Thread (19 messages) 19 messages, 3 authors, 2021-07-10
DORMANTno replies

[PATCH v2 1/2] Makefile: Support building with Clang and LLVM binutils

From: Bin Meng <hidden>
Date: 2021-07-10 05:37:11

On Fri, Jul 9, 2021 at 11:08 PM Jessica Clarke [off-list ref] wrote:
On 9 Jul 2021, at 15:42, Bin Meng [off-list ref] wrote:
quoted
On Fri, Jul 9, 2021 at 10:34 PM Jessica Clarke [off-list ref] wrote:
quoted
On 9 Jul 2021, at 11:30, Bin Meng [off-list ref] wrote:
quoted
On Fri, Jul 9, 2021 at 3:37 PM Bin Meng [off-list ref] wrote:
quoted
On Fri, Jul 9, 2021 at 2:40 PM Bin Meng [off-list ref] wrote:
quoted
On Fri, Jul 9, 2021 at 11:31 AM Jessica Clarke [off-list ref] wrote:
quoted
On 9 Jul 2021, at 04:11, Bin Meng [off-list ref] wrote:
quoted
On Fri, Jul 9, 2021 at 10:35 AM Jessica Clarke [off-list ref] wrote:
quoted
On 9 Jul 2021, at 02:50, Bin Meng [off-list ref] wrote:
quoted
On Fri, Jul 9, 2021 at 1:51 AM Jessica Clarke [off-list ref] wrote:
quoted
This is intended to mirror the Linux kernel. Building with CC=clang will
use Clang as the compiler but default to using the existing binutils.
Building with LLVM=1 will default to using Clang and LLVM binutils.

Whilst GCC will accept the -N linker option and forward it on to the
linker, Clang will not, and so in order to support both compilers we
must use -Wl, to forward it to the linker as is required for most other
linker options.

Signed-off-by: Jessica Clarke <redacted>
---
Makefile  | 57 +++++++++++++++++++++++++++++++++++++++++++++++++------
README.md | 39 +++++++++++++++++++++++++++++++++++--
2 files changed, 88 insertions(+), 8 deletions(-)
diff --git a/Makefile b/Makefile
index 6b64205..3fe8153 100644
--- a/Makefile
+++ b/Makefile
@@ -76,26 +76,54 @@ OPENSBI_VERSION_MINOR=`grep "define OPENSBI_VERSION_MINOR" $(include_dir)/sbi/sb
OPENSBI_VERSION_GIT=$(shell if [ -d $(src_dir)/.git ]; then git describe 2> /dev/null; fi)

# Setup compilation commands
+ifneq ($(LLVM),)
+CC             =       clang
+AR             =       llvm-ar
+LD             =       ld.lld
+OBJCOPY                =       llvm-objcopy
+else
ifdef CROSS_COMPILE
CC             =       $(CROSS_COMPILE)gcc
-CPP            =       $(CROSS_COMPILE)cpp
AR             =       $(CROSS_COMPILE)ar
LD             =       $(CROSS_COMPILE)ld
OBJCOPY                =       $(CROSS_COMPILE)objcopy
else
CC             ?=      gcc
-CPP            ?=      cpp
AR             ?=      ar
LD             ?=      ld
OBJCOPY                ?=      objcopy
endif
+endif
+CPP            =       $(CC) -E
AS             =       $(CC)
DTC            =       dtc

