[Matrix] Add a row\col major toggle in the clang driver (#167628)

fixes #167621

- define the new options in `Options.td` limit the naming to row-major
or column-major.
- In `ToolChains/Clang.cpp` limit the opt usage to only when
`-fenable-matrix` is used.

---------

Co-authored-by: Florian Hahn <[email protected]>

Author: farzonl

clang/docs/LanguageExtensions.rst

@@ -1073,6 +1073,12 @@ The matrix type extension supports explicit casts. Implicit type conversion betw
     i = static_cast<matrix_5_5<int>>(d);
   }
 
+The matrix type extension supports column and row major memory layouts, but not
+all builtins are supported with row-major layout. The layout defaults to column
+major and can be specified using `-fmatrix-memory-layout`. To enable column 
+major layout, use `-fmatrix-memory-layout=column-major`, and for row major
+layout use `-fmatrix-memory-layout=row-major`
+
 Half-Precision Floating Point
 =============================
 

clang/docs/MatrixTypes.rst

@@ -287,6 +287,10 @@ part of the draft specification.
 The elements of a  value of a matrix type are laid out in column-major order
 without padding.
 
+To change memory layout to row major use the `-fmatrix-memory-layout` flag.
+This flag supports two flag argument values either `column-major` or
+`row-major` used like so `-fmatrix-memory-layout=column-major`.` 
+
 We propose to provide a Clang option to override this behavior and allow
 contraction of those operations (e.g. *-ffp-contract=matrix*).
 

clang/docs/ReleaseNotes.rst

@@ -331,6 +331,7 @@ New Compiler Flags
 - New option ``-fsanitize-debug-trap-reasons=`` added to control emitting trap reasons into the debug info when compiling with trapping UBSan (e.g. ``-fsanitize-trap=undefined``).
 - New options for enabling allocation token instrumentation: ``-fsanitize=alloc-token``, ``-falloc-token-max=``, ``-fsanitize-alloc-token-fast-abi``, ``-fsanitize-alloc-token-extended``.
 - The ``-resource-dir`` option is now displayed in the list of options shown by ``--help``.
+- New option ``-fmatrix-memory-layout`` added to control the memory layout of Clang matrix types. (e.g. ``-fmatrix-memory-layout=column-major`` or ``-fmatrix-memory-layout=row-major``).
 
 Lanai Support
 ^^^^^^^^^^^^^^

clang/include/clang/Basic/DiagnosticDriverKinds.td

@@ -792,6 +792,9 @@ def err_cc1_round_trip_mismatch : Error<
 def err_cc1_unbounded_vscale_min : Error<
   "minimum vscale must be an unsigned integer greater than 0">;
 
+def err_conflicting_matrix_layout_flags: Error<
+  "-fmatrix-memory-layout=%0 conflicts with -mllvm -matrix-default-layout=%1">;
+
 def err_drv_using_omit_rtti_component_without_no_rtti : Error<
   "-fexperimental-omit-vtable-rtti call only be used with -fno-rtti">;
 

clang/include/clang/Basic/DiagnosticSemaKinds.td

@@ -13041,6 +13041,9 @@ def err_builtin_trivially_relocate_invalid_arg_type: Error <
 
 def err_builtin_matrix_disabled: Error<
   "matrix types extension is disabled. Pass -fenable-matrix to enable it">;
+def err_builtin_matrix_major_order_disabled: Error<
+  "matrix %select{row|column}0 major %select{load|store}1 is not supported with "
+  "-fmatrix-memory-layout=%select{column|row}0-major. Pass -fmatrix-memory-layout=%select{row|column}0-major to enable it">;
 def err_matrix_index_not_integer: Error<
   "matrix %select{row|column}0 index is not an integer">;
 def err_matrix_index_outside_range: Error<

clang/include/clang/Basic/LangOptions.def

@@ -434,6 +434,7 @@ ENUM_LANGOPT(RegisterStaticDestructors, RegisterStaticDestructorsKind, 2,
 LANGOPT(RegCall4, 1, 0, NotCompatible, "Set __regcall4 as a default calling convention to respect __regcall ABI v.4")
 
 LANGOPT(MatrixTypes, 1, 0, NotCompatible, "Enable or disable the builtin matrix type")
+ENUM_LANGOPT(DefaultMatrixMemoryLayout, MatrixMemoryLayout, 1, MatrixMemoryLayout::MatrixColMajor, NotCompatible, "Defines the default memory Layout for matrices")
 VALUE_LANGOPT(MaxMatrixDimension, 32, (1 << 20) - 1, NotCompatible, "maximum allowed matrix dimension")
 
 LANGOPT(CXXAssumptions, 1, 1, NotCompatible, "Enable or disable codegen and compile-time checks for C++23's [[assume]] attribute")

clang/include/clang/Basic/LangOptions.h

@@ -251,6 +251,13 @@ class LangOptionsBase {
     FEM_UnsetOnCommandLine = 3
   };
 
+  enum class MatrixMemoryLayout : unsigned {
+    // Use column-major layout for matrices
+    MatrixColMajor = 0,
+    // Use row-major layout for matrices
+    MatrixRowMajor = 1,
+  };
+
   enum ExcessPrecisionKind { FPP_Standard, FPP_Fast, FPP_None };
 
   enum class LaxVectorConversionKind {

clang/include/clang/Options/Options.td

@@ -4692,6 +4692,12 @@ def fenable_matrix : Flag<["-"], "fenable-matrix">, Group<f_Group>,
     HelpText<"Enable matrix data type and related builtin functions">,
     MarshallingInfoFlag<LangOpts<"MatrixTypes">, hlsl.KeyPath>;
 
+def fmatrix_memory_layout_EQ : Joined<["-"], "fmatrix-memory-layout=">,
+    Visibility<[ClangOption, CC1Option]>,
+    HelpText<"Sets the matrix memory layout (row-major or column-major)">,
+    Values<"row-major,column-major">,
+    Group<f_Group>;
+
 defm raw_string_literals : BoolFOption<"raw-string-literals",
     LangOpts<"RawStringLiterals">, Default<std#".hasRawStringLiterals()">,
     PosFlag<SetTrue, [], [], "Enable">,

clang/lib/Driver/ToolChains/Clang.cpp

@@ -5703,6 +5703,18 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
     CmdArgs.push_back("-fenable-matrix");
     CmdArgs.push_back("-mllvm");
     CmdArgs.push_back("-enable-matrix");
+    // Only handle default layout if matrix is enabled
+    if (const Arg *A = Args.getLastArg(options::OPT_fmatrix_memory_layout_EQ)) {
+      StringRef Val = A->getValue();
+      if (Val == "row-major" || Val == "column-major") {
+        CmdArgs.push_back(Args.MakeArgString("-fmatrix-memory-layout=" + Val));
+        CmdArgs.push_back("-mllvm");
+        CmdArgs.push_back(Args.MakeArgString("-matrix-default-layout=" + Val));
+
+      } else {
+        D.Diag(diag::err_drv_invalid_value) << A->getAsString(Args) << Val;
+      }
+    }
   }
 
   CodeGenOptions::FramePointerKind FPKeepKind =

clang/lib/Frontend/CompilerInvocation.cpp

@@ -3951,6 +3951,15 @@ void CompilerInvocationBase::GenerateLangArgs(const LangOptions &Opts,
     StringRef S = llvm::getAllocTokenModeAsString(*Opts.AllocTokenMode);
     GenerateArg(Consumer, OPT_falloc_token_mode_EQ, S);
   }
+  // Generate args for matrix types.
+  if (Opts.MatrixTypes) {
+    if (Opts.getDefaultMatrixMemoryLayout() ==
+        LangOptions::MatrixMemoryLayout::MatrixColMajor)
+      GenerateArg(Consumer, OPT_fmatrix_memory_layout_EQ, "column-major");
+    if (Opts.getDefaultMatrixMemoryLayout() ==
+        LangOptions::MatrixMemoryLayout::MatrixRowMajor)
+      GenerateArg(Consumer, OPT_fmatrix_memory_layout_EQ, "row-major");
+  }
 }
 
 bool CompilerInvocation::ParseLangArgs(LangOptions &Opts, ArgList &Args,
@@ -4545,6 +4554,27 @@ bool CompilerInvocation::ParseLangArgs(LangOptions &Opts, ArgList &Args,
       Diags.Report(diag::err_drv_invalid_value) << Arg->getAsString(Args) << S;
   }
 
+  // Enable options for matrix types.
+  if (Opts.MatrixTypes) {
+    if (const Arg *A = Args.getLastArg(OPT_fmatrix_memory_layout_EQ)) {
+      StringRef ClangValue = A->getValue();
+      if (ClangValue == "row-major")
+        Opts.setDefaultMatrixMemoryLayout(
+            LangOptions::MatrixMemoryLayout::MatrixRowMajor);
+      else
+        Opts.setDefaultMatrixMemoryLayout(
+            LangOptions::MatrixMemoryLayout::MatrixColMajor);
+
+      for (Arg *A : Args.filtered(options::OPT_mllvm)) {
+        StringRef OptValue = A->getValue();
+        if (OptValue.consume_front("-matrix-default-layout=") &&
+            ClangValue != OptValue)
+          Diags.Report(diag::err_conflicting_matrix_layout_flags)
+              << ClangValue << OptValue;
+      }
+    }
+  }
+
   // Validate options for HLSL
   if (Opts.HLSL) {
     // TODO: Revisit restricting SPIR-V to logical once we've figured out how to