assuming that all offsets are legal vector accesses, and thus trying to access
the float member of { <2 x float>, float } as the 3rd element of the first
member.
llvm-svn: 132766
former was using the size of the entire alloca, whereas the latter was correctly using
the allocated size of the immediate type being converted (which may differ from the size
of the alloca). This fixes PR10082.
llvm-svn: 132759
the same allocation size but different primitive sizes(e.g., <3xi32> and
<4xi32>). When ScalarRepl promotes them, it can't use a bit cast but
should use a shuffle vector instead.
llvm-svn: 129472
chose is having a non-memcpy/memset use and being larger than any native integer
type. Originally I chose having an access of a size smaller than the total size
of the alloca, but this caused some minor issues on the spirit benchmark where
SRoA runs again after some inlining.
This fixes <rdar://problem/8613163>.
llvm-svn: 127718
alloca as both integer and floating-point vectors of the same size. Bugpoint is
not cooperating with me, but I'll try to find a manual testcase tomorrow.
llvm-svn: 127320
a union of a float, <2 x float>, and <4 x float>. This mostly comes up with the
use of vector intrinsics, especially in NEON when programmers know the layout of
the register file. This enables codegen to eliminate a lot of the subregister
traffic it would otherwise generate.
This commit only enables this for a small number of floating-point cases, but a
lot more integer cases. I assume this is okay for all ports, but I did not do
extensive testing of the quality of code involving i512 vectors and the like. If
there is a use case where this generates worse code than before, let me know and
we can scale it back.
This fixes <rdar://problem/9036264>.
llvm-svn: 127317
with BasicAA's DecomposeGEPExpression, which recently began
using a TargetData. This fixes PR8968, though the testcase
is awkward to reduce.
Also, update several off GetUnderlyingObject's users
which happen to have a TargetData handy to pass it in.
llvm-svn: 124134
occurs because instcombine sinks loads and inserts phis. This kicks in
on such apps as 175.vpr, eon, 403.gcc, xalancbmk and a bunch of times in
spec2006 in some app that uses std::deque.
This resolves the last of rdar://7339113.
llvm-svn: 124090
common cases. This triggers a surprising number of times in SPEC2K6
because min/max idioms end up doing this. For example, code from the
STL ends up looking like this to SRoA:
%202 = load i64* %__old_size, align 8, !tbaa !3
%203 = load i64* %__old_size, align 8, !tbaa !3
%204 = load i64* %__n, align 8, !tbaa !3
%205 = icmp ult i64 %203, %204
%storemerge.i = select i1 %205, i64* %__n, i64* %__old_size
%206 = load i64* %storemerge.i, align 8, !tbaa !3
We can now promote both the __n and the __old_size allocas.
This addresses another chunk of rdar://7339113, poor codegen on
stringswitch.
llvm-svn: 124088
that have PHI or select uses of their element pointers. This can often happen
when instcombine sinks two loads into a successor, inserting a phi or select.
With this patch, we can scalarize the alloca, but the pinned elements are not
yet promoted. This is still a win for large aggregates where only one element
is used. This fixes rdar://8904039 and part of rdar://7339113 (poor codegen
on stringswitch).
llvm-svn: 124070
handle the "Transformation preventing inst" printing,
so that -scalarrepl -debug will always print the rejected
instruction. No functionality change.
llvm-svn: 124066
checks enabled:
1) Use '<' to compare integers in a comparison function rather than '<='.
2) Use the uniqued set DefBlocks rather than Info.DefiningBlocks to initialize
the priority queue.
The speedup of scalarrepl on test-suite + SPEC2000 + SPEC2006 is a bit less, at
just under 16% rather than 17%.
llvm-svn: 123662
eliminating a potentially quadratic data structure, this also gives a 17%
speedup when running -scalarrepl on test-suite + SPEC2000 + SPEC2006. My initial
experiment gave a greater speedup around 25%, but I moved the dominator tree
level computation from dominator tree construction to PromoteMemToReg.
