Make necessary changes to support DIV -> [SUF]Div. This changes llvm to
have three division instructions: signed, unsigned, floating point. The
bytecode and assembler are bacwards compatible, however.
llvm-svn: 31195
1. Better document what is going on here.
2. Only hack on one branch per iteration, making the results less conservative.
3. Handle the problematic case by marking edges executable instead of by
playing with value lattice states. This is far less pessimistic, and fixes
SCCP/ipsccp-gvar.ll.
llvm-svn: 31106
This patch implements the first increment for the Signless Types feature.
All changes pertain to removing the ConstantSInt and ConstantUInt classes
in favor of just using ConstantInt.
llvm-svn: 31063
SimplifyDemandedBits. The idea is that some operations can be simplified if
not all of the computed elements are needed. Some targets (like x86) have a
large number of intrinsics that operate on a single element, but pass other
elts through unmodified. If those other elements are not needed, the
intrinsics can be simplified to scalar operations, and insertelement ops can
be removed.
This turns (f.e.):
ushort %Convert_sse(float %f) {
%tmp = insertelement <4 x float> undef, float %f, uint 0 ; <<4 x float>> [#uses=1]
%tmp10 = insertelement <4 x float> %tmp, float 0.000000e+00, uint 1 ; <<4 x float>> [#uses=1]
%tmp11 = insertelement <4 x float> %tmp10, float 0.000000e+00, uint 2 ; <<4 x float>> [#uses=1]
%tmp12 = insertelement <4 x float> %tmp11, float 0.000000e+00, uint 3 ; <<4 x float>> [#uses=1]
%tmp28 = tail call <4 x float> %llvm.x86.sse.sub.ss( <4 x float> %tmp12, <4 x float> < float 1.000000e+00, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00 > ) ; <<4 x float>> [#uses=1]
%tmp37 = tail call <4 x float> %llvm.x86.sse.mul.ss( <4 x float> %tmp28, <4 x float> < float 5.000000e-01, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00 > ) ; <<4 x float>> [#uses=1]
%tmp48 = tail call <4 x float> %llvm.x86.sse.min.ss( <4 x float> %tmp37, <4 x float> < float 6.553500e+04, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00 > ) ; <<4 x float>> [#uses=1]
%tmp59 = tail call <4 x float> %llvm.x86.sse.max.ss( <4 x float> %tmp48, <4 x float> zeroinitializer ) ; <<4 x float>> [#uses=1]
%tmp = tail call int %llvm.x86.sse.cvttss2si( <4 x float> %tmp59 ) ; <int> [#uses=1]
%tmp69 = cast int %tmp to ushort ; <ushort> [#uses=1]
ret ushort %tmp69
}
into:
ushort %Convert_sse(float %f) {
entry:
%tmp28 = sub float %f, 1.000000e+00 ; <float> [#uses=1]
%tmp37 = mul float %tmp28, 5.000000e-01 ; <float> [#uses=1]
%tmp375 = insertelement <4 x float> undef, float %tmp37, uint 0 ; <<4 x float>> [#uses=1]
%tmp48 = tail call <4 x float> %llvm.x86.sse.min.ss( <4 x float> %tmp375, <4 x float> < float 6.553500e+04, float undef, float undef, float undef > ) ; <<4 x float>> [#uses=1]
%tmp59 = tail call <4 x float> %llvm.x86.sse.max.ss( <4 x float> %tmp48, <4 x float> < float 0.000000e+00, float undef, float undef, float undef > ) ; <<4 x float>> [#uses=1]
%tmp = tail call int %llvm.x86.sse.cvttss2si( <4 x float> %tmp59 ) ; <int> [#uses=1]
%tmp69 = cast int %tmp to ushort ; <ushort> [#uses=1]
ret ushort %tmp69
}
which improves codegen from:
_Convert_sse:
movss LCPI1_0, %xmm0
movss 4(%esp), %xmm1
subss %xmm0, %xmm1
movss LCPI1_1, %xmm0
mulss %xmm0, %xmm1
movss LCPI1_2, %xmm0
minss %xmm0, %xmm1
xorps %xmm0, %xmm0
maxss %xmm0, %xmm1
cvttss2si %xmm1, %eax
andl $65535, %eax
ret
to:
_Convert_sse:
movss 4(%esp), %xmm0
subss LCPI1_0, %xmm0
mulss LCPI1_1, %xmm0
movss LCPI1_2, %xmm1
minss %xmm1, %xmm0
xorps %xmm1, %xmm1
maxss %xmm1, %xmm0
cvttss2si %xmm0, %eax
andl $65535, %eax
ret
This is just a first step, it can be extended in many ways. Testcase here:
Transforms/InstCombine/vec_demanded_elts.ll
llvm-svn: 30752
The critical edge block dominates the dest block if the destblock dominates
all edges other than the one incoming from the critical edge.
llvm-svn: 30696
or when splitting loops with a common header into multiple loops. In particular
the old code would always insert the preheader before the old loop header. This
is disasterous in cases where the loop hasn't been rotated. For example, it can
produce code like:
.. outside the loop...
jmp LBB1_2 #bb13.outer
LBB1_1: #bb1
movsd 8(%esp,%esi,8), %xmm1
mulsd (%edi), %xmm1
addsd %xmm0, %xmm1
addl $24, %edi
incl %esi
jmp LBB1_3 #bb13
LBB1_2: #bb13.outer
leal (%edx,%eax,8), %edi
pxor %xmm1, %xmm1
xorl %esi, %esi
LBB1_3: #bb13
movapd %xmm1, %xmm0
cmpl $4, %esi
jl LBB1_1 #bb1
Note that the loop body is actually LBB1_1 + LBB1_3, which means that the
loop now contains an uncond branch WITHIN it to jump around the inserted
loop header (LBB1_2). Doh.
This patch changes the preheader insertion code to insert it in the right
spot, producing this code:
... outside the loop, fall into the header ...
LBB1_1: #bb13.outer
leal (%edx,%eax,8), %esi
pxor %xmm0, %xmm0
xorl %edi, %edi
jmp LBB1_3 #bb13
LBB1_2: #bb1
movsd 8(%esp,%edi,8), %xmm0
mulsd (%esi), %xmm0
addsd %xmm1, %xmm0
addl $24, %esi
incl %edi
LBB1_3: #bb13
movapd %xmm0, %xmm1
cmpl $4, %edi
jl LBB1_2 #bb1
Totally crazy, no branch in the loop! :)
llvm-svn: 30587
reachable, making it general purpose enough for use by InsertPreheaderForLoop.
Eliminate custom dominfo updating code in InsertPreheaderForLoop, using
UpdateDomInfoForRevectoredPreds instead.
llvm-svn: 30586
DLL* linkages got full (I hope) codegeneration support in C & both x86
assembler backends.
External weak linkage added for future use, we don't provide any
codegeneration, etc. support for it.
llvm-svn: 30374
operations (like findProperties) should be faster, at the expense of
unionSets being slower in cases that are rare in practise.
Don't erase a dead Instruction. This fixes a memory corruption issue.
llvm-svn: 30235
Reorder operations to remove duplicated work.
Fix to leave floating-point types out of the optimization.
Add tests to predsimplify.ll for SwitchInst and SelectInst handling.
llvm-svn: 30055
If a branch's condition has become a ConstantBool, simplify it immediately.
Removing the edge saves work and exposes up more optimization opportunities
in the pass.
Add support for SelectInst.
llvm-svn: 29970
Not only will this take huge amounts of compile time, the resultant loop nests
won't be useful for optimization. This reduces loopsimplify time on
Transforms/LoopSimplify/2006-08-11-LoopSimplifyLongTime.ll from ~32s to ~0.4s
with a debug build of llvm on a 2.7Ghz G5.
llvm-svn: 29647
blocks that target loop blocks.
Before, the code was run once per loop, and depended on the number of
predecessors each block in the loop had. Unfortunately, scanning preds can
be really slow when huge numbers of phis exist or when phis with huge numbers
of inputs exist.
Now, the code is run once per function and scans successors instead of preds,
which is far faster. In addition, the new code is simpler and is goto free,
woo.
This change speeds up a nasty testcase Duraid provided me from taking hours to
taking ~72s with a debug build. The functionality this implements is already
tested in the testsuite as Transforms/CodeExtractor/2004-03-13-LoopExtractorCrash.ll.
llvm-svn: 29644
SlowOperatingInfo, Statistics). Besides providing an example of how to
use these facilities, it also serves to debug problems with runtime linking
when dlopening a loadable module. These three support facilities exercise
different combinations of Text/Weak Weak/Text and Text/Text linking
between the executable and the module.
llvm-svn: 29552
1. Change the usage of LOADABLE_MODULE so that it implies all the things
necessary to make a loadable module. This reduces the user's burdern to
get a loadable module correctly built.
2. Document the usage of LOADABLE_MODULE in the MakefileGuide
3. Adjust the makefile for lib/Transforms/Hello to use the new specification
for building loadable modules
4. Adjust the sample project to not attempt to build a shared library for
its little library. This was just wasteful and not instructive at all.
llvm-svn: 29551
1. Update an obsolete comment.
2. Make the sorting by base an explicit (though still N^2) step, so
that the code is more clear on what it is doing.
3. Partition uses so that uses inside the loop are handled before uses
outside the loop.
Note that none of these changes currently changes the code inserted by LSR,
but they are a stepping stone to getting there.
This code is the result of some crazy pair programming with Nate. :)
llvm-svn: 29493
down approach, inspired by discussions with Tanya.
This approach is significantly faster, because it does not need dominator
frontiers and it does not insert extraneous unused PHI nodes. For example, on
252.eon, in a release-asserts build, this speeds up LCSSA (which is the slowest
pass in gccas) from 9.14s to 0.74s on my G5. This code is also slightly smaller
and significantly simpler than the old code.
Amusingly, in a normal Release build (which includes the
"assert(L->isLCSSAForm());" assertion), asserting that the result of LCSSA
is in LCSSA form is actually slower than the LCSSA transformation pass
itself on 252.eon. I will see if Loop::isLCSSAForm can be sped up next.
llvm-svn: 29463
target CG node. This allows the inliner to properly update the callgraph
when using the pruning inliner. The pruning inliner may not copy over all
call sites from a callee to a caller, so the edges corresponding to those
call sites should not be copied over either.
