------- Additional Comments From danglin at gcc dot gnu dot org 2005-02-27 19:37 ------- This appears to be a reload problem. The reason the insn isn't recognized is that the REG_POINTER flag (f/frame-related) has been lost. Under HP-UX the base register must be marked with the REG_POINTER flag because the base register is used to determine which space register to use for the insn. This matters under HP-UX because the space registers forthe four memory quadrants contain different values.
After the lreg pass, we have: (note 1056 1055 1057 114 NOTE_INSN_DELETED) (insn 1057 1056 1855 114 (set (reg/v:DF 237 [ x ]) (mem/s/j:DF (plus:SI (reg:SI 241) (reg/v/f:SI 213 [ rp ])) [0 <variable>.data S8 A64])) 67 {*pa.md :3791} (insn_list 1055 (nil)) (nil)) The ICE occurs in the greg pass where we now have the following rtl: (note 1056 2195 2197 114 NOTE_INSN_DELETED) (insn 2197 1056 2198 114 (set (reg:SI 1 %r1) (mem:SI (plus:SI (reg/f:SI 30 %r30) (const_int -292 [0xfffffedc])) [0 rp+0 S4 A32])) 37 {*pa.md:2265} (nil) (nil)) (insn 2198 2197 2196 114 (set (reg:DF 68 %fr22) (mem/s/j:DF (plus:SI (reg:SI 21 %r21) (reg:SI 1 %r1)) [0 <variable>.data S8 A64])) -1 (nil) (nil)) At the moment, it's not clear where the flag is lost. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19819