Define -march string #26

cmuellner · 2022-11-08T13:30:26Z

This is essentially a collection of all recent change requests for the -march string:

Define the -march string, as it is not exactly the RISC-V ISA naming string, and highlight the lower-case requirement (https://gcc.gnu.org/pipermail/gcc-patches/2022-November/605249.html)
Relax the ISA string order ([RFC] Relax the canonical order checking of arch string for -march option #11 and Relax the ISA string order checking for -march #14)
Add custom extensions (Specifying custom extensions in the -march string #1)
Add profiles support (https://lists.riscv.org/g/sig-toolchains/message/379 and Naming: Adding note to clarify compatiblity with ISA naming strings riscv/riscv-profiles#78)

Most of this patch is based on proposals and inputs from Kito and Palmer (see linked resources above).

It might not be perfect, but it could serve as a reasonable initial draft for further discussions.

Comments are welcome!

CC: @palmer-dabbelt @kito-cheng @asb @jrtc27

cmuellner · 2022-11-09T09:54:42Z

@palmer-dabbelt commented the following on the GCC mailing list:

IMO trying to handle this with another RISC-V spec is a waste of time:
we've spent many years trying to follow the specs here, it's pretty
clear they're just not meant to be read in that level of detail. This
sort of problem is all over the place in RISC-V land, moving to a
different spec doesn't fix the problem.

As I wrote in my reply message, I think -march should be documented here (as it is already) because this repo is the common place to document things like that for GCC and LLVM.

Also, I don't think a common -march documentation/specification is a waste of time. But of course, I don't consider my proposal perfect. Technical comments will help to improve it. If someone else wants to drive this issue instead, feel free to do so.

kito-cheng · 2022-11-10T01:59:21Z

README.mkd

+* The next part consists of single-letter ISA extensions, which may be in
+  non-canonical order.
+* The next part consists of multi-letter ISA extensions, which may be in
+  non-canonical order.


I would like to relax the order between single-letter / multi-letter ext, see #14

Changes: 2021/08/16: Remove order constraint between single-letter and multi-letter, also add 2 more NOTE.

kito-cheng · 2022-11-10T01:59:32Z

README.mkd

+* `rv64gc_xtheadba`: Valid `-march` string.
+* `rv64gc_Zicsr`: Invalid because of the uppercase character.
+* `rv32mai`: Invalid because base ISA does not follow `rv32`.
+* `rv32i_zicsr_m`: Invalid because `m` after multi-letter extension.


So this will become valid.

anton-afanasyev · 2022-11-21T18:46:16Z

README.mkd

+
+The following interpretation differences exist:
+
+* The expansion of `g` with versions before `20190609` is `imafd`.


typo (here and below): 20190609 -> 20190608

anton-afanasyev · 2022-11-21T18:54:56Z

README.mkd

+* `rv32i_zicsrzifencei`: Invalid because multi-letter extensions need to be
+   separated by an underscore (`zicsr` and `zifencei` are existing individual
+   extensions).
+


Last specification defines "dependency" notion for extensions. We need to either mention that -march= respects this convention, or state that -march= doesn't follow it.
See 29.3 of 20191213: Some ISA extensions depend on the presence of other extensions, e.g., "D" depends on "F" and "F" depends on "Zicsr". These dependences may be implicit in the ISA name: for example, RV32IF is equivalent to RV32IFZicsr, and RV32ID is equivalent to RV32IFD and RV32IFDZicsr.

Thanks for highlighting this.
I think it is expected by users that the march string follows this implicit enablement of extensions.
I will add a sentence to clarify this.

eopXD · 2022-11-22T08:23:53Z

Thank you for the work.

Trying to understand how the interpretation works so maybe I can help on the LLVM side. Is there an existing table for interpretation(s) to the ISA? If so this will help me clarify the confusion that I have to both "dates" and "version number" coexists (based on this PR - 2.2 and 20190608) as an identifier to an interpretation.

cmuellner · 2022-12-01T14:57:49Z

Trying to understand how the interpretation works so maybe I can help on the LLVM side. Is there an existing table for interpretation(s) to the ISA? If so this will help me clarify the confusion that I have to both "dates" and "version number" coexists (based on this PR - 2.2 and 20190608) as an identifier to an interpretation.