This fixes PR827 and Transforms/Inline/2006-07-12-InlinePruneCGUpdate.ll
llvm-svn: 29120
will be profitable. This is mainly to remove some cases where excessive
unswitching would result in long compile times and/or huge generated code.
Once someone comes up with a better heuristic that avoids these cases, this
should be switched out.
llvm-svn: 28962
Remove the Function pointer cast in these calls, converting it to
a cast of argument.
%tmp60 = tail call int cast (int (ulong)* %str to int (int)*)( int 10 )
%tmp60 = tail call int cast (int (ulong)* %str to int (int)*)( uint %tmp51 )
llvm-svn: 28953
"LCSSA" phi node causes indvars to break dominance properties. This fixes
causes indvars to avoid inserting aggressive code in this case, instead
indvars should be fixed to be more aggressive in the face of lcssa phi's.
llvm-svn: 28850
not handling PHI nodes correctly when determining if a value was live-out.
This patch reduces the number of detected live-out variables in the testcase
from 6565 to 485.
llvm-svn: 28771
If a single exit block has multiple predecessors within the loop, it will
appear in the exit blocks list more than once. LCSSA needs to take that into
account so that it doesn't double process that exit block.
llvm-svn: 28750
post-increment value, should be first cast to the appropriated type (to the
type of the common expr). Otherwise, the rewrite of a use based on (common +
iv) may end up with an incorrect type.
llvm-svn: 28735
code (while cloning) it often gets the branch/switch instructions. Since it
knows that edges of the CFG are dead, it need not clone (or even look) at
the obviously dead blocks. This should speed up the inliner substantially on
code where there are lots of inlinable calls to functions with constant
arguments. On C++ code in particular, this kicks in.
llvm-svn: 28641
the iterated Dominance Frontier of the loop-closure Phi's. This is the
second phase of the LCSSA pass. The third phase (coming soon) will be to
update all uses of loop variables to use the loop-closure Phi's instead.
llvm-svn: 28524
makes it so that it constant folds instructions on the fly. This is good
for several reasons:
0. Many instructions are constant foldable after inlining, particularly if
inlining a call with constant arguments.
1. Without this, the inliner has to allocate memory for all of the instructions
that can be constant folded, then a subsequent pass has to delete them. This
gets the job done without this extra work.
2. This makes the inliner *pass* a bit more aggressive: in particular, it
partially solves a phase order issue where the inliner would inline lots
of code that folds away to nothing, but think that the resultant function
is big because of this code that will be gone. Now the code never exists.
This is the first part of a 2-step process. The second part will be smart
enough to see when this implicit constant folding propagates a constant into
a branch or switch instruction, making CFG edges dead.
This implements Transforms/Inline/inline_constprop.ll
llvm-svn: 28521
When doing the initial pass of constant folding, if we get a constantexpr,
simplify the constant expr like we would do if the constant is folded in the
normal loop.
This fixes the missed-optimization regression in
Transforms/InstCombine/getelementptr.ll last night.
llvm-svn: 28224
1. Implement InstCombine/deadcode.ll by not adding instructions in unreachable
blocks (due to constants in conditional branches/switches) to the worklist.
This causes them to be deleted before instcombine starts up, leading to
better optimization.
2. In the prepass over instructions, do trivial constprop/dce as we go. This
has the effect of improving the effectiveness of #1. In addition, it
*significantly* speeds up instcombine on test cases with large amounts of
constant folding code (for example, that produced by code specialization
or partial evaluation). In one example, it speeds up instcombine from
0.0589s to 0.0224s with a release build (a 2.6x speedup).
llvm-svn: 28215
Make the "fold (and (cast A), (cast B)) -> (cast (and A, B))" transformation
only apply when both casts really will cause code to be generated. If one or
both doesn't, then this xform doesn't remove a cast.
This fixes Transforms/InstCombine/2006-05-06-Infloop.ll
llvm-svn: 28141
nondeterminism being bad) could cause some trivial missed optimizations (dead
phi nodes being left around for later passes to clean up).
With this, llvm-gcc4 now bootstraps and correctly compares. I don't know
why I never tried to do it before... :)
llvm-svn: 27984
are visible to analysis as intrinsics. That is, make sure someone doesn't pass
free around by address in some struct (as happens in say 176.gcc).
