[clang] [clang][HIP] Apply clang-cl optimization options to device code (PR #228096)

Yaxun Liu via cfe-commits cfe-commits at lists.llvm.org
Sat Oct 3 04:18:35 PDT 2026


https://github.com/yxsamliu updated https://github.com/llvm/llvm-project/pull/228096

>From f4d578919ca8e4a0c9eebfb27f65ad637e4ffda1 Mon Sep 17 00:00:00 2001
From: "Yaxun (Sam) Liu" <yaxun.liu at amd.com>
Date: Thu, 1 Oct 2026 10:17:45 -0400
Subject: [PATCH 1/4] [clang][HIP] Apply clang-cl optimization options to
 device code

clang-cl options such as /O1, /O2, and /Ox currently optimize only
host code in HIP builds, leaving device code unoptimized.

Translate these options before separating host and device arguments so
both compilations receive the requested optimization settings.
---
 clang/docs/ReleaseNotes.md                |   3 +
 clang/include/clang/Driver/Driver.h       |   3 +-
 clang/lib/Driver/Driver.cpp               | 130 +++++++++++++++++++++-
 clang/lib/Driver/ToolChains/MSVC.cpp      | 130 +---------------------
 clang/test/Driver/cl-options.cu           |   7 ++
 clang/test/Driver/hip-cl-optimization.hip | 106 ++++++++++++++++++
 6 files changed, 247 insertions(+), 132 deletions(-)
 create mode 100644 clang/test/Driver/hip-cl-optimization.hip

diff --git a/clang/docs/ReleaseNotes.md b/clang/docs/ReleaseNotes.md
index d22874d4130eec..4fdefafccea14d 100644
--- a/clang/docs/ReleaseNotes.md
+++ b/clang/docs/ReleaseNotes.md
@@ -815,6 +815,9 @@ features cannot lower the translation-unit ABI level;
 
 #### Miscellaneous Bug Fixes
 
+- Fixed `clang-cl` optimization options such as `/O1`, `/O2`, and
+  `/Ox` being ignored when compiling HIP device code.
+
 #### Miscellaneous Clang Crashes Fixed
 
 - Fixed a crash in CTAD for type alias templates when the aggregate deduction guide could not be resolved. (#GH206994)
diff --git a/clang/include/clang/Driver/Driver.h b/clang/include/clang/Driver/Driver.h
index 15b6fdcb8a5746..914ef62cd62fa9 100644
--- a/clang/include/clang/Driver/Driver.h
+++ b/clang/include/clang/Driver/Driver.h
@@ -360,7 +360,8 @@ class Driver {
   /// TranslateInputArgs - Create a new derived argument list from the input
   /// arguments, after applying the standard argument translations.
   llvm::opt::DerivedArgList *
-  TranslateInputArgs(const llvm::opt::InputArgList &Args) const;
+  TranslateInputArgs(const llvm::opt::InputArgList &Args,
+                     const llvm::Triple &Triple) const;
 
   // handleArguments - All code related to claiming and printing diagnostics
   // related to arguments to the driver are done here.
diff --git a/clang/lib/Driver/Driver.cpp b/clang/lib/Driver/Driver.cpp
index 7649941a68b1cc..07f70fbbe0128c 100644
--- a/clang/lib/Driver/Driver.cpp
+++ b/clang/lib/Driver/Driver.cpp
@@ -459,10 +459,128 @@ Arg *clang::driver::makeInputArg(DerivedArgList &Args, const OptTable &Opts,
   return A;
 }
 
-DerivedArgList *Driver::TranslateInputArgs(const InputArgList &Args) const {
+static void translateMSVCOptArg(Arg *A, llvm::opt::DerivedArgList &DAL,
+                                bool SupportsForcingFramePointer,
+                                const char *ExpandChar, const OptTable &Opts) {
+  assert(A->getOption().matches(options::OPT__SLASH_O));
+
+  StringRef OptStr = A->getValue();
+  for (size_t I = 0, E = OptStr.size(); I != E; ++I) {
+    const char &OptChar = *(OptStr.data() + I);
+    switch (OptChar) {
+    default:
+      break;
+    case '1':
+    case '2':
+    case 'x':
+    case 'd':
+      // Ignore /O[12xd] flags that aren't the last one on the command line.
+      // Only the last one gets expanded.
+      if (&OptChar != ExpandChar) {
+        A->claim();
+        break;
+      }
+      if (OptChar == 'd') {
+        DAL.AddFlagArg(A, Opts.getOption(options::OPT_O0));
+      } else {
+        if (OptChar == '1') {
+          DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "s");
+        } else if (OptChar == '2' || OptChar == 'x') {
+          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fbuiltin));
+          DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "3");
+        }
+        if (SupportsForcingFramePointer &&
+            !DAL.hasArgNoClaim(options::OPT_fno_omit_frame_pointer))
+          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fomit_frame_pointer));
+        if (OptChar == '1' || OptChar == '2')
+          DAL.AddFlagArg(A, Opts.getOption(options::OPT_ffunction_sections));
+      }
+      break;
+    case 'b':
+      if (I + 1 != E && isdigit(OptStr[I + 1])) {
+        switch (OptStr[I + 1]) {
+        case '0':
+          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fno_inline));
+          break;
+        case '1':
+          DAL.AddFlagArg(A,
+                         Opts.getOption(options::OPT_finline_hint_functions));
+          break;
+        case '2':
+        case '3':
+          DAL.AddFlagArg(A, Opts.getOption(options::OPT_finline_functions));
+          break;
+        }
+        ++I;
+      }
+      break;
+    case 'g':
+      A->claim();
+      break;
+    case 'i':
+      if (I + 1 != E && OptStr[I + 1] == '-') {
+        ++I;
+        DAL.AddFlagArg(A, Opts.getOption(options::OPT_fno_builtin));
+      } else {
+        DAL.AddFlagArg(A, Opts.getOption(options::OPT_fbuiltin));
+      }
+      break;
+    case 's':
+      DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "s");
+      break;
+    case 't':
+      DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "3");
+      break;
+    case 'y': {
+      bool OmitFramePointer = true;
+      if (I + 1 != E && OptStr[I + 1] == '-') {
+        OmitFramePointer = false;
+        ++I;
+      }
+      if (SupportsForcingFramePointer) {
+        if (OmitFramePointer)
+          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fomit_frame_pointer));
+        else
+          DAL.AddFlagArg(A,
+                         Opts.getOption(options::OPT_fno_omit_frame_pointer));
+      } else {
+        // Silently accept /Oy- on x86-64 for portable clang-cl build flags.
+        A->claim();
+      }
+      break;
+    }
+    }
+  }
+}
+
+DerivedArgList *Driver::TranslateInputArgs(const InputArgList &Args,
+                                           const llvm::Triple &Triple) const {
   const llvm::opt::OptTable &Opts = getOpts();
   DerivedArgList *DAL = new DerivedArgList(Args);
 
+  // Normalize MSVC optimization options before host and device arguments split.
+  bool TranslateMSVCOpts = Triple.isWindowsMSVCEnvironment();
+  // /Oy and /Oy- do not affect the x86-64 host.
+  bool SupportsForcingFramePointer = Triple.getArch() != llvm::Triple::x86_64;
+  // Expand only the last /O[12xd], preserving overrides such as /O2 /Oy-.
+  const char *ExpandChar = nullptr;
+  if (TranslateMSVCOpts) {
+    for (Arg *A : Args.filtered(options::OPT__SLASH_O)) {
+      StringRef OptStr = A->getValue();
+      for (size_t I = 0, E = OptStr.size(); I != E; ++I) {
+        char OptChar = OptStr[I];
+        char PrevChar = I > 0 ? OptStr[I - 1] : '0';
+        if (PrevChar == 'b') {
+          // OptChar does not expand; it's an argument to the previous char.
+          continue;
+        }
+        if (OptChar == '1' || OptChar == '2' || OptChar == 'x' ||
+            OptChar == 'd')
+          ExpandChar = OptStr.data() + I;
+      }
+    }
+  }
+
   bool HasNostdlib = Args.hasArg(options::OPT_nostdlib);
   bool HasNostdlibxx = Args.hasArg(options::OPT_nostdlibxx);
   bool HasNodefaultlib = Args.hasArg(options::OPT_nodefaultlibs);
@@ -480,6 +598,14 @@ DerivedArgList *Driver::TranslateInputArgs(const InputArgList &Args) const {
       continue;
     }
 
+    if (TranslateMSVCOpts && A->getOption().matches(options::OPT__SLASH_O)) {
+      // Keep the original argument for unused-option diagnostics.
+      DAL->append(A);
+      translateMSVCOptArg(A, *DAL, SupportsForcingFramePointer, ExpandChar,
+                          Opts);
+      continue;
+    }
+
     // Unfortunately, we have to parse some forwarding options (-Xassembler,
     // -Xlinker, -Xpreprocessor) because we either integrate their functionality
     // (assembler and preprocessor), or bypass a previous driver ('collect2').
@@ -1783,7 +1909,7 @@ Compilation *Driver::BuildCompilation(ArrayRef<const char *> ArgList) {
   }
 
   // Perform the default argument translations.
-  DerivedArgList *TranslatedArgs = TranslateInputArgs(*UArgs);
+  DerivedArgList *TranslatedArgs = TranslateInputArgs(*UArgs, TC.getTriple());
 
   // Check if the environment version is valid except wasm case.
   llvm::Triple Triple = TC.getTriple();
diff --git a/clang/lib/Driver/ToolChains/MSVC.cpp b/clang/lib/Driver/ToolChains/MSVC.cpp
index 25ea64a4ba5055..d32d5b9a52fc94 100644
--- a/clang/lib/Driver/ToolChains/MSVC.cpp
+++ b/clang/lib/Driver/ToolChains/MSVC.cpp
@@ -1023,103 +1023,6 @@ SanitizerMask MSVCToolChain::getSupportedSanitizers(
   return Res;
 }
 
-static void TranslateOptArg(Arg *A, llvm::opt::DerivedArgList &DAL,
-                            bool SupportsForcingFramePointer,
-                            const char *ExpandChar, const OptTable &Opts) {
-  assert(A->getOption().matches(options::OPT__SLASH_O));
-
-  StringRef OptStr = A->getValue();
-  for (size_t I = 0, E = OptStr.size(); I != E; ++I) {
-    const char &OptChar = *(OptStr.data() + I);
-    switch (OptChar) {
-    default:
-      break;
-    case '1':
-    case '2':
-    case 'x':
-    case 'd':
-      // Ignore /O[12xd] flags that aren't the last one on the command line.
-      // Only the last one gets expanded.
-      if (&OptChar != ExpandChar) {
-        A->claim();
-        break;
-      }
-      if (OptChar == 'd') {
-        DAL.AddFlagArg(A, Opts.getOption(options::OPT_O0));
-      } else {
-        if (OptChar == '1') {
-          DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "s");
-        } else if (OptChar == '2' || OptChar == 'x') {
-          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fbuiltin));
-          DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "3");
-        }
-        if (SupportsForcingFramePointer &&
-            !DAL.hasArgNoClaim(options::OPT_fno_omit_frame_pointer))
-          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fomit_frame_pointer));
-        if (OptChar == '1' || OptChar == '2')
-          DAL.AddFlagArg(A, Opts.getOption(options::OPT_ffunction_sections));
-      }
-      break;
-    case 'b':
-      if (I + 1 != E && isdigit(OptStr[I + 1])) {
-        switch (OptStr[I + 1]) {
-        case '0':
-          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fno_inline));
-          break;
-        case '1':
-          DAL.AddFlagArg(A, Opts.getOption(options::OPT_finline_hint_functions));
-          break;
-        case '2':
-        case '3':
-          DAL.AddFlagArg(A, Opts.getOption(options::OPT_finline_functions));
-          break;
-        }
-        ++I;
-      }
-      break;
-    case 'g':
-      A->claim();
-      break;
-    case 'i':
-      if (I + 1 != E && OptStr[I + 1] == '-') {
-        ++I;
-        DAL.AddFlagArg(A, Opts.getOption(options::OPT_fno_builtin));
-      } else {
-        DAL.AddFlagArg(A, Opts.getOption(options::OPT_fbuiltin));
-      }
-      break;
-    case 's':
-      DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "s");
-      break;
-    case 't':
-      DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "3");
-      break;
-    case 'y': {
-      bool OmitFramePointer = true;
-      if (I + 1 != E && OptStr[I + 1] == '-') {
-        OmitFramePointer = false;
-        ++I;
-      }
-      if (SupportsForcingFramePointer) {
-        if (OmitFramePointer)
-          DAL.AddFlagArg(A,
-                         Opts.getOption(options::OPT_fomit_frame_pointer));
-        else
-          DAL.AddFlagArg(
-              A, Opts.getOption(options::OPT_fno_omit_frame_pointer));
-      } else {
-        // Don't warn about /Oy- in x86-64 builds (where
-        // SupportsForcingFramePointer is false).  The flag having no effect
-        // there is a compiler-internal optimization, and people shouldn't have
-        // to special-case their build files for x86-64 clang-cl.
-        A->claim();
-      }
-      break;
-    }
-    }
-  }
-}
-
 static void TranslateDArg(Arg *A, llvm::opt::DerivedArgList &DAL,
                           const OptTable &Opts) {
   assert(A->getOption().matches(options::OPT_D));
@@ -1154,39 +1057,8 @@ MSVCToolChain::TranslateArgs(const llvm::opt::DerivedArgList &Args,
   DerivedArgList *DAL = new DerivedArgList(Args.getBaseArgs());
   const OptTable &Opts = getDriver().getOpts();
 
