[Bug middle-end/52306] ICE in cselib_record_set, at cselib.c:2158

[mikpe at it dot uu.se]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52306



--- Comment #17 from Mikael Pettersson <mikpe at it dot uu.se> 2013-02-06 
23:23:58 UTC ---

Created attachment 29376

  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=29376

reduced test case



Please disregard my last two comments, I misread the insn dump and mistook a

note for the invalid concurrent modification of a register.



Thorsten's initial xslt test case, attachment 27325, ICEs 4.6 and 4.7 -with -O2

-fPIC, and 4.8 with just -O2.  This new test case is substantially reduced from

the initial one, but it only ICEs 4.8.  The ICE occurs when cselib processes

insn 67.



IRA produced



(insn 67 66 69 10 (set (reg/f:SI 52 [ D.1521 ])

        (mem/f:SI (post_inc:SI (reg:SI 59 [ ivtmp.11 ])) [3 MEM[base: 0B,

index: ivtmp.11_66, offset: 0B]+0 S4 A16])) pr52306-4.c:44 38 {*movsi_m68k2}

     (expr_list:REG_INC (reg:SI 59 [ ivtmp.11 ])

        (nil)))



with, if I read the dump correctly, reg 52 mapped to %a1 and reg 59 spilled.



Reload then sees fit to reload reg 59 into %a1, resulting in



(insn 67 253 69 10 (set (reg:SI 9 %a1)

        (mem/f:SI (post_inc:SI (reg:SI 9 %a1)) [3 MEM[base: 0B, index:

ivtmp.11_66, offset: 0B]+0 S4 A16])) pr52306-4.c:44 38 {*movsi_m68k2}

     (expr_list:REG_INC (reg:SI 9 %a1)

        (nil)))



which uses %a1 both as an auto-inc source pointer and as the dest of the load,

causing the assertion failure in cselib.



The post_inc of reg 59 (%a1) is actually dead, reload added insns to load %a1

from reg 59's spill slot and to increment that memory cell directly.  I'm

guessing either reload should not have coalesced reg 59 with reg 52 (due to the

post_inc on reg 59), or it should have cancelled the post_inc when it added the

new insn to increment the spill slot directly.