[LLVMdev] The definition of getTypeSize

[Duncan Sands]

Now that I'm working on codegen support for arbitrary precision
integers (think i36 or i129), I've hit the problem of what
getTypeSize and friends should return for such integers (the
current implementations seem to me to be wrong).  However it's
not clear to me how getTypeSize and friends are defined.

There seem to be several possible meanings for the size of a type
(only talking about primitive types here):

(1) The minimum number of bits needed to hold all values of the type.
(2) The minimum number of bits read by a load (maybe a better way of
saying this: if you load a value and store it somewhere else, how many
bits are correctly copied?)
(3) The maximum number of bits that may be overwritten by a store.
(4) The amount of memory allocated by an alloca or malloc for a
variable of this type.
(5) The spacing between successive variables of this type in an
array or struct.

For example, take i36.  For this type (1) is 36; (2) is also 36
(a load typically expands to two 32 bit loads, but any bits beyond 36
are discarded); (3) is 64 given my current implementation (a store writes
two 32 bit values; bits beyond bit 36 hold some rubbish which overwrites
whatever was originally at that memory location); (4) needs to be at
least 64; (5) will also be 64.

In general (1) and (2) will be the same.  (4) needs to be at least
as big as (3).  (5) needs to be at least as big as (4).

Another example is 80-bit floating point types.  Here (1), (2)
and (3) are presumably 80 bits.  On my machine (5) is 96 bits.
I'm not sure what (4) is, presumably 80 or 96.

Which (if any) of these should getTypeSize, getABITypeSize, getTypeSizeInBits
and getABITypeSizeInBits correspond to?

It seems clear that getTypeSizeInBits corresponds to (1) and (2), as
shown by it returning 36 for i36.  This is like gcc's TYPE_PRECISION,
and is a useful concept - but I think the name should be changed, since
right now it implicitly suggests it returns 8*getTypeSize.  If no one
objects, I will rename it to getBitsUsedByType.

Currently getTypeSize doesn't seem to correspond to any of these possibilities,
at least for APInt's: the current implementation returns the APInt bitwidth rounded
up to a multiple of the alignment.  That makes it sound like it's trying to be (5).
I think getTypeSize should be defined to be (3), the maximum number of bits that
may be overwritten by a store [except that it's in bytes].  This means changing the
implementation for APInts, but not for other types.

Clearly getABITypeSize corresponds to (5) [except that it's in bytes].  This
corresponds to gcc's TYPE_SIZE.

Currently getABITypeSizeInBits returns 36 for i36, and otherwise 8*getABITypeSize.
It seems to me that this is clearly wrong for APInts, and that it should always
return 8*getABITypeSize (in which case it can be eliminated).  If no one objects,
I will delete it as redundant.

Finally, looking through various users of getTypeSize, it seems that they assume
that getTypeSize returns (4), the amount of memory allocated for a variable of
this type.  That seems reasonable.  If everyone agrees, I will document that
for LLVM (3) and (4) coincide, and is what getTypeSize returns [except that it
returns the number of bytes, rather than the number of bits].

Best wishes,

Duncan.