-  // /Oy and /Oy- don't have an effect on X86-64
-  bool SupportsForcingFramePointer = getArch() != llvm::Triple::x86_64;
-
-  // The -O[12xd] flag actually expands to several flags.  We must desugar the
-  // flags so that options embedded can be negated.  For example, the '-O2' flag
-  // enables '-Oy'.  Expanding '-O2' into its constituent flags allows us to
-  // correctly handle '-O2 -Oy-' where the trailing '-Oy-' disables a single
-  // aspect of '-O2'.
-  //
-  // Note that this expansion logic only applies to the *last* of '[12xd]'.
-
-  // First step is to search for the character we'd like to expand.
-  const char *ExpandChar = nullptr;
-  for (Arg *A : Args.filtered(options::OPT__SLASH_O)) {
-    StringRef OptStr = A->getValue();
-    for (size_t I = 0, E = OptStr.size(); I != E; ++I) {
-      char OptChar = OptStr[I];
-      char PrevChar = I > 0 ? OptStr[I - 1] : '0';
-      if (PrevChar == 'b') {
-        // OptChar does not expand; it's an argument to the previous char.
-        continue;
-      }
-      if (OptChar == '1' || OptChar == '2' || OptChar == 'x' || OptChar == 'd')
-        ExpandChar = OptStr.data() + I;
-    }
-  }
-
   for (Arg *A : Args) {
-    if (A->getOption().matches(options::OPT__SLASH_O)) {
-      // The -O flag actually takes an amalgam of other options.  For example,
-      // '/Ogyb2' is equivalent to '/Og' '/Oy' '/Ob2'.
-      TranslateOptArg(A, *DAL, SupportsForcingFramePointer, ExpandChar, Opts);
-    } else if (A->getOption().matches(options::OPT_D)) {
+    if (A->getOption().matches(options::OPT_D)) {
       // Translate -Dfoo#bar into -Dfoo=bar.
       TranslateDArg(A, *DAL, Opts);
     } else if (A->getOption().matches(options::OPT__SLASH_permissive)) {
diff --git a/clang/test/Driver/cl-options.cu b/clang/test/Driver/cl-options.cu
index b241ec6672d851..2517e7d77d4a9e 100644
--- a/clang/test/Driver/cl-options.cu
+++ b/clang/test/Driver/cl-options.cu
@@ -25,3 +25,10 @@
 // Gd-NOT: "-fdefault-calling-conv=cdecl"
 // Gd: "-cc1" "-triple"
 // Gd: "-fdefault-calling-conv=cdecl"
+
+// Optimization options must continue to apply to both CUDA compilation jobs.
+// RUN: not %clang_cl /c /O2 -### -nocudalib -nocudainc -- %s 2>&1 | FileCheck -check-prefix=O2 %s
+// O2: "-cc1" "-triple" "nvptx{{(64)?}}-nvidia-cuda"
+// O2-SAME: "-O3"
+// O2: "-cc1" "-triple"
+// O2-SAME: "-O3"
diff --git a/clang/test/Driver/hip-cl-optimization.hip b/clang/test/Driver/hip-cl-optimization.hip
new file mode 100644
index 00000000000000..3a3720f0a5c67f
--- /dev/null
+++ b/clang/test/Driver/hip-cl-optimization.hip
@@ -0,0 +1,106 @@
+// Check that clang-cl optimization options apply to both host and device.
+// Device-specific options must still override the shared optimization level.
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /O1 -- %s 2>&1 | FileCheck %s --check-prefix=OS
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /O2 -- %s 2>&1 | FileCheck %s --check-prefix=O3
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /Ox -- %s 2>&1 | FileCheck %s --check-prefix=O3
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /Od -- %s 2>&1 | FileCheck %s --check-prefix=O0
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /O2 /Od -- %s 2>&1 | FileCheck %s --check-prefix=O0
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /Od /O2 -- %s 2>&1 | FileCheck %s --check-prefix=O3
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /O2 -clang:-O1 -- %s 2>&1 | FileCheck %s --check-prefix=O1
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /O2 -clang:-Xarch_device -clang:-O1 -- %s 2>&1 | FileCheck %s --check-prefix=DEVICE-O1
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /O2 /Ob0 -- %s 2>&1 | FileCheck %s --check-prefix=NOINLINE
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --no-offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /O1 -- %s 2>&1 | FileCheck %s --check-prefix=OS
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --no-offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /O2 -- %s 2>&1 | FileCheck %s --check-prefix=O3
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --no-offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /Ox -- %s 2>&1 | FileCheck %s --check-prefix=O3
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --no-offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /Od -- %s 2>&1 | FileCheck %s --check-prefix=O0
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --no-offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /O2 /Od -- %s 2>&1 | FileCheck %s --check-prefix=O0
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --no-offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /Od /O2 -- %s 2>&1 | FileCheck %s --check-prefix=O3
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --no-offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /O2 -clang:-O1 -- %s 2>&1 | FileCheck %s --check-prefix=O1
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --no-offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /O2 -clang:-Xarch_device -clang:-O1 -- %s 2>&1 | FileCheck %s --check-prefix=DEVICE-O1
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --no-offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /O2 /Ob0 -- %s 2>&1 | FileCheck %s --check-prefix=NOINLINE
+
+// OS: "-cc1" "-triple" "amdgpu11.00-amd-amdhsa"
+// OS-SAME: "-Os"
+// OS: "-cc1" "-triple" "x86_64-pc-windows-msvc{{[^"]*}}"
+// OS-SAME: "-Os"
+
+// O3: "-cc1" "-triple" "amdgpu11.00-amd-amdhsa"
+// O3-SAME: "-O3"
+// O3: "-cc1" "-triple" "x86_64-pc-windows-msvc{{[^"]*}}"
+// O3-SAME: "-O3"
+
+// O0: "-cc1" "-triple" "amdgpu11.00-amd-amdhsa"
+// O0-SAME: "-O0"
+// O0: "-cc1" "-triple" "x86_64-pc-windows-msvc{{[^"]*}}"
+// O0-SAME: "-O0"
+
+// O1: "-cc1" "-triple" "amdgpu11.00-amd-amdhsa"
+// O1-SAME: "-O1"
+// O1: "-cc1" "-triple" "x86_64-pc-windows-msvc{{[^"]*}}"
+// O1-SAME: "-O1"
+
+// DEVICE-O1: "-cc1" "-triple" "amdgpu11.00-amd-amdhsa"
+// DEVICE-O1-SAME: "-O1"
+// DEVICE-O1: "-cc1" "-triple" "x86_64-pc-windows-msvc{{[^"]*}}"
+// DEVICE-O1-SAME: "-O3"
+
+// NOINLINE: "-cc1" "-triple" "amdgpu11.00-amd-amdhsa"
+// NOINLINE-SAME: "-O3"
+// NOINLINE-SAME: "-fno-inline"
+// NOINLINE: "-cc1" "-triple" "x86_64-pc-windows-msvc{{[^"]*}}"
+// NOINLINE-SAME: "-O3"
+// NOINLINE-SAME: "-fno-inline"

>From c437113e897dce7106e7a3fa7ca21dbd8e19b781 Mon Sep 17 00:00:00 2001
From: "Yaxun (Sam) Liu" <yaxun.liu at amd.com>
Date: Fri, 2 Oct 2026 10:21:12 -0400
Subject: [PATCH 2/4] [clang] Extract shared clang-cl optimization translation

Keep clang-cl optimization parsing in an internal helper while the driver
normalizes options before host and device arguments are split.

Select translation by clang-cl mode so Linux and MinGW host toolchains
also receive the canonical optimization options.
---
 clang/docs/ReleaseNotes.md                |   3 +-
 clang/lib/Driver/CMakeLists.txt           |   1 +
 clang/lib/Driver/ClangCLArgs.cpp          | 139 ++++++++++++++++++++++
 clang/lib/Driver/ClangCLArgs.h            |  41 +++++++
 clang/lib/Driver/Driver.cpp               | 127 +-------------------
 clang/test/Driver/cl-options.c            |   9 ++
 clang/test/Driver/hip-cl-optimization.hip |  13 ++
 7 files changed, 210 insertions(+), 123 deletions(-)
 create mode 100644 clang/lib/Driver/ClangCLArgs.cpp
 create mode 100644 clang/lib/Driver/ClangCLArgs.h

diff --git a/clang/docs/ReleaseNotes.md b/clang/docs/ReleaseNotes.md
index 4fdefafccea14d..4d46ed3b646afc 100644
--- a/clang/docs/ReleaseNotes.md
+++ b/clang/docs/ReleaseNotes.md
@@ -816,7 +816,8 @@ features cannot lower the translation-unit ABI level;
 #### Miscellaneous Bug Fixes
 
 - Fixed `clang-cl` optimization options such as `/O1`, `/O2`, and
-  `/Ox` being ignored when compiling HIP device code.
+  `/Ox` being ignored when compiling HIP device code or using a non-MSVC
+  host toolchain.
 
 #### Miscellaneous Clang Crashes Fixed
 
diff --git a/clang/lib/Driver/CMakeLists.txt b/clang/lib/Driver/CMakeLists.txt
index 506536cdc04f5f..c33dfd2471f35a 100644
--- a/clang/lib/Driver/CMakeLists.txt
+++ b/clang/lib/Driver/CMakeLists.txt
@@ -24,6 +24,7 @@ endif()
 
