diff mbox series

[1/3] RISC-V: build: Pass `-mattr` to `lld`

Message ID 20210719205247.1023289-1-twd2.me@gmail.com (mailing list archive)
State New, archived
Headers show
Series RISC-V: build: Allow LTO to be selected | expand

Commit Message

twd2 July 19, 2021, 8:52 p.m. UTC
Pass `-mattr` to `lld` so that it can generate proper machine code.
By the way, it seems that `-march` does not work well now.

Signed-off-by: Wende Tan <twd2.me@gmail.com>
---
 arch/riscv/Makefile | 7 +++++++
 1 file changed, 7 insertions(+)

Comments

Nick Desaulniers July 19, 2021, 9:14 p.m. UTC | #1
On Mon, Jul 19, 2021 at 1:53 PM Wende Tan <twd2.me@gmail.com> wrote:
>
> Pass `-mattr` to `lld` so that it can generate proper machine code.
> By the way, it seems that `-march` does not work well now.
>
> Signed-off-by: Wende Tan <twd2.me@gmail.com>
> ---
>  arch/riscv/Makefile | 7 +++++++
>  1 file changed, 7 insertions(+)
>
> diff --git a/arch/riscv/Makefile b/arch/riscv/Makefile
> index bc74afdbf31e..bbc186a3ab45 100644
> --- a/arch/riscv/Makefile
> +++ b/arch/riscv/Makefile
> @@ -52,9 +52,16 @@ riscv-march-$(CONFIG_ARCH_RV32I)     := rv32ima
>  riscv-march-$(CONFIG_ARCH_RV64I)       := rv64ima
>  riscv-march-$(CONFIG_FPU)              := $(riscv-march-y)fd
>  riscv-march-$(CONFIG_RISCV_ISA_C)      := $(riscv-march-y)c
> +riscv-mattr-y                          := +m,+a
> +riscv-mattr-$(CONFIG_FPU)              := $(riscv-mattr-y),+f,+d
> +riscv-mattr-$(CONFIG_RISCV_ISA_C)      := $(riscv-mattr-y),+c
>  KBUILD_CFLAGS += -march=$(subst fd,,$(riscv-march-y))
>  KBUILD_AFLAGS += -march=$(riscv-march-y)
>
> +ifeq ($(CONFIG_LD_IS_LLD),y)
> +       KBUILD_LDFLAGS += -mllvm -mattr=$(riscv-mattr-y)

We should prevent the proliferation of the compiler internal `-mllvm`
flags from Kbuild Makefiles; these are indicative of LLVM IR specific
issues with regards to LTO.  Can you please file a bug against LLVM
that `-mattr` should be encoded in the IR for LTO's sake?

See also:
commit 0236526d76b8 ("Makefile: lto: Pass -warn-stack-size only on LLD
< 13.0.0")
commit 2398ce80152a ("x86, lto: Pass -stack-alignment only on LLD < 13.0.0")

> +endif
> +
>  KBUILD_CFLAGS += -mno-save-restore
>  KBUILD_CFLAGS += -DCONFIG_PAGE_OFFSET=$(CONFIG_PAGE_OFFSET)
>
> --


--
Thanks,
~Nick Desaulniers
twd2 July 22, 2021, 8:23 a.m. UTC | #2
Thanks for your advice.  After some investigations, I find minimal code
that can reproduce the issue I have had.  The bug is filed at
https://bugs.llvm.org/show_bug.cgi?id=51161.


