[fpc-devel] Currency * Double

LacaK lacak at zoznam.sk
Mon Oct 25 08:29:44 CEST 2021

>> Hi ,
>> what data type is result of multiplication of: (Currency * Double) on
>> Win64 (and on Win32)?
>>     c:=1115;
>>     d:=100000000;
>>     writeln(Round(c*d)); // gives -72967440737
> 	32 bit version gives always 111500000000 as a result (and
> floating point math should always have extended precision intermeiate
> there) . But even with double intermediate result on 64 bit versions
> there should not be any problems. 1.1150000000000000E+011 is well in
> range of that type.
Let's look at disassembly on Win32:
(FPU here uses "st registers" 80 bits wide)

test_curr_int64.lpr:77            c1:=1115;
00401EFC a178a14200               mov    0x42a178,%eax
00401F01 a320204300               mov    %eax,0x432020
00401F06 a17ca14200               mov    0x42a17c,%eax
00401F0B a324204300               mov    %eax,0x432024
test_curr_int64.lpr:78            d:=100000000;
00401F10 a180a14200               mov    0x42a180,%eax
00401F15 a330204300               mov    %eax,0x432030
00401F1A a184a14200               mov    0x42a184,%eax
00401F1F a334204300               mov    %eax,0x432034
test_curr_int64.lpr:79            i1:=Round(c1*d);
00401F24 dd0530204300             fldl   0x432030                 // 
00401F2A dc0d08a14200             fmull  0x42a108                 // 
00401F30 df2d20204300             fildll 0x432020                 // 
st0:=11150000, st1:=1000000000000
00401F36 dec9                     fmulp  %st,%st(1)               // 
00401F38 df2d28a04200             fildll 0x42a028                 // 
st0:=10000, st1:=1.115e+019
00401F3E def9                     fdivrp %st,%st(1)               // 
00401F40 df2d28a04200             fildll 0x42a028                 // 
st0:=10000, st1:=1115000000000000
00401F46 def9                     fdivrp %st,%st(1)               // 
00401F48 df7df8                   fistpll -0x8(%ebp)
00401F4B 9b                       fwait
00401F4C 8b45f8                   mov    -0x8(%ebp),%eax
00401F4F a340204300               mov    %eax,0x432040
00401F54 8b45fc                   mov    -0x4(%ebp),%eax
00401F57 a344204300               mov    %eax,0x432044

And now on Win64:
test_curr_int64.lpr:77                    c1:=1115;
0000000100001EAC 48c70569410400b022aa00   movq 
$0xaa22b0,0x44169(%rip)        # 0x100046020
test_curr_int64.lpr:78                    d:=100000000;
0000000100001EB7 488b05f2f20200           mov 0x2f2f2(%rip),%rax        
# 0x1000311b0 <_$TEST_CURR_INT64$_Ld16>
0000000100001EBE 4889056b410400           mov %rax,0x4416b(%rip)        
# 0x100046030
test_curr_int64.lpr:79                    i1:=Round(c1*d);
0000000100001EC5 f20f100563410400         movsd 
0x44163(%rip),%xmm0        // 100000000
0000000100001ECD f20f590553f20200         mulsd 
0x2f253(%rip),%xmm0        // 1000000000000
0000000100001ED5 f2480f2dc8               cvtsd2si 
%xmm0,%rcx               // 1000000000000
0000000100001EDA 488b053f410400           mov 0x4413f(%rip),%rax         
// 11150000
0000000100001EE1 480fafc8                 imul 
%rax,%rcx                  // -7296744073709551616 <---OVERFLOW HERE
0000000100001EE5 48b84b598638d6c56d34     movabs $0x346dc5d63886594b,%rax
0000000100001EEF 48f7e9                   imul   %rcx
0000000100001EF2 48c1fa0b                 sar    $0xb,%rdx
0000000100001EF6 48c1e93f                 shr    $0x3f,%rcx
0000000100001EFA 4801ca                   add    %rcx,%rdx
0000000100001EFD 4889d1                   mov    %rdx,%rcx
0000000100001F00 e8fb110000               callq  0x100003100 
0000000100001F05 48890534410400           mov %rax,0x44134(%rip)        
# 0x100046040

If I understand correctly multipllication "imul" is performed as 
multiplication of two currency values (int64)
(11150000 * 1000000000000 = 1115 * 10^16) which is over Max Int64 ...
So on Win64 there is currency (int64) intermediate ...


More information about the fpc-devel mailing list