https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85116
Bug ID: 85116
Summary: std::min_element does not optimize well with inlined
predicate
Product: gcc
Version: 7.3.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: christopher.schell at oculus dot com
Target Milestone: ---
According to godbolt (https://godbolt.org/g/igzsnL), the following code:
#define SIZE 1000
std::array<double, SIZE> testArray;
int getMinIdxCPPStyle(double offset)
{
auto minElement = std::min_element(std::cbegin(testArray),
std::cend(testArray), [offset](auto a, auto b) { return std::abs(a - offset) <
std::abs(b - offset); });
return std::distance(std::cbegin(testArray), minElement );
}
generates as the following under -O3
getMinIdxCPPStyle(double):
movq xmm3, QWORD PTR .LC1[rip]
mov eax, OFFSET FLAT:testArray
mov edx, OFFSET FLAT:testArray+8
.L11:
movsd xmm1, QWORD PTR [rdx]
movsd xmm2, QWORD PTR [rax]
subsd xmm1, xmm0
subsd xmm2, xmm0
andpd xmm1, xmm3
andpd xmm2, xmm3
ucomisd xmm2, xmm1
cmova rax, rdx
add rdx, 8
cmp rdx, OFFSET FLAT:testArray+8000
jne .L11
sub rax, OFFSET FLAT:testArray
sar rax, 3
ret
The problem being that the typical c-style loop beats this easily due to
caching the minimum value and not fetching it and recomputing it. Is there a
reason that the generated code should not cache the minimum value in a register
instead of probably causing a cache miss by fetching it and then unnecessarily
running the computations on it again?