I tried to cover exactly this by the amount of examples included in the PR.
If there are unclear cases, then I think we should list them in this document anyway.

Regarding the ISA versions: we will just have to support all existing versioning schemes: so versions are well-ordered, but not lexicographic sortable, strings.

cmuellner · 2022-12-01T14:58:26Z

I've updated the PR according to the provided feedback.
Thanks for the comments!

This is essentially a collection of all recent change requests for the -march string: * Relax the ISA string order ([1]) * Add custom extensions ([2]) * Add profiles support ([3]) Most of this patch is based on proposals from Kito Cheng <[email protected]> (see linked resources below). [1] riscv-non-isa#14 [2] riscv-non-isa#1 [3] https://lists.riscv.org/g/sig-toolchains/message/379 [4] riscv-non-isa#11 Signed-off-by: Christoph Müllner <[email protected]>

cmuellner · 2022-12-01T17:44:38Z

Kito mentioned that there should be more details regarding the versioning of extensions. I've added some details about that to the PR.

jrtc27 · 2022-12-01T17:57:42Z

We should have -march=isa+ext for completeness (yes, _ works in that case, but then you have to care about whether it’s an ISA or profile as your “base”). We also need -march=+foo so you can add to whatever the default is (whether that’s from a flag in user-provided CFLAGS or the system default when not specified). Maybe also -march=-foo, though that gets a bit weird (but that already needs to work in the assembler due to .option arch, so give it the same semantics).

jrtc27 · 2022-12-01T17:58:52Z

I also worry about what happens if a user gives CFLAGS=“-march=isa -misa-spec=…” and a build system then adds -march=profile given you’ve said it’s an error to combine those.

nick-knight

I support this effort. But I think it is worth calling out explicitly where it breaks compatibility with the ISA string as defined in Vol. I Ch. 29, since existing software and build systems are following those conventions already for march. I give two examples below of compatibility breaks. These are just the first two that immediately jumped out at me (I didn't go through the whole PR carefully).

nick-knight · 2022-12-01T18:16:17Z

README.mkd

+  which may be in non-canonical order.
+* Single-letter ISA extensions may be separated from other single-letter extensions
+  by an underscore (`_`).
+* Multi-letter ISA extensions must be separated from other extensions


Vol. I Ch. 29 says:

Extensions with the "Z" prefix must be separated from other multi-letter extensions by an under-
score, e.g., "RV32IMACZicsr_Zifencei".

Note there is no underscore between C and Zicsr.

I properly described what I wanted to express, but I was not aware that -march=rv64imafdczbb is valid in GCC and clang. I will update this to match the currently implemented behavior to reduce the differences to the ISA strings.

Fun fact: GCC currently accepts -march=rv64imafdczbbzba and clang does not.
But I don't think that's an issue.

nick-knight · 2022-12-01T18:17:43Z

README.mkd

+
+A valid `-march` string in the ISA-string-based format must follow these rules:
+
+* All characters of the string must be lowercase.


Vol. I Ch. 29 says:

The ISA naming strings are case insensitive.

Indeed, the examples given therein use mixed case.

This is listed as the first bullet "All characters of the string must be lowercase".
So I think this is already properly documented.

I realize now that this is already enforced by existing toolchains, so in this case I was wrong to say that

existing software and build systems are following those conventions already for march.