Since this approach to computing IDFs has a much lower overhead than the old
code using precomputed DFs, it is worth looking at using this new code for the
second scalarrepl pass as well.
llvm-svn: 123609
then don't try to decimate it into its individual pieces. This will just make a mess of the
IR and is pointless if none of the elements are individually accessed. This was generating
really terrible code for std::bitset (PR8980) because it happens to be lowered by clang
as an {[8 x i8]} structure instead of {i64}.
The testcase now is optimized to:
define i64 @test2(i64 %X) {
br label %L2
L2: ; preds = %0
ret i64 %X
}
before we generated:
define i64 @test2(i64 %X) {
%sroa.store.elt = lshr i64 %X, 56
%1 = trunc i64 %sroa.store.elt to i8
%sroa.store.elt8 = lshr i64 %X, 48
%2 = trunc i64 %sroa.store.elt8 to i8
%sroa.store.elt9 = lshr i64 %X, 40
%3 = trunc i64 %sroa.store.elt9 to i8
%sroa.store.elt10 = lshr i64 %X, 32
%4 = trunc i64 %sroa.store.elt10 to i8
%sroa.store.elt11 = lshr i64 %X, 24
%5 = trunc i64 %sroa.store.elt11 to i8
%sroa.store.elt12 = lshr i64 %X, 16
%6 = trunc i64 %sroa.store.elt12 to i8
%sroa.store.elt13 = lshr i64 %X, 8
%7 = trunc i64 %sroa.store.elt13 to i8
%8 = trunc i64 %X to i8
br label %L2
L2: ; preds = %0
%9 = zext i8 %1 to i64
%10 = shl i64 %9, 56
%11 = zext i8 %2 to i64
%12 = shl i64 %11, 48
%13 = or i64 %12, %10
%14 = zext i8 %3 to i64
%15 = shl i64 %14, 40
%16 = or i64 %15, %13
%17 = zext i8 %4 to i64
%18 = shl i64 %17, 32
%19 = or i64 %18, %16
%20 = zext i8 %5 to i64
%21 = shl i64 %20, 24
%22 = or i64 %21, %19
%23 = zext i8 %6 to i64
%24 = shl i64 %23, 16
%25 = or i64 %24, %22
%26 = zext i8 %7 to i64
%27 = shl i64 %26, 8
%28 = or i64 %27, %25
%29 = zext i8 %8 to i64
%30 = or i64 %29, %28
ret i64 %30
}
In this case, instcombine was able to eliminate the nonsense, but in PR8980 enough
PHIs are in play that instcombine backs off. It's better to not generate this stuff
in the first place.
llvm-svn: 123571
multiple uses. In some cases, all the uses are the same operation,
so instcombine can go ahead and promote the phi. In the testcase
this pushes an add out of the loop.
llvm-svn: 123568
instead of DomTree/DomFrontier. This may be interesting for reducing compile
time. This is currently disabled, but seems to work just fine.
When this is enabled, we eliminate two runs of dominator frontier, one in the
"early per-function" optimizations and one in the "interlaced with inliner"
function passes.
llvm-svn: 123434
This is a minor extension of SROA to handle a special case that is
important for some ARM NEON operations. Some of the NEON intrinsics
return multiple values, which are handled as struct types containing
multiple elements of the same vector type. The corresponding return
types declared in the arm_neon.h header have equivalent arrays. We
need SROA to recognize that it can split up those arrays and structs
into separate vectors, even though they are not always accessed with
the same type. SROA already handles loads and stores of an entire
alloca by using insertvalue/extractvalue to access the individual
pieces, and that code works the same regardless of whether the type
is a struct or an array. So, all that needs to be done is to check
for compatible arrays and homogeneous structs.
llvm-svn: 123381
SROA only split up structs and arrays one level at a time, so padding can
only cause trouble if it is located in between the struct or array elements.
llvm-svn: 123380
if it is passed as a byval argument. The byval argument will just be a
read, so it is safe to read from the original global instead. This allows
us to promote away the %agg.tmp alloca in PR8582
llvm-svn: 119686
must be called in the pass's constructor. This function uses static dependency declarations to recursively initialize
the pass's dependencies.
