Excerpts from Arnd Bergmann's message of February 24, 2022 6:55 pm:
On Thu, Feb 24, 2022 at 6:05 AM Nicholas Piggin npiggin@gmail.com wrote:
Excerpts from Nicholas Piggin's message of February 24, 2022 12:54 pm:
Not sure on the outlook for GCC fix. Either way unfortunately we have toolchains in the wild now that will explode, so we might have to take your patches for the time being.
Perhaps not... Here's a hack that seems to work around the problem.
The issue of removing -many from the kernel and replacing it with appropriate architecture versions is an orthogonal one (that we should do). Either way this hack should be able to allow us to do that as well, on these problem toolchains.
But for now it just uses -many as the trivial regression fix to get back to previous behaviour.
I don't think the previous behavior is what you want to be honest.
-many isn't good but that's what we're using and that is still what we're using upstream on any other toolchain that doesn't have these issues. Including the next binutils version that will ignore the initial .machine directive for 64s.
Neither of these approaches solves that. At least for 64s that is passing -Wa,-many down already. (Although Anders' series gets almost there).
So this is the minimal fix that brings the toolchians in to line with others and behaves how it previously did and fixes immediate build regressions. Removing -many is somewhat independent of that.
We had the same thing on Arm a few years ago when binutils started enforcing this more strictly, and it does catch actual bugs. I think annotating individual inline asm statements is the best choice here, as that documents what the intention is.
A few cases where there are differences in privileged instructions (that won't be compiler generated), that will be done anyway.
For new instructions added to the ISA though? I think it's ugly and unecesaary. There is no ambiguity about the intention when you see a lharx instruction is there?
It would delinate instructions that can't be used on all processors but I don't see much advantage there, it's not an exhaustive check because we have other restrictions on instructions in the kernel environment. And why would inline asm be special but not the rest of the asm? Would you propose to put these .machine directives everywhere in thousands of lines of asm code in the kernel? I don't know that it's an improvement. And inline asm is a small fraction of instructions.
There is one more bug in this series that I looked at with Anders, but he did not send a patch for that so far:
static void dummy_perf(struct pt_regs *regs) { #if defined(CONFIG_FSL_EMB_PERFMON) mtpmr(PMRN_PMGC0, mfpmr(PMRN_PMGC0) & ~PMGC0_PMIE); #elif defined(CONFIG_PPC64) || defined(CONFIG_PPC_BOOK3S_32) if (cur_cpu_spec->pmc_type == PPC_PMC_IBM) mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) & ~(MMCR0_PMXE|MMCR0_PMAO)); #else mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) & ~MMCR0_PMXE); #endif }
Here, the assembler correctly flags the mtpmr/mfpmr as an invalid instruction for a combined 6xx kernel: As far as I can tell, these are only available on e300 but not the others, and instead of the compile-time check for CONFIG_FSL_EMB_PERFMON, there needs to be some runtime check to use the first method on 83xx but the #elif one on the other 6xx machines.
Right that should be caught if you just pass -m<superset> architecture to the assembler that does not include the mtpmr. 32-bit is a lot more complicated than 64s like this though, so it's pssible in some cases you will want more checking and -m<subset> + some .machine directives will work better.
Once you add the .machine directive to your inline asm though, you lose *all* such static checking for the instruction. So it's really not a panacea and has its own downsides.
Thanks, Nick