 add_clang_library(clangDriver
   Action.cpp
+  ClangCLArgs.cpp
   Compilation.cpp
   CreateASTUnitFromArgs.cpp
   CreateInvocationFromArgs.cpp
diff --git a/clang/lib/Driver/ClangCLArgs.cpp b/clang/lib/Driver/ClangCLArgs.cpp
new file mode 100644
index 00000000000000..64ff6adefd2567
--- /dev/null
+++ b/clang/lib/Driver/ClangCLArgs.cpp
@@ -0,0 +1,139 @@
+//===--- ClangCLArgs.cpp - clang-cl arguments -----------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "ClangCLArgs.h"
+#include "clang/Options/Options.h"
+#include "llvm/ADT/StringExtras.h"
+#include "llvm/ADT/StringRef.h"
+#include "llvm/Option/Arg.h"
+#include "llvm/Option/ArgList.h"
+#include "llvm/Option/OptTable.h"
+#include "llvm/Option/Option.h"
+#include "llvm/TargetParser/Triple.h"
+#include <cstddef>
+
+using namespace clang;
+using namespace clang::driver;
+using namespace llvm::opt;
+
+ClangCLArgs::ClangCLArgs(const ArgList &Args, const llvm::Triple &HostTriple)
+    : SupportsForcingFramePointer(HostTriple.getArch() !=
+                                  llvm::Triple::x86_64) {
+  // Expand only the last /O[12xd], preserving overrides such as /O2 /Oy-.
+  for (Arg *A : Args.filtered(options::OPT__SLASH_O)) {
+    llvm::StringRef OptStr = A->getValue();
+    for (size_t I = 0, E = OptStr.size(); I != E; ++I) {
+      char OptChar = OptStr[I];
+      char PrevChar = I > 0 ? OptStr[I - 1] : '0';
+      if (PrevChar == 'b') {
+        // OptChar does not expand; it's an argument to the previous char.
+        continue;
+      }
+      if (OptChar == '1' || OptChar == '2' || OptChar == 'x' || OptChar == 'd')
+        ExpandChar = OptStr.data() + I;
+    }
+  }
+}
+
+bool ClangCLArgs::translateArg(Arg *A, DerivedArgList &DAL) const {
+  if (!A->getOption().matches(options::OPT__SLASH_O))
+    return false;
+
+  // Keep the original argument for unused-option diagnostics.
+  DAL.append(A);
+  const OptTable &Opts = getDriverOptTable();
+
+  llvm::StringRef OptStr = A->getValue();
+  for (size_t I = 0, E = OptStr.size(); I != E; ++I) {
+    const char &OptChar = *(OptStr.data() + I);
+    switch (OptChar) {
+    default:
+      break;
+    case '1':
+    case '2':
+    case 'x':
+    case 'd':
+      // Ignore /O[12xd] flags that aren't the last one on the command line.
+      // Only the last one gets expanded.
+      if (&OptChar != ExpandChar) {
+        A->claim();
+        break;
+      }
+      if (OptChar == 'd') {
+        DAL.AddFlagArg(A, Opts.getOption(options::OPT_O0));
+      } else {
+        if (OptChar == '1') {
+          DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "s");
+        } else if (OptChar == '2' || OptChar == 'x') {
+          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fbuiltin));
+          DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "3");
+        }
+        if (SupportsForcingFramePointer &&
+            !DAL.hasArgNoClaim(options::OPT_fno_omit_frame_pointer))
+          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fomit_frame_pointer));
+        if (OptChar == '1' || OptChar == '2')
+          DAL.AddFlagArg(A, Opts.getOption(options::OPT_ffunction_sections));
+      }
+      break;
+    case 'b':
+      if (I + 1 != E && llvm::isDigit(OptStr[I + 1])) {
+        switch (OptStr[I + 1]) {
+        case '0':
+          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fno_inline));
+          break;
+        case '1':
+          DAL.AddFlagArg(A,
+                         Opts.getOption(options::OPT_finline_hint_functions));
+          break;
+        case '2':
+        case '3':
+          DAL.AddFlagArg(A, Opts.getOption(options::OPT_finline_functions));
+          break;
+        }
+        ++I;
+      }
+      break;
+    case 'g':
+      A->claim();
+      break;
+    case 'i':
+      if (I + 1 != E && OptStr[I + 1] == '-') {
+        ++I;
+        DAL.AddFlagArg(A, Opts.getOption(options::OPT_fno_builtin));
+      } else {
+        DAL.AddFlagArg(A, Opts.getOption(options::OPT_fbuiltin));
+      }
+      break;
+    case 's':
+      DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "s");
+      break;
+    case 't':
+      DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "3");
+      break;
+    case 'y': {
+      bool OmitFramePointer = true;
+      if (I + 1 != E && OptStr[I + 1] == '-') {
+        OmitFramePointer = false;
+        ++I;
+      }
+      if (SupportsForcingFramePointer) {
+        if (OmitFramePointer)
+          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fomit_frame_pointer));
+        else
+          DAL.AddFlagArg(A,
+                         Opts.getOption(options::OPT_fno_omit_frame_pointer));
+      } else {
+        // Silently accept /Oy- on x86-64 for portable clang-cl build flags.
+        A->claim();
+      }
+      break;
+    }
+    }
+  }
+  return true;
+}
diff --git a/clang/lib/Driver/ClangCLArgs.h b/clang/lib/Driver/ClangCLArgs.h
new file mode 100644
index 00000000000000..fbfd24064fa814
--- /dev/null
+++ b/clang/lib/Driver/ClangCLArgs.h
@@ -0,0 +1,41 @@
+//===--- ClangCLArgs.h - clang-cl arguments ---------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_CLANG_LIB_DRIVER_CLANGCLARGS_H
+#define LLVM_CLANG_LIB_DRIVER_CLANGCLARGS_H
+
+#include "llvm/Support/Compiler.h"
+
+namespace llvm {
+class Triple;
+namespace opt {
+class Arg;
+class ArgList;
+class DerivedArgList;
+} // namespace opt
+} // namespace llvm
+
+namespace clang::driver {
+
+/// Translate clang-cl options before host and device arguments are split.
+class LLVM_LIBRARY_VISIBILITY ClangCLArgs {
+  const char *ExpandChar = nullptr;
+  bool SupportsForcingFramePointer;
+
+public:
+  /// The argument values must outlive this translator.
+  ClangCLArgs(const llvm::opt::ArgList &Args, const llvm::Triple &HostTriple);
+
+  /// Append A and its canonical expansion to DAL. Return false without
+  /// modifying DAL if A is not handled.
+  bool translateArg(llvm::opt::Arg *A, llvm::opt::DerivedArgList &DAL) const;
+};
+
+} // namespace clang::driver
+
+#endif // LLVM_CLANG_LIB_DRIVER_CLANGCLARGS_H
diff --git a/clang/lib/Driver/Driver.cpp b/clang/lib/Driver/Driver.cpp
index 07f70fbbe0128c..20d5520fa03b3d 100644
--- a/clang/lib/Driver/Driver.cpp
+++ b/clang/lib/Driver/Driver.cpp
@@ -7,6 +7,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "clang/Driver/Driver.h"
+#include "ClangCLArgs.h"
 #include "ToolChains/AIX.h"
 #include "ToolChains/AMDGPU.h"
 #include "ToolChains/AVR.h"
@@ -459,127 +460,14 @@ Arg *clang::driver::makeInputArg(DerivedArgList &Args, const OptTable &Opts,
   return A;
 }
 
-static void translateMSVCOptArg(Arg *A, llvm::opt::DerivedArgList &DAL,
-                                bool SupportsForcingFramePointer,
-                                const char *ExpandChar, const OptTable &Opts) {
-  assert(A->getOption().matches(options::OPT__SLASH_O));
-
-  StringRef OptStr = A->getValue();
-  for (size_t I = 0, E = OptStr.size(); I != E; ++I) {
-    const char &OptChar = *(OptStr.data() + I);
-    switch (OptChar) {
-    default:
-      break;
-    case '1':
-    case '2':
-    case 'x':
-    case 'd':
-      // Ignore /O[12xd] flags that aren't the last one on the command line.
-      // Only the last one gets expanded.
-      if (&OptChar != ExpandChar) {
-        A->claim();
-        break;
-      }
-      if (OptChar == 'd') {
-        DAL.AddFlagArg(A, Opts.getOption(options::OPT_O0));
-      } else {
-        if (OptChar == '1') {
-          DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "s");
-        } else if (OptChar == '2' || OptChar == 'x') {
-          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fbuiltin));
-          DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "3");
-        }
-        if (SupportsForcingFramePointer &&
-            !DAL.hasArgNoClaim(options::OPT_fno_omit_frame_pointer))
-          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fomit_frame_pointer));
-        if (OptChar == '1' || OptChar == '2')
-          DAL.AddFlagArg(A, Opts.getOption(options::OPT_ffunction_sections));
-      }
-      break;
-    case 'b':
-      if (I + 1 != E && isdigit(OptStr[I + 1])) {
-        switch (OptStr[I + 1]) {
-        case '0':
-          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fno_inline));
-          break;
-        case '1':
-          DAL.AddFlagArg(A,
-                         Opts.getOption(options::OPT_finline_hint_functions));
-          break;
-        case '2':
-        case '3':
-          DAL.AddFlagArg(A, Opts.getOption(options::OPT_finline_functions));
-          break;
-        }
-        ++I;
-      }
-      break;
-    case 'g':
-      A->claim();
-      break;
-    case 'i':
-      if (I + 1 != E && OptStr[I + 1] == '-') {
-        ++I;
-        DAL.AddFlagArg(A, Opts.getOption(options::OPT_fno_builtin));
-      } else {
-        DAL.AddFlagArg(A, Opts.getOption(options::OPT_fbuiltin));
-      }
-      break;
-    case 's':
-      DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "s");
-      break;
-    case 't':
-      DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "3");
-      break;
-    case 'y': {
-      bool OmitFramePointer = true;
-      if (I + 1 != E && OptStr[I + 1] == '-') {
-        OmitFramePointer = false;
-        ++I;
-      }
-      if (SupportsForcingFramePointer) {
-        if (OmitFramePointer)
-          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fomit_frame_pointer));
-        else
-          DAL.AddFlagArg(A,
-                         Opts.getOption(options::OPT_fno_omit_frame_pointer));
-      } else {
-        // Silently accept /Oy- on x86-64 for portable clang-cl build flags.
-        A->claim();
-      }
-      break;
-    }
-    }
-  }
-}
-
 DerivedArgList *Driver::TranslateInputArgs(const InputArgList &Args,
                                            const llvm::Triple &Triple) const {
   const llvm::opt::OptTable &Opts = getOpts();
   DerivedArgList *DAL = new DerivedArgList(Args);
 
-  // Normalize MSVC optimization options before host and device arguments split.
-  bool TranslateMSVCOpts = Triple.isWindowsMSVCEnvironment();
-  // /Oy and /Oy- do not affect the x86-64 host.
-  bool SupportsForcingFramePointer = Triple.getArch() != llvm::Triple::x86_64;
-  // Expand only the last /O[12xd], preserving overrides such as /O2 /Oy-.
-  const char *ExpandChar = nullptr;
-  if (TranslateMSVCOpts) {
-    for (Arg *A : Args.filtered(options::OPT__SLASH_O)) {
-      StringRef OptStr = A->getValue();
-      for (size_t I = 0, E = OptStr.size(); I != E; ++I) {
-        char OptChar = OptStr[I];
-        char PrevChar = I > 0 ? OptStr[I - 1] : '0';
-        if (PrevChar == 'b') {
-          // OptChar does not expand; it's an argument to the previous char.
-          continue;
-        }
-        if (OptChar == '1' || OptChar == '2' || OptChar == 'x' ||
-            OptChar == 'd')
-          ExpandChar = OptStr.data() + I;
-      }
-    }
-  }
+  std::optional<ClangCLArgs> CLArgs;
+  if (IsCLMode())
+    CLArgs.emplace(Args, Triple);
 
   bool HasNostdlib = Args.hasArg(options::OPT_nostdlib);
   bool HasNostdlibxx = Args.hasArg(options::OPT_nostdlibxx);
@@ -598,13 +486,8 @@ DerivedArgList *Driver::TranslateInputArgs(const InputArgList &Args,
       continue;
     }
 
-    if (TranslateMSVCOpts && A->getOption().matches(options::OPT__SLASH_O)) {
-      // Keep the original argument for unused-option diagnostics.
-      DAL->append(A);
-      translateMSVCOptArg(A, *DAL, SupportsForcingFramePointer, ExpandChar,
-                          Opts);
+    if (CLArgs && CLArgs->translateArg(A, *DAL))
       continue;
-    }
 
     // Unfortunately, we have to parse some forwarding options (-Xassembler,
     // -Xlinker, -Xpreprocessor) because we either integrate their functionality
diff --git a/clang/test/Driver/cl-options.c b/clang/test/Driver/cl-options.c
index 57d82622ef4de6..b1d0ccbf60f0de 100644
--- a/clang/test/Driver/cl-options.c
+++ b/clang/test/Driver/cl-options.c
@@ -184,6 +184,15 @@
 // RUN: %clang_cl /Od -### -- %s 2>&1 | FileCheck -check-prefix=Od %s
 // Od: -O0
 
+// clang-cl optimization options are independent of the host toolchain.
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu /O1 -### -- %s 2>&1 | FileCheck %s --check-prefix=CL-LINUX-OS
+// CL-LINUX-OS: "-cc1"
+// CL-LINUX-OS-SAME: "-Os"
+
+// RUN: %clang_cl --target=x86_64-w64-windows-gnu /O2 -### -- %s 2>&1 | FileCheck %s --check-prefix=CL-MINGW-O3
+// CL-MINGW-O3: "-cc1"
+// CL-MINGW-O3-SAME: "-O3"
+
 // RUN: %clang_cl /Oi- /Oi -### -- %s 2>&1 | FileCheck -check-prefix=Oi %s
 // Oi-NOT: -fno-builtin
 
diff --git a/clang/test/Driver/hip-cl-optimization.hip b/clang/test/Driver/hip-cl-optimization.hip
index 3a3720f0a5c67f..1fa80de854e566 100644
--- a/clang/test/Driver/hip-cl-optimization.hip
+++ b/clang/test/Driver/hip-cl-optimization.hip
@@ -5,6 +5,10 @@
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /O1 -- %s 2>&1 | FileCheck %s --check-prefix=OS
 
