https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66148
Thomas Preud'homme <thopre01 at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2015-05-25
Assignee|unassigned at gcc dot gnu.org |thopre01 at gcc dot
gnu.org
Ever confirmed|0 |1
--- Comment #4 from Thomas Preud'homme <thopre01 at gcc dot gnu.org> ---
The patch set out to use information from REG_EQUAL notes to get more accurate
nonzero_bits value. When compiling the testcase with commit r223033 (the one
just before the bug is seen) and options -O2 -fno-exceptions -fno-rtti
-fasynchronous-unwind-tables I can see:
(insn 360 358 361 27 (set (reg:DI 549)
(const_int -822083584 [0xffffffffcf000000]))
../../gcc/gcc/genpreds.c:983 123 {*pa.md:4196}
(expr_list:REG_EQUAL (const_int 3472883712 [0xcf000000])
(nil)))
Here the REG_EQUAL does not match the value that is set: the 32 most
significant bits are zero. Which is why combine decide that the set of
zero_extract can be removed. Looking at the dumps of the various pass,
everything looks fine until cse1 where instruction 360 is replaced from:
(insn 360 359 361 32 (set (reg:DI 549)
(lo_sum:DI (reg:DI 550)
(const_int -822083584 [0xffffffffcf000000])))
../../gcc/gcc/genpreds.c:983 93 {*pa.md:2773}
(expr_list:REG_EQUAL (const_int 3472883712 [0xcf000000])
(nil)))
to:
(insn 360 359 361 32 (set (reg:DI 549)
(const_int -822083584 [0xffffffffcf000000]))
../../gcc/gcc/genpreds.c:983 123 {*pa.md:4196}
(expr_list:REG_DEAD (reg:DI 550)
(expr_list:REG_EQUAL (const_int 3472883712 [0xcf000000])
(nil))))
without the REG_EQUAL note being updated. The replacement is made by the else
if in cse.c with the following comment:
/* Look for a substitution that makes a valid insn. */
I'm not sure what component is at fault here (my guess is that cse is not
expected to update the note directly but probably some other function) and I am
thus not very sure who to reassign this bug to.