Thanks.
Wende


On 2021/7/20 5:14, Nick Desaulniers wrote:
> On Mon, Jul 19, 2021 at 1:53 PM Wende Tan <twd2.me@gmail.com> wrote:
>> Pass `-mattr` to `lld` so that it can generate proper machine code.
>> By the way, it seems that `-march` does not work well now.
>>
>> Signed-off-by: Wende Tan <twd2.me@gmail.com>
>> ---
>>  arch/riscv/Makefile | 7 +++++++
>>  1 file changed, 7 insertions(+)
>>
>> diff --git a/arch/riscv/Makefile b/arch/riscv/Makefile
>> index bc74afdbf31e..bbc186a3ab45 100644
>> --- a/arch/riscv/Makefile
>> +++ b/arch/riscv/Makefile
>> @@ -52,9 +52,16 @@ riscv-march-$(CONFIG_ARCH_RV32I)     := rv32ima
>>  riscv-march-$(CONFIG_ARCH_RV64I)       := rv64ima
>>  riscv-march-$(CONFIG_FPU)              := $(riscv-march-y)fd
>>  riscv-march-$(CONFIG_RISCV_ISA_C)      := $(riscv-march-y)c
>> +riscv-mattr-y                          := +m,+a
>> +riscv-mattr-$(CONFIG_FPU)              := $(riscv-mattr-y),+f,+d
>> +riscv-mattr-$(CONFIG_RISCV_ISA_C)      := $(riscv-mattr-y),+c
>>  KBUILD_CFLAGS += -march=$(subst fd,,$(riscv-march-y))
>>  KBUILD_AFLAGS += -march=$(riscv-march-y)
>>
>> +ifeq ($(CONFIG_LD_IS_LLD),y)
>> +       KBUILD_LDFLAGS += -mllvm -mattr=$(riscv-mattr-y)
> We should prevent the proliferation of the compiler internal `-mllvm`
> flags from Kbuild Makefiles; these are indicative of LLVM IR specific
> issues with regards to LTO.  Can you please file a bug against LLVM
> that `-mattr` should be encoded in the IR for LTO's sake?
>
> See also:
> commit 0236526d76b8 ("Makefile: lto: Pass -warn-stack-size only on LLD
> < 13.0.0")
> commit 2398ce80152a ("x86, lto: Pass -stack-alignment only on LLD < 13.0.0")
>
>> +endif
>> +
>>  KBUILD_CFLAGS += -mno-save-restore
>>  KBUILD_CFLAGS += -DCONFIG_PAGE_OFFSET=$(CONFIG_PAGE_OFFSET)
>>
>> --
>
> --
> Thanks,
> ~Nick Desaulniers
Nick Desaulniers July 27, 2021, 8:22 p.m. UTC | #3
On Thu, Jul 22, 2021 at 1:23 AM twd2 <twd2.me@gmail.com> wrote:
>
> Thanks for your advice.  After some investigations, I find minimal code
> that can reproduce the issue I have had.  The bug is filed at
> https://bugs.llvm.org/show_bug.cgi?id=51161.

Perfect! Thank you for filing that. Let's get that fixed up, then
revisit LTO for RISCV.
Palmer Dabbelt Aug. 4, 2021, 8:59 p.m. UTC | #4
On Mon, 19 Jul 2021 14:14:59 PDT (-0700), ndesaulniers@google.com wrote:
> On Mon, Jul 19, 2021 at 1:53 PM Wende Tan <twd2.me@gmail.com> wrote:
>>
>> Pass `-mattr` to `lld` so that it can generate proper machine code.
>> By the way, it seems that `-march` does not work well now.
>>
>> Signed-off-by: Wende Tan <twd2.me@gmail.com>
>> ---
>>  arch/riscv/Makefile | 7 +++++++
>>  1 file changed, 7 insertions(+)
>>
>> diff --git a/arch/riscv/Makefile b/arch/riscv/Makefile
>> index bc74afdbf31e..bbc186a3ab45 100644
>> --- a/arch/riscv/Makefile
>> +++ b/arch/riscv/Makefile
>> @@ -52,9 +52,16 @@ riscv-march-$(CONFIG_ARCH_RV32I)     := rv32ima
>>  riscv-march-$(CONFIG_ARCH_RV64I)       := rv64ima
>>  riscv-march-$(CONFIG_FPU)              := $(riscv-march-y)fd
>>  riscv-march-$(CONFIG_RISCV_ISA_C)      := $(riscv-march-y)c
>> +riscv-mattr-y                          := +m,+a
>> +riscv-mattr-$(CONFIG_FPU)              := $(riscv-mattr-y),+f,+d
>> +riscv-mattr-$(CONFIG_RISCV_ISA_C)      := $(riscv-mattr-y),+c
>>  KBUILD_CFLAGS += -march=$(subst fd,,$(riscv-march-y))
>>  KBUILD_AFLAGS += -march=$(riscv-march-y)
>>
>> +ifeq ($(CONFIG_LD_IS_LLD),y)
>> +       KBUILD_LDFLAGS += -mllvm -mattr=$(riscv-mattr-y)
>
> We should prevent the proliferation of the compiler internal `-mllvm`
> flags from Kbuild Makefiles; these are indicative of LLVM IR specific
> issues with regards to LTO.  Can you please file a bug against LLVM
> that `-mattr` should be encoded in the IR for LTO's sake?

Works for me.  I'll hold off on taking this until the LLVM bug gets 
sorted out and expect a v2 (ideally with the other textual changes Nick 
suggested).

Thanks!