I still think it is worth a remark about incompatibilities with the ISA spec, but such nonnormative text can always be added later.

asb · 2022-12-09T13:20:58Z

I do worry that profile strings might be rather cryptic and easily confusable (by humans) with ISA naming strings. If it's agreed this is problematic, it's not clear if it's best handled in the profiles spec or in how the profile names are represented in -march arguments.

i2020u32 is cryptic but perhaps less easily confusable than rvi20u32. As we get closer to 2032, rvi32u32 vs the ISA naming string rv32i... seems potentially confusing. As Andrew argues in riscv/riscv-profiles#82 it's possible that the confusability won't matter as almost nobody is using ISA naming strings.

cmuellner · 2022-12-09T13:26:42Z

As mentioned in the profiles ticket, would it be ok, to mention in a note that tools/compilers may impose restrictions about which march string format is accepted (e.g. via prefix or cmdln flag) and that users may look in to the compiler manual for details? So we would just define the formats here and leave some room for further restrictions for tools/compilers (if any).

cmuellner · 2022-12-09T16:35:01Z

We should have -march=isa+ext for completeness (yes, _ works in that case, but then you have to care about whether it’s an ISA or profile as your “base”). We also need -march=+foo so you can add to whatever the default is (whether that’s from a flag in user-provided CFLAGS or the system default when not specified). Maybe also -march=-foo, though that gets a bit weird (but that already needs to work in the assembler due to .option arch, so give it the same semantics).

I'm not sure if we should define this here (unless enough people want this). The idea is good, but it feels a bit like an added-value feature of a tool/compiler.

I also worry about what happens if a user gives CFLAGS=“-march=isa -misa-spec=…” and a build system then adds -march=profile given you’ve said it’s an error to combine those.

Same for this. I would let this behavior be defined by the tool as I fear that different tools prefer different behavior (e.g. accepting the last provided flag vs. throwing an error vs. ./configure flag to define the policy...).

I would prefer to restrict the description to a minimum that's usable by a user. Compilers and tools can then derive from here and extend more features.

Would this be ok?

jrtc27 · 2022-12-09T17:54:17Z

We should have -march=isa+ext for completeness (yes, _ works in that case, but then you have to care about whether it’s an ISA or profile as your “base”). We also need -march=+foo so you can add to whatever the default is (whether that’s from a flag in user-provided CFLAGS or the system default when not specified). Maybe also -march=-foo, though that gets a bit weird (but that already needs to work in the assembler due to .option arch, so give it the same semantics).

I'm not sure if we should define this here (unless enough people want this). The idea is good, but it feels a bit like an added-value feature of a tool/compiler.

I also worry about what happens if a user gives CFLAGS=“-march=isa -misa-spec=…” and a build system then adds -march=profile given you’ve said it’s an error to combine those.

Same for this. I would let this behavior be defined by the tool as I fear that different tools prefer different behavior (e.g. accepting the last provided flag vs. throwing an error vs. ./configure flag to define the policy...).

I would prefer to restrict the description to a minimum that's usable by a user. Compilers and tools can then derive from here and extend more features.

Would this be ok?

If GCC and LLVM do different things for any of these it will be a pain point, so no, I think they really do need to be documented conventions. Otherwise either nobody will use them or they'll use them and be locked into one compiler for no good reason.

kito-cheng · 2022-12-21T14:22:43Z

Following content has send in another private mail thread, but I think it could be also post here:

I agree the naming of profile can be improved to prevent the confusion, but actually I am less concern about the confusion on the -march string in practice.

Consider the following two cases (and use the most confused possible
name rvi32u32):

-march= with profile name only without extra extension
-march=rvi32u32 - the possible and legal misspell version should
be rv32iu32 which is rv32i base plus U extension with version 32,
that's generally a useless configuration, should got error soon when
trying to build/compile/assemble anything.
-march= with profile name only with extra extensions
-march=rvi32u32+xsffoo profile rvi32u32 with xsffoo vendor
extension, and that's apparently a profile based arch string, since
there is a plus sign there, it's only used in profile style -march
string
Arguable point might has 2:
first is plus sign not included in the ISA string now, but
used/defined in future
- Low probability, but I guess we could...try to stop when that happened.
  second is we might also want to accept the plus sign on the
  normal ISA-string-like -march option.
- This could be prevent when we relax the canonical order
  requirement of -march string

So based on the above two situations I think that should not be
confusing (or misuse with misspellings) in practice.