This doesn't get rid of any indirect calls, just ensure calls to free and malloc
are always direct.
llvm-svn: 27560
%tmp = cast <4 x uint> %tmp to <4 x int> ; <<4 x int>> [#uses=1]
%tmp = cast <4 x int> %tmp to <4 x float> ; <<4 x float>> [#uses=1]
into:
%tmp = cast <4 x uint> %tmp to <4 x float> ; <<4 x float>> [#uses=1]
llvm-svn: 27355
%tmp = cast <4 x uint>* %testData to <4 x int>* ; <<4 x int>*> [#uses=1]
%tmp = load <4 x int>* %tmp ; <<4 x int>> [#uses=1]
to this:
%tmp = load <4 x uint>* %testData ; <<4 x uint>> [#uses=1]
%tmp = cast <4 x uint> %tmp to <4 x int> ; <<4 x int>> [#uses=1]
llvm-svn: 27353
stride. For a set of uses of the IV of a stride which is a multiple
of another stride, do not insert a new IV expression. Rather, reuse the
previous IV and rewrite the uses as uses of IV expression multiplied by
the factor.
e.g.
x = 0 ...; x ++
y = 0 ...; y += 4
then use of y can be rewritten as use of 4*x for x86.
llvm-svn: 26803
the pointer is known to come from either a global variable, alloca or
malloc. This allows us to compile this:
P = malloc(28);
memset(P, 0, 28);
into explicit stores on PPC instead of a memset call.
llvm-svn: 26577
Make this code more powerful by using ComputeMaskedBits instead of looking
for an AND operand. This lets us fold this:
int %test23(int %a) {
%tmp.1 = and int %a, 1
%tmp.2 = seteq int %tmp.1, 0
%tmp.3 = cast bool %tmp.2 to int ;; xor tmp1, 1
ret int %tmp.3
}
into: xor (and a, 1), 1
llvm-svn: 26396
caused SPASS to fail building last night.
We can't trivially unswitch a loop if the exit block has phi nodes in it,
because we don't know which predecessor to use.
llvm-svn: 26320
This fixes a testcase that nate reduced from spass.
Also included are a couple minor code changes that don't affect the generated
code at all.
llvm-svn: 26235
unswitch this loop on 2 before sweating to unswitch on 1/3.
void test4(int N, int i, int C, int*P, int*Q) {
int j;
for (j = 0; j < N; ++j) {
switch (C) { // general unswitching.
default: P[i+j] = 0; break;
case 1: Q[i+j] = 0; break;
case 3: P[i+j] = Q[i+j]; break;
case 2: break; // TRIVIAL UNSWITCH on C==2
}
}
}
llvm-svn: 26223
this for example:
for (j = 0; j < N; ++j) { // trivial unswitch
if (C)
P[i+j] = 0;
}
turning it into the obvious code without bothering to duplicate an empty loop.
llvm-svn: 26220
1. Teach GetConstantInType to handle boolean constants.
2. Teach instcombine to fold (compare X, CST) when X has known 0/1 bits.
Testcase here: set.ll:test22
3. Improve the "(X >> c1) & C2 == 0" folding code to allow a noop cast
between the shift and and. More aggressive bitfolding for other reasons
was turning signed shr's into unsigned shr's, leaving the noop cast in
the way.
llvm-svn: 26131
This allows us to simplify on conditions where bits are not known, but they
are not demanded either! This also fixes a couple of bugs in
ComputeMaskedBits that were exposed during this work.
In the future, swaths of instcombine should be removed, as this code
subsumes a bunch of ad-hockery.
llvm-svn: 26122
1. Teach it new tricks: in particular how to propagate through signed shr and sexts.
2. Teach it to return a bitset of known-1 and known-0 bits, instead of just zero.
3. Teach instcombine (AND X, C) to fold when we know all C bits of X.
This implements Regression/Transforms/InstCombine/bittest.ll, and allows
future things to be simplified.
llvm-svn: 26087
instruction onto the worklist (in case they are now dead).
Add a really trivial local DSE implementation to help out bitfield code.
We now fold this:
struct S {
unsigned char a : 1, b : 1, c : 1, d : 2, e : 3;
S();
};
S::S() : a(0), b(0), c(1), d(0), e(6) {}
to this:
void %_ZN1SC1Ev(%struct.S* %this) {
entry:
%tmp.1 = getelementptr %struct.S* %this, int 0, uint 0
store ubyte 38, ubyte* %tmp.1
ret void
}
much earlier (in gccas instead of only in gccld after DSE runs).
llvm-svn: 26050
mask. This allows the code to be simpler and more efficient.
Also, generalize some of the cases in MVIZ a bit, making it slightly more aggressive.
llvm-svn: 26035
'demanded bits', inspired by Nate's work in the dag combiner. This isn't
complete, but needs to unrelated instcombiner changes to continue.
llvm-svn: 26033
1. When rewriting code in outer loops, sometimes we would insert code into
inner loops that is invariant in that loop.
2. Notice that 4*(2+x) is 8+4*x and use that to simplify expressions.
This is a performance neutral change.
llvm-svn: 25964
1. Do not statically construct a map when the program starts up, this
is expensive and cannot be optimized. Instead, create a list.
2. Do not insert entries for all function in the module into a hashmap
that lives the full life of the compiler.
llvm-svn: 25512
1. Use the varargs version of getOrInsertFunction to simplify code.
2. remove #include
3. Reduce the number of #ifdef's.
4. remove extraneous vertical whitespace.
llvm-svn: 25508
Don't do floor->floorf conversion if floorf is not available. This checks
the compiler's host, not its target, which is incorrect for cross-compilers
Not sure that's important as we don't build many cross-compilers.
llvm-svn: 25456
This patch is an incremental step towards supporting a flat symbol table.
