On 04/09/17 18:07, Wilco Dijkstra wrote:
Kyrill Tkachov wrote:-(define_insn_and_split "*iordi_notsesidi_di" - [(set (match_operand:DI 0 "s_register_operand" "=&r,&r") - (ior:DI (not:DI (sign_extend:DI - (match_operand:SI 2 "s_register_operand" "r,r"))) - (match_operand:DI 1 "s_register_operand" "0,r")))] - "TARGET_THUMB2" - "#" - "TARGET_THUMB2 && reload_completed" - [(set (match_dup 0) (ior:SI (not:SI (match_dup 2)) (match_dup 1))) - (set (match_dup 3) (ior:SI (not:SI - (ashiftrt:SI (match_dup 2) (const_int 31))) - (match_dup 4)))] - " - { - operands[3] = gen_highpart (SImode, operands[0]); - operands[0] = gen_lowpart (SImode, operands[0]); - operands[4] = gen_highpart (SImode, operands[1]); - operands[1] = gen_lowpart (SImode, operands[1]); - }" - [(set_attr "length" "8") - (set_attr "predicable" "yes") - (set_attr "predicable_short_it" "no") - (set_attr "type" "multiple")] -) - ... but why can we delete all these? As far as I can see they still perform useful splitting and don't have a NEON-specific pattern. Am I missing something?After Bernd's change almost all DI mode instructions are split before register allocation. So instructions using DI mode no longer exist and thus these extend variants can never be matched and are thus redundant.
Bernd's patch splits them when we don't have NEON. When NEON is available though they still maintain the DImode so we'd still benefit from these transformations, no?
Kyrill
Note that the SI mode instructions have zero/sign extend optimized after being split from DI mode using existing patterns. Wilco
