https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91594
Bug ID: 91594
Summary: Missing horizontal addition auto-vectorization
Product: gcc
Version: 9.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: diegoandres91b at hotmail dot com
Target Milestone: ---
The next code (with -O3 -ffast-math -msse3):
float a2[4], b2[4], c2[4];
void hadd2() {
c2[0] = a2[0] + a2[1];
c2[1] = a2[2] + a2[3];
c2[2] = b2[0] + b2[1];
c2[3] = b2[2] + b2[3];
}
Compiles without auto-vectorization:
hadd2():
movss xmm0, DWORD PTR a2[rip]
addss xmm0, DWORD PTR a2[rip+4]
movss DWORD PTR c2[rip], xmm0
movss xmm0, DWORD PTR a2[rip+8]
addss xmm0, DWORD PTR a2[rip+12]
movss DWORD PTR c2[rip+4], xmm0
movss xmm0, DWORD PTR b2[rip]
addss xmm0, DWORD PTR b2[rip+4]
movss DWORD PTR c2[rip+8], xmm0
movss xmm0, DWORD PTR b2[rip+8]
addss xmm0, DWORD PTR b2[rip+12]
movss DWORD PTR c2[rip+12], xmm0
ret
The expected code with HADDPS instruction (which does not compile):
hadd2():
movaps xmm0, XMMWORD PTR a1[rip]
haddps xmm0, XMMWORD PTR b1[rip]
movaps XMMWORD PTR c1[rip], xmm0
ret
In contrast, the normal addition code:
void add2() {
c2[0] = a2[0] + b2[0];
c2[1] = a2[1] + b2[1];
c2[2] = a2[2] + b2[2];
c2[3] = a2[3] + b2[3];
}
Compiles with auto-vectorization:
add2():
movaps xmm0, XMMWORD PTR a2[rip]
addps xmm0, XMMWORD PTR b2[rip]
movaps XMMWORD PTR c2[rip], xmm0
ret
Compiler Explorer Code: https://gcc.godbolt.org/z/9Hs9su
