------- Additional Comments From sje at cup dot hp dot com 2005-01-10 23:43
-------
I looked into this bug some, it is fixed on 4.0, but the patch that fixes it is
the integration of tree-ssa into the mainline. I believe that the underlying
combine bug still exists but the tree-ssa changes result in different RTL being
generated so that the combine bug isn't hit. I looked at try_combine using the
3.4 sources. Attached is the i1, i2, i3 values we enter try_combine with and
what it generates. I have no idea how it came up with the result that it did
but that seems to be the heart of the problem.
In try_combine:
I1 is
(nil)
I2 is
(insn 53 51 54 2 (set (reg:BI 360)
(eq:BI (reg/v:SI 341 [ rnd_mode ])
(const_int 2 [0x2]))) 209 {*cmpsi_normal} (nil)
(expr_list:REG_DEAD (reg/v:SI 341 [ rnd_mode ])
(nil)))
I3 is
(insn 57 56 58 2 (set (reg:BI 363)
(and:BI (gt:BI (reg/v:SI 342 [ signx ])
(const_int 0 [0x0]))
(reg:BI 360))) 66 {*cmpsi_and_1} (insn_list 53 (nil))
(expr_list:REG_DEAD (reg/v:SI 342 [ signx ])
(expr_list:REG_DEAD (reg:BI 360)
(nil))))
==========================
Returns:
(insn 57 56 58 2 (set (reg:BI 363)
(const_int 0 [0x0])) 1 {movbi} (nil)
(nil))
==========================
--
What |Removed |Added
----------------------------------------------------------------------------
CC| |sje at cup dot hp dot com
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=17169
