* Claudiu Zissulescu <claudiu.zissule...@synopsys.com> [2017-11-27 12:09:59
+0100]:
> From: Claudiu Zissulescu <claz...@gmail.com>
>
> The new implementation attempts to clean up the existing trampoline
> implementation for ARC making it to work for linux type of systems.
>
> gcc/
> 2017-11-10 Claudiu Zissulescu <claz...@synopsys.com>
>
> * config/arc/arc.c (TARGET_TRAMPOLINE_ADJUST_ADDRESS): Delete.
> (emit_store_direct): Likewise.
> (arc_trampoline_adjust_address): Likewise.
> (arc_asm_trampoline_template): New function.
> (arc_initialize_trampoline): Use asm_trampoline_template.
> (TARGET_ASM_TRAMPOLINE_TEMPLATE): Define.
> * config/arc/arc.h (TRAMPOLINE_SIZE): Adjust to 16.
> *config/arc/arc.md (flush_icache): Delete pattern.
^-- Missing space here.
Otherwise, looks fine.
Thanks,
Andrew
> ---
> gcc/config/arc/arc.c | 89
> +++++++++++++++++++++++++--------------------------
> gcc/config/arc/arc.h | 2 +-
> gcc/config/arc/arc.md | 9 ------
> 3 files changed, 44 insertions(+), 56 deletions(-)
>
> diff --git a/gcc/config/arc/arc.c b/gcc/config/arc/arc.c
> index 0eeeb42..053f3c2 100644
> --- a/gcc/config/arc/arc.c
> +++ b/gcc/config/arc/arc.c
> @@ -588,8 +588,6 @@ static void arc_finalize_pic (void);
>
> #define TARGET_TRAMPOLINE_INIT arc_initialize_trampoline
>
> -#define TARGET_TRAMPOLINE_ADJUST_ADDRESS arc_trampoline_adjust_address
> -
> #define TARGET_CAN_ELIMINATE arc_can_eliminate
>
> #define TARGET_FRAME_POINTER_REQUIRED arc_frame_pointer_required
> @@ -3727,69 +3725,65 @@ output_shift (rtx *operands)
>
> /* Nested function support. */
>
> -/* Directly store VALUE into memory object BLOCK at OFFSET. */
> -
> -static void
> -emit_store_direct (rtx block, int offset, int value)
> -{
> - emit_insn (gen_store_direct (adjust_address (block, SImode, offset),
> - force_reg (SImode,
> - gen_int_mode (value, SImode))));
> -}
> +/* Output assembler code for a block containing the constant parts of
> + a trampoline, leaving space for variable parts.
>
> -/* Emit RTL insns to initialize the variable parts of a trampoline.
> - FNADDR is an RTX for the address of the function's pure code.
> - CXT is an RTX for the static chain value for the function. */
> -/* With potentially multiple shared objects loaded, and multiple stacks
> - present for multiple thereds where trampolines might reside, a simple
> - range check will likely not suffice for the profiler to tell if a callee
> - is a trampoline. We a speedier check by making the trampoline start at
> - an address that is not 4-byte aligned.
> A trampoline looks like this:
>
> - nop_s 0x78e0
> -entry:
> ld_s r12,[pcl,12] 0xd403
> ld r11,[pcl,12] 0x170c 700b
> j_s [r12] 0x7c00
> - nop_s 0x78e0
> + .word function's address
> + .word static chain value
> +
> +*/
> +
> +static void
> +arc_asm_trampoline_template (FILE *f)
> +{
> + asm_fprintf (f, "\tld_s\t%s,[pcl,8]\n", ARC_TEMP_SCRATCH_REG);
> + asm_fprintf (f, "\tld\t%s,[pcl,12]\n", reg_names[STATIC_CHAIN_REGNUM]);
> + asm_fprintf (f, "\tj_s\t[%s]\n", ARC_TEMP_SCRATCH_REG);
> + assemble_aligned_integer (UNITS_PER_WORD, const0_rtx);
> + assemble_aligned_integer (UNITS_PER_WORD, const0_rtx);
> +}
> +
> +/* Emit RTL insns to initialize the variable parts of a trampoline.
> + FNADDR is an RTX for the address of the function's pure code. CXT
> + is an RTX for the static chain value for the function.
>
> The fastest trampoline to execute for trampolines within +-8KB of CTX
> would be:
> +
> add2 r11,pcl,s12
> j [limm] 0x20200f80 limm
> - and that would also be faster to write to the stack by computing the
> offset
> - from CTX to TRAMP at compile time. However, it would really be better to
> - get rid of the high cost of cache invalidation when generating
> trampolines,
> - which requires that the code part of trampolines stays constant, and
> - additionally either
> - - making sure that no executable code but trampolines is on the stack,
> - no icache entries linger for the area of the stack from when before the
> - stack was allocated, and allocating trampolines in trampoline-only
> - cache lines
> - or
> - - allocate trampolines fram a special pool of pre-allocated trampolines.
