From: Liao Shihua <shihua@iscas.ac.cn>
To: Nelson Chu <nelson@rivosinc.com>
Cc: binutils@sourceware.org, kito.cheng@gmail.com,
shiyulong@iscas.ac.cn, jiawei@iscas.ac.cn,
chenyixuan@iscas.ac.cn, guoren@kernel.org, wuwei2016@iscas.ac.cn
Subject: Re: [PATCH] Support N32(32-bit ABI on 64-bit ISA) in riscv
Date: Mon, 18 Nov 2024 19:03:24 +0800 [thread overview]
Message-ID: <747b20ae-8b77-432e-b03f-2b3c84c88811@iscas.ac.cn> (raw)
In-Reply-To: <CAPpQWtBtZiaBsjZR6Ngk5oGQ5Z7PRuHKhmF1KUWz8iLTBuA76w@mail.gmail.com>
[-- Attachment #1: Type: text/plain, Size: 18147 bytes --]
Hello, Nelson
Thanks for your advice, I will try to do this.
Liao Shihua
在 2024/11/18 10:40, Nelson Chu 写道:
> There are some comments as follows. Not quite sure what to do with
> experimental features in the toolchain, we used to have assembler/link
> options or configuration options to make it different from regular
> users. Besides, it is probably easier to remove the features if they
> are experimental and controlled by experimental options when they
> cause problems or are not useful as expected. So if N32 is, then
> ideally we should have some experimental options. For example, if
> people still want the link errors when linking rv32 and rv64, then
> they just don't enable the experimental options.
>
> Thanks
> Nelson
>
> On Sat, Nov 16, 2024 at 11:16 AM Liao Shihua <shihua@iscas.ac.cn> wrote:
>
> RISC-V N32 ABI means using 32-bit ABI on 64-bit ISA, the discussion in
> https://github.com/riscv-non-isa/riscv-elf-psabi-doc/pull/381 .
> At this moment, N32 is supported batemental toolchain.
> Three OpenSource RTOS using this feature and have been merged in
> upstream.
> You can see them in
> EasyXem (AUTOSAR CP R19-11):
> https://atomgit.com/easyxmen/XMen/tree/rv64ilp32-dev
> Nuttx: https://github.com/apache/nuttx
> RT-Thread:https://github.com/RT-Thread/rt-thread/pull/9194
>
> This patch add a new bfd_mach bfd_mach_riscv64n32 and a new
> e_flags N32.
> bfd_mach_riscv64n32 has the same bits in a word/address and
> ARCH_SIZE with rv32,
> but use rv64's PRSTATUS. N32 use the 6th bit of e_flags layout.
> In addition, this patch replace xlen with abi_xlen in
> riscv_target_format().
>
> bfd/ChangeLog:
>
> * archures.c: Add bfd_mach_riscv64n32.
> * bfd-in2.h (bfd_mach_riscv64n32): Ditto.
> * cpu-riscv.c: Ditto.
> * elfnn-riscv.c (ABI_N32_P):Add ABI_N32_P
> (perform_relocation):Ditto.
> (riscv_merge_arch_attr_info):Remove elf check when using N32.
> (_bfd_riscv_elf_merge_private_bfd_data):Ditto.
> (_bfd_riscv_relax_section):Ditto.
> (riscv_elf_object_p):Using bfd_mach_riscv64n32 with N32.
>
> binutils/ChangeLog:
>
> * readelf.c (decode_RISCV_machine_flags):Add N32 Flag.
>
> gas/ChangeLog:
>
> * config/tc-riscv.c (riscv_set_n32):Set N32 flag.
> (riscv_set_abi_by_arch):Removed check.
> (riscv_target_format):Ditto
> (md_begin):Ditto
> (s_riscv_attribute):Ditto
> * testsuite/gas/riscv/mabi-fail-rv64iq-ilp32.d: Removed.
> * testsuite/gas/riscv/mabi-fail-rv64iq-ilp32.l: Removed.
>
> include/ChangeLog:
>
> * elf/riscv.h (EF_RISCV_N32):Add N32 Flag.