cmuellner · 2023-03-27T15:01:01Z

In today's Toolchain SIG call, we agreed to split this PR into multiple ones not to mix different issues and to move this forward:

document current -march string
document -march string relaxation
add custom extensions
add profiles support

kito-cheng reviewed Nov 10, 2022

View reviewed changes

anton-afanasyev reviewed Nov 21, 2022

View reviewed changes

cmuellner force-pushed the march-profiles branch from e8d5a82 to 5570a72 Compare December 1, 2022 14:52

cmuellner force-pushed the march-profiles branch from 5570a72 to 82ac13d Compare December 1, 2022 17:21

cmuellner force-pushed the march-profiles branch from 82ac13d to 252f1c3 Compare December 1, 2022 17:43

nick-knight reviewed Dec 1, 2022

View reviewed changes

asb mentioned this pull request Dec 8, 2022

Profile names are easily confused with ISA naming strings riscv/riscv-profiles#82

Closed

pz9115 mentioned this pull request May 2, 2023

Add RISC-V Profiles format #36

Open

jjscheel mentioned this pull request Aug 28, 2023

Profiles Compilers riscv-admin/dev-partners#16

Open

13 tasks

cmuellner mentioned this pull request Dec 7, 2023

[RFC] Defining an ELF attribute for RISC-V target profiles riscv-non-isa/riscv-elf-psabi-doc#409

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Define -march string #26

Define -march string #26

cmuellner commented Nov 8, 2022 •

edited

Loading

cmuellner commented Nov 9, 2022

kito-cheng Nov 10, 2022

cmuellner Dec 1, 2022

kito-cheng Nov 10, 2022

cmuellner Dec 1, 2022

anton-afanasyev Nov 21, 2022

cmuellner Dec 1, 2022

anton-afanasyev Nov 21, 2022

cmuellner Dec 1, 2022

eopXD commented Nov 22, 2022

cmuellner commented Dec 1, 2022

cmuellner commented Dec 1, 2022

cmuellner commented Dec 1, 2022

jrtc27 commented Dec 1, 2022

jrtc27 commented Dec 1, 2022

nick-knight left a comment •

edited

Loading

nick-knight Dec 1, 2022 •

edited

Loading

cmuellner Dec 9, 2022 •

edited

Loading

nick-knight Dec 1, 2022 •

edited

Loading

cmuellner Dec 9, 2022

nick-knight Dec 9, 2022

asb commented Dec 9, 2022

cmuellner commented Dec 9, 2022

cmuellner commented Dec 9, 2022

jrtc27 commented Dec 9, 2022

kito-cheng commented Dec 21, 2022

cmuellner commented Mar 27, 2023


		The following interpretation differences exist:

		* The expansion of `g` with versions before `20190609` is `imafd`.


		A valid `-march` string in the ISA-string-based format must follow these rules:

		* All characters of the string must be lowercase.

Define -march string #26

Are you sure you want to change the base?

Define -march string #26

Conversation

cmuellner commented Nov 8, 2022 • edited Loading

cmuellner commented Nov 9, 2022

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

eopXD commented Nov 22, 2022

cmuellner commented Dec 1, 2022

cmuellner commented Dec 1, 2022

cmuellner commented Dec 1, 2022

jrtc27 commented Dec 1, 2022

jrtc27 commented Dec 1, 2022

nick-knight left a comment • edited Loading

Choose a reason for hiding this comment

nick-knight Dec 1, 2022 • edited Loading

Choose a reason for hiding this comment

cmuellner Dec 9, 2022 • edited Loading

Choose a reason for hiding this comment

nick-knight Dec 1, 2022 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

asb commented Dec 9, 2022

cmuellner commented Dec 9, 2022

cmuellner commented Dec 9, 2022

jrtc27 commented Dec 9, 2022

kito-cheng commented Dec 21, 2022

cmuellner commented Mar 27, 2023

cmuellner commented Nov 8, 2022 •

edited

Loading

nick-knight left a comment •

edited

Loading

nick-knight Dec 1, 2022 •

edited

Loading

cmuellner Dec 9, 2022 •

edited

Loading

nick-knight Dec 1, 2022 •

edited

Loading