drm-kmod: Build and run on ARM64 with 16.0-CURRENT #402
Conversation
I've successfully built and run graphics/drm-66-kmod with those patches on my Honeycomb LX2 with Radeon RX 460. With these patches 6.6-lts could be a substitute for anyone who relied upon retired graphics/drm-510-kmod.
|
|
||
| .if ${MACHINE_ARCH} == "aarch64" | ||
|
|
||
| dcn_calcs.o: dcn_calcs.c |
There was a problem hiding this comment.
Can we avoid these duplicate rules? This is to remove -mgeneral-regs-only from all of these?
There was a problem hiding this comment.
I guess so. I decided to act safely and removed -mgeneral-regs-only from those file clang complained about + dcn302_fpu.c as a special case explained above.
There was a problem hiding this comment.
I mean just arrange to have CFLAGS set appropriately rather than repeating the rules. I suspect that all of these correspond to the ${FPUFLAGS} additions for x86 above, like
CFLAGS.dcn_calcs.c= ${FPUFLAGS} -mstack-alignment=4
E.g. is there -mno-general-regs-only that we could set in ${FPUFLAGS} on arm64 to override -mgeneral-regs-only? Or something along those lines.
|
Btw, as @bzfbd pointed out in a private discussion, Honeycomb LX2 has Cortex-A72 which are cache coherent which simplifies testing. But I don't have any other ARM64 with PCIe for a discrete GPU to test at the moment. |
| CFLAGS.rc_calc_dpi.c= ${FPUFLAGS} -mstack-alignment=4 | ||
| CFLAGS.rc_calc_fpu.c= ${FPUFLAGS} -mstack-alignment=4 | ||
|
|
||
| .if ${MACHINE_ARCH} == "aarch64" |
There was a problem hiding this comment.
Almost all of those files below use "double" as a type, i.e. clang blames ABI:
requires 'double' type support, but ABI 'aapcs' does not support it
but it's probably misleading diagnostics: llvm/llvm-project#90001
| dcn301_fpu.o: dcn301_fpu.c | ||
| ${CC} ${STATIC_CFLAGS} ${CFLAGS:N-mgeneral-regs-only} -c ${.IMPSRC} \ | ||
| -o ${.TARGET} | ||
| dcn302_fpu.o: dcn302_fpu.c |
There was a problem hiding this comment.
Compiling this one with -mgeneral-regs-only causes clang to produce machine code which requires __floatundidf to convert integer to double if I understand correctly, i.e.
kldload amdgpu
complains early about the undefined symbol.
| #if defined(CONFIG_X86) | ||
| return boot_cpu_has(X86_FEATURE_HYPERVISOR); | ||
| #elif defined(CONFIG_ARM64) | ||
| return !is_kernel_in_hyp_mode(); |
There was a problem hiding this comment.
is_kernel_in_hyp_mode isn't available in the LinuxKPI. I didn't know how to substitute it properly and removed for now. Any suggestions are welcome.
| #include <xen/xen.h> | ||
|
|
||
| /* XXX-DSL: For FreeBSD and LinuxKPI only! */ | ||
| #if defined(CONFIG_ARM64) |
There was a problem hiding this comment.
It'd be even more strict than "CONFIG_ARM64", I guess. I'm not sure.
|
|
||
| .if ${MACHINE_ARCH} == "aarch64" | ||
|
|
||
| dcn_calcs.o: dcn_calcs.c |
There was a problem hiding this comment.
I guess so. I decided to act safely and removed -mgeneral-regs-only from those file clang complained about + dcn302_fpu.c as a special case explained above.
|
Most (all ?) those are in the master branch. I started to fix stuff for arm64 a few weeks ago and if we want to have arm64 support again we should start by fixing the master branch. |
I've tried to apply the same patches on drm-latest-kmod, but |
Yes it behave the same on me, I couldn't debug more as my honeycomb decided to died on me (still need to have a look at what's happening). But at least the patches should be cherry-picked from the main branch to 6.6 |
3 participants
I've successfully built and run graphics/drm-66-kmod with those patches on my Honeycomb LX2 with Radeon RX 460. With these patches 6.6-lts could be a substitute for anyone who relied upon retired graphics/drm-510-kmod.
This is how KDE Plasma desktop looks like on my Honeycomb LX2: