On Wed, Dec 20, 2023 at 05:45:05PM -0500, Jason Merrill wrote:
> On 12/20/23 14:20, Jakub Jelinek wrote:
> > + if ((warn_sizeof_pointer_memaccess || alloc_size_attr)
> > && (complain & tf_warning)
> > && !vec_safe_is_empty (*args)
> > && !processing_template_decl)
> > {
> > - location_t sizeof_arg_loc[3];
> > - tree sizeof_arg[3];
> > + location_t sizeof_arg_loc[6];
> > + tree sizeof_arg[6];
>
> Why do we need to check 6 args for calloc? The patch is OK, just wondering.
The 3 for warn_sizeof_pointer_memaccess comes from the fact that we only
warn on certain builtins and don't really need it on 4th and later
arguments.
For the calloc case it warns about any 2 argument alloc_size attribute, so
generally it could be even INT_MAX or so; while for the C++ FE which
still has SIZEOF_EXPRs around it could be implementable if the function is
adjusted to be called with different arguments, for the C FE we need to
remember whether each argument was a sizeof or not because we fold
everything immediately.
I've grepped my /usr/include/ with what arguments alloc_size attribute is
used and found some 1,2 and 2,3 and 3,4 cases, so I went for 6 as a
compromise between not wasting too much compile time on it vs. covering
majority of functions with alloc_size 2 argument attributes.
It is unlikely allocation functions would need dozens of arguments...
Anyway, bootstrap fails with the patch due to:
../../gcc/gimple-fold.cc:7089:15: error: allocation of insufficient size ‘72’
for type ‘tree_node’ with size ‘216’ [-Werror=alloc-size]
../../gcc/gimple-fold.cc:7096:15: error: allocation of insufficient size ‘72’
for type ‘tree_node’ with size ‘216’ [-Werror=alloc-size]
This is on
/* Allocate SSA names(lhs1) on the stack. */
tree lhs1 = (tree)XALLOCA (tree_ssa_name);
memset (lhs1, 0, sizeof (tree_ssa_name));
TREE_SET_CODE (lhs1, SSA_NAME);
TREE_TYPE (lhs1) = type;
init_ssa_name_imm_use (lhs1);
where tree is a union and we don't allocate enough memory for the
whole union, just for one member of it.
Guess that is a case the warning is meant to warn about,
so we need some workaround. I find it really weird to allocate
constant 72 bytes using alloca...
Then there is also:
../../gcc/collect2.cc:625:39: error: ‘void* xcalloc(size_t, size_t)’ sizes
specified with ‘sizeof’ in the earlier argument and not in the later argument
[-Werror=calloc-transposed-args]
which is what the warning is meant to warn about.
So perhaps incremental:
2023-12-21 Jakub Jelinek <ja...@redhat.com>
* gimple-fold.cc (maybe_fold_comparisons_from_match_pd):
Use unsigned char buffers for lhs1 and lhs2 instead of allocating
them through XALLOCA.
* collect2.cc (maybe_run_lto_and_relink): Swap xcalloc arguments.
--- gcc/gimple-fold.cc.jj 2023-12-17 22:50:01.077589250 +0100
+++ gcc/gimple-fold.cc 2023-12-21 00:25:26.254807466 +0100
@@ -7086,14 +7086,16 @@ maybe_fold_comparisons_from_match_pd (tr
gimple_set_bb (stmt2, NULL);
/* Allocate SSA names(lhs1) on the stack. */
- tree lhs1 = (tree)XALLOCA (tree_ssa_name);
+ alignas (tree_node) unsigned char lhs1buf[sizeof (tree_ssa_name)];
+ tree lhs1 = (tree) &lhs1buf[0];
memset (lhs1, 0, sizeof (tree_ssa_name));
TREE_SET_CODE (lhs1, SSA_NAME);
TREE_TYPE (lhs1) = type;
init_ssa_name_imm_use (lhs1);
/* Allocate SSA names(lhs2) on the stack. */
- tree lhs2 = (tree)XALLOCA (tree_ssa_name);
+ alignas (tree_node) unsigned char lhs2buf[sizeof (tree_ssa_name)];
+ tree lhs2 = (tree) &lhs2buf[0];
memset (lhs2, 0, sizeof (tree_ssa_name));
TREE_SET_CODE (lhs2, SSA_NAME);
TREE_TYPE (lhs2) = type;
--- gcc/collect2.cc.jj 2023-11-10 22:48:01.356867230 +0100
+++ gcc/collect2.cc 2023-12-21 00:18:31.821159336 +0100
@@ -622,7 +622,7 @@ maybe_run_lto_and_relink (char **lto_ld_
LTO object replaced by all partitions and other LTO
objects removed. */
- lto_c_argv = (char **) xcalloc (sizeof (char *), num_lto_c_args);
+ lto_c_argv = (char **) xcalloc (num_lto_c_args, sizeof (char *));
lto_c_ptr = CONST_CAST2 (const char **, char **, lto_c_argv);
*lto_c_ptr++ = lto_wrapper;
Jakub
