canonical order for Z extensions

[jim-wilson]

The GNU toolchain requires that the single letter extensions come first, followed by S*, H*, Z*, and X*. However, the table 27.1 lists 4 Z extensions as coming before the S* extensions. Since the GNU toolchain does not officially support any S* or H* extensions currently, there is no problem yet. However, as soon as we add our first S* or H* extension the GNU toolchain will not be compatible with the ISA spec. There are also a number of new Z extensions expected soon, B and K add a number of them, plus Zfh, Zfinx, etc. What about those? Will they go after S and H or before? A consistent rule makes things easier for the toolchain. I would prefer that all Z extensions go next to each other to make parsing canonical order easier.

rohan:2406$ riscv32-unknown-elf-gcc -march=rv32i_sfoo_zifencei -mabi=ilp32 -S tmp.c rohan:2407$ riscv32-unknown-elf-gcc -march=rv32i_zifencei_sfoo -mabi=ilp32 -S tmp.c riscv32-unknown-elf-gcc: error: '-march=rv32i_zifencei_sfoo': unexpected ISA string at end: 'sfoo' rohan:2408$

With binutils modified to accept sfoo as an extension rohan:2007$ riscv32-unknown-elf-as -march=rv32i_sfoo_zifencei -c tmp.s rohan:2007$ riscv32-unknown-elf-as -march=rv32i_zifencei_sfoo -c tmp.s Assembler messages: Error: -march=rv32i_zifencei_sfoo: unexpected ISA string at end: sfoo rohan:2008$

So both GCC and Binutils have incorrect parsing of architecture strings if zifencei is supposed to come before S*.

Someone should check LLVM.

[Nelson1225]

I look at the llvm source here, https://llvm.org/doxygen/RISCVTargetStreamer_8cpp_source.html#l00040

In the emitTargetAttributes, it seems like llvm does not support any S* or H* extensions either. I think the orders of if statement means the orders of the extensions, so the orders of S*, H* and Z* depend on where do we want to put them in the future patches.

[Nelson1225]

I have a patch that fixes the prefixed order problem, and also support to parse the multi-letter prefixes like zxm. https://sourceware.org/pipermail/binutils/2021-April/116054.html

After applying the patch, and modifying the binutils to accept sfoo, hfoo and zxmfoo, I can get the expected result as spec,

$ riscv64-unknown-elf-as -misa-spec=20191213 -march=rv64i_zicsr_sfoo_hfoo_zxmfoo_xriscv1p0 empty.s -o empty.o $ riscv64-unknown-elf-readelf -A empty.o Attribute Section: riscv File Attributes Tag_RISCV_arch: "rv64i2p1_zicsr2p0_sfoo1p0_hfoo1p0_zxmfoo1p0_xriscv1p0"

$ riscv64-unknown-elf-as -march=rv64i_sfoo_hfoo_zifencei empty.s -o empty.o Assembler messages: Error: -march=rv64i_sfoo_hfoo_zifencei: prefixed ISA extension zifencei' is not in expected order. It must come before hfoo'

We used to get "unexpected ISA string at end: sfoo" error for the above case, but now we can give the more friendly errors to users. Though it may be better to report "It must come before sfoo" rather than the hfoo. I think it can be improved in the future patches.