Matthew Malcomson <[email protected]> writes:
> --- a/gcc/config/aarch64/aarch64-protos.h
> +++ b/gcc/config/aarch64/aarch64-protos.h
> @@ -780,6 +780,7 @@ extern const atomic_ool_names aarch64_ool_ldeor_names;
>
> tree aarch64_resolve_overloaded_builtin_general (location_t, tree, void *);
>
> +const char * aarch64_sls_barrier (int);
Should be no space after the “*”.
> extern bool aarch64_harden_sls_retbr_p (void);
> extern bool aarch64_harden_sls_blr_p (void);
>
> diff --git a/gcc/config/aarch64/aarch64.h b/gcc/config/aarch64/aarch64.h
> index
> 24767c747bab0d711627c5c646937c42f210d70b..5da3d94e335fc315e1d90e6a674f2f09cf1a4529
> 100644
> --- a/gcc/config/aarch64/aarch64.h
> +++ b/gcc/config/aarch64/aarch64.h
> @@ -281,6 +281,7 @@ extern unsigned aarch64_architecture_version;
> #define AARCH64_ISA_F32MM (aarch64_isa_flags & AARCH64_FL_F32MM)
> #define AARCH64_ISA_F64MM (aarch64_isa_flags & AARCH64_FL_F64MM)
> #define AARCH64_ISA_BF16 (aarch64_isa_flags & AARCH64_FL_BF16)
> +#define AARCH64_ISA_SB (aarch64_isa_flags & AARCH64_FL_SB)
>
> /* Crypto is an optional extension to AdvSIMD. */
> #define TARGET_CRYPTO (TARGET_SIMD && AARCH64_ISA_CRYPTO)
> @@ -378,6 +379,9 @@ extern unsigned aarch64_architecture_version;
> #define TARGET_FIX_ERR_A53_835769_DEFAULT 1
> #endif
>
> +/* SB instruction is enabled through +sb. */
> +#define TARGET_SB (AARCH64_ISA_SB)
> +
> /* Apply the workaround for Cortex-A53 erratum 835769. */
> #define TARGET_FIX_ERR_A53_835769 \
> ((aarch64_fix_a53_err835769 == 2) \
> @@ -1058,8 +1062,11 @@ typedef struct
>
> #define RETURN_ADDR_RTX aarch64_return_addr
>
> -/* BTI c + 3 insns + 2 pointer-sized entries. */
> -#define TRAMPOLINE_SIZE (TARGET_ILP32 ? 24 : 32)
> +/* BTI c + 3 insns
> + + sls barrier of DSB + ISB.
> + + 2 pointer-sized entries. */
> +#define TRAMPOLINE_SIZE (24 \
> + + (TARGET_ILP32 ? 8 : 16))
Personal taste, sorry, but IMO this is easier to read on one line.
>
> /* Trampolines contain dwords, so must be dword aligned. */
> #define TRAMPOLINE_ALIGNMENT 64
> diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
> index
> 775f49991e5f599a843d3ef490b8cd044acfe78f..9356937fe266c68196392a1589b3cf96607de104
> 100644
> --- a/gcc/config/aarch64/aarch64.c
> +++ b/gcc/config/aarch64/aarch64.c
> @@ -10822,8 +10822,8 @@ aarch64_return_addr (int count, rtx frame
> ATTRIBUTE_UNUSED)
> static void
> aarch64_asm_trampoline_template (FILE *f)
> {
> - int offset1 = 16;
> - int offset2 = 20;
> + int offset1 = 24;
> + int offset2 = 28;
Huh, the offset handling in this function is a bit twisty, but that's
not your fault :-)
> […]
> @@ -11054,6 +11065,7 @@ aarch64_output_casesi (rtx *operands)
> output_asm_insn (buf, operands);
> output_asm_insn (patterns[index][1], operands);
> output_asm_insn ("br\t%3", operands);
> + output_asm_insn (aarch64_sls_barrier (aarch64_harden_sls_retbr_p ()),
> operands);
Long line.
