Re: [PATCH] arm64: tlb: fix ARM64_TLB_RANGE with LLVM's integrated assembler

From: Catalin Marinas
Date: Thu Aug 06 2020 - 13:12:43 EST


On Wed, Aug 05, 2020 at 11:19:20AM -0700, Sami Tolvanen wrote:
> diff --git a/arch/arm64/include/asm/tlbflush.h b/arch/arm64/include/asm/tlbflush.h
> index d493174415db..66c2aab5e9cb 100644
> --- a/arch/arm64/include/asm/tlbflush.h
> +++ b/arch/arm64/include/asm/tlbflush.h
> @@ -16,6 +16,16 @@
> #include <asm/cputype.h>
> #include <asm/mmu.h>
>
> +/*
> + * Enable ARMv8.4-TLBI instructions with ARM64_TLB_RANGE. Note that binutils
> + * doesn't support .arch_extension tlb-rmi, so use .arch armv8.4-a instead.
> + */
> +#ifdef CONFIG_ARM64_TLB_RANGE
> +#define __TLBI_PREAMBLE ".arch armv8.4-a\n"
> +#else
> +#define __TLBI_PREAMBLE
> +#endif
> +
> /*
> * Raw TLBI operations.
> *
> @@ -28,14 +38,16 @@
> * not. The macros handles invoking the asm with or without the
> * register argument as appropriate.
> */
> -#define __TLBI_0(op, arg) asm ("tlbi " #op "\n" \
> +#define __TLBI_0(op, arg) asm (__TLBI_PREAMBLE \
> + "tlbi " #op "\n" \
> ALTERNATIVE("nop\n nop", \
> "dsb ish\n tlbi " #op, \
> ARM64_WORKAROUND_REPEAT_TLBI, \
> CONFIG_ARM64_WORKAROUND_REPEAT_TLBI) \
> : : )
>
> -#define __TLBI_1(op, arg) asm ("tlbi " #op ", %0\n" \
> +#define __TLBI_1(op, arg) asm (__TLBI_PREAMBLE \
> + "tlbi " #op ", %0\n" \
> ALTERNATIVE("nop\n nop", \
> "dsb ish\n tlbi " #op ", %0", \
> ARM64_WORKAROUND_REPEAT_TLBI, \

A potential problem here is that for gas (not sure about the integrated
assembler), .arch overrides any other .arch. So if we end up with two
preambles included in the same generated .S files in the future, it will
lead to some random behaviour.

Does the LLVM integrated assembler have the same behaviour on .arch
overriding a prior .arch?

Maybe a better solution is for all inline asm on arm64 to have a
standard preamble which is the maximum supported architecture version.
We can add individual .arch_extension as those are not overriding.

--
Catalin