>
> opcodes/ChangeLog:
>
> * riscv-dis.c (print_insn_args):
> (riscv_disassemble_insn):
>
> ---
> bfd/archures.c | 1 +
> bfd/bfd-in2.h | 1 +
> bfd/cpu-riscv.c | 2 ++
> bfd/elfnn-riscv.c | 32
> +++++++++++++------
> binutils/readelf.c | 3 ++
> gas/config/tc-riscv.c | 23 ++++++++++---
> .../gas/riscv/mabi-fail-rv64iq-ilp32.d | 3 --
> .../gas/riscv/mabi-fail-rv64iq-ilp32.l | 2 --
> include/elf/riscv.h | 3 ++
> opcodes/riscv-dis.c | 6 ++--
> 10 files changed, 53 insertions(+), 23 deletions(-)
> delete mode 100644 gas/testsuite/gas/riscv/mabi-fail-rv64iq-ilp32.d
> delete mode 100644 gas/testsuite/gas/riscv/mabi-fail-rv64iq-ilp32.l
>
> diff --git a/bfd/archures.c b/bfd/archures.c
> index c4decc59e4a..063659631fa 100644
> --- a/bfd/archures.c
> +++ b/bfd/archures.c
> @@ -448,6 +448,7 @@ DESCRIPTION
> . bfd_arch_riscv,
> .#define bfd_mach_riscv32 132
> .#define bfd_mach_riscv64 164
> +.#define bfd_mach_riscv64n32 16432
> . bfd_arch_rl78,
> .#define bfd_mach_rl78 0x75
> . bfd_arch_rx, {* Renesas RX. *}
> diff --git a/bfd/bfd-in2.h b/bfd/bfd-in2.h
> index 3b047d922f3..0ee6bbc2a73 100644
> --- a/bfd/bfd-in2.h
> +++ b/bfd/bfd-in2.h
> @@ -1714,6 +1714,7 @@ enum bfd_architecture
> bfd_arch_riscv,
> #define bfd_mach_riscv32 132
> #define bfd_mach_riscv64 164
> +#define bfd_mach_riscv64n32 16432
> bfd_arch_rl78,
> #define bfd_mach_rl78 0x75
> bfd_arch_rx, /* Renesas RX. */
> diff --git a/bfd/cpu-riscv.c b/bfd/cpu-riscv.c
> index 58cbdd846be..1df99dc6e2d 100644
> --- a/bfd/cpu-riscv.c
> +++ b/bfd/cpu-riscv.c
> @@ -86,6 +86,7 @@ riscv_scan (const struct bfd_arch_info *info,
> const char *string)
> enum
> {
> I_riscv64,
> + I_riscv64n32,
> I_riscv32
> };
>
> @@ -96,6 +97,7 @@ enum
> static const bfd_arch_info_type arch_info_struct[] =
> {
> N (64, bfd_mach_riscv64, "riscv:rv64", false, NN (I_riscv64)),
> + N (32, bfd_mach_riscv64n32, "riscv:rv64", false, NN
> (I_riscv64n32)),
> N (32, bfd_mach_riscv32, "riscv:rv32", false, NULL)
> };
>
> diff --git a/bfd/elfnn-riscv.c b/bfd/elfnn-riscv.c
> index c8bf45f4293..b05f4fde433 100644
> --- a/bfd/elfnn-riscv.c
> +++ b/bfd/elfnn-riscv.c
> @@ -138,6 +138,11 @@
>
> #define RISCV_ELF_WORD_BYTES (1 << RISCV_ELF_LOG_WORD_BYTES)
>
> +#define ABI_N32_P(abfd) \
> + ((elf_elfheader (abfd)->e_flags & EF_RISCV_N32) != 0)
> +
> +static bool ABI_N32 = false;
> +
> /* The name of the dynamic interpreter. This is put in the .interp
> section. */
>
> @@ -1823,7 +1828,8 @@ perform_relocation (const reloc_howto_type
> *howto,
>
> case R_RISCV_CALL:
> case R_RISCV_CALL_PLT:
> - if (ARCH_SIZE > 32 && !VALID_UTYPE_IMM
> (RISCV_CONST_HIGH_PART (value)))
> + if ((ARCH_SIZE > 32 || ABI_N32_P (input_bfd))
> + && !VALID_UTYPE_IMM (RISCV_CONST_HIGH_PART (value)))
> return bfd_reloc_overflow;
> value = ENCODE_UTYPE_IMM (RISCV_CONST_HIGH_PART (value))
> | (ENCODE_ITYPE_IMM (value) << 32);
> @@ -3927,7 +3933,7 @@ riscv_merge_arch_attr_info (bfd *ibfd, char
> *in_arch, char *out_arch)
> return NULL;
>
> /* Checking XLEN. */
> - if (xlen_out != xlen_in)
> + if (xlen_out != xlen_in && !ABI_N32_P (ibfd))
