https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83202
Bug ID: 83202
Summary: Try joining operations on consecutive array elements
during tree vectorization
Product: gcc
Version: 7.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
Assignee: unassigned at gcc dot gnu.org
Reporter: [email protected]
Target Milestone: ---
void test(double data[4][4])
{
for (int i = 0; i < 4; i++)
{
for (int j = i; j < 4; j+=2)
{
data[i][j] = data[i][j] * data[i][j];
data[i][j+1] = data[i][j+1] * data[i][j+1];
}
}
}
gcc creates this:
test(double (*) [4]):
vmovsd xmm0, QWORD PTR [rdi]
vmulsd xmm0, xmm0, xmm0
vmovsd QWORD PTR [rdi], xmm0
vmovsd xmm0, QWORD PTR [rdi+8]
vmulsd xmm0, xmm0, xmm0
vmovsd QWORD PTR [rdi+8], xmm0
vmovsd xmm0, QWORD PTR [rdi+16]
vmulsd xmm0, xmm0, xmm0
vmovsd QWORD PTR [rdi+16], xmm0
vmovsd xmm0, QWORD PTR [rdi+24]
vmulsd xmm0, xmm0, xmm0
vmovsd QWORD PTR [rdi+24], xmm0
vmovsd xmm0, QWORD PTR [rdi+40]
vmulsd xmm0, xmm0, xmm0
vmovsd QWORD PTR [rdi+40], xmm0
vmovsd xmm0, QWORD PTR [rdi+48]
vmulsd xmm0, xmm0, xmm0
vmovsd QWORD PTR [rdi+48], xmm0
vmovsd xmm0, QWORD PTR [rdi+56]
vmulsd xmm0, xmm0, xmm0
vmovsd QWORD PTR [rdi+56], xmm0
vmovsd xmm0, QWORD PTR [rdi+64]
vmulsd xmm0, xmm0, xmm0
vmovsd QWORD PTR [rdi+64], xmm0
vmovsd xmm0, QWORD PTR [rdi+80]
vmulsd xmm0, xmm0, xmm0
vmovsd QWORD PTR [rdi+80], xmm0
vmovsd xmm0, QWORD PTR [rdi+88]
vmulsd xmm0, xmm0, xmm0
vmovsd QWORD PTR [rdi+88], xmm0
vmovsd xmm0, QWORD PTR [rdi+120]
vmulsd xmm0, xmm0, xmm0
vmovsd QWORD PTR [rdi+120], xmm0
vmovsd xmm0, QWORD PTR [rdi+128]
vmulsd xmm0, xmm0, xmm0
vmovsd QWORD PTR [rdi+128], xmm0
ret
clang detects that it is possible to use packed operations instead of scalar
ones, and produces this. Please implement similar optimization in gcc too.
test(double (*) [4]): # @test(double (*) [4])
vmovupd xmm0, xmmword ptr [rdi]
vmovupd xmm1, xmmword ptr [rdi + 16]
vmovupd xmm2, xmmword ptr [rdi + 40]
vmovupd xmm3, xmmword ptr [rdi + 56]
vmulpd xmm0, xmm0, xmm0
vmovupd xmmword ptr [rdi], xmm0
vmulpd xmm0, xmm1, xmm1
vmovupd xmmword ptr [rdi + 16], xmm0
vmulpd xmm0, xmm2, xmm2
vmovupd xmmword ptr [rdi + 40], xmm0
vmulpd xmm0, xmm3, xmm3
vmovupd xmmword ptr [rdi + 56], xmm0
vmovupd xmm0, xmmword ptr [rdi + 80]
vmulpd xmm0, xmm0, xmm0
vmovupd xmmword ptr [rdi + 80], xmm0
vmovupd xmm0, xmmword ptr [rdi + 120]
vmulpd xmm0, xmm0, xmm0
vmovupd xmmword ptr [rdi + 120], xmm0
ret