It de-overloads the intrinsic functions by providing type-specific intrinsics
and arranging for automatically upgrading from the old overloaded name to
the new non-overloaded name. Specifically:
llvm.isunordered -> llvm.isunordered.f32, llvm.isunordered.f64
llvm.sqrt -> llvm.sqrt.f32, llvm.sqrt.f64
llvm.ctpop -> llvm.ctpop.i8, llvm.ctpop.i16, llvm.ctpop.i32, llvm.ctpop.i64
llvm.ctlz -> llvm.ctlz.i8, llvm.ctlz.i16, llvm.ctlz.i32, llvm.ctlz.i64
llvm.cttz -> llvm.cttz.i8, llvm.cttz.i16, llvm.cttz.i32, llvm.cttz.i64
New code should not use the overloaded intrinsic names. Warnings will be
emitted if they are used.
llvm-svn: 25366
it doesn't contain any calls. This is a fairly common case for C++ code,
so it will probably speed up the inliner marginally in these cases.
llvm-svn: 25284
function was not an alloca, we wouldn't check the entry block for any allocas,
leading to increased stack space in some cases. In practice, allocas are almost
always at the top of the block, so this was never noticed.
llvm-svn: 25280
the shifts.
This allows us to fold this (which is the 'integer add a constant' sequence
from cozmic's scheme compmiler):
int %x(uint %anf-temporary776) {
%anf-temporary777 = shr uint %anf-temporary776, ubyte 1
%anf-temporary800 = cast uint %anf-temporary777 to int
%anf-temporary804 = shl int %anf-temporary800, ubyte 1
%anf-temporary805 = add int %anf-temporary804, -2
%anf-temporary806 = or int %anf-temporary805, 1
ret int %anf-temporary806
}
into this:
int %x(uint %anf-temporary776) {
%anf-temporary776 = cast uint %anf-temporary776 to int
%anf-temporary776.mask1 = add int %anf-temporary776, -2
%anf-temporary805 = or int %anf-temporary776.mask1, 1
ret int %anf-temporary805
}
note that instcombine already knew how to eliminate the AND that the two
shifts fold into. This is tested by InstCombine/shift.ll:test26
-Chris
llvm-svn: 25128
a) use better local variable names (OldMT -> OldFT) where "M" is used to
mean "Function" (perhaps it was previously "Method"?)
b) print out the module identifier in a warning message so that it is
possible to track down in which module the error occurred.
llvm-svn: 24698
186.crafty by about 16% (from 15.109s to 13.045s) on my system.
This turns allocas with unions/casts into scalars. For example crafty has
something like this:
union doub {
unsigned short i[4];
long long d;
};
int f(long long a) {
return ((union doub){.d=a}).i[1];
}
Instead of generating loads and stores to an alloca, we now promote the
whole thing to a scalar long value.
This implements: Transforms/ScalarRepl/AggregatePromote.ll
llvm-svn: 24667
The code is organized into 3 parts (2 passes)
1) a linked set of profiling passes, which implement an analysis group (linked, like alias analysis are). These insert profiling into the program, and remember what they inserted, so that at a later time they can be queried about any instruction.
2) a pass that handles inserting the random sampling framework. This also has options to control how random samples are choosen. Currently implemented are Global counters, register allocated global counters, and read cycle counter (see? there was a reason for it).
The profiling passes are almost identical to the existing ones (block, function, and null profiling is supported right now), and they are valid passes without the sampling framework (hence the existing passes can be unified with the new ones, not done yet).
Some things are a bit ugly still, but that should be fixed up soon enough.
Other todo? making the counter values not "magic 2^16 -1" values, but dynamically choosable.
llvm-svn: 24493
has a single def. In this case, look for uses that are dominated by the def
and attempt to rewrite them to directly use the stored value.
This speeds up mem2reg on these values and reduces the number of phi nodes
inserted. This should address PR665.
llvm-svn: 24411
Add support for specifying alignment and size of setjmp jmpbufs.
No targets currently do anything with this information, nor is it presrved
in the bytecode representation. That's coming up next.
llvm-svn: 24196
a few times in crafty:
OLD: %tmp.36 = div int %tmp.35, 8 ; <int> [#uses=1]
NEW: %tmp.36 = div uint %tmp.35, 8 ; <uint> [#uses=0]
OLD: %tmp.19 = div int %tmp.18, 8 ; <int> [#uses=1]
NEW: %tmp.19 = div uint %tmp.18, 8 ; <uint> [#uses=0]
OLD: %tmp.117 = div int %tmp.116, 8 ; <int> [#uses=1]
NEW: %tmp.117 = div uint %tmp.116, 8 ; <uint> [#uses=0]
OLD: %tmp.92 = div int %tmp.91, 8 ; <int> [#uses=1]
NEW: %tmp.92 = div uint %tmp.91, 8 ; <uint> [#uses=0]
Which all turn into shrs.
