The meaning of getTypeSize was not clear - clarifying it is important
now that we have x86 long double and arbitrary precision integers.
The issue with long double is that it requires 80 bits, and this is
not a multiple of its alignment. This gives a primitive type for
which getTypeSize differed from getABITypeSize. For arbitrary precision
integers it is even worse: there is the minimum number of bits needed to
hold the type (eg: 36 for an i36), the maximum number of bits that will
be overwriten when storing the type (40 bits for i36) and the ABI size
(i.e. the storage size rounded up to a multiple of the alignment; 64 bits
for i36).
This patch removes getTypeSize (not really - it is still there but
deprecated to allow for a gradual transition). Instead there is:
(1) getTypeSizeInBits - a number of bits that suffices to hold all
values of the type. For a primitive type, this is the minimum number
of bits. For an i36 this is 36 bits. For x86 long double it is 80.
This corresponds to gcc's TYPE_PRECISION.
(2) getTypeStoreSizeInBits - the maximum number of bits that is
written when storing the type (or read when reading it). For an
i36 this is 40 bits, for an x86 long double it is 80 bits. This
is the size alias analysis is interested in (getTypeStoreSize
returns the number of bytes). There doesn't seem to be anything
corresponding to this in gcc.
(3) getABITypeSizeInBits - this is getTypeStoreSizeInBits rounded
up to a multiple of the alignment. For an i36 this is 64, for an
x86 long double this is 96 or 128 depending on the OS. This is the
spacing between consecutive elements when you form an array out of
this type (getABITypeSize returns the number of bytes). This is
TYPE_SIZE in gcc.
Since successive elements in a SequentialType (arrays, pointers
and vectors) need to be aligned, the spacing between them will be
given by getABITypeSize. This means that the size of an array
is the length times the getABITypeSize. It also means that GEP
computations need to use getABITypeSize when computing offsets.
Furthermore, if an alloca allocates several elements at once then
these too need to be aligned, so the size of the alloca has to be
the number of elements multiplied by getABITypeSize. Logically
speaking this doesn't have to be the case when allocating just
one element, but it is simpler to also use getABITypeSize in this
case. So alloca's and mallocs should use getABITypeSize. Finally,
since gcc's only notion of size is that given by getABITypeSize, if
you want to output assembler etc the same as gcc then getABITypeSize
is the size you want.
Since a store will overwrite no more than getTypeStoreSize bytes,
and a read will read no more than that many bytes, this is the
notion of size appropriate for alias analysis calculations.
In this patch I have corrected all type size uses except some of
those in ScalarReplAggregates, lib/Codegen, lib/Target (the hard
cases). I will get around to auditing these too at some point,
but I could do with some help.
Finally, I made one change which I think wise but others might
consider pointless and suboptimal: in an unpacked struct the
amount of space allocated for a field is now given by the ABI
size rather than getTypeStoreSize. I did this because every
other place that reserves memory for a type (eg: alloca) now
uses getABITypeSize, and I didn't want to make an exception
for unpacked structs, i.e. I did it to make things more uniform.
This only effects structs containing long doubles and arbitrary
precision integers. If someone wants to pack these types more
tightly they can always use a packed struct.
llvm-svn: 43620
In the old way, we computed and inserted phi nodes for the whole IDF of
the definitions of the alloca, then computed which ones were dead and
removed them.
In the new method, we first compute the region where the value is live,
and use that information to only insert phi nodes that are live. This
eliminates the need to compute liveness later, and stops the algorithm
from inserting a bunch of phis which it then later removes.
This speeds up the testcase in PR1432 from 2.00s to 0.15s (14x) in a
release build and 6.84s->0.50s (14x) in a debug build.
llvm-svn: 40825
stored value was a non-instruction value. Doh.
This increase the # single store allocas from 8982 to 9026, and
speeds up mem2reg on the testcase in PR1432 from 2.17 to 2.13s.
llvm-svn: 40813
1. Check for revisiting a block before checking domination, which is faster.
2. If the stored value isn't an instruction, we don't have to check for domination.
3. If we have a value used in the same block more than once, make sure to remove the
block from the UsingBlocks vector. Not doing so forces us to go through the slow
path for the alloca.
The combination of these improvements increases the number of allocas on the fastpath
from 8935 to 8982 on PR1432. This speeds it up from 2.90s to 2.20s (31%)
llvm-svn: 40811
a using block from the list if we handle it. Not doing this caused us
to not be able to promote (with the fast path) allocas which have uses (whoops).
This increases the # allocas hitting this fastpath from 4042 to 8935 on the
testcase in PR1432, speeding up mem2reg by 2.6x
llvm-svn: 40809
