https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111097
Bug ID: 111097
Summary: mistifying bug triggered by trivial code change
Product: gcc
Version: 13.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: fabio at cannizzo dot net
Target Milestone: ---
Created attachment 55774
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=55774&action=edit
reproducible test case
I come across a weird bug. If I instrument the code with traces, the bug
disappears. If I remove some force inline attributes, the bug also disappears.
Eventually I manage to reduce to a minimalist example.
The attached file contains a class with a member state vector which is used as
a circular buffer
alignas(64) __m128i m_state[156];
The constructor fills it with some chosen numbers.
The first time the method `genrand_uint32` is called, it invokes the `refill`
function, which is supposed to modify all members of the `m_state` vector. Then
we extract one by one the element in the vector as `uint32_t`.
We print the 4 and last 4 elements of the state vector, reinterpreted as
`uint32_t[]` before refill and after refill, and the content is correct. We
also print the first and last 2 elements of the array extracted by
`genrand_uint32`, which should match exactly the state vector after refill, but
the first element is wrong.
pre-refill: 3403902862 1263040418 ... 2214668603 3759392492
post-refill: 1643499593 2393179749 ... 1152747345 4215414382
extracted: 3403902862 2393179749 ... 1152747345 4215414382
To be clear, the first element extracted should be 1643499593, i.e. the correct
output should be:
pre-refill: 3403902862 1263040418 ... 2214668603 3759392492
post-refill: 1643499593 2393179749 ... 1152747345 4215414382
extracted: 1643499593 2393179749 ... 1152747345 4215414382
If I remove a few of the FORCE_INLINE macros, e.g. if I change
FORCE_INLINE void refill()
to
NO_INLINE void refill()
the error disappears.
If I change 1 to 0 in this `#ifdef`, the error disappears.
#if 1 // here p is of type __m128i* and tmp has type __m128i
struct XV {
__m128i m_v;
XV(__m128i v) : m_v(v) {}
};
((XV *)p)[0] = tmp;
#else
p[0] = tmp;
#endif
If I reorder the `genrand_uint32` from
if (m_prnd != end()) { // most likely case
return *m_prnd++;
}
else {
refill();
m_prnd = begin();
return *m_prnd++;
}
to
if (m_prnd == end()) {
refill();
m_prnd = begin();
}
return *m_prnd++;
the error disappears.
A godbolt link to this code is also available:
https://godbolt.org/z/v9b7z8c66