Re: [patch][version 6] add -ftrivial-auto-var-init and variable attribute "uninitialized" to gcc

[Richard Biener <rguenther@suse.de>]

On Thu, 12 Aug 2021, Qing Zhao wrote:

> 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.

Hmm, I think you can use use_register_for_decl(lhs) to decide to use an
alternate type to generate the pattern (and feed to 
can_native_interpret_type_p) by using
lang_hooks.type_for_mode (TYPE_MODE (TREE_TYPE (lhs))).  You can then
build the assignment from the pattern as

 VIEW_CONVERT <reg-type> (lhs) = pattern_cst;

note that more RTL-expand-ish would be to simply expand 'lhs' and
see whether it's a REG_P or a MEM_P and decide based on that.  Of course
that no longer allows you to use the high-level expand_assignment
or memset but you'd need to work on a lower level then.

Richard.

> Thanks a lot for the help.
> 
> Qing
> 
> 
> > On Aug 9, 2021, at 9:09 AM, Richard Biener <rguenther@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.
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH, Maxfeldstrasse 5, 90409 Nuernberg,
Germany; GF: Felix Imendörffer; HRB 36809 (AG Nuernberg)