https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113560
Bug ID: 113560
Summary: Strange code generated when optimizing a
multiplication on x86_64
Product: gcc
Version: 13.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: accelerator0099 at gmail dot com
Target Milestone: ---
Code:
#include <immintrin.h>
auto f(char *buf, unsigned long long in) noexcept
{
unsigned long long hi{};
auto lo{_mulx_u64(in, 0x2af31dc462ull, &hi)};
lo = _mulx_u64(lo, 100, &hi);
__builtin_memcpy(buf + 2, &hi, 2);
return buf + 10;
}
auto g(char *buf, unsigned long long in) noexcept
{
unsigned long long hi{};
_mulx_u64(in, 100, &hi);
__builtin_memcpy(buf + 2, &hi, 2);
return buf + 10;
}
Compile with:
-Ofast -std=c++23 -march=znver4
GCC 13.2 and truck generate:
f(char*, unsigned long long):
movabs rdx, 184467440738
mov rax, rdi
mulx r9, r8, rsi
xor r9d, r9d
mov rsi, r8
mov rdi, r9
add rsi, r8
shld rdi, r8, 1
add rsi, r8
adc rdi, r9
shld rdi, rsi, 3
sal rsi, 3
add rsi, r8
adc rdi, r9
add rax, 10
shld rdi, rsi, 2
mov WORD PTR [rax-8], di
ret
g(char*, unsigned long long):
mov eax, 100
mul rsi
lea rax, [rdi+10]
mov WORD PTR [rdi+2], dx
ret
GCC 12 generates:
f(char*, unsigned long long):
movabs rdx, 184467440738
mov rax, rsi
imul rax, rdx
mov edx, 100
mulx rdx, rax, rax
lea rax, [rdi+10]
mov WORD PTR [rdi+2], dx
ret
g(char*, unsigned long long):
mov eax, 100
mul rsi
lea rax, [rdi+10]
mov WORD PTR [rdi+2], dx
ret
Clang:
f(char*, unsigned long long):
unsigned long long)
movabs rdx, 184467440738
mov eax, 100
imul rdx, rsi
mulx rax, rax, rax
mov word ptr [rdi + 2], ax
lea rax, [rdi + 10]
ret
g(char*, unsigned long long):
unsigned long long)
mov eax, 100
mov rdx, rsi
mulx rax, rax, rax
mov word ptr [rdi + 2], ax
lea rax, [rdi + 10]
ret
See also:
https://gcc.godbolt.org/z/df7Gr1MKo
