https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77503
--- Comment #12 from amker at gcc dot gnu.org --- ICE is because prologue peeling changes code (to be specific, the initial argument of reduction phi node in loop header), as a result, the statement to be vectorized is not the statement that was previously analyzed. One possible fix is be conservative and fall back to generic conditional reduction if loop needs to be peeled for prologue. Another (better) fix is we don't update/inherit reduction value from prologue loop for vectorized loop, and after vectorized loop, we combine the result manually. Code would be like: prolog_loop: c_1 = PHI<c_2, 0> c_2 = (cond) ? 1 : c_1; if (prolog_cond) goto prolog_header else goto prolog_exit prolog_exit: c_3 = PHI<c_2>; vector_loop: c_4 = PHI<c_5, 0>; c_5 = (cond) ? 1 : c_4 if (vector_cond) goto vector_header else goto vector_exit vector_exit: c_6 = PHI<c_5>; c_7 = (c_3 == 1 || c_5 == 1) ? 1 : 0; We need to change how loop peeling updates SSA(PHI) nodes. Since I have other patches rewriting loop peeling, better to check this method after loop peeling patches. For now, I think we'd better go with the first method which disables the optimization in peeling cases.