-# Guess the compillers xlen
-OPENSBI_CC_XLEN := $(shell TMP=`$(CC) -dumpmachine | sed 's/riscv\([0-9][0-9]\).*/\1/'`; echo $${TMP})
+ifneq ($(shell $(CC) --version 2>&1 | head -n 1 | grep clang),)
+CC_IS_CLANG    =       y
+else
+CC_IS_CLANG    =       n
+endif
+
+ifneq ($(shell $(LD) --version 2>&1 | head -n 1 | grep LLD),)
+LD_IS_LLD      =       y
+else
+LD_IS_LLD      =       n
+endif
+
+ifeq ($(CC_IS_CLANG),y)
+ifneq ($(CROSS_COMPILE),)
+CLANG_TARGET   =       -target $(notdir $(CROSS_COMPILE:%-=%))
It's odd that when we use full LLVM toolchains we still need to
specify CROSS_COMPILE in order to only guess the "--target" value.
This can be written directly to --target=riscv64 / riscv32 depending
OPENSBI_CC_XLEN
Hm, that?s true. To be honest, the fact that OpenSBI defaults to an
unprefixed GCC is rather dubious, that?ll likely be for either
riscv64-linux-gnu or riscv64-unknown-freebsd and thus not suitable
(there can be subtle issues with using the wrong one, though RISC-V is
more uniform than some other architectures) so should probably grab
XLEN from the default GCC and then look for riscvXLEN-unknown-elf-gcc
if CROSS_COMPILE wasn?t specified. I can at least make it do something
like that for Clang (keep this code for CROSS_COMPILE not empty, then
add a -target riscvXLEN-unknown-elf after guessing the XLEN if
CROSS_COMPILE was empty).
I don't think we should over-complicate things. Passing riscv64 /
riscv32 to --target is enough for OpenSBI when building with clang.
We should use the right triple though, and doing so requires knowing XLEN.
quoted
quoted
quoted
quoted
+endif
+endif
+
+# Guess the compiler's XLEN
+OPENSBI_CC_XLEN := $(shell TMP=`$(CC) $(CLANG_TARGET) -dumpmachine | sed 's/riscv\([0-9][0-9]\).*/\1/'`; echo $${TMP})
+
+# Guess the compiler's ABI and ISA
+ifneq ($(CC_IS_CLANG),y)
OPENSBI_CC_ABI := $(shell TMP=`$(CC) -v 2>&1 | sed -n 's/.*\(with\-abi=\([a-zA-Z0-9]*\)\).*/\2/p'`; echo $${TMP})
OPENSBI_CC_ISA := $(shell TMP=`$(CC) -v 2>&1 | sed -n 's/.*\(with\-arch=\([a-zA-Z0-9]*\)\).*/\2/p'`; echo $${TMP})
+endif

# Setup platform XLEN
ifndef PLATFORM_RISCV_XLEN
@@ -194,7 +222,11 @@ else
endif

# Setup compilation commands flags
-GENFLAGS       =       -I$(platform_src_dir)/include
+ifeq ($(CC_IS_CLANG),y)
+GENFLAGS       +=      $(CLANG_TARGET)
+GENFLAGS       +=      -Wno-unused-command-line-argument
+endif
+GENFLAGS       +=      -I$(platform_src_dir)/include
GENFLAGS       +=      -I$(include_dir)
ifneq ($(OPENSBI_VERSION_GIT),)
GENFLAGS       +=      -DOPENSBI_VERSION_GIT="\"$(OPENSBI_VERSION_GIT)\""
@@ -208,6 +240,9 @@ CFLAGS              +=      -fno-omit-frame-pointer -fno-optimize-sibling-calls
CFLAGS         +=      -mno-save-restore -mstrict-align
CFLAGS         +=      -mabi=$(PLATFORM_RISCV_ABI) -march=$(PLATFORM_RISCV_ISA)
CFLAGS         +=      -mcmodel=$(PLATFORM_RISCV_CODE_MODEL)
+ifeq ($(LD_IS_LLD),y)
+CFLAGS         +=      -mno-relax
+endif
CFLAGS         +=      $(GENFLAGS)
CFLAGS         +=      $(platform-cflags-y)
CFLAGS         +=      -fno-pie -no-pie
@@ -222,18 +257,28 @@ ASFLAGS           +=      -fno-omit-frame-pointer -fno-optimize-sibling-calls
ASFLAGS                +=      -mno-save-restore -mstrict-align
ASFLAGS                +=      -mabi=$(PLATFORM_RISCV_ABI) -march=$(PLATFORM_RISCV_ISA)
ASFLAGS                +=      -mcmodel=$(PLATFORM_RISCV_CODE_MODEL)
+ifeq ($(LD_IS_LLD),y)
+ASFLAGS                +=      -mno-relax
+endif
ASFLAGS                +=      $(GENFLAGS)
ASFLAGS                +=      $(platform-asflags-y)
ASFLAGS                +=      $(firmware-asflags-y)

