 ! RUN:     -P \
 ! RUN:   | FileCheck %s
 
 ! CHECK: "-P"
 ! CHECK: "-finput-charset=utf-8"
 ! CHECK: "-fdefault-double-8"
 ! CHECK: "-fdefault-integer-8"
 ! CHECK: "-fdefault-real-8"
 ! CHECK: "-flarge-sizes"
 ! CHECK: "-ffp-contract=fast"
 ! CHECK: "-menable-no-infs"
 ! CHECK: "-menable-no-nans"
 ! CHECK: "-fapprox-func"
 ! CHECK: "-fno-signed-zeros"
+! CHECK: "-mreassociate"
 ! CHECK: "-fconvert=little-endian"
 ! CHECK:  "-mllvm" "-print-before-all"