[DRAFT][RISCV] Emit arch string macro to facilitate ASM programming

[nemanjai]

In order to make assembly programming more convenient, emit macro __riscv_cmdline_arch_string that will be defined to the arch string based on the command line arguments. This string may differ from the actual string that is added to the object or assembly file. This provides a convenient mechanism for the programmer to add the .attribute directive into pre-processed asm files.

Example file.S:
.attribute 5, __riscv_cmdline_arch_string

[llvmbot]

@llvm/pr-subscribers-clang

@llvm/pr-subscribers-backend-risc-v

Author: Nemanja Ivanovic (nemanjai)

Changes

In order to make assembly programming more convenient, emit macro __riscv_cmdline_arch_string that will be defined to the arch string based on the command line arguments. This string may differ from the actual string that is added to the object or assembly file. This provides a convenient mechanism for the programmer to add the .attribute directive into pre-processed asm files.

Example file.S:
.attribute 5, __riscv_cmdline_arch_string

---

Full diff: https://github.com/llvm/llvm-project/pull/85063.diff

2 Files Affected:

• (modified) clang/lib/Basic/Targets/RISCV.cpp (+2)
• (modified) clang/test/Preprocessor/riscv-target-features.c (+8)

diff --git a/clang/lib/Basic/Targets/RISCV.cpp b/clang/lib/Basic/Targets/RISCV.cpp
index a6d4af2b88111a..718d94f2d2621e 100644
--- a/clang/lib/Basic/Targets/RISCV.cpp
+++ b/clang/lib/Basic/Targets/RISCV.cpp
@@ -139,6 +139,7 @@ void RISCVTargetInfo::getTargetDefines(const LangOptions &Opts,
   unsigned MinVLen = ISAInfo->getMinVLen();
   unsigned MaxELen = ISAInfo->getMaxELen();
   unsigned MaxELenFp = ISAInfo->getMaxELenFp();
+  std::string ArchString = "\"" + ISAInfo->toString() + "\"";
   if (CodeModel == "default")
     CodeModel = "small";
 
@@ -222,6 +223,7 @@ void RISCVTargetInfo::getTargetDefines(const LangOptions &Opts,
     else
       Builder.defineMacro("__riscv_32e");
   }
+  Builder.defineMacro("__riscv_cmdline_arch_string", ArchString);
 }
 
 static constexpr Builtin::Info BuiltinInfo[] = {
diff --git a/clang/test/Preprocessor/riscv-target-features.c b/clang/test/Preprocessor/riscv-target-features.c
index 1a15be1c6e4dc1..6023abe2665c2a 100644
--- a/clang/test/Preprocessor/riscv-target-features.c
+++ b/clang/test/Preprocessor/riscv-target-features.c
@@ -185,8 +185,10 @@
 // RUN: %clang --target=riscv64-unknown-linux-gnu \
 // RUN:   -march=rv64ia -E -dM %s \
 // RUN:   -o - | FileCheck --check-prefix=CHECK-A-EXT %s
+// CHECK-A-EXT: __INTPTR_WIDTH__ [[WIDTH:[1-9]+]]
 // CHECK-A-EXT: __riscv_a 2001000{{$}}
 // CHECK-A-EXT: __riscv_atomic 1
+// CHECK-A-EXT: __riscv_cmdline_arch_string "rv[[WIDTH]]i2p1_a2p1"
 
 // RUN: %clang --target=riscv32-unknown-linux-gnu \
 // RUN:   -march=rv32ic -E -dM %s \
@@ -203,6 +205,8 @@
 // RUN: %clang --target=riscv64-unknown-linux-gnu \
 // RUN:   -march=rv64ifd -E -dM %s \
 // RUN:   -o - | FileCheck --check-prefix=CHECK-D-EXT %s
+// CHECK-D-EXT: __INTPTR_WIDTH__ [[WIDTH:[1-9]+]]
+// CHECK-D-EXT: __riscv_cmdline_arch_string "rv[[WIDTH]]i2p1_f2p2_d2p2_zicsr2p0"
 // CHECK-D-EXT: __riscv_d 2002000{{$}}
 // CHECK-D-EXT: __riscv_fdiv 1
 // CHECK-D-EXT: __riscv_flen 64
@@ -217,6 +221,8 @@
 // CHECK-RV32E: __riscv_32e 1
 // CHECK-RV64E: __riscv_64e 1
 // CHECK-E-EXT: __riscv_abi_rve 1
+// CHECK-RV32E: __riscv_cmdline_arch_string "rv32e2p0"
+// CHECK-RV64E: __riscv_cmdline_arch_string "rv64e2p0"
 // CHECK-E-EXT: __riscv_e 2000000{{$}}
 
 // RUN: %clang --target=riscv32-unknown-linux-gnu \
@@ -225,6 +231,8 @@
 // RUN: %clang --target=riscv64-unknown-linux-gnu \
 // RUN:   -march=rv64if -E -dM %s \
 // RUN:   -o - | FileCheck --check-prefix=CHECK-F-EXT %s
+// CHECK-F-EXT: __INTPTR_WIDTH__ [[WIDTH:[1-9]+]]
+// CHECK-F-EXT: __riscv_cmdline_arch_string "rv[[WIDTH]]i2p1_f2p2_zicsr2p0"
 // CHECK-F-EXT: __riscv_f 2002000{{$}}
 // CHECK-F-EXT: __riscv_fdiv 1
 // CHECK-F-EXT: __riscv_flen 32

[nemanjai]

This is a suggestion born out of a request from a user to have a way to emit the ISA attributes into objects produced from pre-processed asm files. Perhaps there is already a method to do this that I'm not aware of, but if not, I think this is a convenient and lightweight way to do so.

[wangpc-pp]

I think we will add attributes automatically?

~/workspace# cat a.S
        .globl  foo 
        .p2align        1
        .type   foo,@function
foo:
   ret
~/workspace# clang -march=rv64gcv -c a.S
~/workspace# llvm-readobj -A a.o
File: a.o
Format: elf64-littleriscv
Arch: riscv64
AddressSize: 64bit
LoadName: <Not found>
BuildAttributes {
  FormatVersion: 0x41
  Section 1 {
    SectionLength: 157
    Vendor: riscv
    Tag: Tag_File (0x1)
    Size: 147
    FileAttributes {
      Attribute {
        Tag: 5
        TagName: arch
        Value: rv64i2p1_m2p0_a2p1_f2p2_d2p2_c2p0_v1p0_zicsr2p0_zifencei2p0_zve32f1p0_zve32x1p0_zve64d1p0_zve64f1p0_zve64x1p0_zvl128b1p0_zvl32b1p0_zvl64b1p0                                                                            
      }
    }
  }
}

[jrtc27]

Attributes are added automatically, unless you override it with your own .attribute arch.

[nemanjai]

Ah, I missed the fact that there is an option -riscv-add-build-attributes that the clang driver passes when invoking cc1as. Perhaps that option should default to true? Thanks and sorry for the noise.