llvm-svn: 24190
8 times in vortex, allowing the srems to be turned into shrs:
OLD: %tmp.104 = rem int %tmp.5.i37, 16 ; <int> [#uses=1]
NEW: %tmp.104 = rem uint %tmp.5.i37, 16 ; <uint> [#uses=0]
OLD: %tmp.98 = rem int %tmp.5.i24, 16 ; <int> [#uses=1]
NEW: %tmp.98 = rem uint %tmp.5.i24, 16 ; <uint> [#uses=0]
OLD: %tmp.91 = rem int %tmp.5.i19, 8 ; <int> [#uses=1]
NEW: %tmp.91 = rem uint %tmp.5.i19, 8 ; <uint> [#uses=0]
OLD: %tmp.88 = rem int %tmp.5.i14, 8 ; <int> [#uses=1]
NEW: %tmp.88 = rem uint %tmp.5.i14, 8 ; <uint> [#uses=0]
OLD: %tmp.85 = rem int %tmp.5.i9, 1024 ; <int> [#uses=2]
NEW: %tmp.85 = rem uint %tmp.5.i9, 1024 ; <uint> [#uses=0]
OLD: %tmp.82 = rem int %tmp.5.i, 512 ; <int> [#uses=2]
NEW: %tmp.82 = rem uint %tmp.5.i1, 512 ; <uint> [#uses=0]
OLD: %tmp.48.i = rem int %tmp.5.i.i161, 4 ; <int> [#uses=1]
NEW: %tmp.48.i = rem uint %tmp.5.i.i161, 4 ; <uint> [#uses=0]
OLD: %tmp.20.i2 = rem int %tmp.5.i.i, 4 ; <int> [#uses=1]
NEW: %tmp.20.i2 = rem uint %tmp.5.i.i, 4 ; <uint> [#uses=0]
it also occurs 9 times in gcc, but with odd constant divisors (1009 and 61)
so the payoff isn't as great.
llvm-svn: 24189
into the LLVMAnalysis library.
This allows LLVMTranform and LLVMTransformUtils to be archives and linked
with LLVMAnalysis.a, which provides any missing definitions.
llvm-svn: 24036
SparcV9 JIT.
2. Make LLVMTransformUtils a relinked object file and always link it before
LLVMAnalysis.a. These two libraries have circular dependencies on each
other which creates problem when building the SparcV9 JIT. This change
fixes the dependency on all platforms problems with a minimum of fuss.
llvm-svn: 24023
one use (but one is a cast). This handles the very common case of:
X = alloc [n x byte]
Y = cast X to somethingbetter
seteq X, null
In order to avoid infinite looping when there are multiple casts, we only
allow this if the xform is strictly increasing the alignment of the
allocation.
llvm-svn: 23961
where the second has less alignment required. If we had explicit alignment
support in the IR, we could handle this case, but we can't until we do.
llvm-svn: 23960
pointer marking the end of the list, the zero *must* be cast to the pointer
type. An un-cast zero is a 32-bit int, and at least on x86_64, gcc will
not extend the zero to 64 bits, thus allowing the upper 32 bits to be
random junk.
The new END_WITH_NULL macro may be used to annotate a such a function
so that GCC (version 4 or newer) will detect the use of un-casted zero
at compile time.
llvm-svn: 23888
check the presplit pred, not the post-split pred. This was causing us
to make the wrong decision in some cases, leaving the critical edge block
in the loop.
llvm-svn: 23601
is performed so it is only at most once per function that contains an invoke
instead of once per invoke in the function. This patch has the following perks:
1. It fixes PR631, which complains about slowness.
2. If fixes PR240, which complains about non-volatile vars being live across
setjmp/longjmps.
3. It improves (but does not fix) the jmpbuf alignment issue on itanium by not
forcing the jmpbufs to always be 8-bytes off the alignment of the structure.
4. It speeds up 253.perlbmk from 338s to 13.70s (a 25x improvement!), making us
now about 4% faster than GCC.
Further improvements are also possible.
llvm-svn: 23477
Implement the start of global ctor optimization. It is currently smart
enough to remove the global ctor for cases like this:
struct foo {
foo() {}
} x;
... saving a bit of startup time for the program.
llvm-svn: 23433
not define a value that is used outside of it's block. This catches many
more simplifications, e.g. 854 in 176.gcc, 137 in vpr, etc.
This implements branch-phi-thread.ll:test3.ll
llvm-svn: 23397
if () { store A -> P; } else { store B -> P; }
into a PHI node with one store, in the most trival case. This implements
load.ll:test10.
llvm-svn: 23324
load are exactly consequtive. This is picked up by other passes, but this
triggers thousands of times in fortran programs that use static locals
(and is thus a compile-time speedup).
llvm-svn: 23320
code for IV uses outside of loops that are not dominated by the latch block.
We should only convert these uses to use the post-inc value if they ARE
dominated by the latch block.
Also use a new LoopInfo method to simplify some code.
This fixes Transforms/LoopStrengthReduce/2005-09-12-UsesOutOutsideOfLoop.ll
llvm-svn: 23318
in building maximal expressions before simplifying them. In particular, i
cases like this:
X-(A+B+X)
the code would consider A+B+X to be a maximal expression (not understanding
that the single use '-' would be turned into a + later), simplify it (a noop)
then later get simplified again.
Each of these simplify steps is where the cost of reassociation comes from,
so this patch should speed up the already fast pass a bit.