Clients that only create passes through the createFooPass() APIs will require no changes. Clients that want to use the
CommandLine options for passes will need to manually call the appropriate initialization functions in PassInitialization.h
before parsing commandline arguments.
I have tested this with all standard configurations of clang and llvm-gcc on Darwin. It is possible that there are problems
with the static dependencies that will only be visible with non-standard options. If you encounter any crash in pass
registration/creation, please send the testcase to me directly.
llvm-svn: 116820
perform initialization without static constructors AND without explicit initialization
by the client. For the moment, passes are required to initialize both their
(potential) dependencies and any passes they preserve. I hope to be able to relax
the latter requirement in the future.
llvm-svn: 116334
The x86_mmx type is used for MMX intrinsics, parameters and
return values where these use MMX registers, and is also
supported in load, store, and bitcast.
Only the above operations generate MMX instructions, and optimizations
do not operate on or produce MMX intrinsics.
MMX-sized vectors <2 x i32> etc. are lowered to XMM or split into
smaller pieces. Optimizations may occur on these forms and the
result casted back to x86_mmx, provided the result feeds into a
previous existing x86_mmx operation.
The point of all this is prevent optimizations from introducing
MMX operations, which is unsafe due to the EMMS problem.
llvm-svn: 115243
on llvmdev: SRoA is introducing MMX datatypes like <1 x i64>,
which then cause random problems because the X86 backend is
producing mmx stuff without inserting proper emms calls.
In the short term, force off MMX datatypes. In the long term,
the X86 backend should not select generic vector types to MMX
registers. This is being worked on, but won't be done in time
for 2.8. rdar://8380055
llvm-svn: 112696
large integers, the first inserted value would always create
an 'or X, 0'. Even though this is trivially zapped by
instcombine, don't bother creating this pointless instruction.
llvm-svn: 106979
with a fix for self-hosting
rotate CallInst operands, i.e. move callee to the back
of the operand array
the motivation for this patch are laid out in my mail to llvm-commits:
more efficient access to operands and callee, faster callgraph-construction,
smaller compiler binary
llvm-svn: 101465
with a fix
rotate CallInst operands, i.e. move callee to the back
of the operand array
the motivation for this patch are laid out in my mail to llvm-commits:
more efficient access to operands and callee, faster callgraph-construction,
smaller compiler binary
llvm-svn: 101397
of the operand array
the motivation for this patch are laid out in my mail to llvm-commits:
more efficient access to operands and callee, faster callgraph-construction,
smaller compiler binary
llvm-svn: 101364
Added support for address spaces and added a isVolatile field to memcpy, memmove, and memset,
e.g., llvm.memcpy.i32(i8*, i8*, i32, i32) -> llvm.memcpy.p0i8.p0i8.i32(i8*, i8*, i32, i32, i1)
llvm-svn: 100304
Added support for address spaces and added a isVolatile field to memcpy, memmove, and memset,
e.g., llvm.memcpy.i32(i8*, i8*, i32, i32) -> llvm.memcpy.p0i8.p0i8.i32(i8*, i8*, i32, i32, i1)
llvm-svn: 100191
e.g., llvm.memcpy.i32(i8*, i8*, i32, i32) -> llvm.memcpy.p0i8.p0i8.i32(i8*, i8*, i32, i32, i1)
A update of langref will occur in a subsequent checkin.
llvm-svn: 99928
The SRThreshold value makes perfect sense for checking if an entire aggregate
should be promoted to a scalar integer, but it is not so good for splitting
an aggregate into its separate elements. A struct may contain a large embedded
array along with some scalar fields that would benefit from being split apart
by SROA. Even if the total aggregate size is large, it may still be good to
perform SROA. Thus, the most important piece of this patch is simply moving
the aggregate size comparison vs. SRThreshold so that it guards only the
aggregate promotion.
