[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <d75ef570-c0ad-cea4-687a-d02b560aa676@iscas.ac.cn>
Date: Sun, 30 Jul 2023 01:36:49 +0800
From: Mingzheng Xing <xingmingzheng@...as.ac.cn>
To: Conor Dooley <conor@...nel.org>
Cc: Paul Walmsley <paul.walmsley@...ive.com>,
Palmer Dabbelt <palmer@...belt.com>,
Albert Ou <aou@...s.berkeley.edu>,
Nathan Chancellor <nathan@...nel.org>,
Nick Desaulniers <ndesaulniers@...gle.com>,
Tom Rix <trix@...hat.com>, Bin Meng <bmeng@...ylab.org>,
linux-riscv@...ts.infradead.org, linux-kernel@...r.kernel.org,
llvm@...ts.linux.dev, stable@...r.kernel.org,
Guo Ren <guoren@...nel.org>
Subject: Re: [PATCH v2] riscv: Handle zicsr/zifencei issue between gcc and
binutils
On 7/27/23 15:53, Conor Dooley wrote:
> On Wed, Jul 26, 2023 at 08:41:55PM +0100, Conor Dooley wrote:
>> On Thu, Jul 27, 2023 at 03:34:16AM +0800, Mingzheng Xing wrote:
>>> On 7/27/23 02:02, Conor Dooley wrote:
>>>> This is still broken for:
>>>> CONFIG_CLANG_VERSION=0
>>>> CONFIG_AS_IS_GNU=y
>>>> CONFIG_AS_VERSION=23500
>>>> CONFIG_LD_IS_BFD=y
>>>> CONFIG_LD_VERSION=23500
>>> Do you mean that these CONFIG_* will cause kernel
>>> compilation errors when paired with certain versions of GCC?
>>> Or perhaps I misunderstood your meaning.
>> No, this section is generated by kconfig, although I messed up my
>> trimming of the list & accidentally removed the gcc version, rather
>> than the clang version. Here's the full thing:
>>
>> CONFIG_CC_VERSION_TEXT="riscv64-unknown-linux-gnu-gcc (g2ee5e430018) 12.2.0"
>> CONFIG_CC_IS_GCC=y
>> CONFIG_GCC_VERSION=120200
>> CONFIG_CLANG_VERSION=0
>> CONFIG_AS_IS_GNU=y
>> CONFIG_AS_VERSION=23500
>> CONFIG_LD_IS_BFD=y
>> CONFIG_LD_VERSION=23500
>> CONFIG_LLD_VERSION=0
>> CONFIG_CC_CAN_LINK=y
>> CONFIG_CC_CAN_LINK_STATIC=y
>> CONFIG_CC_HAS_ASM_GOTO_OUTPUT=y
>> CONFIG_CC_HAS_ASM_GOTO_TIED_OUTPUT=y
>> CONFIG_CC_HAS_ASM_INLINE=y
>> CONFIG_CC_HAS_NO_PROFILE_FN_ATTR=y
>> CONFIG_PAHOLE_VERSION=0
>> CONFIG_CONSTRUCTORS=y
>> CONFIG_IRQ_WORK=y
>> CONFIG_BUILDTIME_TABLE_SORT=y
> I think this should sort things out for the even-older binutils case. I
> took the opportunity to fix some grammatical issues that seem to have
> snuck into the help text in your patch & to drop the \, since the
> depends on fits in one line.
hi, Conor.
I reproduced the error with gcc-12.2.0 and binutils-2.35. I tried a
different solution, which I think makes the logic easier. Showing
the new patch code:
diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig
index 4c07b9189c86..a6fa1eed895c 100644
--- a/arch/riscv/Kconfig
+++ b/arch/riscv/Kconfig
@@ -569,25 +569,24 @@ config TOOLCHAIN_HAS_ZIHINTPAUSE
config TOOLCHAIN_NEEDS_EXPLICIT_ZICSR_ZIFENCEI
def_bool y
- # https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=aed44286efa8ae8717a77d94b51ac3614e2ca6dc
- depends on AS_IS_GNU && AS_VERSION >= 23800
+ depends on AS_IS_GNU && AS_VERSION >= 23600
help
- Newer binutils versions default to ISA spec version 20191213 which
- moves some instructions from the I extension to the Zicsr and Zifencei
- extensions.