> {
> _bfd_error_handler
> (_("error: %pB: ISA string of input (%s) doesn't match "
> @@ -3947,7 +3953,7 @@ riscv_merge_arch_attr_info (bfd *ibfd, char
> *in_arch, char *out_arch)
> if (!riscv_merge_multi_letter_ext (&in, &out))
> return NULL;
>
> - if (xlen_in != xlen_out)
> + if (xlen_in != xlen_out && !ABI_N32_P (ibfd))
> {
> _bfd_error_handler
> (_("error: %pB: XLEN of input (%u) doesn't match "
> @@ -3955,7 +3961,7 @@ riscv_merge_arch_attr_info (bfd *ibfd, char
> *in_arch, char *out_arch)
> return NULL;
> }
>
> - if (xlen_in != ARCH_SIZE)
> + if (xlen_in != ARCH_SIZE && !ABI_N32_P (ibfd))
> {
> _bfd_error_handler
> (_("error: %pB: unsupported XLEN (%u), you might be "
> @@ -3963,7 +3969,7 @@ riscv_merge_arch_attr_info (bfd *ibfd, char
> *in_arch, char *out_arch)
> return NULL;
> }
>
> - merged_arch_str = riscv_arch_str (ARCH_SIZE, &merged_subsets);
> + merged_arch_str = riscv_arch_str (xlen_in, &merged_subsets);
>
>
> The xlen of n32 seems to be rv64. If we link n32 with normal rv64
> objects, it looks fine since xlen will always be rv64. But how about
> linking n32 with normal rv32 objects? The merged xlen will be changed
> to the latest input xlen, so it might be rv32 or rv64 when we link n32
> with other rv32 objects, it looks weird but is it expected? This
> might also have a merged output with rv32 + q, which looks like a
> conflict but there won't be any alerts since we used not to call
> riscv_parse_check_conflicts by linker.
>
> /* Release the subset lists. */
> riscv_release_subset_list (&in_subsets);
> @@ -4233,6 +4239,9 @@ _bfd_riscv_elf_merge_private_bfd_data (bfd
> *ibfd, struct bfd_link_info *info)
>
> /* Allow linking TSO and non-TSO, and keep the TSO flag. */
> elf_elfheader (obfd)->e_flags |= new_flags & EF_RISCV_TSO;
> +
> + /* Allow linking N32 and non-N32, and keep the N32 flag. */
> + elf_elfheader (obfd)->e_flags |= new_flags & EF_RISCV_N32;
>
> return true;
>
> @@ -5400,7 +5409,7 @@ _bfd_riscv_relax_section (bfd *abfd,
> asection *sec,
> return ret;
> }
>
> -#if ARCH_SIZE == 32
> +#if ARCH_SIZE == 32 && !ABI_N32
> # define PRSTATUS_SIZE 204
> # define PRSTATUS_OFFSET_PR_CURSIG 12
> # define PRSTATUS_OFFSET_PR_PID 24
> @@ -5570,10 +5579,13 @@ riscv_elf_grok_psinfo (bfd *abfd,
> Elf_Internal_Note *note)
> static bool
> riscv_elf_object_p (bfd *abfd)
> {
> - /* There are only two mach types in RISCV currently. */
> - if (strcmp (abfd->xvec->name, "elf32-littleriscv") == 0
> - || strcmp (abfd->xvec->name, "elf32-bigriscv") == 0)
> - bfd_default_set_arch_mach (abfd, bfd_arch_riscv,
> bfd_mach_riscv32);
> + ABI_N32 = ABI_N32_P (abfd);
> + /* There are only three mach types in RISCV currently. */
> + if (ABI_N32)
> + bfd_default_set_arch_mach (abfd, bfd_arch_riscv,
> bfd_mach_riscv64n32);
> + else if (strcmp (abfd->xvec->name, "elf32-littleriscv") == 0
> + || strcmp (abfd->xvec->name, "elf32-bigriscv") == 0)
> + bfd_default_set_arch_mach (abfd, bfd_arch_riscv,
> bfd_mach_riscv32);
> else
> bfd_default_set_arch_mach (abfd, bfd_arch_riscv,
> bfd_mach_riscv64);
