https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116170
Kewen Lin <linkw at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |bergner at gcc dot gnu.org,
| |segher at gcc dot gnu.org,
| |vmakarov at gcc dot gnu.org
--- Comment #3 from Kewen Lin <linkw at gcc dot gnu.org> ---
The culprit commit offers more chances on float constant copying, but it leads
LRA to generate one insn which is an unrecognizable insn:
(insn 181 0 0 (set (const_vector:V1TI [
(const_int 0 [0])
])
(reg:V1TI 179)) -1
(nil))
Initially at function curr_insn_transform curr_insn is
(insn 106 102 160 12 (set (mem:V1TI (and:DI (plus:DI (plus:DI (reg/f:DI 110
sfp)
(reg:DI 162))
(const_int 144 [0x90]))
(const_int -16 [0xfffffffffffffff0])) [2 S16 A128])
(subreg:V1TI (reg:KF 133) 0)) "test.i":9:5 discrim 2 1178
{*vsx_le_perm_store_v1ti}
(expr_list:REG_DEAD (reg:DI 162)
(expr_list:REG_DEAD (reg:KF 133)
(nil))))
with equiv substition, it becomes to
(insn 106 102 160 12 (set (mem:V1TI (and:DI (plus:DI (plus:DI (reg/f:DI 110
sfp)
(reg:DI 162))
(const_int 144 [0x90]))
(const_int -16 [0xfffffffffffffff0])) [2 S16 A128])
(const_vector:V1TI [
(const_int 0 [0])
])) "test.i":9:5 discrim 2 1178 {*vsx_le_perm_store_v1ti}
(expr_list:REG_DEAD (reg:DI 162)
(expr_list:REG_DEAD (reg:KF 133)
(nil))))
the corresponding pattern is:
(define_insn "*vsx_le_perm_store_<mode>"
[(set (match_operand:VSX_LE_128 0 "memory_operand" "=Z,Q")
(match_operand:VSX_LE_128 1 "vsx_register_operand" "+wa,r"))]
"!BYTES_BIG_ENDIAN && TARGET_VSX && !TARGET_P9_VECTOR
&& !altivec_indexed_or_indirect_operand (operands[0], <MODE>mode)"
"@
#
#"
[(set_attr "type" "vecstore,store")
(set_attr "length" "12,8")
(set_attr "isa" "<VSisa>,*")])
, alt 0 is selected, LRA fixes it up with:
Creating newreg=179, assigning class VSX_REGS to r179
(insn 180 179 0 (set (reg:V1TI 179)
(const_vector:V1TI [
(const_int 0 [0])
])) 1193 {vsx_movv1ti_64bit}
(nil))
which is sensible, but the unexpected thing is that is also generates
(insn 181 0 0 (set (const_vector:V1TI [
(const_int 0 [0])
])
(reg:V1TI 179)) -1
(nil))
...
via the code:
*loc = new_reg;
if (type != OP_IN
&& find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
{
start_sequence ();
lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
emit_insn (after);
after = get_insns ();
end_sequence ();
*loc = new_reg;
}
...
the reason here is the type isn't OP_IN so it wants to assign back to the OUT
OP, but the old here is a const_vector which is replaced with equiv, I think we
should stop move emission for this case.
The below patch makes it pass:
diff --git a/gcc/lra-constraints.cc b/gcc/lra-constraints.cc
index 92b343fa99a..024c85c87d9 100644
--- a/gcc/lra-constraints.cc
+++ b/gcc/lra-constraints.cc
@@ -4742,7 +4742,8 @@ curr_insn_transform (bool check_only_p)
}
*loc = new_reg;
if (type != OP_IN
- && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
+ && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX
+ && nonimmediate_operand (old, GET_MODE (old)))
{
start_sequence ();
lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old,
new_reg);
