I have a fragment of code that looks like this:

VAR     z, a, alpha, b, c, d, e: Int64;
         f: DOUBLE;
BEGIN
   td:= td + 0.5;                       // astronomical to civil
   z:= Floor(td);
   f:= td - z;
   IF z < 2299161.0 THEN
     a:= z
   ELSE BEGIN
     alpha:= Floor((z - 1867216.25) / 36524.25);
     a:= z + 1 + alpha - Floor(alpha / 4)
   END;

Using 2.4.4 with -a -al -s this compiles to a .s which looks like this:

.Ll34:
# [439] IF z < 2299161.0 THEN
ld [%i6-32],%o0
ld [%i6-28],%o1
call fpc_int64_to_double
nop
sethi %hi(_$MOONPHASE$_Ld8),%o0
ld [%lo(_$MOONPHASE$_Ld8)+%o0],%f1
fcmpd %f0,%f1
nop
fbl .Lj87
nop
ba .Lj88
nop
.Lj87:
.stabn 68,0,441,.Ll35 - MOONPHASE_JYEAR$DOUBLE$INT64$INT64$DOUBLE .Ll35:
# [441] a:= z

This appears to be the only place in the program where fcmpd %f0,%f1 is generated, although the same opcode is generated with other registers as opeands. Assembling with as (2.17 or 2.18), I get this error:

moonphase.s:629: Error: Illegal operands

If I change the code to look like this:

   f:= td - z;
   IF z - 2299161.0 < 0 THEN
     a:= z
   ELSE BEGIN

it generates this:

.Ll34:
# [440] IF z - 2299161.0 < 0 THEN
ld [%i6-32],%o0
ld [%i6-28],%o1
call fpc_int64_to_double
nop
sethi %hi(_$MOONPHASE$_Ld8),%o0
ld [%lo(_$MOONPHASE$_Ld8)+%o0],%f1
fsubs %f0,%f1,%f0
sethi %hi(_$MOONPHASE$_Ld9),%o0
ld [%lo(_$MOONPHASE$_Ld9)+%o0],%f1
fcmps %f0,%f1
nop
fbl .Lj87

Having a workaround makes this not particularly crucial, but I'd welcome any comments.

The sparc only has single precision floating point registers. Double precision values are stored in two consecutive single precision registers, and hence arguments to double precision opcodes must always be even registers. The above demonstrates a bug in the register allocator. Please file a bug report.

Jonas

Jonas Maebe wrote

The sparc only has single precision floating point registers. Double precision values are stored in two consecutive single precision registers, and hence arguments to double precision opcodes must always be even registers. The above demonstrates a bug in the register allocator. Please file a bug report.

Thanks Jonas, will do in the morning.

I've worked around it by noticing that I only need integer operations at that point- I'm transcribing Perl (MoonPhase.pm) based on C (moontool.c), and it's not the first instance of bad type handling that I've found.

Submitted as 20072