+ Binutils has supported zicsr and zifencei extensions since version 2.36,
+ try to adapt to the changes by using explicit zicsr and zifencei via
+ -march. For two special cases, where clang<17 or gcc<11.1.0, we will
+ deal with them in CONFIG_TOOLCHAIN_NEEDS_OLD_ISA_SPEC.
config TOOLCHAIN_NEEDS_OLD_ISA_SPEC
def_bool y
depends on TOOLCHAIN_NEEDS_EXPLICIT_ZICSR_ZIFENCEI
# https://github.com/llvm/llvm-project/commit/22e199e6afb1263c943c0c0d4498694e15bf8a16
- depends on CC_IS_CLANG && CLANG_VERSION < 170000
- help
- Certain versions of clang do not support zicsr and zifencei via -march
- but newer versions of binutils require it for the reasons noted in the
- help text of CONFIG_TOOLCHAIN_NEEDS_EXPLICIT_ZICSR_ZIFENCEI. This
- option causes an older ISA spec compatible with these older versions
- of clang to be passed to GAS, which has the same result as passing zicsr
- and zifencei to -march.
+ # https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=b03be74bad08c382da47e048007a78fa3fb4ef49
+ depends on (CC_IS_CLANG && CLANG_VERSION < 170000) || (CC_IS_GCC && GCC_VERSION < 110100)
+ help
+ Certain versions of clang and GCC do not support zicsr and zifencei via
+ -march. This option causes an older ISA spec compatible with these older
+ versions of clang and GCC to be passed to GAS, which has the same result
+ as passing zicsr and zifencei to -march.
config FPU
bool "FPU support"
diff --git a/arch/riscv/kernel/compat_vdso/Makefile b/arch/riscv/kernel/compat_vdso/Makefile
index 189345773e7e..b86e5e2c3aea 100644
--- a/arch/riscv/kernel/compat_vdso/Makefile
+++ b/arch/riscv/kernel/compat_vdso/Makefile
@@ -11,7 +11,13 @@ compat_vdso-syms += flush_icache
COMPAT_CC := $(CC)
COMPAT_LD := $(LD)
-COMPAT_CC_FLAGS := -march=rv32g -mabi=ilp32
+# binutils 2.35 does not support the zifencei extension, but in the ISA
+# spec 20191213, G stands for IMAFD_ZICSR_ZIFENCEI.
+ifdef CONFIG_TOOLCHAIN_NEEDS_EXPLICIT_ZICSR_ZIFENCEI
+ COMPAT_CC_FLAGS := -march=rv32g -mabi=ilp32
+else
+ COMPAT_CC_FLAGS := -march=rv32imafd -mabi=ilp32
+endif
COMPAT_LD_FLAGS := -melf32lriscv
# Disable attributes, as they're useless and break the build.
--
2.34.1
Here are the results of my tests:
gcc binutils patched no patch(test on master)
11.4.0 2.35 ok ok
11.4.0 2.36 ok ok
11.4.0 2.38 ok ok
12.2.0 2.35 ok error[1]
12.2.0 2.36 ok error[2]
12.2.0 2.38 ok ok
10.5.0 2.35 ok ok
10.5.0 2.36 ok ok
10.5.0 2.38 ok error[3]
11.1.0 2.35 ok ok
11.1.0 2.36 ok ok
11.1.0 2.38 ok ok
11.2.0 2.35 ok ok
11.2.0 2.36 ok ok
11.2.0 2.38 ok ok
[1]
Assembler messages:
Fatal error: -march=rv32imafd_zicsr_zifencei: Invalid or unknown z ISA extension: 'zifencei'
make[2]: *** [arch/riscv/kernel/compat_vdso/Makefile:47: arch/riscv/kernel/compat_vdso/rt_sigreturn.o] Error 1
[2]
./arch/riscv/include/asm/vdso/gettimeofday.h: Assembler messages:
./arch/riscv/include/asm/vdso/gettimeofday.h:79: Error: unrecognized opcode `csrr a5,0xc01'
./arch/riscv/include/asm/vdso/gettimeofday.h:79: Error: unrecognized opcode `csrr a5,0xc01'
./arch/riscv/include/asm/vdso/gettimeofday.h:79: Error: unrecognized opcode `csrr a5,0xc01'
./arch/riscv/include/asm/vdso/gettimeofday.h:79: Error: unrecognized opcode `csrr a5,0xc01'
make[2]: *** [scripts/Makefile.build:243: arch/riscv/kernel/vdso/vgettimeofday.o] Error 1
[3]
cc1: error: '-march=rv64imac_zicsr_zifencei': unsupported ISA subset 'z'
cc1: error: ABI requires '-march=rv64'
make[2]: *** [scripts/Makefile.build:243: scripts/mod/empty.o] Error 1
make[2]: *** Waiting for unfinished jobs....
