https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91674
Bug ID: 91674
Summary: [ARM/thumb] redundant memcpy does not get optimized
away on thumb
Product: gcc
Version: 8.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: andij.cr at gmail dot com
Target Milestone: ---
consider this c++ function
#include <cstring>
#include <array>
#include <cstdint>
auto to_bytes(uint32_t arg){
std::array<uint8_t, sizeof(arg)> out{};
std::memcpy(out.data(), &arg, sizeof(arg));
return out;
}
on a little endian arch this function could be no-op.
compiled with g++ -Os we get:
to_bytes(unsigned int):
mov eax, edi
ret
on arm this somewhat works:
compiled with arm-none-eabi-g++ -Os
to_bytes(unsigned int):
sub sp, sp, #8
add sp, sp, #8
bx lr
notice the redundant sub followed by an add
but if if thumb is forced, the full optimization is not performed
compiled with arm-none-eabi-g++ -Os -march=armv7-m -mtune=cortex-m3
to_bytes(unsigned int):
mov r3, r0
movs r0, #0
uxtb r2, r3
bfi r0, r2, #0, #8
ubfx r2, r3, #8, #8
bfi r0, r2, #8, #8
ubfx r2, r3, #16, #8
bfi r0, r2, #16, #8
lsrs r3, r3, #24
sub sp, sp, #8
bfi r0, r3, #24, #8
add sp, sp, #8
bx lr
in contrast, cross compiling with clang7 produces the desired optimization:
compiled with clang++7 --target=arm-none-eabi -march=armv7-m -mtune=cortex-m3
to_bytes(unsigned int):
bx lr
notice also how there is no redundant stack pointer manipulation
