Add support for building AArch64 with Clang #374
Conversation
|
Unlike #322 this does not add a LLVM= flag to change the default it just allows building if CC/LD are explicitly set to LLVM tools. |
arichardson
changed the title
|
ping? |
|
These look pretty good, hoping to get a look at it this weekend. Mostly
need to make sure it works well with GCC and see about trying to make it
more generic. There have been multiple attempts to compile with clang over
the years but they're generally fairly limited.
… Message ID: ***@***.***>
|
Thanks! I did build with GCC as well, and ran the build all script so hopefully it all works. Once these changes are in I'll submit another PR with changes to allow building for RISC-V with clang. |
make/macros.mk
Outdated
| @@ -58,3 +58,17 @@ define MAKECONFIGHEADER | |||
| echo $3 >> $1.tmp; \ | |||
| $(call TESTANDREPLACEFILE,$1.tmp,$1) | |||
| endef | |||
|
|
|||
| check_compiler_flag = $(shell $(CC) -c -xc /dev/null -o /dev/null $(1) 2>/dev/null && echo yes || echo no) | |||
| check_warning_supported = $(call check_compiler_flag,-Werror=unknown-warning-option $(1)) | |||
There was a problem hiding this comment.
This seems problematic, since gcc doesn't seem to support -Werror=unknown-warning-option. It causes it always to choose that the switch is not supported.
There was a problem hiding this comment.
Hmm I thought I tested this with GCC but it does look like the docs don't mention it. Will try to update this to only add the flag for clang
There was a problem hiding this comment.
Oh thanks. Been trying to figure out a workaround, but seems that the correct solution is to either only test for clang or to have two paths, based on whether or not the -Werror=unknown-warning-option switch seems to work.
If it's not functional it seems that gcc has some strange behavior with regards to 'no-' switches. If you pass a -Wno- switch that it doesn't recognize it'll accept it unless it gets a compile failure, in which case it then tells you of the weird switch. This seems to be intentional, to allow for backwards compatibility with older versions of the compiler. documented at https://gcc.gnu.org/onlinedocs/gcc/Warning-Options.html#:~:text=When%20an%20unrecognized%20warning%20option%20is%20requested
There was a problem hiding this comment.
I started to go down the route of just fixing the code that the -Wno- switches were working around, but they're kinda annoying. The big warning failure is in some apis that have the NONNULL attribute set for particular arguments. Inside the function if you test for null anyway (which is a generally good idea, IMO) you get a warning. Kinda annoying, the main reason i had the NONNULL set was to catch obvious null cases at the caller, but not to cause it to elide all checks in the function itself. Really probably means the nonnull attribute is largely a waste of time.
There was a problem hiding this comment.
Sorry for the delay here - I've pushed a new version that works for both GCC and Clang in my testing.
There was a problem hiding this comment.
Regarding the attribute((nonnull)) that warning probably does make sense since the way that attribute works is that the compiler will elide all NULL checks. Unlike this attribute the Clang _Nonnull/_Nullable annotations only affect compiler warnings and not code generation.
This GCC extension is not supported by clang and in this case is not necessary. Move the function to the top level to avoid the syntax error.
Clang sets the defines for GCC 4.2.1, so we have to check __clang__ for these macros in addition to the GCC version.
Clang does not accept this .if condition since phys_offset is a register alias and not an absolute expression. We can keep these two instructions here if the argument is zero since the result will be the same. Additionally, this macro is only called once and always passes a non-zero argument. If more calls are added in the future and avoiding these two instructions just before a loop is really important, we could use `.ifnc \phys_offset,0` instead, but that looks rather obscure to me.
The LLD linker does not allow joined short arguments, so split -dT <script> into -d -T <script>.
Currently, clang does not support the -Wno-nonnull-compare and -Wmaybe-uninitialized warning flags so this adds lots of unknown warning flag output for each compile job when not using GCC. This commit adds a makefile macro to check for supported warning flags and only adds them if the compiler actually supports them.
Clang incorrectly diagnoses msr operations as need a 64-bit operand even if the underlying register is actually 32 bits. Silence this warning. There are quite a few occurrences of this warning so I opted to add the -Wno-flag instead of wrapping all callsites in `#pragma clang diagnostic ignored -Wasm-operand-widths`.
|
LGTM, sorry for the delay! |
These changes allow me to build with
make CC="/usr/bin/clang -target aarch64-unknown-elf" LD=/usr/bin/ld.lld TOOLCHAIN_PREFIX=/usr/bin/ qemu-virt-arm64-test