cc1: error: '-march=rv64imac_zicsr_zifencei': unsupported ISA subset 'z'
cc1: error: ABI requires '-march=rv64'
> diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig
> index e1b66ee88323..2d0d89213c97 100644
> --- a/arch/riscv/Kconfig
> +++ b/arch/riscv/Kconfig
> @@ -571,25 +571,27 @@ config TOOLCHAIN_NEEDS_EXPLICIT_ZICSR_ZIFENCEI
> def_bool y
> # https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=aed44286efa8ae8717a77d94b51ac3614e2ca6dc
> # https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=98416dbb0a62579d4a7a4a76bab51b5b52fec2cd
> - depends on GCC_VERSION >= 120100 || (AS_IS_GNU && AS_VERSION >= 23800)
> + depends on AS_IS_GNU
> + depends on (GCC_VERSION >= 120100 && AS_VERSION >= 23600) || AS_VERSION >= 23800
Tests verified that explicit _ZICSR_ZIFENCEI via -march is required
for gcc>=12.1.0, but this only happens for binutils>=2.36,
binutils 2.35 + gcc>=12.1.0 does not need that. Considering
binutils 2.35 together complicates things. So what do you think
of the above new version patch?
Some more info:
- The commit[4] for patch changes.
- binutils 2.36 supports the zifencei extension[5] and splits
zifencei and zicsr from I[6].
[4] commit 0715372a06ce ("riscv: compat: vdso: Add COMPAT_VDSO base code implementation")
[5] https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=5a1b31e1e1cee6e9f1c92abff59cdcfff0dddf30
[6] https://sourceware.org/git/?p=binutils-gdb.git;a=commit;h=729a53530e86972d1143553a415db34e6e01d5d2
Thanks,
Mingzheng.
> help
> - Binutils-2.38 and GCC-12.1.0 bump default ISA spec to newer version
> + Binutils-2.38 and GCC-12.1.0 bump the default ISA spec to version
> 20191213 which moves some instructions from the I extension to the
> - Zicsr and Zifencei extensions.
> + Zicsr and Zifencei extensions. On the other hand, Binutils prior to
> + 2.35 does not understand these arguments and will error if they are
> + passed.
>
> config TOOLCHAIN_NEEDS_OLD_ISA_SPEC
> def_bool y
> depends on TOOLCHAIN_NEEDS_EXPLICIT_ZICSR_ZIFENCEI
> # https://github.com/llvm/llvm-project/commit/22e199e6afb1263c943c0c0d4498694e15bf8a16
> # https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=b03be74bad08c382da47e048007a78fa3fb4ef49
> - depends on (CC_IS_CLANG && CLANG_VERSION < 170000) || \
> - (CC_IS_GCC && GCC_VERSION < 110100)
> + depends on (CC_IS_CLANG && CLANG_VERSION < 170000) || (CC_IS_GCC && GCC_VERSION < 110100)
> help
> - Certain versions of clang (or GCC) do not support zicsr and zifencei via
> + Certain versions of clang and GCC do not support zicsr and zifencei via
> -march but newer versions of binutils require it for the reasons noted
> in the help text of CONFIG_TOOLCHAIN_NEEDS_EXPLICIT_ZICSR_ZIFENCEI. This
> option causes an older ISA spec compatible with these older versions
> - of clang (or GCC) to be passed to GAS, which has the same result as
> + of clang and GCC to be passed to GAS, which has the same result as
> passing zicsr and zifencei to -march.
>
> config FPU
>
>
> _______________________________________________
> linux-riscv mailing list
> linux-riscv@...ts.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-riscv
Powered by blists - more mailing lists