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /O1 -- %s 2>&1 | FileCheck %s --check-prefix=LINUX
+
 // RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /O2 -- %s 2>&1 | FileCheck %s --check-prefix=O3
@@ -41,6 +45,10 @@
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /O1 -- %s 2>&1 | FileCheck %s --check-prefix=OS
 
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c --no-offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /O1 -- %s 2>&1 | FileCheck %s --check-prefix=LINUX
+
 // RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --no-offload-new-driver \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /O2 -- %s 2>&1 | FileCheck %s --check-prefix=O3
@@ -78,6 +86,11 @@
 // OS: "-cc1" "-triple" "x86_64-pc-windows-msvc{{[^"]*}}"
 // OS-SAME: "-Os"
 
+// LINUX: "-cc1" "-triple" "amdgpu11.00-amd-amdhsa"
+// LINUX-SAME: "-Os"
+// LINUX: "-cc1" "-triple" "x86_64-unknown-linux-gnu"
+// LINUX-SAME: "-Os"
+
 // O3: "-cc1" "-triple" "amdgpu11.00-amd-amdhsa"
 // O3-SAME: "-O3"
 // O3: "-cc1" "-triple" "x86_64-pc-windows-msvc{{[^"]*}}"

>From 21a546f5f4ea283212ebd9c708ede29e65fb4158 Mon Sep 17 00:00:00 2001
From: "Yaxun (Sam) Liu" <yaxun.liu at amd.com>
Date: Fri, 2 Oct 2026 12:34:42 -0400
Subject: [PATCH 3/4] [clang] Move remaining MSVC argument translation to
 ClangCLArgs

Normalize macro separators and /permissive options for both host and
device compilations, and remove MSVCToolChain::TranslateArgs.

Preserve late OpenMP and Xarch arguments, optimization bundle precedence,
and synthesized argument ownership during per-toolchain filtering.
---
 clang/docs/ReleaseNotes.md                    |   4 +-
 clang/lib/Driver/ClangCLArgs.cpp              |  99 +++++++++++++---
 clang/lib/Driver/ClangCLArgs.h                |  24 +++-
 clang/lib/Driver/Compilation.cpp              |  12 ++
 clang/lib/Driver/Driver.cpp                   |  12 ++
 clang/lib/Driver/ToolChain.cpp                |  15 +++
 clang/lib/Driver/ToolChains/MSVC.cpp          |  53 ---------
 clang/lib/Driver/ToolChains/MSVC.h            |   4 -
 clang/test/Driver/cl-argument-translation.c   |  58 +++++++++
 clang/test/Driver/cl-options.cu               |  16 +++
 .../Driver/hip-cl-argument-translation.hip    | 107 +++++++++++++++++
 .../Driver/openmp-cl-argument-translation.cpp | 111 ++++++++++++++++++
 12 files changed, 434 insertions(+), 81 deletions(-)
 create mode 100644 clang/test/Driver/cl-argument-translation.c
 create mode 100644 clang/test/Driver/hip-cl-argument-translation.hip
 create mode 100644 clang/test/Driver/openmp-cl-argument-translation.cpp

diff --git a/clang/docs/ReleaseNotes.md b/clang/docs/ReleaseNotes.md
index 4d46ed3b646afc..f6f0e9bb88d639 100644
--- a/clang/docs/ReleaseNotes.md
+++ b/clang/docs/ReleaseNotes.md
@@ -817,7 +817,9 @@ features cannot lower the translation-unit ABI level;
 
 - Fixed `clang-cl` optimization options such as `/O1`, `/O2`, and
   `/Ox` being ignored when compiling HIP device code or using a non-MSVC
-  host toolchain.
+  host toolchain. Macro definitions using `#` as the value separator and
+  `/permissive` options are also translated consistently for host and device
+  compilation.
 
 #### Miscellaneous Clang Crashes Fixed
 
diff --git a/clang/lib/Driver/ClangCLArgs.cpp b/clang/lib/Driver/ClangCLArgs.cpp
index 64ff6adefd2567..a06d1983da3d54 100644
--- a/clang/lib/Driver/ClangCLArgs.cpp
+++ b/clang/lib/Driver/ClangCLArgs.cpp
@@ -16,6 +16,7 @@
 #include "llvm/Option/Option.h"
 #include "llvm/TargetParser/Triple.h"
 #include <cstddef>
+#include <string>
 
 using namespace clang;
 using namespace clang::driver;
@@ -40,13 +41,33 @@ ClangCLArgs::ClangCLArgs(const ArgList &Args, const llvm::Triple &HostTriple)
   }
 }
 
-bool ClangCLArgs::translateArg(Arg *A, DerivedArgList &DAL) const {
+bool ClangCLArgs::translateArg(Arg *A, DerivedArgList &DAL,
+                               const DerivedArgList *Owner) const {
+  const DerivedArgList &SynthesizedArgs = Owner ? *Owner : DAL;
+  const Arg *BaseArg = &A->getBaseArg();
+  const OptTable &Opts = getDriverOptTable();
+  switch (A->getOption().getID()) {
+  case options::OPT__SLASH_permissive:
+    DAL.append(SynthesizedArgs.MakeFlagArg(
+        BaseArg, Opts.getOption(options::OPT_fdelayed_template_parsing)));
+    DAL.append(SynthesizedArgs.MakeFlagArg(
+        BaseArg, Opts.getOption(options::OPT_fno_operator_names)));
+    return true;
+  case options::OPT__SLASH_permissive_:
+    DAL.append(SynthesizedArgs.MakeFlagArg(
+        BaseArg, Opts.getOption(options::OPT_fno_delayed_template_parsing)));
+    DAL.append(SynthesizedArgs.MakeFlagArg(
+        BaseArg, Opts.getOption(options::OPT_foperator_names)));
+    return true;
+  default:
+    break;
+  }
+
   if (!A->getOption().matches(options::OPT__SLASH_O))
     return false;
 
   // Keep the original argument for unused-option diagnostics.
   DAL.append(A);
-  const OptTable &Opts = getDriverOptTable();
 
   llvm::StringRef OptStr = A->getValue();
   for (size_t I = 0, E = OptStr.size(); I != E; ++I) {
@@ -65,34 +86,42 @@ bool ClangCLArgs::translateArg(Arg *A, DerivedArgList &DAL) const {
         break;
       }
       if (OptChar == 'd') {
-        DAL.AddFlagArg(A, Opts.getOption(options::OPT_O0));
+        DAL.append(SynthesizedArgs.MakeFlagArg(
+            BaseArg, Opts.getOption(options::OPT_O0)));
       } else {
         if (OptChar == '1') {
-          DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "s");
+          DAL.append(SynthesizedArgs.MakeJoinedArg(
+              BaseArg, Opts.getOption(options::OPT_O), "s"));
         } else if (OptChar == '2' || OptChar == 'x') {
-          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fbuiltin));
-          DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "3");
+          DAL.append(SynthesizedArgs.MakeFlagArg(
+              BaseArg, Opts.getOption(options::OPT_fbuiltin)));
+          DAL.append(SynthesizedArgs.MakeJoinedArg(
+              BaseArg, Opts.getOption(options::OPT_O), "3"));
         }
         if (SupportsForcingFramePointer &&
             !DAL.hasArgNoClaim(options::OPT_fno_omit_frame_pointer))
-          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fomit_frame_pointer));
+          DAL.append(SynthesizedArgs.MakeFlagArg(
+              BaseArg, Opts.getOption(options::OPT_fomit_frame_pointer)));
         if (OptChar == '1' || OptChar == '2')
-          DAL.AddFlagArg(A, Opts.getOption(options::OPT_ffunction_sections));
+          DAL.append(SynthesizedArgs.MakeFlagArg(
+              BaseArg, Opts.getOption(options::OPT_ffunction_sections)));
       }
       break;
     case 'b':
       if (I + 1 != E && llvm::isDigit(OptStr[I + 1])) {
         switch (OptStr[I + 1]) {
         case '0':
-          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fno_inline));
+          DAL.append(SynthesizedArgs.MakeFlagArg(
+              BaseArg, Opts.getOption(options::OPT_fno_inline)));
           break;
         case '1':
-          DAL.AddFlagArg(A,
-                         Opts.getOption(options::OPT_finline_hint_functions));
+          DAL.append(SynthesizedArgs.MakeFlagArg(
+              BaseArg, Opts.getOption(options::OPT_finline_hint_functions)));
           break;
         case '2':
         case '3':
-          DAL.AddFlagArg(A, Opts.getOption(options::OPT_finline_functions));
+          DAL.append(SynthesizedArgs.MakeFlagArg(
+              BaseArg, Opts.getOption(options::OPT_finline_functions)));
           break;
         }
         ++I;
@@ -104,16 +133,20 @@ bool ClangCLArgs::translateArg(Arg *A, DerivedArgList &DAL) const {
     case 'i':
       if (I + 1 != E && OptStr[I + 1] == '-') {
         ++I;
-        DAL.AddFlagArg(A, Opts.getOption(options::OPT_fno_builtin));
+        DAL.append(SynthesizedArgs.MakeFlagArg(
+            BaseArg, Opts.getOption(options::OPT_fno_builtin)));
       } else {
-        DAL.AddFlagArg(A, Opts.getOption(options::OPT_fbuiltin));
+        DAL.append(SynthesizedArgs.MakeFlagArg(
+            BaseArg, Opts.getOption(options::OPT_fbuiltin)));
       }
       break;
     case 's':
-      DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "s");
+      DAL.append(SynthesizedArgs.MakeJoinedArg(
+          BaseArg, Opts.getOption(options::OPT_O), "s"));
       break;
     case 't':
-      DAL.AddJoinedArg(A, Opts.getOption(options::OPT_O), "3");
+      DAL.append(SynthesizedArgs.MakeJoinedArg(
+          BaseArg, Opts.getOption(options::OPT_O), "3"));
       break;
     case 'y': {
       bool OmitFramePointer = true;
@@ -123,10 +156,11 @@ bool ClangCLArgs::translateArg(Arg *A, DerivedArgList &DAL) const {
       }
       if (SupportsForcingFramePointer) {
         if (OmitFramePointer)
-          DAL.AddFlagArg(A, Opts.getOption(options::OPT_fomit_frame_pointer));
+          DAL.append(SynthesizedArgs.MakeFlagArg(
+              BaseArg, Opts.getOption(options::OPT_fomit_frame_pointer)));
         else
-          DAL.AddFlagArg(A,
-                         Opts.getOption(options::OPT_fno_omit_frame_pointer));
+          DAL.append(SynthesizedArgs.MakeFlagArg(
+              BaseArg, Opts.getOption(options::OPT_fno_omit_frame_pointer)));
       } else {
         // Silently accept /Oy- on x86-64 for portable clang-cl build flags.
         A->claim();
@@ -137,3 +171,30 @@ bool ClangCLArgs::translateArg(Arg *A, DerivedArgList &DAL) const {
   }
   return true;
 }
+
+DerivedArgList *ClangCLArgs::translateArgs(const DerivedArgList &Args,
+                                           const llvm::Triple &HostTriple,
+                                           const DerivedArgList &Owner) {
+  ClangCLArgs Translator(Args, HostTriple);
+  auto *DAL = new DerivedArgList(Args.getBaseArgs());
+  for (Arg *A : Args) {
+    if (!A->getOption().matches(options::OPT__SLASH_O) &&
+        A->getBaseArg().getOption().matches(options::OPT__SLASH_O))
+      continue;
+    if (!Translator.translateArg(A, *DAL, &Owner))
+      DAL->append(A);
+  }
+  return DAL;
+}
+
+const char *ClangCLArgs::translateMacroDefinition(const char *Value,
+                                                  const ArgList &Args) {
+  llvm::StringRef Val = Value;
+  size_t Hash = Val.find('#');
+  if (Hash == llvm::StringRef::npos || Hash > Val.find('='))
+    return Value;
+
+  std::string NewVal = std::string(Val);
+  NewVal[Hash] = '=';
+  return Args.MakeArgString(NewVal);
+}
diff --git a/clang/lib/Driver/ClangCLArgs.h b/clang/lib/Driver/ClangCLArgs.h
index fbfd24064fa814..c713be83a31e13 100644
--- a/clang/lib/Driver/ClangCLArgs.h
+++ b/clang/lib/Driver/ClangCLArgs.h
@@ -22,7 +22,7 @@ class DerivedArgList;
 
 namespace clang::driver {
 
-/// Translate clang-cl options before host and device arguments are split.
+/// Translate clang-cl options before and after offload argument filtering.
 class LLVM_LIBRARY_VISIBILITY ClangCLArgs {
   const char *ExpandChar = nullptr;
   bool SupportsForcingFramePointer;
@@ -31,9 +31,25 @@ class LLVM_LIBRARY_VISIBILITY ClangCLArgs {
   /// The argument values must outlive this translator.
   ClangCLArgs(const llvm::opt::ArgList &Args, const llvm::Triple &HostTriple);
 