> */
> +
> + and that would also be faster to write to the stack by computing
> + the offset from CTX to TRAMP at compile time. However, it would
> + really be better to get rid of the high cost of cache invalidation
> + when generating trampolines, which requires that the code part of
> + trampolines stays constant, and additionally either making sure
> + that no executable code but trampolines is on the stack, no icache
> + entries linger for the area of the stack from when before the stack
> + was allocated, and allocating trampolines in trampoline-only cache
> + lines or allocate trampolines fram a special pool of pre-allocated
> + trampolines. */
>
> static void
> arc_initialize_trampoline (rtx tramp, tree fndecl, rtx cxt)
> {
> rtx fnaddr = XEXP (DECL_RTL (fndecl), 0);
>
> - emit_store_direct (tramp, 0, TARGET_BIG_ENDIAN ? 0x78e0d403 : 0xd40378e0);
> - emit_store_direct (tramp, 4, TARGET_BIG_ENDIAN ? 0x170c700b : 0x700b170c);
> - emit_store_direct (tramp, 8, TARGET_BIG_ENDIAN ? 0x7c0078e0 : 0x78e07c00);
> - emit_move_insn (adjust_address (tramp, SImode, 12), fnaddr);
> - emit_move_insn (adjust_address (tramp, SImode, 16), cxt);
> - emit_insn (gen_flush_icache (adjust_address (tramp, SImode, 0)));
> -}
> + emit_block_move (tramp, assemble_trampoline_template (),
> + GEN_INT (TRAMPOLINE_SIZE), BLOCK_OP_NORMAL);
>
> -/* Allow the profiler to easily distinguish trampolines from normal
> - functions. */
> + emit_move_insn (adjust_address (tramp, SImode, 8), fnaddr);
> + emit_move_insn (adjust_address (tramp, SImode, 12), cxt);
>
> -static rtx
> -arc_trampoline_adjust_address (rtx addr)
> -{
> - return plus_constant (Pmode, addr, 2);
> + emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__clear_cache"),
> + LCT_NORMAL, VOIDmode, 2, XEXP (tramp, 0), Pmode,
> + plus_constant (Pmode, XEXP (tramp, 0), TRAMPOLINE_SIZE),
> + Pmode);
> }
>
> /* Add the given function declaration to emit code in JLI section. */
> @@ -11412,6 +11406,9 @@ arc_cannot_substitute_mem_equiv_p (rtx)
> #undef TARGET_CANNOT_SUBSTITUTE_MEM_EQUIV_P
> #define TARGET_CANNOT_SUBSTITUTE_MEM_EQUIV_P
> arc_cannot_substitute_mem_equiv_p
>
> +#undef TARGET_ASM_TRAMPOLINE_TEMPLATE
> +#define TARGET_ASM_TRAMPOLINE_TEMPLATE arc_asm_trampoline_template
> +
> struct gcc_target targetm = TARGET_INITIALIZER;
>
> #include "gt-arc.h"
> diff --git a/gcc/config/arc/arc.h b/gcc/config/arc/arc.h
> index 8c31fb2..317a653 100644
> --- a/gcc/config/arc/arc.h
> +++ b/gcc/config/arc/arc.h
> @@ -829,7 +829,7 @@ extern int arc_initial_elimination_offset(int from, int
> to);
> /* Trampolines. */
>
> /* Length in units of the trampoline for entering a nested function. */
> -#define TRAMPOLINE_SIZE 20
> +#define TRAMPOLINE_SIZE 16
>
> /* Alignment required for a trampoline in bits . */
> /* For actual data alignment we just need 32, no more than the stack;
> diff --git a/gcc/config/arc/arc.md b/gcc/config/arc/arc.md
> index 155ee6c..e1418a9 100644
> --- a/gcc/config/arc/arc.md
> +++ b/gcc/config/arc/arc.md
> @@ -4345,15 +4345,6 @@ archs4xd, archs4xd_slow, core_3"
> (set_attr "iscompact" "true")
> (set_attr "length" "2")])
>
> -;; Special pattern to flush the icache.
> -;; ??? Not sure what to do here. Some ARC's are known to support this.
> -
> -(define_insn "flush_icache"
> - [(unspec_volatile [(match_operand:SI 0 "memory_operand" "m")] 0)]
> - ""
> - "* return \"\";"
> - [(set_attr "type" "misc")])
> -
> ;; Split up troublesome insns for better scheduling.
>
> ;; Peepholes go at the end.
> --
> 1.9.1
>