>
> See also:
> commit 0236526d76b8 ("Makefile: lto: Pass -warn-stack-size only on LLD
> < 13.0.0")
> commit 2398ce80152a ("x86, lto: Pass -stack-alignment only on LLD < 13.0.0")
>
>> +endif
>> +
>>  KBUILD_CFLAGS += -mno-save-restore
>>  KBUILD_CFLAGS += -DCONFIG_PAGE_OFFSET=$(CONFIG_PAGE_OFFSET)
>>
>> --
twd2 Aug. 5, 2021, 8:12 p.m. UTC | #5
Sure. I'll send a v2 after the LLVM issue is resolved.

Thanks!
Wende


On 2021/8/5 4:59, Palmer Dabbelt wrote:
> On Mon, 19 Jul 2021 14:14:59 PDT (-0700), ndesaulniers@google.com wrote:
>> On Mon, Jul 19, 2021 at 1:53 PM Wende Tan <twd2.me@gmail.com> wrote:
>>>
>>> Pass `-mattr` to `lld` so that it can generate proper machine code.
>>> By the way, it seems that `-march` does not work well now.
>>>
>>> Signed-off-by: Wende Tan <twd2.me@gmail.com>
>>> ---
>>>  arch/riscv/Makefile | 7 +++++++
>>>  1 file changed, 7 insertions(+)
>>>
>>> diff --git a/arch/riscv/Makefile b/arch/riscv/Makefile
>>> index bc74afdbf31e..bbc186a3ab45 100644
>>> --- a/arch/riscv/Makefile
>>> +++ b/arch/riscv/Makefile
>>> @@ -52,9 +52,16 @@ riscv-march-$(CONFIG_ARCH_RV32I)     := rv32ima
>>>  riscv-march-$(CONFIG_ARCH_RV64I)       := rv64ima
>>>  riscv-march-$(CONFIG_FPU)              := $(riscv-march-y)fd
>>>  riscv-march-$(CONFIG_RISCV_ISA_C)      := $(riscv-march-y)c
>>> +riscv-mattr-y                          := +m,+a
>>> +riscv-mattr-$(CONFIG_FPU)              := $(riscv-mattr-y),+f,+d
>>> +riscv-mattr-$(CONFIG_RISCV_ISA_C)      := $(riscv-mattr-y),+c
>>>  KBUILD_CFLAGS += -march=$(subst fd,,$(riscv-march-y))
>>>  KBUILD_AFLAGS += -march=$(riscv-march-y)
>>>
>>> +ifeq ($(CONFIG_LD_IS_LLD),y)
>>> +       KBUILD_LDFLAGS += -mllvm -mattr=$(riscv-mattr-y)
>>
>> We should prevent the proliferation of the compiler internal `-mllvm`
>> flags from Kbuild Makefiles; these are indicative of LLVM IR specific
>> issues with regards to LTO.  Can you please file a bug against LLVM
>> that `-mattr` should be encoded in the IR for LTO's sake?
>
> Works for me.  I'll hold off on taking this until the LLVM bug gets sorted out and expect a v2 (ideally with the other textual changes Nick suggested).
>
> Thanks!
>
>>
>> See also:
>> commit 0236526d76b8 ("Makefile: lto: Pass -warn-stack-size only on LLD
>> < 13.0.0")
>> commit 2398ce80152a ("x86, lto: Pass -stack-alignment only on LLD < 13.0.0")
>>
>>> +endif
>>> +
>>>  KBUILD_CFLAGS += -mno-save-restore
>>>  KBUILD_CFLAGS += -DCONFIG_PAGE_OFFSET=$(CONFIG_PAGE_OFFSET)
>>>
>>> --
Fangrui Song Aug. 5, 2021, 8:18 p.m. UTC | #6
On Thu, Aug 5, 2021 at 1:12 PM twd2 <twd2.me@gmail.com> wrote:
>
> Sure. I'll send a v2 after the LLVM issue is resolved.
>
> Thanks!
> Wende
>
>
> On 2021/8/5 4:59, Palmer Dabbelt wrote:
> > On Mon, 19 Jul 2021 14:14:59 PDT (-0700), ndesaulniers@google.com wrote:
> >> On Mon, Jul 19, 2021 at 1:53 PM Wende Tan <twd2.me@gmail.com> wrote:
> >>>
> >>> Pass `-mattr` to `lld` so that it can generate proper machine code.
> >>> By the way, it seems that `-march` does not work well now.
> >>>
> >>> Signed-off-by: Wende Tan <twd2.me@gmail.com>
> >>> ---
> >>>  arch/riscv/Makefile | 7 +++++++
> >>>  1 file changed, 7 insertions(+)
> >>>
> >>> diff --git a/arch/riscv/Makefile b/arch/riscv/Makefile
> >>> index bc74afdbf31e..bbc186a3ab45 100644
> >>> --- a/arch/riscv/Makefile
> >>> +++ b/arch/riscv/Makefile
> >>> @@ -52,9 +52,16 @@ riscv-march-$(CONFIG_ARCH_RV32I)     := rv32ima
> >>>  riscv-march-$(CONFIG_ARCH_RV64I)       := rv64ima
> >>>  riscv-march-$(CONFIG_FPU)              := $(riscv-march-y)fd
> >>>  riscv-march-$(CONFIG_RISCV_ISA_C)      := $(riscv-march-y)c
> >>> +riscv-mattr-y                          := +m,+a
> >>> +riscv-mattr-$(CONFIG_FPU)              := $(riscv-mattr-y),+f,+d
> >>> +riscv-mattr-$(CONFIG_RISCV_ISA_C)      := $(riscv-mattr-y),+c
> >>>  KBUILD_CFLAGS += -march=$(subst fd,,$(riscv-march-y))
> >>>  KBUILD_AFLAGS += -march=$(riscv-march-y)
> >>>
> >>> +ifeq ($(CONFIG_LD_IS_LLD),y)
> >>> +       KBUILD_LDFLAGS += -mllvm -mattr=$(riscv-mattr-y)
> >>
> >> We should prevent the proliferation of the compiler internal `-mllvm`
> >> flags from Kbuild Makefiles; these are indicative of LLVM IR specific
> >> issues with regards to LTO.  Can you please file a bug against LLVM
> >> that `-mattr` should be encoded in the IR for LTO's sake?
> >
> > Works for me.  I'll hold off on taking this until the LLVM bug gets sorted out and expect a v2 (ideally with the other textual changes Nick suggested).
> >
> > Thanks!