-  /// Append A and its canonical expansion to DAL. Return false without
-  /// modifying DAL if A is not handled.
-  bool translateArg(llvm::opt::Arg *A, llvm::opt::DerivedArgList &DAL) const;
+  /// Append the canonical translation of A to DAL. Return false without
+  /// modifying DAL if A is not handled. If supplied, Owner owns synthesized
+  /// arguments instead of DAL and must outlive its consumers.
+  bool translateArg(llvm::opt::Arg *A, llvm::opt::DerivedArgList &DAL,
+                    const llvm::opt::DerivedArgList *Owner = nullptr) const;
+
+  /// Translate a filtered list that may contain new clang-cl options, replacing
+  /// earlier /O expansions. The caller owns the returned list; Owner owns its
+  /// synthesized arguments and must outlive its consumers.
+  static llvm::opt::DerivedArgList *
+  translateArgs(const llvm::opt::DerivedArgList &Args,
+                const llvm::Triple &HostTriple,
+                const llvm::opt::DerivedArgList &Owner);
+
+  /// Normalize MSVC's '#' macro separator, also used by Clang targeting MSVC
+  /// and by arguments parsed after host/device filtering. Any replacement
+  /// string has the lifetime of Args; otherwise return Value unchanged.
+  static const char *translateMacroDefinition(const char *Value,
+                                              const llvm::opt::ArgList &Args);
 };
 
 } // namespace clang::driver
diff --git a/clang/lib/Driver/Compilation.cpp b/clang/lib/Driver/Compilation.cpp
index c81c4445a29f95..3ccfe60c520d92 100644
--- a/clang/lib/Driver/Compilation.cpp
+++ b/clang/lib/Driver/Compilation.cpp
@@ -7,6 +7,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "clang/Driver/Compilation.h"
+#include "ClangCLArgs.h"
 #include "clang/Basic/LLVM.h"
 #include "clang/Driver/Action.h"
 #include "clang/Driver/CommonArgs.h"
@@ -91,6 +92,17 @@ Compilation::getArgsForToolChain(const ToolChain *TC, BoundArch BA,
         delete OpenMPArgs;
     }
 
+    if (NewDAL && DeviceOffloadKind == Action::OFK_OpenMP &&
+        (TheDriver.IsCLMode() ||
+         DefaultToolChain.getTriple().isWindowsMSVCEnvironment())) {
+      // OpenMP forwarding can introduce clang-cl options after the shared
+      // translation. Keep synthesized arguments in the compilation-owned list.
+      auto *DAL = ClangCLArgs::translateArgs(
+          *NewDAL, DefaultToolChain.getTriple(), *TranslatedArgs);
+      delete NewDAL;
+      NewDAL = DAL;
+    }
+
     if (!NewDAL) {
       Entry = TC->TranslateArgs(*TranslatedArgs, BA, DeviceOffloadKind);
       if (!Entry)
diff --git a/clang/lib/Driver/Driver.cpp b/clang/lib/Driver/Driver.cpp
index 20d5520fa03b3d..6b901118f9c7b1 100644
--- a/clang/lib/Driver/Driver.cpp
+++ b/clang/lib/Driver/Driver.cpp
@@ -489,6 +489,18 @@ DerivedArgList *Driver::TranslateInputArgs(const InputArgList &Args,
     if (CLArgs && CLArgs->translateArg(A, *DAL))
       continue;
 
+    // Preserve MSVC macro syntax for Clang's other driver modes as well.
+    if ((CLArgs || Triple.isWindowsMSVCEnvironment()) &&
+        A->getOption().matches(options::OPT_D)) {
+      const char *Value =
+          ClangCLArgs::translateMacroDefinition(A->getValue(), Args);
+      if (Value == A->getValue())
+        DAL->append(A);
+      else
+        DAL->AddJoinedArg(A, Opts.getOption(options::OPT_D), Value);
+      continue;
+    }
+
     // Unfortunately, we have to parse some forwarding options (-Xassembler,
     // -Xlinker, -Xpreprocessor) because we either integrate their functionality
     // (assembler and preprocessor), or bypass a previous driver ('collect2').
diff --git a/clang/lib/Driver/ToolChain.cpp b/clang/lib/Driver/ToolChain.cpp
index 255a9c9602b540..2d7dc6e63534b2 100644
--- a/clang/lib/Driver/ToolChain.cpp
+++ b/clang/lib/Driver/ToolChain.cpp
@@ -7,6 +7,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "clang/Driver/ToolChain.h"
+#include "ClangCLArgs.h"
 #include "ToolChains/Arch/AArch64.h"
 #include "ToolChains/Arch/AMDGPU.h"
 #include "ToolChains/Arch/ARM.h"
@@ -2074,6 +2075,12 @@ llvm::opt::DerivedArgList *ToolChain::TranslateOpenMPTargetArgs(
     }
     XOpenMPTargetArg->setBaseArg(A);
     A = XOpenMPTargetArg.release();
+    if (A->getOption().matches(options::OPT_D) &&
+        (getDriver().IsCLMode() ||
+         llvm::Triple(llvm::Triple::normalize(getDriver().getTargetTriple()))
+             .isWindowsMSVCEnvironment()))
+      A->getValues()[0] =
+          ClangCLArgs::translateMacroDefinition(A->getValue(), Args);
     AllocatedArgs.push_back(A);
     DAL->append(A);
     Modified = true;
@@ -2154,6 +2161,14 @@ void ToolChain::TranslateXarchArgs(
     DAL->AddSynthesizedArg(A);
   else
     AllocatedArgs->push_back(A);
+
+  // Forwarded macros are parsed after the shared input translation.
+  if (A->getOption().matches(options::OPT_D) &&
+      (getDriver().IsCLMode() ||
+       llvm::Triple(llvm::Triple::normalize(getDriver().getTargetTriple()))
+           .isWindowsMSVCEnvironment()))
+    A->getValues()[0] =
+        ClangCLArgs::translateMacroDefinition(A->getValue(), Args);
 }
 
 /// Match any triple recognized arch aliases.
diff --git a/clang/lib/Driver/ToolChains/MSVC.cpp b/clang/lib/Driver/ToolChains/MSVC.cpp
index d32d5b9a52fc94..8f4c8f89ce58f9 100644
--- a/clang/lib/Driver/ToolChains/MSVC.cpp
+++ b/clang/lib/Driver/ToolChains/MSVC.cpp
@@ -1023,59 +1023,6 @@ SanitizerMask MSVCToolChain::getSupportedSanitizers(
   return Res;
 }
 
-static void TranslateDArg(Arg *A, llvm::opt::DerivedArgList &DAL,
-                          const OptTable &Opts) {
-  assert(A->getOption().matches(options::OPT_D));
-
-  StringRef Val = A->getValue();
-  size_t Hash = Val.find('#');
-  if (Hash == StringRef::npos || Hash > Val.find('=')) {
-    DAL.append(A);
-    return;
-  }
-
-  std::string NewVal = std::string(Val);
-  NewVal[Hash] = '=';
-  DAL.AddJoinedArg(A, Opts.getOption(options::OPT_D), NewVal);
-}
-
-static void TranslatePermissive(Arg *A, llvm::opt::DerivedArgList &DAL,
-                                const OptTable &Opts) {
-  DAL.AddFlagArg(A, Opts.getOption(options::OPT__SLASH_Zc_twoPhase_));
-  DAL.AddFlagArg(A, Opts.getOption(options::OPT_fno_operator_names));
-}
-
-static void TranslatePermissiveMinus(Arg *A, llvm::opt::DerivedArgList &DAL,
-                                     const OptTable &Opts) {
-  DAL.AddFlagArg(A, Opts.getOption(options::OPT__SLASH_Zc_twoPhase));
-  DAL.AddFlagArg(A, Opts.getOption(options::OPT_foperator_names));
-}
-
-llvm::opt::DerivedArgList *
-MSVCToolChain::TranslateArgs(const llvm::opt::DerivedArgList &Args,
-                             BoundArch BA, Action::OffloadKind OFK) const {
-  DerivedArgList *DAL = new DerivedArgList(Args.getBaseArgs());
-  const OptTable &Opts = getDriver().getOpts();
-
-  for (Arg *A : Args) {
-    if (A->getOption().matches(options::OPT_D)) {
-      // Translate -Dfoo#bar into -Dfoo=bar.
-      TranslateDArg(A, *DAL, Opts);
-    } else if (A->getOption().matches(options::OPT__SLASH_permissive)) {
-      // Expand /permissive
-      TranslatePermissive(A, *DAL, Opts);
-    } else if (A->getOption().matches(options::OPT__SLASH_permissive_)) {
-      // Expand /permissive-
-      TranslatePermissiveMinus(A, *DAL, Opts);
-    } else if (OFK != Action::OFK_HIP) {
-      // HIP Toolchain translates input args by itself.
-      DAL->append(A);
-    }
-  }
-
-  return DAL;
-}
-
 void MSVCToolChain::addClangTargetOptions(
     const ArgList &DriverArgs, ArgStringList &CC1Args, BoundArch BA,
     Action::OffloadKind DeviceOffloadKind) const {
diff --git a/clang/lib/Driver/ToolChains/MSVC.h b/clang/lib/Driver/ToolChains/MSVC.h
index 8ff8b2815d8df7..90461a1443da31 100644
--- a/clang/lib/Driver/ToolChains/MSVC.h
+++ b/clang/lib/Driver/ToolChains/MSVC.h
@@ -61,10 +61,6 @@ class LLVM_LIBRARY_VISIBILITY MSVCToolChain : public ToolChain {
   MSVCToolChain(const Driver &D, const llvm::Triple &Triple,
                 const llvm::opt::ArgList &Args);
 
