[Bug rtl-optimization/85300] [6/7/8 Regression] ICE in exact_int_to_float_conversion_p, at simplify-rtx.c:895

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85300

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |segher at gcc dot gnu.org

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
This breaks during combine.
We have:
(gdb) p debug_rtx (i3)
(insn 68 67 69 7 (set (reg:SI 89 [ _3 ])
        (fix:SI (reg/v:SF 90 [ a ]))) "pr85300.c":4 174 {fix_truncsfsi_sse}
     (expr_list:REG_DEAD (reg/v:SF 90 [ a ])
        (nil)))
(gdb) p debug_rtx (i2)
(insn 63 62 64 7 (set (reg/v:SF 90 [ a ])
        (float:SF (reg:SI 88 [ _2 ]))) "pr85300.c":8 203 {*floatsisf2_mixed}
     (expr_list:REG_DEAD (reg:SI 88 [ _2 ])
        (nil)))
(gdb) p debug_rtx (i1)
(insn 62 61 63 7 (set (reg:SI 88 [ _2 ])
        (const_int 256 [0x100])) "pr85300.c":8 86 {*movsi_internal}
     (nil))
and first subst it so that we get:
(set (reg:SI 89 [ _3 ])
    (fix:SI (float:SF (reg:SI 88 [ _2 ]))))
and then subst r88 for 256 and get:
(set (reg:SI 89 [ _3 ])
    (const_int 256 [0x100]))
That is fine.  But, during that substitution we also modify the i2 pattern,
making it invalid:
(insn 63 62 64 7 (set (reg/v:SF 90 [ a ])
        (float:SF (const_int 256 [0x100]))) "pr85300.c":8 203
{*floatsisf2_mixed}
     (expr_list:REG_DEAD (reg:SI 88 [ _2 ])
        (nil)))
because the FLOAT rtx was shared between PATTERN (i2) and newpat.