https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116533
Bug ID: 116533
Summary: Possibly missing SAVE_EXPR in
get_member_function_from_ptrfunc()
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: sirl at gcc dot gnu.org
Target Milestone: ---
This came up during the discussion of PR 116449.
get_member_function_from_ptrfunc() generates a statement with repeated
expressions, but the repeated expressions are not wrapped into SAVE_EXPRs.
This code
class C
{
public:
void P(int);
void IP();
int parr[16];
};
typedef void (C::*fp)();
typedef struct arr_t
{
fp func;
} arr_t;
static arr_t farr[1] =
{
{ &C::IP },
};
void C::P(int c)
{
((*this).*farr[parr[c]].func)();
}
results in this statement for C::P() in .original:
<<cleanup_point <<< Unknown tree: expr_stmt
(((long int) farr[((struct C *) this)->parr[c]].func.__pfn & 1) != 0)
? (void C::<T3fd> (struct C *) *) *(*((int (*) () * *) this
+ (sizetype) farr[((struct C *) this)->parr[c]].func.__delta)
+ (sizetype) ((long int) farr[((struct C *) this)->parr[c]].func.__pfn +
-1))
: farr[((struct C *) this)->parr[c]].func.__pfn
((struct C *) this + (sizetype) farr[((struct C *)
this)->parr[c]].func.__delta) >>>>>;
As you can see the terms "farr[((struct C *) this)->parr[c]].func" and
"farr[((struct C *) this)->parr[c]].func.__pfn" are used everywhere (condition
and both arms of the ternary).
Compiling results in the following .gimple:
void C::P (struct C * const this, int c)
{
void C::<T3fd> (struct C *) * iftmp.0;
# DEBUG BEGIN_STMT
_1 = this->parr[c];
_2 = farr[_1].func.__pfn;
_3 = (long int) _2;
_4 = _3 & 1;
if (_4 != 0) goto <D.2821>; else goto <D.2822>;
<D.2821>:
_5 = this->parr[c];
_6 = farr[_5].func.__delta;
_7 = (sizetype) _6;
_8 = this + _7;
_9 = *_8;
_10 = this->parr[c];
_11 = farr[_10].func.__pfn;
_12 = (long int) _11;
_13 = _12 + -1;
_14 = (sizetype) _13;
_15 = _9 + _14;
iftmp.0 = *_15;
goto <D.2823>;
<D.2822>:
_16 = this->parr[c];
iftmp.0 = farr[_16].func.__pfn;
<D.2823>:
iftmp.1_17 = iftmp.0;
_18 = this->parr[c];
_19 = farr[_18].func.__delta;
_20 = (sizetype) _19;
_21 = this + _20;
iftmp.1_17 (_21);
}
Now with this tentative patch:
@@ -4191,11 +4191,10 @@ get_member_function_from_ptrfunc (tree
*instance_ptrptr, tree function,
if (TREE_SIDE_EFFECTS (instance_ptr))
instance_ptr = instance_save_expr = save_expr (instance_ptr);
- if (TREE_SIDE_EFFECTS (function))
- function = save_expr (function);
+ function = save_expr(function);
/* Start by extracting all the information from the PMF itself. */
- e3 = pfn_from_ptrmemfunc (function);
+ e3 = save_expr(pfn_from_ptrmemfunc (function));
delta = delta_from_ptrmemfunc (function);
idx = build1 (NOP_EXPR, vtable_index_type, e3);
switch (TARGET_PTRMEMFUNC_VBIT_LOCATION)
results in this statement:
<<cleanup_point <<< Unknown tree: expr_stmt
(((long int) SAVE_EXPR <SAVE_EXPR <farr[((struct C *)
this)->parr[c]].func>.__pfn> & 1) != 0)
? (void C::<T40a> (struct C *) *) *(*((int (*) () * *) this
+ (sizetype) SAVE_EXPR <farr[((struct C *) this)->parr[c]].func>.__delta)
+ (sizetype) ((long int) SAVE_EXPR <SAVE_EXPR <farr[((struct C *)
this)->parr[c]].func>.__pfn> + -1))
: SAVE_EXPR <SAVE_EXPR <farr[((struct C *) this)->parr[c]].func>.__pfn>
((struct C *) this + (sizetype) SAVE_EXPR <farr[((struct C *)
this)->parr[c]].func>.__delta) >>>>>;
and gimple like this:
void C::P (struct C * const this, int c)
{
void C::<T40a> (struct C *) * iftmp.0;
struct
{
void C::<T40a> (struct C *) * __pfn;
long int __delta;
} D.2853;
void C::<T40a> (struct C *) * D.2854;
# DEBUG BEGIN_STMT
_1 = this->parr[c];
D.2853 = farr[_1].func;
D.2854 = D.2853.__pfn;
_2 = (long int) D.2854;
_3 = _2 & 1;
if (_3 != 0) goto <D.2855>; else goto <D.2856>;
<D.2855>:
_4 = D.2853.__delta;
_5 = (sizetype) _4;
_6 = this + _5;
_7 = *_6;
_8 = (long int) D.2854;
_9 = _8 + -1;
_10 = (sizetype) _9;
_11 = _7 + _10;
iftmp.0 = *_11;
goto <D.2857>;
<D.2856>:
iftmp.0 = D.2854;
<D.2857>:
iftmp.1_12 = iftmp.0;
_13 = D.2853.__delta;
_14 = (sizetype) _13;
_15 = this + _14;
iftmp.1_12 (_15);
}
In my eyes this gimple looks much better, but there maybe valid reasons why the
SAVE_EXPRs were not added before.
But a lot of the code in get_member_function_from_ptrfunc() is decades old
("75th Cygnus merge" for example), so it's hard
to find anything related in the ChangeLogs.
The tentative patch maybe incomplete and/or too broad, but it nicely shows
what's going on.
So, what do you think? Should more SAVE_EXPR be added in
get_member_function_from_ptrfunc()?