-  llvm::opt::DerivedArgList *
-  TranslateArgs(const llvm::opt::DerivedArgList &Args, BoundArch BA,
-                Action::OffloadKind DeviceOffloadKind) const override;
-
   UnwindTableLevel
   getDefaultUnwindTableLevel(const llvm::opt::ArgList &Args) const override;
   bool isPICDefault() const override;
diff --git a/clang/test/Driver/cl-argument-translation.c b/clang/test/Driver/cl-argument-translation.c
new file mode 100644
index 00000000000000..e1758265fa8890
--- /dev/null
+++ b/clang/test/Driver/cl-argument-translation.c
@@ -0,0 +1,58 @@
+// clang-cl syntax is normalized independently of the target toolchain.
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc /c -### /DFOO#7 /DBAR=a#b /DBAZ#a=b /DEMPTY# \
+// RUN:   /DQUX#a#b -- %s 2>&1 | FileCheck %s --check-prefix=MACROS
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu /c -### /DFOO#7 /DBAR=a#b /DBAZ#a=b /DEMPTY# \
+// RUN:   /DQUX#a#b -- %s 2>&1 | FileCheck %s --check-prefix=MACROS
+// RUN: %clang_cl --target=x86_64-w64-windows-gnu /c -### /DFOO#7 /DBAR=a#b /DBAZ#a=b /DEMPTY# \
+// RUN:   /DQUX#a#b -- %s 2>&1 | FileCheck %s --check-prefix=MACROS
+// MACROS: "-D" "FOO=7" "-D" "BAR=a#b" "-D" "BAZ=a=b" "-D" "EMPTY=" "-D" "QUX=a#b"
+
+// Preserve GNU-mode Clang's existing MSVC macro compatibility.
+// RUN: %clang --target=x86_64-pc-windows-msvc -c -### -DFOO#7 -- %s 2>&1 | FileCheck %s --check-prefix=MSVC
+// RUN: %clang --target=x86_64-unknown-linux-gnu -c -### -DFOO#7 -- %s 2>&1 | FileCheck %s --check-prefix=GNU
+// MSVC: "-D" "FOO=7"
+// GNU: "-D" "FOO#7"
+
+// Last-option precedence must include both slash and canonical options.
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc /c -x c++ -### /permissive- /permissive -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=PERMISSIVE
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu /c -x c++ -### /permissive- /permissive -- %s 2>&1 \
+// RUN:   | FileCheck %s --check-prefix=PERMISSIVE
+// RUN: %clang_cl --target=x86_64-w64-windows-gnu /c -x c++ -### /permissive -- %s 2>&1 | FileCheck %s \
+// RUN:   --check-prefix=PERMISSIVE
+// PERMISSIVE: "-fno-operator-names"
+// PERMISSIVE-SAME: "-fdelayed-template-parsing"
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc /c -x c++ -### /permissive /permissive- -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=CONFORMING
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu /c -x c++ -### /permissive /permissive- -- %s 2>&1 \
+// RUN:   | FileCheck %s --check-prefix=CONFORMING
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc /c -x c++ -### /permissive /Zc:twoPhase \
+// RUN:   -clang:-foperator-names -- %s 2>&1 | FileCheck %s --check-prefix=CONFORMING
+// /clang: arguments are appended after the other command-line arguments.
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu /c -x c++ -### \
+// RUN:   -clang:-fno-delayed-template-parsing -clang:-foperator-names /permissive -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=CONFORMING
+// CONFORMING: "-cc1"
+// CONFORMING-NOT: "-fno-operator-names"
+// CONFORMING-NOT: "-fdelayed-template-parsing"
+
+// RUN: %clang --target=x86_64-pc-windows-msvc -c -### -Xarch_host -DFOO#7 -- %s 2>&1 | FileCheck %s \
+// RUN:   --check-prefix=MSVC
+// RUN: %clang --target=x86_64-windows-msvc -c -### -Xarch_host -DFOO#7 -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=MSVC
+// RUN: %clang --target=x86_64-win32 -c -### -Xarch_host -DFOO#7 -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=MSVC
+// RUN: %clang --target=x86_64-unknown-linux-gnu -c -### -Xarch_host -DFOO#7 -- %s 2>&1 | FileCheck %s \
+// RUN:   --check-prefix=GNU
+// RUN: %clang --target=x86_64-w64-windows-gnu -c -### -DFOO#7 -- %s 2>&1 | FileCheck %s --check-prefix=GNU
+
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu /EP /DFOO#7 /DTRANSLATION_CHECK -- %s | FileCheck \
+// RUN:   %s --check-prefix=PREPROCESS
+// PREPROCESS: 7
+#ifdef TRANSLATION_CHECK
+#if FOO != 7
+#error macro separator was not translated
+#endif
+FOO
+#endif
diff --git a/clang/test/Driver/cl-options.cu b/clang/test/Driver/cl-options.cu
index 2517e7d77d4a9e..b373f4c7d0a031 100644
--- a/clang/test/Driver/cl-options.cu
+++ b/clang/test/Driver/cl-options.cu
@@ -32,3 +32,19 @@
 // O2-SAME: "-O3"
 // O2: "-cc1" "-triple"
 // O2-SAME: "-O3"
+
+// Shared macros and language settings also reach CUDA without reintroducing
+// the original macro alongside its translated value.
+// RUN: not %clang_cl --target=x86_64-pc-windows-msvc -### /c \
+// RUN:   --cuda-gpu-arch=sm_35 -nocudainc -nocudalib \
+// RUN:   /DFOO#7 /permissive- -- %s 2>&1 | FileCheck %s --check-prefix=CL-COMMON
+// CL-COMMON: "-cc1" "-triple" "nvptx64-nvidia-cuda"
+// CL-COMMON-SAME: "-D" "FOO=7"
+// CL-COMMON-NOT: "FOO#7"
+// CL-COMMON-NOT: "-fno-operator-names"
+// CL-COMMON-NOT: "-fdelayed-template-parsing"
+// CL-COMMON: "-cc1" "-triple" "x86_64-pc-windows-msvc{{[^"]*}}"
+// CL-COMMON-SAME: "-D" "FOO=7"
+// CL-COMMON-NOT: "FOO#7"
+// CL-COMMON-NOT: "-fno-operator-names"
+// CL-COMMON-NOT: "-fdelayed-template-parsing"
diff --git a/clang/test/Driver/hip-cl-argument-translation.hip b/clang/test/Driver/hip-cl-argument-translation.hip
new file mode 100644
index 00000000000000..e35201d37c2351
--- /dev/null
+++ b/clang/test/Driver/hip-cl-argument-translation.hip
@@ -0,0 +1,107 @@
+// Shared source-language options must reach both HIP jobs.
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /DFOO#7 /permissive- -- %s 2>&1 | FileCheck %s --check-prefix=COMMON
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /permissive -- %s 2>&1 | FileCheck %s --check-prefix=PERMISSIVE
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /DFOO#1 -clang:-Xarch_host -clang:-DFOO#2 -clang:-Xarch_device -clang:-DDEV#3 -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=SCOPED
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /permissive -clang:-Xarch_device -clang:-fno-delayed-template-parsing -clang:-Xarch_device \
+// RUN:   -clang:-foperator-names -- %s 2>&1 | FileCheck %s --check-prefix=DEVICE-CONFORMING
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --no-offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /DFOO#7 /permissive- -- %s 2>&1 | FileCheck %s --check-prefix=COMMON
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --no-offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /permissive -- %s 2>&1 | FileCheck %s --check-prefix=PERMISSIVE
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --no-offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /DFOO#1 -clang:-Xarch_host -clang:-DFOO#2 -clang:-Xarch_device -clang:-DDEV#3 -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=SCOPED
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --no-offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /permissive -clang:-Xarch_device -clang:-fno-delayed-template-parsing -clang:-Xarch_device \
+// RUN:   -clang:-foperator-names -- %s 2>&1 | FileCheck %s --check-prefix=DEVICE-CONFORMING
+
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /DFOO#7 /permissive- -- %s 2>&1 | FileCheck %s --check-prefix=COMMON
+
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /permissive -- %s 2>&1 | FileCheck %s --check-prefix=PERMISSIVE
+
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /DFOO#1 -clang:-Xarch_host -clang:-DFOO#2 -clang:-Xarch_device -clang:-DDEV#3 -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=SCOPED
+
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /permissive -clang:-Xarch_device -clang:-fno-delayed-template-parsing -clang:-Xarch_device \
+// RUN:   -clang:-foperator-names -- %s 2>&1 | FileCheck %s --check-prefix=DEVICE-CONFORMING
+
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c --no-offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /DFOO#7 /permissive- -- %s 2>&1 | FileCheck %s --check-prefix=COMMON
+
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c --no-offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /permissive -- %s 2>&1 | FileCheck %s --check-prefix=PERMISSIVE
+
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c --no-offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /DFOO#1 -clang:-Xarch_host -clang:-DFOO#2 -clang:-Xarch_device -clang:-DDEV#3 -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=SCOPED
+
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c --no-offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
+// RUN:   /permissive -clang:-Xarch_device -clang:-fno-delayed-template-parsing -clang:-Xarch_device \
+// RUN:   -clang:-foperator-names -- %s 2>&1 | FileCheck %s --check-prefix=DEVICE-CONFORMING
+
+// COMMON: "-cc1" "-triple" "amdgpu11.00-amd-amdhsa"
+// COMMON-SAME: "-D" "FOO=7"
+// COMMON-NOT: "-fdelayed-template-parsing"
+// COMMON-NOT: "-fno-operator-names"
+// COMMON: "-cc1" "-triple" "x86_64-{{[^"]*}}"
+// COMMON-SAME: "-D" "FOO=7"
+// COMMON-NOT: "-fdelayed-template-parsing"
+// COMMON-NOT: "-fno-operator-names"
+
+// PERMISSIVE: "-cc1" "-triple" "amdgpu11.00-amd-amdhsa"
+// PERMISSIVE-SAME: "-fno-operator-names"
+// PERMISSIVE-SAME: "-fdelayed-template-parsing"
+// PERMISSIVE: "-cc1" "-triple" "x86_64-{{[^"]*}}"
+// PERMISSIVE-SAME: "-fno-operator-names"
+// PERMISSIVE-SAME: "-fdelayed-template-parsing"
+
+// SCOPED: "-cc1" "-triple" "amdgpu11.00-amd-amdhsa"
+// SCOPED-SAME: "-D" "FOO=1" "-D" "DEV=3"
+// SCOPED-NOT: "FOO=2"
+// SCOPED: "-cc1" "-triple" "x86_64-{{[^"]*}}"
+// SCOPED-SAME: "-D" "FOO=1" "-D" "FOO=2"
+// SCOPED-NOT: "DEV=3"
+
+// DEVICE-CONFORMING: "-cc1" "-triple" "amdgpu11.00-amd-amdhsa"
+// DEVICE-CONFORMING-NOT: "-fno-operator-names"
+// DEVICE-CONFORMING-NOT: "-fdelayed-template-parsing"
+// DEVICE-CONFORMING: "-cc1" "-triple" "x86_64-{{[^"]*}}"
+// DEVICE-CONFORMING-SAME: "-fno-operator-names"
+// DEVICE-CONFORMING-SAME: "-fdelayed-template-parsing"
+
+// GNU-mode offloading retains MSVC macro syntax when selected by the host.
+// RUN: %clang --target=x86_64-pc-windows-msvc -### -c --offload-new-driver \
+// RUN:   -x hip --offload-arch=gfx1100 -nogpuinc -nogpulib \
+// RUN:   -DFOO#1 -Xarch_host -DFOO#2 -Xarch_device -DDEV#3 -- %s 2>&1 | FileCheck %s --check-prefix=SCOPED
diff --git a/clang/test/Driver/openmp-cl-argument-translation.cpp b/clang/test/Driver/openmp-cl-argument-translation.cpp
new file mode 100644
index 00000000000000..583db4718c9247
--- /dev/null
+++ b/clang/test/Driver/openmp-cl-argument-translation.cpp
@@ -0,0 +1,111 @@
+// OpenMP target arguments are parsed after the shared input translation.
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c \
+// RUN:   -clang:-fopenmp -clang:-fopenmp-targets=amdgcn-amd-amdhsa \
+// RUN:   -clang:--offload-arch=gfx1100 -clang:-nogpulib \
+// RUN:   -clang:-Xopenmp-target=amdgcn-amd-amdhsa -clang:-DFOO#7 -- %s 2>&1 | FileCheck %s
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c \
+// RUN:   -clang:-fopenmp -clang:-fopenmp-targets=amdgcn-amd-amdhsa \
+// RUN:   -clang:--offload-arch=gfx1100 -clang:-nogpulib \
+// RUN:   -clang:-Xopenmp-target=amdgcn-amd-amdhsa -clang:-DFOO#7 -- %s 2>&1 | FileCheck %s
+// RUN: %clang --target=x86_64-windows-msvc -### -c \
+// RUN:   -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa \
+// RUN:   --offload-arch=gfx1100 -nogpulib \
+// RUN:   -Xopenmp-target=amdgcn-amd-amdhsa -DFOO#7 -- %s 2>&1 | FileCheck %s
+// CHECK: "-cc1" "-triple" "x86_64-{{[^"]*}}"
+// CHECK-NOT: "FOO=7"
+// CHECK: "-cc1" "-triple" "amdgpu11.00-amd-amdhsa"
+// CHECK-SAME: "-D" "FOO=7"
+// CHECK: "-cc1" "-triple" "x86_64-{{[^"]*}}"
+// CHECK-NOT: "FOO=7"
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c \
+// RUN:   -clang:-fopenmp -clang:-fopenmp-targets=x86_64-pc-windows-msvc \
+// RUN:   -clang:-nogpulib \
+// RUN:   -clang:-Xopenmp-target=x86_64-pc-windows-msvc -clang:/O2 \
+// RUN:   -clang:-Xopenmp-target=x86_64-pc-windows-msvc -clang:/permissive -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=DEVICE-O3
+
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c \
+// RUN:   -clang:-fopenmp -clang:-fopenmp-targets=amdgcn-amd-amdhsa \
+// RUN:   -clang:--offload-arch=gfx1100 -clang:-nogpulib \
+// RUN:   -clang:-Xopenmp-target=amdgcn-amd-amdhsa -clang:/O2 \
+// RUN:   -clang:-Xopenmp-target=amdgcn-amd-amdhsa -clang:/permissive -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=DEVICE-O3
+
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c \
+// RUN:   -clang:-fopenmp -clang:-fopenmp-targets=amdgcn-amd-amdhsa \
+// RUN:   -clang:--offload-arch=gfx1100 -clang:-nogpulib \
+// RUN:   -clang:-Xopenmp-target=amdgcn-amd-amdhsa -clang:/O2 \
+// RUN:   -clang:-Xopenmp-target=amdgcn-amd-amdhsa -clang:/permissive -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=DEVICE-O3
+
+// DEVICE-O3: "-cc1" "-triple" "x86_64-{{[^"]*}}"
+// DEVICE-O3-NOT: "-O3"
+// DEVICE-O3-NOT: "-fno-operator-names"
+// DEVICE-O3: "-cc1" "-triple" "{{[^"]*}}"
+// DEVICE-O3-SAME: "-O3"
+// DEVICE-O3-SAME: "-fno-operator-names"
+// DEVICE-O3-SAME: "-fdelayed-template-parsing"
+// DEVICE-O3: "-cc1" "-triple" "x86_64-{{[^"]*}}"
+// DEVICE-O3-NOT: "-O3"
+// DEVICE-O3-NOT: "-fno-operator-names"
+// DEVICE-O3-NOT: argument unused during compilation
+
+// A forwarded /Od must remove the other flags implied by shared /O2.
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c /O2 /permissive \
+// RUN:   -clang:-fopenmp -clang:-fopenmp-targets=amdgcn-amd-amdhsa \
+// RUN:   -clang:--offload-arch=gfx1100 -clang:-nogpulib \
+// RUN:   -clang:-Xopenmp-target=amdgcn-amd-amdhsa -clang:/Od \
+// RUN:   -clang:-Xopenmp-target=amdgcn-amd-amdhsa -clang:/permissive- -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=DEVICE-OD
+// DEVICE-OD: "-cc1" "-triple" "x86_64-pc-windows-msvc{{[^"]*}}"
+// DEVICE-OD-SAME: "-O3"
+// DEVICE-OD-SAME: "-fno-operator-names"
+// DEVICE-OD: "-cc1" "-triple" "amdgpu11.00-amd-amdhsa"
+// DEVICE-OD-NOT: "-fbuiltin"
+// DEVICE-OD-NOT: "-ffunction-sections"
+// DEVICE-OD: "-O0"
+// DEVICE-OD-NOT: "-O3"
+// DEVICE-OD-NOT: "-fno-operator-names"
+// DEVICE-OD-NOT: "-fdelayed-template-parsing"
+// DEVICE-OD: "-cc1" "-triple" "x86_64-pc-windows-msvc{{[^"]*}}"
+
+// Repeated forwarded bundles use the last expandable option; explicit
+// suboptions and canonical options still retain their parsed-list order.
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c \
+// RUN:   -clang:-fopenmp -clang:-fopenmp-targets=amdgcn-amd-amdhsa \
+// RUN:   -clang:--offload-arch=gfx1100 -clang:-nogpulib \
+// RUN:   -clang:-Xopenmp-target=amdgcn-amd-amdhsa -clang:/O2 \
+// RUN:   -clang:-Xopenmp-target=amdgcn-amd-amdhsa -clang:/O1 \
+// RUN:   -clang:-Xopenmp-target=amdgcn-amd-amdhsa -clang:/Ob0 \
+// RUN:   -clang:-Xarch_device -clang:-O1 -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=DEVICE-NATIVE
+// DEVICE-NATIVE: "-cc1" "-triple" "x86_64-pc-windows-msvc{{[^"]*}}"
+// DEVICE-NATIVE: "-cc1" "-triple" "amdgpu11.00-amd-amdhsa"
+// DEVICE-NATIVE-NOT: "-Os"
+// DEVICE-NATIVE-NOT: "-O3"
+// DEVICE-NATIVE: "-O1"
+// DEVICE-NATIVE-SAME: "-fno-inline"
+// DEVICE-NATIVE: "-cc1" "-triple" "x86_64-pc-windows-msvc{{[^"]*}}"
+
+// GNU-mode Clang targeting MSVC also accepted these late slash options.
+// RUN: %clang --target=x86_64-windows-msvc -### -c \
+// RUN:   -fopenmp -fopenmp-targets=x86_64-pc-windows-msvc -nogpulib \
+// RUN:   -Xopenmp-target=x86_64-pc-windows-msvc /O2 \
+// RUN:   -Xopenmp-target=x86_64-pc-windows-msvc /permissive -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=DEVICE-O3
+
+// Separate device architecture argument lists retain their own overrides.
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c /O2 \
+// RUN:   -clang:-fopenmp -clang:-fopenmp-targets=amdgcn-amd-amdhsa \
+// RUN:   -clang:--offload-arch=gfx1100,gfx1101 -clang:-nogpulib \
+// RUN:   -clang:-Xopenmp-target=amdgcn-amd-amdhsa -clang:/O1 \
+// RUN:   -clang:-Xarch_gfx1100 -clang:-O0 -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=MULTIARCH
+// MULTIARCH: "-cc1" "-triple" "x86_64-pc-windows-msvc{{[^"]*}}"
+// MULTIARCH-SAME: "-O3"
+// MULTIARCH: "-cc1" "-triple" "amdgpu11.00-amd-amdhsa"
+// MULTIARCH-SAME: "-O0"
+// MULTIARCH: "-cc1" "-triple" "amdgpu11.01-amd-amdhsa"
+// MULTIARCH-SAME: "-Os"
+// MULTIARCH: "-cc1" "-triple" "x86_64-pc-windows-msvc{{[^"]*}}"

