https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85424
Bug ID: 85424
Summary: The __builtin_packlongdouble function might have
issues with the output overlapping the inputs
Product: gcc
Version: 7.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: meissner at gcc dot gnu.org
Target Milestone: ---
When I was working on PR target/85075, I noticed the big endian version of the
compiler would not build the 32-bit library that used the function to create an
explicit IBM extended double value.
The problem is the pattern:
(define_insn_and_split "pack<mode>"
[(set (match_operand:FMOVE128 0 "register_operand" "=d,&d")
(unspec:FMOVE128
[(match_operand:<FP128_64> 1 "register_operand" "0,d")
(match_operand:<FP128_64> 2 "register_operand" "d,d")]
UNSPEC_PACK_128BIT))]
"FLOAT128_2REG_P (<MODE>mode)"
"@
fmr %L0,%2
#"
"&& reload_completed && REGNO (operands[0]) != REGNO (operands[1])"
[(set (match_dup 3) (match_dup 1))
(set (match_dup 4) (match_dup 2))]
{
unsigned dest_hi = REGNO (operands[0]);
unsigned dest_lo = dest_hi + 1;
gcc_assert (!IN_RANGE (REGNO (operands[1]), dest_hi, dest_lo));
gcc_assert (!IN_RANGE (REGNO (operands[2]), dest_hi, dest_lo));
operands[3] = gen_rtx_REG (<FP128_64>mode, dest_hi);
operands[4] = gen_rtx_REG (<FP128_64>mode, dest_lo);
}
[(set_attr "type" "fpsimple,fp")
(set_attr "length" "4,8")])
Generated an insn not found message, when one of the input operands overlapped
the output. In the case I saw, operand1 was the same as the second output
register. It is better to just delete the support for the register overlapping
(and hence trying to save a move), then to see this pop again in somewhat rare
circumstances.
