On Tue, May 31, 2022 at 5:37 AM Takayuki 'January June' Suwa via
Gcc-patches <gcc-patches@gcc.gnu.org> wrote:
>
> Hi all,
>
> In some targets, initialization code for char array may be split into two
> parts even if the initialization consists of all zeros:
>
> /* example */
> extern void foo(char*);
> void test(void) {
> char a[256] = { 0, 0, 0, 0, 0, 0, 0, 0, 0 };
> foo(a);
> }
>
> ;; Xtensa (xtensa-lx106)
> .LC0:
> .string ""
> .string ""
> .string ""
> .string ""
> .string ""
> .string ""
> .string ""
> .string ""
> .string ""
> .string ""
> .zero 246
> test:
> movi a9, 0x110
> sub sp, sp, a9
> l32r a3, .LC1
> movi.n a4, 0xa
> mov.n a2, sp
> s32i a0, sp, 268
> call0 memcpy
> movi a4, 0xf6
> movi.n a3, 0
> addi.n a2, sp, 10
> call0 memset
> mov.n a2, sp
> call0 foo
> l32i a0, sp, 268
> movi a9, 0x110
> add.n sp, sp, a9
> ret.n
>
> ;; H8/300 (-mh -mint32)
> .LC0:
> .string ""
> .string ""
> .string ""
> .string ""
> .string ""
> .string ""
> .string ""
> .string ""
> .string ""
> .string ""
> .zero 246
> _test:
> sub.l #256,er7
> sub.l er2,er2
> add.b #10,r2l
> mov.l #.LC0,er1
> mov.l er7,er0
> jsr @_memcpy
> sub.l er2,er2
> add.b #246,r2l
> sub.l er1,er1
> sub.l er0,er0
> add.b #10,r0l
> add.l er7,er0
> jsr @_memset
> mov.l er7,er0
> jsr @_foo
> add.l #256,er7
> rts
>
> i386 target (both 32 and 64bit) does not show such behavior.
>
> gcc/ChangeLog:
>
> * expr.cc (store_expr): Add check if the initialization content
> consists of all zeros.
> ---
> gcc/expr.cc | 7 +++++++
> 1 file changed, 7 insertions(+)
>
> diff --git a/gcc/expr.cc b/gcc/expr.cc
> index 7197996cec7..f94310dc7b9 100644
> --- a/gcc/expr.cc
> +++ b/gcc/expr.cc
> @@ -6015,6 +6015,7 @@ store_expr (tree exp, rtx target, int call_param_p,
> rtx dest_mem;
> tree str = TREE_CODE (exp) == STRING_CST
> ? exp : TREE_OPERAND (TREE_OPERAND (exp, 0), 0);
> + char ch;
>
> exp_len = int_expr_size (exp);
> if (exp_len <= 0)
> @@ -6032,6 +6033,12 @@ store_expr (tree exp, rtx target, int call_param_p,
> }
>
> str_copy_len = TREE_STRING_LENGTH (str);
> + /* If str contains only zeroes, no need to store to target. */
> + ch = 0;
> + for (HOST_WIDE_INT i = 0; i < str_copy_len; ++i)
> + ch |= TREE_STRING_POINTER (str)[i];
> + if (ch == 0)
> + str_copy_len = 0;
Not sure if I decipher the current code correctly but maybe we instead
want to prune str_copy_len from the end for trailing \0 bytes instead of
just special-casing all-zero initializers?
> if ((STORE_MAX_PIECES & (STORE_MAX_PIECES - 1)) == 0)
> {
> str_copy_len += STORE_MAX_PIECES - 1;
> --
> 2.20.1