ARFLAGS                =       rcs

-ELFFLAGS       +=      -Wl,--build-id=none -N -static-libgcc -lgcc
+ifeq ($(LD_IS_LLD),y)
+ELFFLAGS       +=      -fuse-ld=lld
+endif
+ELFFLAGS       +=      -Wl,--build-id=none -Wl,-N -static-libgcc -lgcc
ELFFLAGS       +=      $(platform-ldflags-y)
ELFFLAGS       +=      $(firmware-ldflags-y)

MERGEFLAGS     +=      -r
+ifeq ($(LD_IS_LLD),y)
+MERGEFLAGS     +=      -b elf
+else
MERGEFLAGS     +=      -b elf$(PLATFORM_RISCV_XLEN)-littleriscv
+endif
MERGEFLAGS     +=      -m elf$(PLATFORM_RISCV_XLEN)lriscv

DTSCPPFLAGS    =       $(CPPFLAGS) -nostdinc -nostdlib -fno-builtin -D__DTS__ -x assembler-with-cpp
diff --git a/README.md b/README.md
index 03c02fb..e97dcc4 100644
--- a/README.md
+++ b/README.md
@@ -96,8 +96,13 @@ Required Toolchain
------------------
OpenSBI can be compiled natively or cross-compiled on a x86 host. For
-cross-compilation, you can build your own toolchain or just download
-a prebuilt one from the [Bootlin toolchain repository].
+cross-compilation, you can build your own toolchain, download a prebuilt one
+from the [Bootlin toolchain repository] or install a distribution-provided
+toolchain; if you opt to use LLVM/Clang, most distribution toolchains will
+support cross-compiling for RISC-V using the same toolchain as your native
+LLVM/Clang toolchain due to LLVM's ability to support multiple backends in the
+same binary, so is often an easy way to obtain a working cross-compilation
+toolchain.

Please note that only a 64-bit version of the toolchain is available in
the Bootlin toolchain repository for now.
@@ -202,6 +207,36 @@ export PLATFORM_RISCV_XLEN=32
will generate 32-bit OpenSBI images. And vice vesa.

+Building with Clang/LLVM
+------------------------
+
+OpenSBI can also be built with Clang/LLVM. To build with just Clang but keep
+the default binutils (which will still use the *CROSS_COMPILE* prefix if
+defined), override the *CC* make variable with:
+```
+make CC=clang
Testing with the pre-built official LLVM 12 release for Ubuntu 20.04
[1], along with bootlin pre-built cross-compile GCC [2]:

There are 3 build warnings when using the default binutils:

AS        platform/generic/firmware/fw_dynamic.o
firmware/fw_base.S:557:2: warning: fw_platform_init changed binding to STB_WEAK
.weak fw_platform_init
^

AS        platform/generic/firmware/fw_jump.o
firmware/fw_base.S:557:2: warning: fw_platform_init changed binding to STB_WEAK
.weak fw_platform_init
^

AS        platform/generic/firmware/fw_payload.o
firmware/fw_base.S:557:2: warning: fw_platform_init changed binding to STB_WEAK
.weak fw_platform_init
^
Yeah, those are known. They?re harmless and easy to fix (just delete
the .globl lines, as the two are mutually exclusive), so I didn?t
include it as part of this patch series, LLVM 12 just got stricter
about this as it?s dodgy code.
Please include a patch to fix these warnings as part of this series.
We should not allow any building warnings to happen.
Ok, that?s a trivial patch to include in v3.
quoted
quoted
quoted
And several warnings from the GNU linker:

/share/toolchains/riscv64/bin/riscv64-linux-ld: warning: library
search path "/lib/../lib64" is unsafe for cross-compilation
/share/toolchains/riscv64/bin/riscv64-linux-ld: warning: library
search path "/usr/lib/../lib64" is unsafe for cross-compilation
/share/toolchains/riscv64/bin/riscv64-linux-ld: warning: library
search path "/lib" is unsafe for cross-compilation
/share/toolchains/riscv64/bin/riscv64-linux-ld: warning: library
search path "/usr/lib" is unsafe for cross-compilation
Hm, indeed, Clang likes to add some host directories to the search
path, seemingly only for RISC-V. RISC-V?s bare-metal toolchain driver
is a bit quirky due to all the multilib stuff the GNU world decided to
introduce so that?s a bug. They should be harmless though given we
don?t pass -lfoo, when linking in SBI libraries we pass the path.
Is it a bug of clang, or GNU ld? Could you please file a bug report to
get this issue tracked (if there isn't one already), and mentioned in
the commit message?
Clang, though I suspect some of these are a result of you using the
wrong triple (see below).
quoted
quoted
quoted
The generated fw_dynamic firmware image does not boot on QEMU 'virt'.
Initial debugging shows that it returns SBI_EINVAL in
sanitize_domain().
My pure LLVM=1-built fw_dynamic binary on FreeBSD boots U-Boot just fine
(-M virt -m 2048 -nographic -bios fw_dynamic.elf -kernel u-boot), as
does an LLVM=1 with LD overridden to be GNU ld 2.33.1 (also on FreeBSD)
so I don?t know what?s going on there. Though I did discover that
-fuse-ld=bfd is needed for the non-LLD case otherwise Clang won?t pick
up an ld.bfd intended to override a system LLD. So my guess is that one
or other of the binaries you downloaded has broken something. Does
using LLD work? If you tell me exactly what you ran I can try it on a
Linux machine myself and see if I can reproduce it.
Switching to full LLVM build does not build for me.

ELF       platform/generic/firmware/fw_dynamic.elf
ld.lld: error: can't create dynamic relocation R_RISCV_64 against
symbol: _fw_start in readonly segment; recompile object files with
-fPIC or pass '-Wl,-z,notext' to allow text relocations in the output
quoted
quoted
quoted
defined in opensbi/build/platform/generic/firmware/fw_dynamic.elf.ld:8
referenced by fw_base.S:502 (opensbi/firmware/fw_base.S:502)
           opensbi/build/platform/generic/firmware/fw_dynamic.o:(.entry+0x3A0)
ld.lld: error: can't create dynamic relocation R_RISCV_64 against
symbol: _fw_reloc_end in readonly segment; recompile object files with
-fPIC or pass '-Wl,-z,notext' to allow text relocations in the output
quoted
quoted
quoted
defined in opensbi/build/platform/generic/firmware/fw_dynamic.elf.ld:92
referenced by fw_base.S:502 (opensbi/firmware/fw_base.S:502)
           opensbi/build/platform/generic/firmware/fw_dynamic.o:(.entry+0x3B0)
clang-12: error: linker command failed with exit code 1 (use -v to see
invocation)
make: *** [Makefile:396:
opensbi/build/platform/generic/firmware/fw_dynamic.elf] Error 1
This is a result of you using the wrong triple, though it?s also a sign
that we do slightly bogus things. For FW_PIC we link with -pie, but
that doesn?t make sense for bare-metal. We also don?t link with
-static, but it?s the default (and only) supported thing for bare-metal
targets. If you use a bare-metal triple then the -pie gets ignored (we
should probably remove it from objects.mk) and -static is implied. If
you use a Linux triple then the -pie gets honoured and the lack of
-static means it defaults to dynamic linking, so LLD rightly complains
about the fact that fw_base.S is creating a pointer in a read-only
section that requires run-time relocation. I don?t know why you don?t
see the same thing with GNU ld but it?s probably just silently allowing
it and leaving it to crash at run time.
I am confused. Did you mean "riscv64-linux-" is a bare-metal triple? I
thought "riscv64-unknown-elf-" is one bare-metal triple, and "-target
riscv64" is too.