For GNU ld compatibility: -mllvm -mattr=xxx => -plugin-opt=-mattr=xxx

Which can allow us to drop the CONFIG_LD_IS_LLD condition.
(gold doesn't support RISC-V yet.)

Note: LLVM LTO + GNU ld support is experimental and has very little testing.
To actually make GNU ld work, we probably also need to teach GNU nm
and ar to specify --plugin LLVMgold.so
https://llvm.org/docs/GoldPlugin.html


> >>
> >> See also:
> >> commit 0236526d76b8 ("Makefile: lto: Pass -warn-stack-size only on LLD
> >> < 13.0.0")
> >> commit 2398ce80152a ("x86, lto: Pass -stack-alignment only on LLD < 13.0.0")
> >>
> >>> +endif
> >>> +
> >>>  KBUILD_CFLAGS += -mno-save-restore
> >>>  KBUILD_CFLAGS += -DCONFIG_PAGE_OFFSET=$(CONFIG_PAGE_OFFSET)
> >>>
> >>> --
>
> --
> You received this message because you are subscribed to the Google Groups "Clang Built Linux" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to clang-built-linux+unsubscribe@googlegroups.com.
> To view this discussion on the web visit https://groups.google.com/d/msgid/clang-built-linux/a6615b69-d99f-220b-d980-1f5f40ee2999%40gmail.com.
twd2 Aug. 5, 2021, 9:05 p.m. UTC | #7
Hi,

I think we even don't need this option after the issue I filed is fixed.

