llvm/docs/ExtendedIntegerResults.txt

//===----------------------------------------------------------------------===//
// Representing sign/zero extension of function results
//===----------------------------------------------------------------------===//

Mar 25, 2009  - Initial Revision

Most ABIs specify that functions which return small integers do so in a
specific integer GPR.  This is an efficient way to go, but raises the question:
if the returned value is smaller than the register, what do the high bits hold?

There are three (interesting) possible answers: undefined, zero extended, or
sign extended.  The number of bits in question depends on the data-type that
the front-end is referencing (typically i1/i8/i16/i32).

Knowing the answer to this is important for two reasons: 1) we want to be able
to implement the ABI correctly.  If we need to sign extend the result according
to the ABI, we really really do need to do this to preserve correctness.  2)
this information is often useful for optimization purposes, and we want the
mid-level optimizers to be able to process this (e.g. eliminate redundant
extensions).

For example, lets pretend that X86 requires the caller to properly extend the
result of a return (I'm not sure this is the case, but the argument doesn't
depend on this).  Given this, we should compile this:

int a();
short b() { return a(); }

into:

_b:
	subl	$12, %esp
	call	L_a$stub
	addl	$12, %esp
	cwtl
	ret

An optimization example is that we should be able to eliminate the explicit
sign extension in this example:

short y();
int z() {
  return ((int)y() << 16) >> 16;
}

_z:
	subl	$12, %esp
	call	_y
	;;  movswl %ax, %eax   -> not needed because eax is already sext'd
	addl	$12, %esp
	ret

//===----------------------------------------------------------------------===//
// What we have right now.
//===----------------------------------------------------------------------===//

Currently, these sorts of things are modelled by compiling a function to return
the small type and a signext/zeroext marker is used.  For example, we compile
Z into:

define i32 @z() nounwind {
entry:
	%0 = tail call signext i16 (...)* @y() nounwind
	%1 = sext i16 %0 to i32
	ret i32 %1
}

and b into:

define signext i16 @b() nounwind {
entry:
	%0 = tail call i32 (...)* @a() nounwind		; <i32> [#uses=1]
	%retval12 = trunc i32 %0 to i16		; <i16> [#uses=1]
	ret i16 %retval12
}

This has some problems: 1) the actual precise semantics are really poorly
defined (see PR3779).  2) some targets might want the caller to extend, some
might want the callee to extend 3) the mid-level optimizer doesn't know the
size of the GPR, so it doesn't know that %0 is sign extended up to 32-bits 
here, and even if it did, it could not eliminate the sext. 4) the code
generator has historically assumed that the result is extended to i32, which is
a problem on PIC16 (and is also probably wrong on alpha and other 64-bit
targets).

//===----------------------------------------------------------------------===//
// The proposal
//===----------------------------------------------------------------------===//

I suggest that we have the front-end fully lower out the ABI issues here to
LLVM IR.  This makes it 100% explicit what is going on and means that there is
no cause for confusion.  For example, the cases above should compile into:

define i32 @z() nounwind {
entry:
        %0 = tail call i32 (...)* @y() nounwind
	%1 = trunc i32 %0 to i16
        %2 = sext i16 %1 to i32
        ret i32 %2
}
define i32 @b() nounwind {
entry:
	%0 = tail call i32 (...)* @a() nounwind
	%retval12 = trunc i32 %0 to i16
	%tmp = sext i16 %retval12 to i32
	ret i32 %tmp
}

In this model, no functions will return an i1/i8/i16 (and on a x86-64 target
that extends results to i64, no i32).  This solves the ambiguity issue, allows us 
to fully describe all possible ABIs, and now allows the optimizers to reason
about and eliminate these extensions.

The one thing that is missing is the ability for the front-end and optimizer to
specify/infer the guarantees provided by the ABI to allow other optimizations.
For example, in the y/z case, since y is known to return a sign extended value,
the trunc/sext in z should be eliminable.

This can be done by introducing new sext/zext attributes which mean "I know
that the result of the function is sign extended at least N bits.  Given this,
and given that it is stuck on the y function, the mid-level optimizer could
easily eliminate the extensions etc with existing functionality.

The major disadvantage of doing this sort of thing is that it makes the ABI
lowering stuff even more explicit in the front-end, and that we would like to
eventually move to having the code generator do more of this work.  However,
the sad truth of the matter is that this is a) unlikely to happen anytime in
the near future, and b) this is no worse than we have now with the existing
attributes.

C compilers fundamentally have to reason about the target in many ways.  
This is ugly and horrible, but a fact of life.