Thanks to Dan for noticing this!
llvm-svn: 23214
Do not claim to not change the CFG. We do change the cfg to split critical
edges. This isn't causing us a problem now, but could likely do so in the
future.
llvm-svn: 22824
edge so that the code is not always executed for both operands. This
prevents LSR from inserting code into loops whose exit blocks contain
PHI uses of IV expressions (which are outside of loops). On gzip, for
example, we turn this ugly code:
.LBB_test_1: ; loopentry
add r27, r3, r28
lhz r27, 3(r27)
add r26, r4, r28
lhz r26, 3(r26)
add r25, r30, r28 ;; Only live if exiting the loop
add r24, r29, r28 ;; Only live if exiting the loop
cmpw cr0, r27, r26
bne .LBB_test_5 ; loopexit
into this:
.LBB_test_1: ; loopentry
or r27, r28, r28
add r28, r3, r27
lhz r28, 3(r28)
add r26, r4, r27
lhz r26, 3(r26)
cmpw cr0, r28, r26
beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2: ; loopentry.loopexit_crit_edge
add r2, r30, r27
add r8, r29, r27
b .LBB_test_9 ; loopexit
.LBB_test_2: ; shortcirc_next.0
...
blt .LBB_test_1
into this:
.LBB_test_1: ; loopentry
or r27, r28, r28
add r28, r3, r27
lhz r28, 3(r28)
add r26, r4, r27
lhz r26, 3(r26)
cmpw cr0, r28, r26
beq .LBB_test_3 ; shortcirc_next.0
.LBB_test_2: ; loopentry.loopexit_crit_edge
add r2, r30, r27
add r8, r29, r27
b .LBB_t_3: ; shortcirc_next.0
.LBB_test_3: ; shortcirc_next.0
...
blt .LBB_test_1
Next step: get the block out of the loop so that the loop is all
fall-throughs again.
llvm-svn: 22766
Instead, just update the BB in-place. This is both faster, and it prevents
split-critical-edges from shuffling the PHI argument list unneccesarily.
llvm-svn: 22765
into just Y. This often occurs when it seperates loops that have collapsed loop
headers. This implements LoopSimplify/phi-node-simplify.ll
llvm-svn: 22746
For code like this:
void foo(float *a, float *b, int n, int stride_a, int stride_b) {
int i;
for (i=0; i<n; i++)
a[i*stride_a] = b[i*stride_b];
}
we now emit:
.LBB_foo2_2: ; no_exit
lfs f0, 0(r4)
stfs f0, 0(r3)
addi r7, r7, 1
add r4, r2, r4
add r3, r6, r3
cmpw cr0, r7, r5
blt .LBB_foo2_2 ; no_exit
instead of:
.LBB_foo_2: ; no_exit
mullw r8, r2, r7 ;; multiply!
slwi r8, r8, 2
lfsx f0, r4, r8
mullw r8, r2, r6 ;; multiply!
slwi r8, r8, 2
stfsx f0, r3, r8
addi r2, r2, 1
cmpw cr0, r2, r5
blt .LBB_foo_2 ; no_exit
loops with variable strides occur pretty often. For example, in SPECFP2K
there are 317 variable strides in 177.mesa, 3 in 179.art, 14 in 188.ammp,
56 in 168.wupwise, 36 in 172.mgrid.
Now we can allow indvars to turn functions written like this:
void foo2(float *a, float *b, int n, int stride_a, int stride_b) {
int i, ai = 0, bi = 0;
for (i=0; i<n; i++)
{
a[ai] = b[bi];
ai += stride_a;
bi += stride_b;
}
}
into code like the above for better analysis. With this patch, they generate
identical code.
llvm-svn: 22740
The termination condition actually wants to use the post-incremented value
of the loop, not a new indvar with an unusual base.
On PPC, for example, this allows us to compile
LoopStrengthReduce/exit_compare_live_range.ll to:
_foo:
li r2, 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r2, r2, 1
cmpw cr0, r2, r4
bne .LBB_foo_1 ; no_exit
blr
instead of:
_foo:
li r2, 1 ;; IV starts at 1, not 0
.LBB_foo_1: ; no_exit
li r5, 0
stw r5, 0(r3)
addi r5, r2, 1
cmpw cr0, r2, r4
or r2, r5, r5 ;; Reg-reg copy, extra live range
bne .LBB_foo_1 ; no_exit
blr
This implements LoopStrengthReduce/exit_compare_live_range.ll
llvm-svn: 22699
* Teach this code to move allocas out of the loop when tail call eliminating
a call marked 'tail'. This implements TailCallElim/move_alloca_for_tail_call.ll
* Do not perform this transformation if a call is marked 'tail' and if there
are allocas that we cannot move out of the loop in #2. Doing so would increase
the stack usage of the function. This implements fixes
PR615 and TailCallElim/dont-tce-tail-marked-call.ll.
llvm-svn: 22690
BasicBlock's removePredecessor routine. This requires shuffling around
the definition and implementation of hasContantValue from Utils.h,cpp into
Instructions.h,cpp
llvm-svn: 22664
that the symbolic evaluator is not always able to use subtraction to remove
expressions. This makes the code faster, and fixes the last crash on 178.galgel.