I changed to pass "-target riscv64" to clang, and now it builds and
boots fine with LLVM=1 case.
I further looked at this one. Even passing "-target riscv64" leads to
a successful build and boot to U-Boot, I checked the generated ELF
image and found the .rela.dyn section is empty.
By hardcoding "-target riscv64-unknown-elf", and

$ make LLVM=1 PLATFORM=generic

I got the same result as "-target riscv64". It builds and boots, but
with an empty.rela.dyn
If unspecified, the vendor and OS default to unknown and elf
respectively, so the two are entirely equivalent, but I felt it best to
be explicit, especially since anyone without an LLVM background reading
the Makefile might be confused.
Ah, thanks! I am not aware of this clang triple convention :)
quoted
This is a bare-metal binary, of course .rela.dyn is going to be empty,
there?s no run-time linker to do any relocations. How on earth is the
FW_PIC support supposed to work?
See commit 0f20e8adcf42 ("firmware: Support position independent
execution") for the FW_PIC changes. Basically OpenSBI relocates itself
at the very beginning of the boot phase if building with -fpie.
Sure, but riscv64-unknown-elf PIEs do not exist, same as
aarch64-unknown-elf or any other triple.
quoted
quoted
That looks utterly broken to me. Is it *intended* to be abusing riscv64-linux-gnu? Because that?s just plain wrong...
I am not sure what you mean by "abusing riscv64-linux-gnu"? But this
-fpie stuff is perfectly okay and commonly used by every architecture
in the U-Boot world.
Ewwwww. That is horrendous. I think other toolchain authors would be
horrified to know this is happening. This has never been supported and
(as an LLVM developer) there are all manner of subtle issues here that
don?t matter until they do. It is *not* ok, it?s an awful awful hack
that someone should have fixed by adding static PIE support for
bare-metal binaries.
U-Boot's utilization of PIE to do self-relocation has been there for
at least 10 years. I believe some toolchain developers are aware of
that.
But then I don?t understand why GNU ld isn?t
giving errors about the relocations in read-only sections. Is its error
checking just broken? GNU ld normal gives errors for that kind of
thing, but maybe it?s just broken on RISC-V. Building myself with a
FreeBSD triple (on FreeBSD) I can reproduce the errors, and using GNU
ld instead of LLD I don?t see them, but inspecting the binary I see
there are relocations against .text and it?s become writeable but
there?s no DT_TEXTREL in the output, so GNU ld is definitely broken
somehow here (even -Wl,-z,text to forcefully disable text relocations
in case it?s just a different default behaves no differently, though
even if it were the default it still must emit a DT_TEXTREL, which it
doesn?t).
If you think this is a bug of GNU ld, will you file a bug report for that?
By ?abuse?, I mean: OpenSBI is neither a Linux kernel nor something
running in Linux userspace, therefore using a riscv64-linux-gnu triple
is categorically wrong.
So "riscv64-linux" triple can be used for Linux kernel, why can't it
be used for OpenSBI? The Linux kernel itself is also "bare-metal" to
some extent.

Using your example in your follow-up thread, "As one example,
(__builtin_)__clear_cache for a bare-metal target will emit a fence.i
but for a Linux target will emit an inline syscall", but this is also
wrong for Linux kernel by emitting an inline syscall.
However, people are lazy so reuse their Linux
toolchains rather than building a whole new bare-metal toolchain to
compile bare-metal applications. FW_PIC support then gets added and
appears to work because people are incorrectly using Linux toolchains,
and anyone using a bare-metal toolchain still sees the build succeeding
and the binary boot just fine *unless they load it to a different
location, at which point there are no relocations for OpenSBI to
process and relocate itself*. Thus, using a Linux toolchain becomes a
*requirement*, despite the fact that the README only documents
riscv64-unknown-elf as an expected prefix and a Linux toolchain should
never have been permitted in the first place.
Yeah, using a bare-metal toolchain is probably broken in this FW_PIC
case, and that's something we should fix.
In my opinion, FW_PIC as it stands is broken by design without proper
toolchain support. If you want to still support it with Linux toolchains
despite the fact that shouldn?t be a thing you?re allowed to use then
fine, but if a bare-metal toolchain is used OpenSBI should error out if
you ever try to enable FW_PIC (and default it to off) because that will
silently produce a non-PIE binary.
Regards,
Bin
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help