Hi, Richard,
For RTL expansion of call to .DEFERRED_INIT, I changed my code per your
suggestions like following:
======================
#define INIT_PATTERN_VALUE 0xFE
static void
expand_DEFERRED_INIT (internal_fn, gcall *stmt)
{
tree lhs = gimple_call_lhs (stmt);
tree var_size = gimple_call_arg (stmt, 0);
enum auto_init_type init_type
= (enum auto_init_type) TREE_INT_CST_LOW (gimple_call_arg (stmt, 1));
bool is_vla = (bool) TREE_INT_CST_LOW (gimple_call_arg (stmt, 2));
tree var_type = TREE_TYPE (lhs);
gcc_assert (init_type > AUTO_INIT_UNINITIALIZED);
if (is_vla || (!can_native_interpret_type_p (var_type)))
{
/* If this is a VLA or the type of the variable cannot be natively
interpreted, expand to a memset to initialize it. */
if (TREE_CODE (lhs) == SSA_NAME)
lhs = SSA_NAME_VAR (lhs);
tree var_addr = NULL_TREE;
if (is_vla)
var_addr = TREE_OPERAND (lhs, 0);
else
{
TREE_ADDRESSABLE (lhs) = 1;
var_addr = build_fold_addr_expr (lhs);
}
tree value = (init_type == AUTO_INIT_PATTERN) ?
build_int_cst (unsigned_char_type_node,
INIT_PATTERN_VALUE) :
build_zero_cst (unsigned_char_type_node);
tree m_call = build_call_expr (builtin_decl_implicit (BUILT_IN_MEMSET),
3, var_addr, value, var_size);
/* Expand this memset call. */
expand_builtin_memset (m_call, NULL_RTX, TYPE_MODE (var_type));
}
else
{
/* If this is not a VLA and the type of the variable can be natively
interpreted, expand to assignment to generate better code. */
tree pattern = NULL_TREE;
unsigned HOST_WIDE_INT total_bytes
= tree_to_uhwi (TYPE_SIZE_UNIT (var_type));
if (init_type == AUTO_INIT_PATTERN)
{
unsigned char *buf = (unsigned char *) xmalloc (total_bytes);
memset (buf, INIT_PATTERN_VALUE, total_bytes);
pattern = native_interpret_expr (var_type, buf, total_bytes);
gcc_assert (pattern);
}
tree init = (init_type == AUTO_INIT_PATTERN) ?
pattern :
build_zero_cst (var_type);
expand_assignment (lhs, init, false);
}
}
===========================
Now, I used “can_native_interpret_type_p (var_type)” instead of
“use_register_for_decl (lhs)” to decide
whether to use “memset” or use “assign” to expand this function.
However, this exposed an bug that is very hard to be addressed:
*******For the testing case: test suite/gcc.dg/uninit-I.c:
/* { dg-do compile } */
/* { dg-options "-O2 -Wuninitialized" } */
int sys_msgctl (void)
{
struct { int mode; } setbuf;
return setbuf.mode; /* { dg-warning "'setbuf\.mode' is used" } */
==
******the above auto var “setbuf” has “struct” type, which
“can_native_interpret_type_p(var_type)” is false, therefore,
Expanding this .DEFERRED_INIT call went down the “memset” expansion route.
However, this structure type can be fitted into a register, therefore cannot be
taken address anymore at this stage, even though I tried:
TREE_ADDRESSABLE (lhs) = 1;
var_addr = build_fold_addr_expr (lhs);
To create an address variable for it, the expansion still failed at expr.c:
line 8412:
during RTL pass: expand
/home/opc/Work/GCC/latest-gcc/gcc/testsuite/gcc.dg/auto-init-uninit-I.c:6:24:
internal compiler error: in expand_expr_addr_expr_1, at expr.c:8412
0xd04104 expand_expr_addr_expr_1
../../latest-gcc/gcc/expr.c:8412
0xd04a95 expand_expr_addr_expr
../../latest-gcc/gcc/expr.c:8525
0xd13592 expand_expr_real_1(tree_node*, rtx_def*, machine_mode,
expand_modifier, rtx_def**, bool)
../../latest-gcc/gcc/expr.c:11741
0xd05142 expand_expr_real(tree_node*, rtx_def*, machine_mode, expand_modifier,
rtx_def**, bool)
../../latest-gcc/gcc/expr.c:8713
0xaed1d3 expand_expr
../../latest-gcc/gcc/expr.h:301
0xaf0d89 get_memory_rtx
../../latest-gcc/gcc/builtins.c:1370
0xafb4fb expand_builtin_memset_args
../../latest-gcc/gcc/builtins.c:4102
0xafacde expand_builtin_memset(tree_node*, rtx_def*, machine_mode)
../../latest-gcc/gcc/builtins.c:3886
0xe97fb3 expand_DEFERRED_INIT
******That’s the major reason why I chose “use_register_for_decl(lhs)” to
decide “memset” expansion or “assign” expansion, “memset” expansion
needs to take address of the variable, if the variable has been decided to fit
into a register, then its address cannot taken anymore at this stage.
******using “can_native_interpret_type_p” did make the “pattern” generation
part much cleaner and simpler, however, looks like it didn’t work correctly.
Based on this, I’d like to keep my previous implementation by using
“use_register_for_decl” to decide whether to take “memset” expansion or
“assign” expansion.
Therefore, I might still need to keep the “UGLY” implementation of generatting
“pattern” constant for different types?
Let me know your opinion on this.
Thanks a lot for the help.
Qing
> On Aug 9, 2021, at 9:09 AM, Richard Biener <rguent...@suse.de> wrote:
>
> On Tue, 27 Jul 2021, Qing Zhao wrote:
>
> + created during gimplification phase. Refer to gimplify_vla_decl
> + for details. */
> + tree var_decl = (TREE_CODE (var) == SSA_NAME) ?
> + SSA_NAME_VAR (var) : var;
> + gcc_assert (DECL_HAS_VALUE_EXPR_P (var_decl));
> + gcc_assert (TREE_CODE (DECL_VALUE_EXPR (var_decl)) ==
> INDIRECT_REF);
> + /* Get the address of this vla variable. */
> + vlaaddr = TREE_OPERAND (DECL_VALUE_EXPR (var_decl), 0);
>
> err - isn't the address of the decl represented by the LHS
> regardless whether this is a VLA or not? Looking at DECL_VALUE_EXPR
> looks quite fragile since that's not sth data dependence honors.
> It looks you only partly gimplify the build init here? All
> DECL_VALUE_EXPRs should have been resolved.
>
> + if (is_vla || (!use_register_for_decl (var)))
> ...
> + else
> + {
> + /* If this variable is in a register, use expand_assignment might
> + generate better code. */
>
> you compute the patter initializer even when not needing it,
> that's wasteful. It's also quite ugly, IMHO you should
> use can_native_interpret_type_p (var_type) and native_interpret
> a char [] array initialized to the pattern and if
> !can_native_interpret_type_p () go the memset route.