We have also been checking the number of elements to decide if an aggregate
should be split up. The limit of "SRThreshold/4" seemed rather arbitrary,
and I don't think it's very useful to derive this limit from SRThreshold
anyway. I've collected some data showing that the current default limit of
32 (since SRThreshold defaults to 128) is a reasonable cutoff for struct
types. One thing suggested by the data is that distinguishing between structs
and arrays might be useful. There are (obviously) a lot more large arrays
than large structs (as measured by the number of elements and not the total
size -- a large array inside a struct still counts as a single element given
the way we do SROA right now). Out of 8377 arrays where we successfully
performed SROA while compiling a large set of benchmarks, only 16 of them had
more than 8 elements. And, for those 16 arrays, it's not at all clear that
SROA was actually beneficial. So, to offset the compile time cost of
investigating more large structs for SROA, the patch lowers the limit on array
elements to 8.
This fixes Apple Radar 7563690.
llvm-svn: 95224
parameter with a default value, instead of just hardcoding it in the
implementation. The limit of MaxLookup = 6 was introduced in r69151 to fix
a performance problem with O(n^2) behavior in instcombine, but the scalarrepl
pass is relying on getUnderlyingObject to go all the way back to an AllocaInst.
Making the limit part of the method signature makes it clear that by default
the result is limited and should help avoid similar problems in the future.
This fixes pr6126.
llvm-svn: 94433
are the same. I had already fixed a similar problem where the source and
destination were different bitcasts derived from the same alloca, but the
previous fix still did not handle the case where both operands are exactly
the same value. Radar 7552893.
llvm-svn: 93848
load is needed when we have a small store into a large alloca (at which
point we get a load/insert/store sequence), but when you do a full-sized
store, this load ends up being dead.
This dead load is bad in really large nasty testcases where the load ends
up causing mem2reg to insert large chains of dependent phi nodes which only
ADCE can delete. Instead of doing this, just don't insert the dead load.
This fixes rdar://6864035
llvm-svn: 91917
missing check that an array reference doesn't go past the end of the array,
and remove some redundant checks for in-bound array and vector references
that are no longer needed.
llvm-svn: 91897
two-element arrays. After restructuring the SROA code, it was not safe to
do this without adding more checking. It is not clear that this special-case
has really been useful, and removing this simplifies the code quite a bit.
llvm-svn: 91828
* change FindElementAndOffset to return a uint64_t instead of unsigned, and
to identify the type to be used for that result in a GEP instruction.
* move "isa<ConstantInt>" to be first in conditional.
* replace some dyn_casts with casts.
* add a comment about handling mem intrinsics.
llvm-svn: 91762
problem", this broke llvm-gcc bootstrap for release builds on
x86_64-apple-darwin10.
This reverts commit db22309800b224a9f5f51baf76071d7a93ce59c9.
llvm-svn: 91534
found last time. Instead of trying to modify the IR while iterating over it,
I've change it to keep a list of WeakVH references to dead instructions, and
then delete those instructions later. I also added some special case code to
detect and handle the situation when both operands of a memcpy intrinsic are
referencing the same alloca.
llvm-svn: 91459
While scanning through the uses of an alloca, keep track of the current offset
relative to the start of the alloca, and check memory references to see if
the offset & size correspond to a component within the alloca. This has the
nice benefit of unifying much of the code from isSafeUseOfAllocation,
isSafeElementUse, and isSafeUseOfBitCastedAllocation. The code to rewrite
the uses of a promoted alloca, after it is determined to be safe, is
reorganized in the same way.
Also, when rewriting GEP instructions, mark them as "in-bounds" since all the
indices are known to be safe.
llvm-svn: 91184
array indexes. The "complex" case of SRoA still handles them, and correctly.
This fixes a weirdness where we'd correctly avoid transforming A[0][42] if
the 42 was too large, but we'd only do it if it was one gep, not two separate
ones.
llvm-svn: 90007
the new predicates I added) instead of going through a context and doing a
pointer comparison. Besides being cheaper, this allows a smart compiler
to turn the if sequence into a switch.
llvm-svn: 83297
a Twine, e.g., for names).
- I am a little ambivalent about this; we don't want the string conversion of
utostr, but using overload '+' mixed with string and integer arguments is
sketchy. On the other hand, this particular usage is something of an idiom.
llvm-svn: 77579
Cameron Zwarich