https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71224
Bug ID: 71224
Summary: Looping over local copy of aggregate invokes undefined
behavior -Waggressive-loop-optimizations
Product: gcc
Version: 7.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: ibuclaw at gdcproject dot org
Target Milestone: ---
This probably is related to PR71177 and PR57199, but unlike them, this test
case doesn't have any external header dependencies, and also affects gcc as
well as g++.
However though it was in my D frontend that I first noticed this problem.
---
typedef __SIZE_TYPE__ size_t;
typedef struct Array
{
size_t length;
size_t* ptr;
} Array;
void setlength(Array* pthis, size_t nlength)
{
Array aggr = {
.length=pthis->length - nlength,
.ptr=pthis->ptr + nlength
};
for (size_t key = 0; key < aggr.length; key++)
aggr.ptr[key] = 1;
}
void insertBack(Array* pthis)
{
setlength(pthis, pthis->length + 1);
}
---
Can even reproduce if I were to lower the loop in the same way that g++ does.
---
void setlength(Array* pthis, size_t nlength)
{
Array aggr = {.length=pthis->length - nlength, .ptr=pthis->ptr + nlength};
size_t key = 0;
while (1)
{
if (aggr.length <= key)
break;
aggr.ptr[key] = 1;
key++;
}
}
---
Just like in PR57199, if I were to add the following inside the for body.
---
if (__builtin_expect(aggr.length > key, 0))
__builtin_unreachable();
---
Then the warning goes away.
