https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127743
Uroš Bizjak <ubizjak at gmail dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
Last reconfirmed| |2026-10-06
Ever confirmed|0 |1
Status|UNCONFIRMED |NEW
Component|target |rtl-optimization
--- Comment #4 from Uroš Bizjak <ubizjak at gmail dot com> ---
_.fwprop1:
148: r106:HF=[r118:DI+0x2]
REG_DEAD r118:DI
...
155: r213:HI=r106:HF#0
REG_DEAD r106:HF
156: r214:SI=zero_extend(r213:HI)
REG_DEAD r213:HI
159: r209:V2HF=r214:SI#0
REG_DEAD r216:SI
160: [r111:DI]=r209:V2HF
REG_DEAD r209:V2HF
_.ext_dce:
Processing insn:
156: r214:SI=zero_extend(r213:HI)
REG_DEAD r213:HI
Trying to simplify pattern:
(zero_extend:SI (reg:HI 213 [ _9 ]))
rescanning insn with uid = 156.
Successfully transformed to:
(subreg:SI (reg:HI 213 [ _9 ]) 0)
148: r106:HF=[r118:DI+0x2]
REG_DEAD r118:DI
...
155: r213:HI=r106:HF#0
REG_DEAD r106:HF
156: r214:SI=r213:HI#0
REG_DEAD r213:HI
159: r209:V2HF=r214:SI#0
REG_DEAD r214:SI
160: [r111:DI]=r209:V2HF
REG_DEAD r209:V2HF
Compilation goes downhill from here.
_.combine:
allowing combination of insns 156 and 159
original costs 4 + 12 = 16
replacement cost 12
deferring deletion of insn with uid = 156.
modifying insn i3 159: r209:V2HF=r213:HI#0
REG_DEAD r213:HI
deferring rescan insn with uid = 159.
148: r106:HF=[r118:DI+0x2]
REG_DEAD r118:DI
...
155: r213:HI=r106:HF#0
REG_DEAD r106:HF
156: NOTE_INSN_DELETED
159: r209:V2HF=r213:HI#0
REG_DEAD r213:HI
160: [r111:DI]=r209:V2HF
REG_DEAD r209:V2HF
So, ext_dce pass is converting zero_extend (insn 156) with defined
high bits to a paradoxical subreg that has undefined high bits.
Confirmed as RTL optimization problem.
