[clang][CodeGen] Introduce `-frecord-command-line` for MachO

[antoniofrighetto]

Summary

Allow clang driver command-line recording when targeting MachO object files as well.

[sgraenitz]

That's an excellent patch! I didn't expect it could be so compact. Thanks for working on this! Please find below 2 inline comments on details.
On July 25th release/17.x will branch away. It would be great to land this before and have it in the upcoming release.

llvm/lib/CodeGen/TargetLoweringObjectFileImpl.cpp

1424

Can we put it in __DATA?

Also the ELF implementation notes that it "attempts to mimic GCC's switch of the same name" and thus calls the section .GCC.command.line. I guess we cannot use the dots in MachO, but should we add a gcc prefix?

llvm/test/CodeGen/AArch64/arm64-command-line-metadata.ll

1 ↗

(On Diff #542041)

The triple doesn't match the name of the test: arm64-command-line-metadata.ll

Why not rename the file to commandline-metadata.ll (as in llvm/test/CodeGen/X86/commandline-metadata.ll) and maybe change the triple into the more common arm64-apple?

Optionally, we could also add RUN and CHECK lines for the existing ELF feature in AArch64 if there isn't one already.

[antoniofrighetto]

Thanks a lot for the reviews and for pointing out clang tests as well, there was a minor update to do there too!

llvm/lib/CodeGen/TargetLoweringObjectFileImpl.cpp

1424

Whilst I agree it should be better to distinguish this from executable data, I think this should live as read-only data, which __TEXT is traditionally for.

Following MachO conventions, I first tried __gcc_command_line, but the section name is restricted to 16 chars, and I'm not sure adding more bytes for the name is worth the change (thus I thought we'd prefer __command_line over __gcc_cmd_line).

[sgraenitz]

Great! This LGTM. Let's keep the review open for a few days and see if there is more feedback.

llvm/lib/CodeGen/TargetLoweringObjectFileImpl.cpp

1424

[...] the section name is restricted to 16 chars, and I'm not sure adding more bytes for the name is worth the change

Ok, then I guess it's fine to keep this as is.

I think this should live as read-only data, which __TEXT is traditionally for.

Oh interesting, I had expected permissions to be R-X for everything in the text segment, but I might be wrong. And yes, the command-line string should be R-- (read-only).

So, I looked at a few similar cases in existing code and it confirms your statement. __cstring e.g. appears to in __TEXT as well even though objdump will show it as type DATA:

Idx Name             Size     VMA              Type
  0 __text           00014029 0000000000000000 TEXT
  1 __StaticInit     0000005f 0000000000014030 TEXT
  ...
  6 __data           00000040 00000000000148e0 DATA
  7 __cstring        0000084b 0000000000014920 DATA

I am probably lacking some MachO expertise to understand the details here. Since you do the same as we do for __cstring in other places, I think this is good.

clang/lib/Driver/ToolChains/Clang.cpp

   // By default, -gno-record-gcc-switches is set on and no recording.
   auto GRecordSwitches =
       Args.hasFlag(options::OPT_grecord_command_line,
                    options::OPT_gno_record_command_line, false);
   auto FRecordSwitches =
       Args.hasFlag(options::OPT_frecord_command_line,
                    options::OPT_fno_record_command_line, false);
   if (FRecordSwitches && !Triple.isOSBinFormatELF() &&
-      !Triple.isOSBinFormatXCOFF())
+      !Triple.isOSBinFormatXCOFF() && !Triple.isOSBinFormatMachO())
     D.Diag(diag::err_drv_unsupported_opt_for_target)
         << Args.getLastArg(options::OPT_frecord_command_line)->getAsString(Args)
         << TripleStr;
   if (TC.UseDwarfDebugFlags() || GRecordSwitches || FRecordSwitches) {
     ArgStringList OriginalArgs;
     for (const auto &Arg : Args)
       Arg->render(Args, OriginalArgs);
 