>
> diff --git a/binutils/readelf.c b/binutils/readelf.c
> index 73163e0ee21..58248d84135 100644
> --- a/binutils/readelf.c
> +++ b/binutils/readelf.c
> @@ -4532,6 +4532,9 @@ decode_RISCV_machine_flags (char *out,
> unsigned e_flags)
> if (e_flags & EF_RISCV_TSO)
> out = stpcpy (out, ", TSO");
>
> + if (e_flags & EF_RISCV_N32)
> + out = stpcpy (out, ", N32");
> +
> switch (e_flags & EF_RISCV_FLOAT_ABI)
> {
> case EF_RISCV_FLOAT_ABI_SOFT:
> diff --git a/gas/config/tc-riscv.c b/gas/config/tc-riscv.c
> index e19142dca19..dde9c52ccc1 100644
> --- a/gas/config/tc-riscv.c
> +++ b/gas/config/tc-riscv.c
> @@ -306,6 +306,13 @@ riscv_set_tso (void)
> elf_flags |= EF_RISCV_TSO;
> }
>
> +/* Turn on the n32 flag for elf_flags once we have enabled n32
> model. */
> +static void
> +riscv_set_n32 (void)
> +{
> + elf_flags |= EF_RISCV_N32;
> +}
> +
> /* The linked list hanging off of .subsets_list records all
> enabled extensions,
> which are parsed from the architecture string. The
> architecture string can
> be set by the -march option, the elf architecture attributes,
> and the
> @@ -409,7 +416,7 @@ riscv_set_abi_by_arch (void)
> gas_assert (abi_xlen != 0 && xlen != 0 && float_abi !=
> FLOAT_ABI_DEFAULT);
> if (abi_xlen > xlen)
> as_bad ("can't have %d-bit ABI on %d-bit ISA", abi_xlen,
> xlen);
> - else if (abi_xlen < xlen)
> + else if (abi_xlen < xlen && (abi_xlen != 32 && xlen != 64))
> as_bad ("%d-bit ABI not yet supported on %d-bit ISA",
> abi_xlen, xlen);
>
> if (riscv_subset_supports (&riscv_rps_as, "e") && !rve_abi)
> @@ -435,6 +442,10 @@ riscv_set_abi_by_arch (void)
>
> if (rve_abi)
> elf_flags |= EF_RISCV_RVE;
> +
> + if (abi_xlen == 32 && xlen == 64)
> + riscv_set_n32 ();
> +
> }
>
> /* Handle of the OPCODE hash table. */
> @@ -734,9 +745,9 @@ const char *
> riscv_target_format (void)
> {
> if (target_big_endian)
> - return xlen == 64 ? "elf64-bigriscv" : "elf32-bigriscv";
> + return abi_xlen == 64 ? "elf64-bigriscv" : "elf32-bigriscv";
> else
> - return xlen == 64 ? "elf64-littleriscv" : "elf32-littleriscv";
> + return abi_xlen == 64 ? "elf64-littleriscv" :
> "elf32-littleriscv";
> }
>
> /* Return the length of instruction INSN. */
> @@ -1876,7 +1887,8 @@ riscv_record_pcrel_fixup (htab_t p, const
> asection *sec, bfd_vma address,
> void
> md_begin (void)
> {
> - unsigned long mach = xlen == 64 ? bfd_mach_riscv64 :
> bfd_mach_riscv32;
> + unsigned long mach = xlen == 64 ?
> + (abi_xlen == 32 ? bfd_mach_riscv64n32 : bfd_mach_riscv64) :
> bfd_mach_riscv32;
>
>
> Indent, exceed 80 char for one line.
>
>
> if (! bfd_set_arch_mach (stdoutput, bfd_arch_riscv, mach))
> as_warn (_("could not set architecture and machine"));
> @@ -5674,7 +5686,8 @@ s_riscv_attribute (int ignored ATTRIBUTE_UNUSED)
> if (old_xlen != xlen)
> {
> /* We must re-init bfd again if xlen is changed. */
> - unsigned long mach = xlen == 64 ? bfd_mach_riscv64 :
> bfd_mach_riscv32;
> + unsigned long mach = xlen == 64 ?
> + (abi_xlen == 32 ? bfd_mach_riscv64n32 : bfd_mach_riscv64) :
> bfd_mach_riscv32;
>
>
> Indent wrong.
