https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103797
Jan Hubicka <hubicka at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- Status|WAITING |NEW --- Comment #7 from Jan Hubicka <hubicka at gcc dot gnu.org> --- OK, here is completely fake testcase that does similar operaitons: #include <math.h> struct test {float x; float y; float z;} test; float f; void t() { float x = test.x; float y = test.y; float z = test.z; x = x * f; y = y * f; z = z * f; x = sqrt (x); y = sqrt (y); z = sqrt (z); x = x / f; y = y / f; z = z / f; test.x=x; test.y=y; test.z=z; } We seem to fail to vectorize it with: t.c:20:9: missed: op not supported by target. t.c:17:5: missed: not vectorized: relevant stmt not supported: x_15 = x_24 / f.0_1; clang seems to use divps happilly, so I am not sure why it is not supported. Even more funny is that with -Ofast it is compiled into multiplication by reciprocal: t: .LFB0: .cfi_startproc movss f(%rip), %xmm4 movss .LC0(%rip), %xmm2 movss test(%rip), %xmm0 movss test+4(%rip), %xmm3 divss %xmm4, %xmm2 movss test+8(%rip), %xmm1 mulss %xmm4, %xmm0 mulss %xmm4, %xmm3 mulss %xmm4, %xmm1 sqrtss %xmm0, %xmm0 sqrtss %xmm3, %xmm3 sqrtss %xmm1, %xmm1 mulss %xmm2, %xmm0 mulss %xmm2, %xmm3 mulss %xmm2, %xmm1 unpcklps %xmm3, %xmm0 movlps %xmm0, test(%rip) movss %xmm1, test+8(%rip) ret and rewriting it that way by hand: #include <math.h> struct test {float x; float y; float z;} test; float f; void t() { float x = test.x; float y = test.y; float z = test.z; float m = 1/f; x = x * f; y = y * f; z = z * f; x = sqrt (x); y = sqrt (y); z = sqrt (z); x = x * m; y = y * m; z = z * m; test.x=x; test.y=y; test.z=z; } gets the expected result: t: .LFB0: .cfi_startproc movss f(%rip), %xmm0 movq test(%rip), %xmm1 movaps %xmm0, %xmm2 shufps $0xe0, %xmm2, %xmm2 mulps %xmm1, %xmm2 movss .LC0(%rip), %xmm1 divss %xmm0, %xmm1 mulss test+8(%rip), %xmm0 sqrtps %xmm2, %xmm2 sqrtss %xmm0, %xmm0 movaps %xmm1, %xmm3 shufps $0xe0, %xmm3, %xmm3 mulss %xmm0, %xmm1 mulps %xmm3, %xmm2 movss %xmm1, test+8(%rip) movlps %xmm2, test(%rip) ret .cfi_endproc Having this however I do not see slp analyzing the divide in the original code at all.