>From 1551eb5afed5f37ba5575fc1bb2b1c7d905c6e29 Mon Sep 17 00:00:00 2001
From: "Yaxun (Sam) Liu" <yaxun.liu at amd.com>
Date: Sat, 3 Oct 2026 07:17:50 -0400
Subject: [PATCH 4/4] [clang] Use clang-cl mode for filtered argument
 translation

Translate filtered argument lists in clang-cl mode regardless of the
host toolchain or offload kind. Keep GNU-mode MSVC forwarded-option
compatibility in the OpenMP parser using the same ClangCLArgs helper.

Pass the resolved host triple to that parser so frame-pointer options
respect -m32 and -m64. Add host-filter, MinGW, and architecture-override
regression coverage, and let HIP tests exercise the default driver.
---
 clang/include/clang/Driver/ToolChain.h        |  3 +-
 clang/lib/Driver/Compilation.cpp              |  9 ++--
 clang/lib/Driver/ToolChain.cpp                | 17 +++++--
 clang/test/Driver/cl-argument-translation.c   | 18 +++++++
 .../Driver/hip-cl-argument-translation.hip    | 18 +++----
 clang/test/Driver/hip-cl-optimization.hip     | 20 ++++----
 .../Driver/openmp-cl-argument-translation.cpp | 48 +++++++++++++++++++
 7 files changed, 102 insertions(+), 31 deletions(-)

diff --git a/clang/include/clang/Driver/ToolChain.h b/clang/include/clang/Driver/ToolChain.h
index 7304ddc43b63f7..b01dbb18c483cf 100644
--- a/clang/include/clang/Driver/ToolChain.h
+++ b/clang/include/clang/Driver/ToolChain.h
@@ -402,8 +402,9 @@ class ToolChain {
   /// TranslateOpenMPTargetArgs - Create a new derived argument list for
   /// that contains the OpenMP target specific flags passed via
   /// -Xopenmp-target -opt=val OR -Xopenmp-target=<triple> -opt=val
+  /// HostTriple is the resolved host target, including architecture overrides.
   virtual llvm::opt::DerivedArgList *TranslateOpenMPTargetArgs(
-      const llvm::opt::DerivedArgList &Args, bool SameTripleAsHost,
+      const llvm::opt::DerivedArgList &Args, const llvm::Triple &HostTriple,
       SmallVectorImpl<llvm::opt::Arg *> &AllocatedArgs) const;
 
   /// Append the argument following \p A to \p DAL assuming \p A is an Xarch
diff --git a/clang/lib/Driver/Compilation.cpp b/clang/lib/Driver/Compilation.cpp
index 3ccfe60c520d92..c7ddf07bedf959 100644
--- a/clang/lib/Driver/Compilation.cpp
+++ b/clang/lib/Driver/Compilation.cpp
@@ -74,9 +74,8 @@ Compilation::getArgsForToolChain(const ToolChain *TC, BoundArch BA,
     // Translate OpenMP toolchain arguments provided via the -Xopenmp-target flags.
     if (DeviceOffloadKind == Action::OFK_OpenMP) {
       const ToolChain *HostTC = getSingleOffloadToolChain<Action::OFK_Host>();
-      bool SameTripleAsHost = (TC->getTriple() == HostTC->getTriple());
       OpenMPArgs = TC->TranslateOpenMPTargetArgs(
-          *TranslatedArgs, SameTripleAsHost, AllocatedArgs);
+          *TranslatedArgs, HostTC->getTriple(), AllocatedArgs);
     }
 
     DerivedArgList *NewDAL = nullptr;
@@ -92,10 +91,8 @@ Compilation::getArgsForToolChain(const ToolChain *TC, BoundArch BA,
         delete OpenMPArgs;
     }
 
-    if (NewDAL && DeviceOffloadKind == Action::OFK_OpenMP &&
-        (TheDriver.IsCLMode() ||
-         DefaultToolChain.getTriple().isWindowsMSVCEnvironment())) {
-      // OpenMP forwarding can introduce clang-cl options after the shared
+    if (NewDAL && TheDriver.IsCLMode()) {
+      // Forwarding can introduce clang-cl options after the shared
       // translation. Keep synthesized arguments in the compilation-owned list.
       auto *DAL = ClangCLArgs::translateArgs(
           *NewDAL, DefaultToolChain.getTriple(), *TranslatedArgs);
diff --git a/clang/lib/Driver/ToolChain.cpp b/clang/lib/Driver/ToolChain.cpp
index 2d7dc6e63534b2..242f75af52d7b0 100644
--- a/clang/lib/Driver/ToolChain.cpp
+++ b/clang/lib/Driver/ToolChain.cpp
@@ -2009,11 +2009,12 @@ ToolChain::computeMSVCVersion(const Driver *D,
 }
 
 llvm::opt::DerivedArgList *ToolChain::TranslateOpenMPTargetArgs(
-    const llvm::opt::DerivedArgList &Args, bool SameTripleAsHost,
+    const llvm::opt::DerivedArgList &Args, const llvm::Triple &HostTriple,
     SmallVectorImpl<llvm::opt::Arg *> &AllocatedArgs) const {
   DerivedArgList *DAL = new DerivedArgList(Args.getBaseArgs());
   const OptTable &Opts = getDriver().getOpts();
   bool Modified = false;
+  bool SameTripleAsHost = getTriple() == HostTriple;
 
   // Handle -Xopenmp-target flags
   for (auto *A : Args) {
@@ -2076,9 +2077,7 @@ llvm::opt::DerivedArgList *ToolChain::TranslateOpenMPTargetArgs(
     XOpenMPTargetArg->setBaseArg(A);
     A = XOpenMPTargetArg.release();
     if (A->getOption().matches(options::OPT_D) &&
-        (getDriver().IsCLMode() ||
-         llvm::Triple(llvm::Triple::normalize(getDriver().getTargetTriple()))
-             .isWindowsMSVCEnvironment()))
+        (getDriver().IsCLMode() || HostTriple.isWindowsMSVCEnvironment()))
       A->getValues()[0] =
           ClangCLArgs::translateMacroDefinition(A->getValue(), Args);
     AllocatedArgs.push_back(A);
@@ -2086,8 +2085,16 @@ llvm::opt::DerivedArgList *ToolChain::TranslateOpenMPTargetArgs(
     Modified = true;
   }
 
