https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110879
Bug ID: 110879
Summary: Unnecessary reread from memory in a loop
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: palevichva at gmail dot com
Target Milestone: ---
Created attachment 55678
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=55678&action=edit
preprocessed file by g++ from revision dd2eb972a
I've found a strange regression in optimization. Trunk version of g++ produces
less optimal assembly. It rereads same memory location in every iteration of a
loop. More specifically, it rereads fields _M_finish and _M_end_of_storage of a
vector from memory every push_back call, although it is not necessary.
Released version 13.2 doesn't do that, and just uses values from registers.
I'm compiling following code:
#include <vector>
std::vector<int> f(std::size_t n) {
std::vector<int> res;
res.reserve(n);
for (std::size_t i = 0; i < n; ++i) {
res.push_back(i*i);
}
return res;
}
The main body of a loop looks like this:
~/.local/gcc/bin/g++ -S -fverbose-asm -O3 -std=c++20 pb.cpp
>.L41:
># /home/scaiper/.local/gcc/include/c++/14.0.0/bits/stl_construct.h:97: {
>return ::new((void*)__location) _Tp(std::forward<_Args>(__args)...); }
> movl %r15d, (%rbx) # _3, *prephitmp_51
># /home/scaiper/.local/gcc/include/c++/14.0.0/bits/vector.tcc:119:
>++this->_M_impl._M_finish;
> addq $4, %rbx #, tmp135
> movq %rbx, 8(%rbp) # tmp135,
> res_8(D)->D.35756._M_impl.D.35067._M_finish
>.L8:
># pb.cpp:6: for (std::size_t i = 0; i < n; ++i) {
> addq $1, %r13 #, i
># pb.cpp:6: for (std::size_t i = 0; i < n; ++i) {
> cmpq %r13, %r12 # i, n
> je .L1 #,
># /home/scaiper/.local/gcc/include/c++/14.0.0/bits/vector.tcc:114: if
>(this->_M_impl._M_finish != this->_M_impl._M_end_of_storage)
> movq 8(%rbp), %rbx # res_8(D)->D.35756._M_impl.D.35067._M_finish,
> prephitmp_51
># /home/scaiper/.local/gcc/include/c++/14.0.0/bits/vector.tcc:114: if
>(this->_M_impl._M_finish != this->_M_impl._M_end_of_storage)
> movq 16(%rbp), %rax #
> res_8(D)->D.35756._M_impl.D.35067._M_end_of_storage, pretmp_52
>.L16:
># pb.cpp:7: res.push_back(i*i);
> movl %r13d, %r15d # i, _3
> imull %r13d, %r15d # i, _3
># /home/scaiper/.local/gcc/include/c++/14.0.0/bits/vector.tcc:114: if
>(this->_M_impl._M_finish != this->_M_impl._M_end_of_storage)
> cmpq %rax, %rbx # pretmp_52, prephitmp_51
> jne .L41 #,
Same loop as produced by 13.2:
~/.local/gcc-13.2/bin/g++ -v -S -fverbose-asm -O3 -std=c++20 pb.cpp
>.L43:
># /home/scaiper/.local/gcc-13.2/include/c++/13.2.0/bits/stl_construct.h:97:
> { return ::new((void*)__location) _Tp(std::forward<_Args>(__args)...); }
> movl %r12d, (%rcx) # _3, *prephitmp_4
># /home/scaiper/.local/gcc-13.2/include/c++/13.2.0/bits/vector.tcc:119:
> ++this->_M_impl._M_finish;
> addq $4, %rcx #, prephitmp_4
> movq %rcx, 8(%rbp) # prephitmp_4,
> res_8(D)->D.35699._M_impl.D.35010._M_finish
>.L8:
># pb.cpp:6: for (std::size_t i = 0; i < n; ++i) {
> addq $1, %rbx #, i
># pb.cpp:6: for (std::size_t i = 0; i < n; ++i) {
> cmpq %rbx, %r13 # i, n
> je .L1 #,
>.L18:
># pb.cpp:7: res.push_back(i*i);
> movl %ebx, %r12d # i, _3
> imull %ebx, %r12d # i, _3
># /home/scaiper/.local/gcc-13.2/include/c++/13.2.0/bits/vector.tcc:114:
> if (this->_M_impl._M_finish != this->_M_impl._M_end_of_storage)
> cmpq %r8, %rcx # prephitmp_74, prephitmp_4
> jne .L43 #,
Notice this extra commands in the first snippet:
movq 8(%rbp), %rbx
movq 16(%rbp), %rax
I've bisected this problem to the commit dd2eb972a (libstdc++: Use RAII in
std::vector::_M_realloc_insert)
(https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=dd2eb972a5b063e10c83878d5c9336a818fa8291).
It doesn't look like commit is the problem. Code looks pretty equivalent. But
for some reason compiler produces different result.
I'm using version built from aforementioned commit dd2eb972a:
Target: x86_64-pc-linux-gnu
Configured with: ../gcc/configure --enable-languages=c++ --disable-multilib
--prefix=/home/scaiper/.local/gcc
gcc version 14.0.0 20230623 (experimental) (GCC)
COLLECT_GCC_OPTIONS='-v' '-S' '-fverbose-asm' '-O3' '-std=c++20'
'-shared-libgcc' '-mtune=generic' '-march=x86-64'
Comparing with 13.2:
Target: x86_64-pc-linux-gnu
Configured with: ../gcc/configure --enable-languages=c++ --disable-multilib
--prefix=/home/scaiper/.local/gcc-13.2
gcc version 13.2.0 (GCC)
COLLECT_GCC_OPTIONS='-v' '-S' '-fverbose-asm' '-O3' '-std=c++20'
'-shared-libgcc' '-mtune=generic' '-march=x86-64'