>
> bfd_find_target (riscv_target_format (), stdoutput);
>
> if (! bfd_set_arch_mach (stdoutput, bfd_arch_riscv, mach))
> diff --git a/gas/testsuite/gas/riscv/mabi-fail-rv64iq-ilp32.d
> b/gas/testsuite/gas/riscv/mabi-fail-rv64iq-ilp32.d
> deleted file mode 100644
> index e3155f48956..00000000000
> --- a/gas/testsuite/gas/riscv/mabi-fail-rv64iq-ilp32.d
> +++ /dev/null
> @@ -1,3 +0,0 @@
> -#as: -march-attr -mabi=ilp32
> -#source: mabi-attr-rv64iq.s
> -#error_output: mabi-fail-rv64iq-ilp32.l
> diff --git a/gas/testsuite/gas/riscv/mabi-fail-rv64iq-ilp32.l
> b/gas/testsuite/gas/riscv/mabi-fail-rv64iq-ilp32.l
> deleted file mode 100644
> index 8d45a07fd36..00000000000
> --- a/gas/testsuite/gas/riscv/mabi-fail-rv64iq-ilp32.l
> +++ /dev/null
> @@ -1,2 +0,0 @@
> -.*Assembler messages:
> -.*Error: 32-bit ABI not yet supported on 64-bit ISA
> diff --git a/include/elf/riscv.h b/include/elf/riscv.h
> index 24903c04d91..f88a1c0701a 100644
> --- a/include/elf/riscv.h
> +++ b/include/elf/riscv.h
> @@ -140,6 +140,9 @@ END_RELOC_NUMBERS (R_RISCV_max)
> /* File uses the TSO model. */
> #define EF_RISCV_TSO 0x0010
>
> +/* File uses the N32 model. */
> +#define EF_RISCV_N32 0x0020
> +
> /* Additional section types. */
> #define SHT_RISCV_ATTRIBUTES (SHT_LOPROC + 3) /* Section holds
> attributes. */
>
> diff --git a/opcodes/riscv-dis.c b/opcodes/riscv-dis.c
> index 80018db837a..09fac0c3cbe 100644
> --- a/opcodes/riscv-dis.c
> +++ b/opcodes/riscv-dis.c
> @@ -349,7 +349,7 @@ print_insn_args (const char *oparg, insn_t l,
> bfd_vma pc, disassemble_info *info
> case 'j':
> if (((l & MASK_C_ADDI) == MATCH_C_ADDI) && rd != 0)
> maybe_print_address (pd, rd, EXTRACT_CITYPE_IMM
> (l), 0);
> - if (info->mach == bfd_mach_riscv64
> + if (info->mach == bfd_mach_riscv64 || info->mach ==
> bfd_mach_riscv64n32
>
>
> Indent, exceed 80 char for one line., and also looks like missing
> brackets.
>
> && ((l & MASK_C_ADDIW) == MATCH_C_ADDIW) && rd != 0)
> maybe_print_address (pd, rd, EXTRACT_CITYPE_IMM
> (l), 1);
> print (info->stream, dis_style_immediate, "%d",
> @@ -555,7 +555,7 @@ print_insn_args (const char *oparg, insn_t l,
> bfd_vma pc, disassemble_info *info
> if (((l & MASK_ADDI) == MATCH_ADDI && rs1 != 0)
> || (l & MASK_JALR) == MATCH_JALR)
> maybe_print_address (pd, rs1, EXTRACT_ITYPE_IMM (l), 0);
> - if (info->mach == bfd_mach_riscv64
> + if (info->mach == bfd_mach_riscv64|| info->mach ==
> bfd_mach_riscv64n32
>
>
> Indent, exceed 80 char for one line, a space before ||, and also looks
> like missing brackets.
>
> && ((l & MASK_ADDIW) == MATCH_ADDIW) && rs1 != 0)
> maybe_print_address (pd, rs1, EXTRACT_ITYPE_IMM (l), 1);
> print (info->stream, dis_style_immediate, "%d",
> @@ -941,7 +941,7 @@ riscv_disassemble_insn (bfd_vma memaddr,
> if (op != NULL)
> {
> /* If XLEN is not known, get its value from the ELF class. */
> - if (info->mach == bfd_mach_riscv64)
> + if (info->mach == bfd_mach_riscv64 || info->mach ==
> bfd_mach_riscv64n32)
> xlen = 64;
> else if (info->mach == bfd_mach_riscv32)
> xlen = 32;
> --
> 2.34.1
>
>
next prev parent reply other threads:[~2024-11-18 11:03 UTC|newest]
Thread overview: 6+ messages / expand[flat|nested] mbox.gz Atom feed top
2024-11-16 3:13 Liao Shihua
2024-11-16 8:58 ` Guo Ren
2024-11-18 2:40 ` Nelson Chu
2024-11-18 11:03 ` Liao Shihua [this message]
2024-11-18 7:53 ` Jan Beulich
2024-11-18 8:48 ` Liao Shihua
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=747b20ae-8b77-432e-b03f-2b3c84c88811@iscas.ac.cn \
--to=shihua@iscas.ac.cn \
--cc=binutils@sourceware.org \
--cc=chenyixuan@iscas.ac.cn \
--cc=guoren@kernel.org \
--cc=jiawei@iscas.ac.cn \
--cc=kito.cheng@gmail.com \
--cc=nelson@rivosinc.com \
--cc=shiyulong@iscas.ac.cn \
--cc=wuwei2016@iscas.ac.cn \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).