-  if (Modified)
+  if (Modified) {
+    // GNU-mode Clang targeting MSVC also accepts forwarded clang-cl options.
+    if (!getDriver().IsCLMode() && HostTriple.isWindowsMSVCEnvironment()) {
+      DerivedArgList *Translated =
+          ClangCLArgs::translateArgs(*DAL, HostTriple, Args);
+      delete DAL;
+      return Translated;
+    }
     return DAL;
+  }
 
   delete DAL;
   return nullptr;
diff --git a/clang/test/Driver/cl-argument-translation.c b/clang/test/Driver/cl-argument-translation.c
index e1758265fa8890..a297ff88b7e2df 100644
--- a/clang/test/Driver/cl-argument-translation.c
+++ b/clang/test/Driver/cl-argument-translation.c
@@ -13,6 +13,24 @@
 // MSVC: "-D" "FOO=7"
 // GNU: "-D" "FOO#7"
 
+// Filtering host arguments preserves explicit overrides of shared /O options.
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc /c -### /O2 /Ob0 \
+// RUN:   -clang:-Xarch_host -clang:-O1 -clang:-Xarch_host -clang:-fno-builtin \
+// RUN:   -- %s 2>&1 | FileCheck %s --check-prefix=FILTERED
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu /c -### /O2 /Ob0 \
+// RUN:   -clang:-Xarch_host -clang:-O1 -clang:-Xarch_host -clang:-fno-builtin \
+// RUN:   -- %s 2>&1 | FileCheck %s --check-prefix=FILTERED
+// RUN: %clang_cl --target=x86_64-w64-windows-gnu /c -### /O2 /Ob0 \
+// RUN:   -clang:-Xarch_host -clang:-O1 -clang:-Xarch_host -clang:-fno-builtin \
+// RUN:   -- %s 2>&1 | FileCheck %s --check-prefix=FILTERED
+// FILTERED: "-cc1"
+// FILTERED-NOT: "-O3"
+// FILTERED: "-O1"
+// FILTERED-NOT: "-O3"
+// FILTERED: "-fno-builtin"
+// FILTERED-SAME: "-fno-inline"
+// FILTERED-NOT: argument unused during compilation
+
 // Last-option precedence must include both slash and canonical options.
 // RUN: %clang_cl --target=x86_64-pc-windows-msvc /c -x c++ -### /permissive- /permissive -- %s 2>&1 | \
 // RUN:   FileCheck %s --check-prefix=PERMISSIVE
diff --git a/clang/test/Driver/hip-cl-argument-translation.hip b/clang/test/Driver/hip-cl-argument-translation.hip
index e35201d37c2351..241d265cce3eca 100644
--- a/clang/test/Driver/hip-cl-argument-translation.hip
+++ b/clang/test/Driver/hip-cl-argument-translation.hip
@@ -1,18 +1,18 @@
 // Shared source-language options must reach both HIP jobs.
-// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /DFOO#7 /permissive- -- %s 2>&1 | FileCheck %s --check-prefix=COMMON
 
-// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /permissive -- %s 2>&1 | FileCheck %s --check-prefix=PERMISSIVE
 
-// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /DFOO#1 -clang:-Xarch_host -clang:-DFOO#2 -clang:-Xarch_device -clang:-DDEV#3 -- %s 2>&1 | \
 // RUN:   FileCheck %s --check-prefix=SCOPED
 
-// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /permissive -clang:-Xarch_device -clang:-fno-delayed-template-parsing -clang:-Xarch_device \
 // RUN:   -clang:-foperator-names -- %s 2>&1 | FileCheck %s --check-prefix=DEVICE-CONFORMING
@@ -35,20 +35,20 @@
 // RUN:   /permissive -clang:-Xarch_device -clang:-fno-delayed-template-parsing -clang:-Xarch_device \
 // RUN:   -clang:-foperator-names -- %s 2>&1 | FileCheck %s --check-prefix=DEVICE-CONFORMING
 
-// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c --offload-new-driver \
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /DFOO#7 /permissive- -- %s 2>&1 | FileCheck %s --check-prefix=COMMON
 
-// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c --offload-new-driver \
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /permissive -- %s 2>&1 | FileCheck %s --check-prefix=PERMISSIVE
 
-// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c --offload-new-driver \
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /DFOO#1 -clang:-Xarch_host -clang:-DFOO#2 -clang:-Xarch_device -clang:-DDEV#3 -- %s 2>&1 | \
 // RUN:   FileCheck %s --check-prefix=SCOPED
 
-// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c --offload-new-driver \
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /permissive -clang:-Xarch_device -clang:-fno-delayed-template-parsing -clang:-Xarch_device \
 // RUN:   -clang:-foperator-names -- %s 2>&1 | FileCheck %s --check-prefix=DEVICE-CONFORMING
@@ -102,6 +102,6 @@
 // DEVICE-CONFORMING-SAME: "-fdelayed-template-parsing"
 
 // GNU-mode offloading retains MSVC macro syntax when selected by the host.
-// RUN: %clang --target=x86_64-pc-windows-msvc -### -c --offload-new-driver \
+// RUN: %clang --target=x86_64-pc-windows-msvc -### -c \
 // RUN:   -x hip --offload-arch=gfx1100 -nogpuinc -nogpulib \
 // RUN:   -DFOO#1 -Xarch_host -DFOO#2 -Xarch_device -DDEV#3 -- %s 2>&1 | FileCheck %s --check-prefix=SCOPED
diff --git a/clang/test/Driver/hip-cl-optimization.hip b/clang/test/Driver/hip-cl-optimization.hip
index 1fa80de854e566..0dc4cdd97e5924 100644
--- a/clang/test/Driver/hip-cl-optimization.hip
+++ b/clang/test/Driver/hip-cl-optimization.hip
@@ -1,43 +1,43 @@
 // Check that clang-cl optimization options apply to both host and device.
 // Device-specific options must still override the shared optimization level.
 
-// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /O1 -- %s 2>&1 | FileCheck %s --check-prefix=OS
 
-// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c --offload-new-driver \
+// RUN: %clang_cl --target=x86_64-unknown-linux-gnu -### /c \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /O1 -- %s 2>&1 | FileCheck %s --check-prefix=LINUX
 
-// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /O2 -- %s 2>&1 | FileCheck %s --check-prefix=O3
 
-// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /Ox -- %s 2>&1 | FileCheck %s --check-prefix=O3
 
-// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /Od -- %s 2>&1 | FileCheck %s --check-prefix=O0
 
-// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /O2 /Od -- %s 2>&1 | FileCheck %s --check-prefix=O0
 
-// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /Od /O2 -- %s 2>&1 | FileCheck %s --check-prefix=O3
 
-// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /O2 -clang:-O1 -- %s 2>&1 | FileCheck %s --check-prefix=O1
 
-// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /O2 -clang:-Xarch_device -clang:-O1 -- %s 2>&1 | FileCheck %s --check-prefix=DEVICE-O1
 
-// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c --offload-new-driver \
+// RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c \
 // RUN:   -x hip --offload-arch=gfx1100 -clang:-nogpuinc -clang:-nogpulib \
 // RUN:   /O2 /Ob0 -- %s 2>&1 | FileCheck %s --check-prefix=NOINLINE
 
diff --git a/clang/test/Driver/openmp-cl-argument-translation.cpp b/clang/test/Driver/openmp-cl-argument-translation.cpp
index 583db4718c9247..5f357472377553 100644
--- a/clang/test/Driver/openmp-cl-argument-translation.cpp
+++ b/clang/test/Driver/openmp-cl-argument-translation.cpp
@@ -39,6 +39,13 @@
 // RUN:   -clang:-Xopenmp-target=amdgcn-amd-amdhsa -clang:/permissive -- %s 2>&1 | \
 // RUN:   FileCheck %s --check-prefix=DEVICE-O3
 
+// RUN: %clang_cl --target=x86_64-w64-windows-gnu -### /c \
+// RUN:   -clang:-fopenmp -clang:-fopenmp-targets=amdgcn-amd-amdhsa \
+// RUN:   -clang:--offload-arch=gfx1100 -clang:-nogpulib \
+// RUN:   -clang:-Xopenmp-target=amdgcn-amd-amdhsa -clang:/O2 \
+// RUN:   -clang:-Xopenmp-target=amdgcn-amd-amdhsa -clang:/permissive -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=DEVICE-O3
+
 // DEVICE-O3: "-cc1" "-triple" "x86_64-{{[^"]*}}"
 // DEVICE-O3-NOT: "-O3"
 // DEVICE-O3-NOT: "-fno-operator-names"
@@ -95,6 +102,47 @@
 // RUN:   -Xopenmp-target=x86_64-pc-windows-msvc /permissive -- %s 2>&1 | \
 // RUN:   FileCheck %s --check-prefix=DEVICE-O3
 
+// Forwarded frame-pointer options use the host architecture after -m32/-m64.
+// /Oy- is silently accepted on x86-64, including when bundled with /O2.
+// RUN: %clang --target=i686-pc-windows-msvc -m64 -### -c -fopenmp \
+// RUN:   -fopenmp-targets=x86_64-pc-windows-msvc -nogpulib \
+// RUN:   -Xopenmp-target=x86_64-pc-windows-msvc /Oy- -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=FRAME64
+// RUN: %clang --target=i686-pc-windows-msvc -m64 -### -c -fopenmp \
+// RUN:   -fopenmp-targets=x86_64-pc-windows-msvc -nogpulib \
+// RUN:   -Xopenmp-target=x86_64-pc-windows-msvc /O2y- -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=FRAME64
+// FRAME64: "-cc1" "-triple" "x86_64-pc-windows-msvc{{[^"]*}}"
+// FRAME64: "-cc1" "-triple" "x86_64-pc-windows-msvc{{[^"]*}}"
+// FRAME64-SAME: "-mframe-pointer=none"
+
+// RUN: %clang --target=x86_64-pc-windows-msvc -m32 -### -c -fopenmp \
+// RUN:   -fopenmp-targets=i386-pc-windows-msvc -nogpulib \
+// RUN:   -Xopenmp-target=i386-pc-windows-msvc /Oy -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=FRAME32-OMIT
+// FRAME32-OMIT: "-cc1" "-triple" "i386-pc-windows-msvc{{[^"]*}}"
+// FRAME32-OMIT: "-cc1" "-triple" "i386-pc-windows-msvc{{[^"]*}}"
+// FRAME32-OMIT-SAME: "-mframe-pointer=none"
+
+// RUN: %clang --target=x86_64-pc-windows-msvc -m32 -### -c -fopenmp \
+// RUN:   -fopenmp-targets=i386-pc-windows-msvc -nogpulib \
+// RUN:   -Xopenmp-target=i386-pc-windows-msvc /O2y- -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=FRAME32-KEEP
+// FRAME32-KEEP: "-cc1" "-triple" "i386-pc-windows-msvc{{[^"]*}}"
+// FRAME32-KEEP: "-cc1" "-triple" "i386-pc-windows-msvc{{[^"]*}}"
+// FRAME32-KEEP-SAME: "-O3"
+// FRAME32-KEEP-SAME: "-mframe-pointer=all"
+
+// A different device architecture does not change the host's /Oy- policy.
+// RUN: %clang --target=i686-pc-windows-msvc -m64 -### -c -fopenmp \
+// RUN:   -fopenmp-targets=i386-pc-windows-msvc -nogpulib \
+// RUN:   -Xopenmp-target=i386-pc-windows-msvc /O2y- -- %s 2>&1 | \
+// RUN:   FileCheck %s --check-prefix=HOST64-DEVICE32
+// HOST64-DEVICE32: "-cc1" "-triple" "x86_64-pc-windows-msvc{{[^"]*}}"
+// HOST64-DEVICE32: "-cc1" "-triple" "i386-pc-windows-msvc{{[^"]*}}"
+// HOST64-DEVICE32-SAME: "-O3"
+// HOST64-DEVICE32-SAME: "-mframe-pointer=none"
+
 // Separate device architecture argument lists retain their own overrides.
 // RUN: %clang_cl --target=x86_64-pc-windows-msvc -### /c /O2 \
 // RUN:   -clang:-fopenmp -clang:-fopenmp-targets=amdgcn-amd-amdhsa \



More information about the cfe-commits mailing list