https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81803
--- Comment #5 from James Cowgill <james410 at cowgill dot org.uk> ---
I have just noticed this which seems curious. Is the 39 -> 40 combine really a
valid transformation? It seems we've lost the sign extension and we're just
putting a 32-bit value into a 64-bit register without trying to clear the upper
bits anymore?
Before
-----
(insn 38 37 39 3 (set (reg:SI 235)
(ne:SI (reg:SI 234 [ h ])
(const_int 0 [0]))) "reduced.c":16 500 {*sne_zero_sisi}
(expr_list:REG_DEAD (reg:SI 234 [ h ])
(nil)))
(insn 39 38 40 3 (set (reg:HI 233)
(subreg:HI (reg:SI 235) 0)) "reduced.c":16 358 {*movhi_internal}
(expr_list:REG_DEAD (reg:SI 235)
(nil)))
(insn 40 39 133 3 (set (reg:DI 207 [ iftmp.3_18 ])
(sign_extend:DI (reg:HI 233))) "reduced.c":16 244 {*extendhidi2_seh}
(expr_list:REG_DEAD (reg:HI 233)
(nil)))
combine
-----
Trying 39 -> 40:
Successfully matched this instruction:
(set (reg:DI 207 [ iftmp.3_18 ])
(subreg:DI (reg:SI 235) 0))
allowing combination of insns 39 and 40
original costs 4 + 4 = 8
replacement cost 4
deferring deletion of insn with uid = 39.
modifying insn i3 40: r207:DI=r235:SI#0
REG_DEAD r235:SI
deferring rescan insn with uid = 40.
Trying 38 -> 40:
Successfully matched this instruction:
(set (subreg:SI (reg:DI 207 [ iftmp.3_18 ]) 0)
(ne:SI (reg:SI 234 [ h ])
(const_int 0 [0])))
allowing combination of insns 38 and 40
original costs 4 + 4 = 8
replacement cost 4
deferring deletion of insn with uid = 38.
modifying insn i3 40: r207:DI#0=r234:SI!=0
REG_DEAD r234:SI
deferring rescan insn with uid = 40.
After
------
(note 38 37 39 3 NOTE_INSN_DELETED)
(note 39 38 40 3 NOTE_INSN_DELETED)
(insn 40 39 133 3 (set (subreg:SI (reg:DI 207 [ iftmp.3_18 ]) 0)
(ne:SI (reg:SI 234 [ h ])
(const_int 0 [0]))) "reduced.c":16 500 {*sne_zero_sisi}
(expr_list:REG_DEAD (reg:SI 234 [ h ])
(nil)))