> […]
> diff --git a/gcc/config/aarch64/aarch64.md b/gcc/config/aarch64/aarch64.md
> index
> ff15505d45546124868d2531b7f4e5b0f1f5bebc..75ef87a3b4674cc73cb42cc82cfb8e782acf77f6
> 100644
> --- a/gcc/config/aarch64/aarch64.md
> +++ b/gcc/config/aarch64/aarch64.md
> @@ -447,8 +447,15 @@
> (define_insn "indirect_jump"
> [(set (pc) (match_operand:DI 0 "register_operand" "r"))]
> ""
> - "br\\t%0"
> - [(set_attr "type" "branch")]
> + {
> + output_asm_insn ("br\\t%0", operands);
> + return aarch64_sls_barrier (aarch64_harden_sls_retbr_p ());
> + }
> + [(set_attr "type" "branch")
> + (set (attr "length")
> + (cond [(match_test "!aarch64_harden_sls_retbr_p ()") (const_int 4)
> + (match_test "TARGET_SB") (const_int 8)]
> + (const_int 12)))]
Rather than duplicating this several times, I think it would be better
to add a new attribute like “sls_mitigation”, set that attribute in the
define_insns, and then use “sls_mitigation” in the default “length”
calculation. See e.g. what rth did with “movprfx”.
> […]
> diff --git
> a/gcc/testsuite/gcc.target/aarch64/sls-mitigation/sls-miti-retbr-pacret.c
> b/gcc/testsuite/gcc.target/aarch64/sls-mitigation/sls-miti-retbr-pacret.c
> new file mode 100644
> index
> 0000000000000000000000000000000000000000..11f614b4ef2eb0fa3707cb46a55583d6685b89d0
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sls-mitigation/sls-miti-retbr-pacret.c
> @@ -0,0 +1,19 @@
> +/* { dg-do compile } */
> +/* { dg-additional-options "-mharden-sls=retbr -mbranch-protection=pac-ret
> -march=armv8.3-a" } */
> +
> +/* Testing the do_return pattern for retaa and retab. */
> +long retbr_subcall(void);
> +long retbr_do_return_retaa(void)
> +{
> + return retbr_subcall()+1;
> +}
> +__attribute__((target("branch-protection=pac-ret+b-key")))
> +long retbr_do_return_retab(void)
> +{
> + return retbr_subcall()+1;
> +}
> +
> +/* Ensure there are no BR or RET instructions which are not directly followed
> + by a speculation barrier. */
> +/* { dg-final { scan-assembler-not
> "\t(br|ret|retaa|retab)\tx\[0-9\]\[0-9\]?\n\t(?!dsb\tsy\n\tisb|sb)" } } */
Isn't the “sb” alternative invalid given the -march option?
Probably slightly easier to read if the regexp is quoted using {…}
rather than "…". Same for the other tests.
> […]
> +/* { dg-final { scan-assembler-not "ret\t" } } */
> diff --git a/gcc/testsuite/gcc.target/aarch64/sls-mitigation/sls-miti-retbr.c
> b/gcc/testsuite/gcc.target/aarch64/sls-mitigation/sls-miti-retbr.c
> new file mode 100644
> index
> 0000000000000000000000000000000000000000..5cd4da6bbb719a5135faa2c9818dc873e3d5af70
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sls-mitigation/sls-miti-retbr.c
> […]
> +/* Testing the indirect_jump pattern. */
> +typedef signed __attribute__((mode(DI))) intptr_t;
Just to check, have you tested this with -mabi=ilp32? Looks like it'll
probably be OK, was just suspicious because this isn't “intptr_t” there.
xp b/gcc/testsuite/gcc.target/aarch64/sls-mitigation/sls-mitigation.exp
> new file mode 100644
> index
> 0000000000000000000000000000000000000000..fb63c6dfe230e64b11919381c30a3a05eee52e16
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/aarch64/sls-mitigation/sls-mitigation.exp
> @@ -0,0 +1,73 @@
> +# Regression driver for SLS mitigation on AArch64.
> +# Copyright (C) 2020-2020 Free Software Foundation, Inc.
Just 2020.
Thanks,
Richard
