[fpc-devel] Inline Constants (was On a port of Free Pascal to the IBM 370)

Mark Morgan Lloyd markMLl.fpc-devel at telemetry.co.uk
Mon Jan 30 18:50:20 CET 2012


steve smithers wrote:

> Firstly, let me explain that there are two different points regarding what has
> been called "literal" values as concerns S/370 architecture and it's Assemblers.
> The first of these is called "immediate" values where the literal is included
> in the actual code generated for that instruction.  The second are called
> "Literals" and describe unnamed constants that are defined on the instruction
> that uses them in the source code, but resolve to storage areas that are built
> into the object deck later.
> 
> The "Porting GCC to System/390" document in section 3.1 referred to and other
> posts state "the original S/390 architecture did not provide instructions that
> could use literal values as immediate operands".  This is untrue.  Since the
> System/360 was introduced there was a class of instructions called SI (Storage
> Immediate) that allowed just that.  The values were however, limited to 1 byte.
> This has applied to it's descendents (370, 370/XA  390, ESA z/OS and z/OS 64)
> The 390 extensions in the mid 1990's defined new instructions and extensions to
> increase this limit to 2 bytes and later to 4 bytes, perhaps, beyond.  I've
> never worked on the latest 64bit machines so I can't comment.
> 
> An example of SI instruction use.
> 
> Code            Source                  Comments
> 92C1 C024       MVI   FIELD,C'A'        Move character A to field.
> A728 0009       LHI   R2,H'9'           Load 9 into register 2 Note H is a
>                                         halfword or 16 bit integer value
> 
> The code generated is 92C1,C024 where 92 is the opcode, C1 is the character 'A'
> and C024 is the address in standard base/displacement form.  Or A7 is the
> opcode, 28 specifies a 32bit load into R2 0009 is the value to load into the
> register.  LHI is S/390 and later.
> 
> An equivalent example of literal instruction use.
> D200 C024 C136  MVC   FIELD,=C'A'       Move character A to field.
> 4820 C134       LH    R2,=H'9'          Load 9 into register 2 Note H is a
>                                         halfword or 16 bit integer value
> 
> The code generated is D200,C024,C136 where D2 is the opcode, C024 is the
> address of FIELD and C136 the address of the literal, both in standard
> base/displacement form.  The 00 is data regarding the length of data to move
> limited from 1 to 256 bytes in this case.  Or 48 is the opcode, 20 specifies
> the target register (R2) and the optional index register (unused) C134 is the
> address of the 16 bit value to load in base/displacement form.  Where are the
> constants?  Well they are generated automatically at the end of the module, or
> if you wish to define them elsewhere you can include a "LTORG" statement which
> tells the assembler to define them.
> 
> What I would like to know is "Why is this a problem?"  So the constants are
> defined elsewhere, what issues does this raise?

Are you saying that IN ALL CASES it is completely transparent, and will 
work with functions of any size, even on e.g. S/390 preceding G5?

My reading of that document- and I stress that I will be very happy to 
be corrected- is that the assembly source would have to be interrupted 
every few K for a literal table.

Bear in mind that since FPC generates assembly source code (for whatever 
target CPU), it might be very inconvenient to interrupt the flow of a 
large function. And please note that I am being very careful to 
distinguish here between generic "assembler source" and "IBM Assembler", 
which as I understand it is a large and complex field with its own 
culture and conventions (like referring to an "output deck" :-)

-- 
Mark Morgan Lloyd
markMLl .AT. telemetry.co .DOT. uk

[Opinions above are the author's, not those of his employers or colleagues]



More information about the fpc-devel mailing list