Thanks!
Wende


On 2021/8/6 4:18, Fāng-ruì Sòng wrote:
> On Thu, Aug 5, 2021 at 1:12 PM twd2 <twd2.me@gmail.com> wrote:
>> Sure. I'll send a v2 after the LLVM issue is resolved.
>>
>> Thanks!
>> Wende
>>
>>
>> On 2021/8/5 4:59, Palmer Dabbelt wrote:
>>> On Mon, 19 Jul 2021 14:14:59 PDT (-0700), ndesaulniers@google.com wrote:
>>>> On Mon, Jul 19, 2021 at 1:53 PM Wende Tan <twd2.me@gmail.com> wrote:
>>>>> Pass `-mattr` to `lld` so that it can generate proper machine code.
>>>>> By the way, it seems that `-march` does not work well now.
>>>>>
>>>>> Signed-off-by: Wende Tan <twd2.me@gmail.com>
>>>>> ---
>>>>>  arch/riscv/Makefile | 7 +++++++
>>>>>  1 file changed, 7 insertions(+)
>>>>>
>>>>> diff --git a/arch/riscv/Makefile b/arch/riscv/Makefile
>>>>> index bc74afdbf31e..bbc186a3ab45 100644
>>>>> --- a/arch/riscv/Makefile
>>>>> +++ b/arch/riscv/Makefile
>>>>> @@ -52,9 +52,16 @@ riscv-march-$(CONFIG_ARCH_RV32I)     := rv32ima
>>>>>  riscv-march-$(CONFIG_ARCH_RV64I)       := rv64ima
>>>>>  riscv-march-$(CONFIG_FPU)              := $(riscv-march-y)fd
>>>>>  riscv-march-$(CONFIG_RISCV_ISA_C)      := $(riscv-march-y)c
>>>>> +riscv-mattr-y                          := +m,+a
>>>>> +riscv-mattr-$(CONFIG_FPU)              := $(riscv-mattr-y),+f,+d
>>>>> +riscv-mattr-$(CONFIG_RISCV_ISA_C)      := $(riscv-mattr-y),+c
>>>>>  KBUILD_CFLAGS += -march=$(subst fd,,$(riscv-march-y))
>>>>>  KBUILD_AFLAGS += -march=$(riscv-march-y)
>>>>>
>>>>> +ifeq ($(CONFIG_LD_IS_LLD),y)
>>>>> +       KBUILD_LDFLAGS += -mllvm -mattr=$(riscv-mattr-y)
>>>> We should prevent the proliferation of the compiler internal `-mllvm`
>>>> flags from Kbuild Makefiles; these are indicative of LLVM IR specific
>>>> issues with regards to LTO.  Can you please file a bug against LLVM
>>>> that `-mattr` should be encoded in the IR for LTO's sake?
>>> Works for me.  I'll hold off on taking this until the LLVM bug gets sorted out and expect a v2 (ideally with the other textual changes Nick suggested).
>>>
>>> Thanks!
> For GNU ld compatibility: -mllvm -mattr=xxx => -plugin-opt=-mattr=xxx
>
> Which can allow us to drop the CONFIG_LD_IS_LLD condition.
> (gold doesn't support RISC-V yet.)
>
> Note: LLVM LTO + GNU ld support is experimental and has very little testing.
> To actually make GNU ld work, we probably also need to teach GNU nm
> and ar to specify --plugin LLVMgold.so
> https://llvm.org/docs/GoldPlugin.html
>
>
>>>> See also:
>>>> commit 0236526d76b8 ("Makefile: lto: Pass -warn-stack-size only on LLD
>>>> < 13.0.0")
>>>> commit 2398ce80152a ("x86, lto: Pass -stack-alignment only on LLD < 13.0.0")
>>>>
>>>>> +endif
>>>>> +
>>>>>  KBUILD_CFLAGS += -mno-save-restore
>>>>>  KBUILD_CFLAGS += -DCONFIG_PAGE_OFFSET=$(CONFIG_PAGE_OFFSET)
>>>>>
>>>>> --
>> --
>> You received this message because you are subscribed to the Google Groups "Clang Built Linux" group.
>> To unsubscribe from this group and stop receiving emails from it, send an email to clang-built-linux+unsubscribe@googlegroups.com.
>> To view this discussion on the web visit https://groups.google.com/d/msgid/clang-built-linux/a6615b69-d99f-220b-d980-1f5f40ee2999%40gmail.com.
>
diff mbox series

Patch

diff --git a/arch/riscv/Makefile b/arch/riscv/Makefile
index bc74afdbf31e..bbc186a3ab45 100644
--- a/arch/riscv/Makefile
+++ b/arch/riscv/Makefile
@@ -52,9 +52,16 @@  riscv-march-$(CONFIG_ARCH_RV32I)	:= rv32ima
 riscv-march-$(CONFIG_ARCH_RV64I)	:= rv64ima
 riscv-march-$(CONFIG_FPU)		:= $(riscv-march-y)fd
 riscv-march-$(CONFIG_RISCV_ISA_C)	:= $(riscv-march-y)c
+riscv-mattr-y				:= +m,+a
+riscv-mattr-$(CONFIG_FPU)		:= $(riscv-mattr-y),+f,+d
+riscv-mattr-$(CONFIG_RISCV_ISA_C)	:= $(riscv-mattr-y),+c
 KBUILD_CFLAGS += -march=$(subst fd,,$(riscv-march-y))
 KBUILD_AFLAGS += -march=$(riscv-march-y)
 
+ifeq ($(CONFIG_LD_IS_LLD),y)
+	KBUILD_LDFLAGS += -mllvm -mattr=$(riscv-mattr-y)
+endif
+
 KBUILD_CFLAGS += -mno-save-restore
 KBUILD_CFLAGS += -DCONFIG_PAGE_OFFSET=$(CONFIG_PAGE_OFFSET)