https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95737
Segher Boessenkool <segher at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- Assignee|unassigned at gcc dot gnu.org |segher at gcc dot gnu.org CC| |segher at gcc dot gnu.org Target|powerpc-*-*-* |powerpc64*-*-* Ever confirmed|0 |1 Status|UNCONFIRMED |ASSIGNED Last reconfirmed| |2020-06-21 --- Comment #5 from Segher Boessenkool <segher at gcc dot gnu.org> --- In gimple this already is negativeLessThan (long long unsigned int a, long long unsigned int b) { _Bool _1; int _2; int _3; long long unsigned int _6; _1 = a_4(D) < b_5(D); _2 = (int) _1; _3 = -_2; _6 = (long long unsigned int) _3; return _6; } Then, it is expanded as a sign_extend:DI of a subreg:SI, and nowhere does it see this isn't necessary (it isn't because that SI cannot be negative). The RTL code isn't optimised very well before combine, and that does Trying 11 -> 12: 11: {r128:SI=ca:SI-0x1;clobber ca:SI;} REG_UNUSED ca:SI 12: r123:DI=sign_extend(r128:SI) REG_DEAD r128:SI Failed to match this instruction: (set (reg:DI 123) (sign_extend:DI (plus:SI (reg:SI 98 ca [+4 ]) (const_int -1 [0xffffffffffffffff])))) (note everything is made SImode in insn 11 before, it absorbed the subreg). Combine cannot keep track of known zero bits of hard regs well, so it fails to see that XER[CA] is only ever 0 or 1 here (it always is, but it doesn't know that either). I'll try to add an extra pattern for this extend, that will do the trick I think.