clang/test/Driver/clang_f_opts.c

 // RUN: %clang -### -S -target x86_64-unknown-linux -fno-record-gcc-switches %s 2>&1 | FileCheck -check-prefix=CHECK-NO-RECORD-GCC-SWITCHES %s
 // RUN: %clang -### -S -target x86_64-unknown-linux -fno-record-gcc-switches -frecord-gcc-switches %s 2>&1 | FileCheck -check-prefix=CHECK-RECORD-GCC-SWITCHES %s
 // RUN: %clang -### -S -target x86_64-unknown-linux -frecord-gcc-switches -fno-record-gcc-switches %s 2>&1 | FileCheck -check-prefix=CHECK-NO-RECORD-GCC-SWITCHES %s
 // RUN: %clang -### -S -target x86_64-unknown-linux -frecord-command-line %s 2>&1 | FileCheck -check-prefix=CHECK-RECORD-GCC-SWITCHES %s
 // RUN: %clang -### -S -target x86_64-unknown-linux -fno-record-command-line %s 2>&1 | FileCheck -check-prefix=CHECK-NO-RECORD-GCC-SWITCHES %s
 // RUN: %clang -### -S -target x86_64-unknown-linux -fno-record-command-line -frecord-command-line %s 2>&1 | FileCheck -check-prefix=CHECK-RECORD-GCC-SWITCHES %s
 // RUN: %clang -### -S -target x86_64-unknown-linux -frecord-command-line -fno-record-command-line %s 2>&1 | FileCheck -check-prefix=CHECK-NO-RECORD-GCC-SWITCHES %s
 // Test with a couple examples of non-ELF object file formats
-// RUN: %clang -### -S -target x86_64-unknown-macosx -frecord-command-line %s 2>&1 | FileCheck -check-prefix=CHECK-RECORD-GCC-SWITCHES-ERROR %s
+// RUN: %clang -### -S -target x86_64-unknown-macosx -frecord-command-line %s 2>&1 | FileCheck -check-prefix=CHECK-RECORD-GCC-SWITCHES %s
 // RUN: %clang -### -S -target x86_64-unknown-windows -frecord-command-line %s 2>&1 | FileCheck -check-prefix=CHECK-RECORD-GCC-SWITCHES-ERROR %s
 // CHECK-RECORD-GCC-SWITCHES: "-record-command-line"
 // CHECK-NO-RECORD-GCC-SWITCHES-NOT: "-record-command-line"
 // CHECK-RECORD-GCC-SWITCHES-ERROR: error: unsupported option '-frecord-command-line' for target
 // Test when clang is in a path containing a space.
 // The initial `rm` is a workaround for https://openradar.appspot.com/FB8914243
 // (Scenario: Run tests once, `clang` gets copied and run at new location and signature
 // is cached at the new clang's inode, then clang is changed, tests run again, old signature

llvm/include/llvm/CodeGen/TargetLoweringObjectFileImpl.h

public:
   const MCExpr *getIndirectSymViaGOTPCRel(const GlobalValue *GV,
                                           const MCSymbol *Sym,
                                           const MCValue &MV, int64_t Offset,
                                           MachineModuleInfo *MMI,
                                           MCStreamer &Streamer) const override;
 
   void getNameWithPrefix(SmallVectorImpl<char> &OutName, const GlobalValue *GV,
                          const TargetMachine &TM) const override;
+
+  MCSection *getSectionForCommandLines() const override;
 };
 
 class TargetLoweringObjectFileCOFF : public TargetLoweringObjectFile {
   mutable unsigned NextUniqueID = 0;
   const TargetMachine *TM = nullptr;
 
 public:
   ~TargetLoweringObjectFileCOFF() override = default;

llvm/lib/CodeGen/TargetLoweringObjectFileImpl.cpp

if (Kind.isMergeableConst4())
     return FourByteConstantSection;
   if (Kind.isMergeableConst8())
     return EightByteConstantSection;
   if (Kind.isMergeableConst16())
     return SixteenByteConstantSection;
   return ReadOnlySection;  // .const
 }
 
+MCSection *TargetLoweringObjectFileMachO::getSectionForCommandLines() const {
+  return getContext().getMachOSection("__TEXT", "__command_line", 0,
+                                      SectionKind::getReadOnly());
+}
+
 const MCExpr *TargetLoweringObjectFileMachO::getTTypeGlobalReference(
     const GlobalValue *GV, unsigned Encoding, const TargetMachine &TM,
     MachineModuleInfo *MMI, MCStreamer &Streamer) const {
   // The mach-o version of this method defaults to returning a stub reference.
 
   if (Encoding & DW_EH_PE_indirect) {
     MachineModuleInfoMachO &MachOMMI =
       MMI->getObjFileInfo<MachineModuleInfoMachO>();

llvm/test/CodeGen/AArch64/commandline-metadata.ll

+; RUN: llc -mtriple=arm64-linux-gnu < %s | FileCheck %s
+; RUN: llc -mtriple=arm64-apple-darwin < %s | FileCheck %s --check-prefix=CHECK-MACHO
+
+; Verify that llvm.commandline metadata is emitted to the corresponding command line section.
+
+; CHECK:              .text
+; CHECK:              .section .GCC.command.line,"MS",@progbits,1
+; CHECK-NEXT:         .zero 1
+; CHECK-NEXT:         .ascii "clang -command1"
+; CHECK-NEXT:         .zero 1
+; CHECK-NEXT:         .ascii "clang -command2"
+; CHECK-NEXT:         .zero 1
+
+; CHECK-MACHO:        .section  __TEXT,__text,regular,pure_instructions
+; CHECK-MACHO-NEXT:   .section  __TEXT,__command_line
+; CHECK-MACHO-NEXT:   .space  1
+; CHECK-MACHO-NEXT:   .ascii  "clang -command1"
+; CHECK-MACHO-NEXT:   .space  1
+; CHECK-MACHO-NEXT:   .ascii  "clang -command2"
+; CHECK-MACHO-NEXT:   .space  1
+
+!llvm.commandline = !{!0, !1}
+!0 = !{!"clang -command1"}
+!1 = !{!"clang -command2"}