Finally, add a statistic to see how many phi nodes are inserted.
On 178.galgel, we get the follow stats:
2562 loop-reduce - Number of PHIs inserted
3927 loop-reduce - Number of GEPs strength reduced
llvm-svn: 22662
method.
* Fix a crash on 178.galgel, where we would insert expressions before PHI
nodes instead of into the PHI node predecessor blocks.
llvm-svn: 22657
for (i = 0; i < N; ++i)
A[i][foo()] = 0;
here we still want to strength reduce the A[i] part, even though foo() is
l-v.
This also simplifies some of the 'CanReduce' logic.
This implements Transforms/LoopStrengthReduce/ops_after_indvar.ll
llvm-svn: 22652
1. We only analyze instructions once, guaranteed
2. AnalyzeGetElementPtrUsers has been ripped apart and replaced with
something much simpler.
The next step is to handle expressions that are not all indvar+loop-invariant
values (e.g. handling indvar+loopvariant).
llvm-svn: 22649
Only emit one PHI node for IV uses with identical bases and strides (after
moving foldable immediates to the load/store instruction).
This implements LoopStrengthReduce/dont_insert_redundant_ops.ll, allowing
us to generate this PPC code for test1:
or r30, r3, r3
.LBB_test1_1: ; Loop
li r2, 0
stw r2, 0(r30)
stw r2, 4(r30)
bl L_pred$stub
addi r30, r30, 8
cmplwi cr0, r3, 0
bne .LBB_test1_1 ; Loop
instead of this code:
or r30, r3, r3
or r29, r3, r3
.LBB_test1_1: ; Loop
li r2, 0
stw r2, 0(r29)
stw r2, 4(r30)
bl L_pred$stub
addi r30, r30, 8 ;; Two iv's with step of 8
addi r29, r29, 8
cmplwi cr0, r3, 0
bne .LBB_test1_1 ; Loop
llvm-svn: 22635
map from instruction* to SCEVHandles. When we delete instructions, we have
to tell it about it. We would run into nasty cases where new instructions
were reallocated at old instruction addresses and get the old map values.
Bad bad bad :(
llvm-svn: 22632
consideration the case where a reference in an unreachable block could
occur. This fixes Transforms/SimplifyCFG/2005-08-01-PHIUpdateFail.ll,
something I ran into while bugpoint'ing another pass.
llvm-svn: 22584
SimplifyLibCalls probably has to be audited to make sure it does not make
this mistake elsewhere. Also, if this code knows that the type will be
unsigned, obviously one arm of this is dead.
Reid, can you take a look into this further?
llvm-svn: 22566
target data to decide which loop induction variables to strength reduce
and how to do so. This work is mostly by Chris Lattner, with tweaks by
me to get it working on some of MultiSource.
llvm-svn: 22558
Because the instcombine has to scan the entire function when it starts up
to begin with, we might as well do it in DFO so we can nuke unreachable code.
This fixes: Transforms/InstCombine/2005-07-07-DeadPHILoop.ll
llvm-svn: 22348
The optimization for locally used allocas was not safe for allocas that
were read before they were written. This change disables that optimization
in that case.
llvm-svn: 22318
is a mismatch in their character type pointers (i.e. fprintf() prints an
array of ubytes while fwrite() takes an array of sbytes).
We can probably do better than this (such as casting the ubyte to an
sbyte).
llvm-svn: 22310
* Check for availability of ffsll call in configure script
* Support ffs, ffsl, and ffsll conversion to constant value if the argument
is constant.
llvm-svn: 22027
This makes reassociate realize that loads should be treated as unmovable, and
gives distinct ranks to distinct values defined in the same basic block, allowing
reassociate to do its thing.
llvm-svn: 21783
of trying to do local reassociation tweaks at each level, only process an expression
tree once (at its root). This does not improve the reassociation pass in any real way.
llvm-svn: 21768
strlen(x) != 0 -> *x != 0
strlen(x) == 0 -> *x == 0
* Change nested statistics to use style of other LLVM statistics so that
only the name of the optimization (simplify-libcalls) is used as the
statistic name, and the description indicates which specific all is
optimized. Cuts down on some redundancy and saves a few bytes of space.
* Make note of stpcpy optimization that could be done.
llvm-svn: 21766
the result, turn signed shift rights into unsigned shift rights if possible.
This leads to later simplification and happens *often* in 176.gcc. For example,
this testcase:
struct xxx { unsigned int code : 8; };
enum codes { A, B, C, D, E, F };
int foo(struct xxx *P) {
if ((enum codes)P->code == A)
bar();
}
used to be compiled to:
int %foo(%struct.xxx* %P) {
%tmp.1 = getelementptr %struct.xxx* %P, int 0, uint 0 ; <uint*> [#uses=1]
%tmp.2 = load uint* %tmp.1 ; <uint> [#uses=1]
%tmp.3 = cast uint %tmp.2 to int ; <int> [#uses=1]
%tmp.4 = shl int %tmp.3, ubyte 24 ; <int> [#uses=1]
%tmp.5 = shr int %tmp.4, ubyte 24 ; <int> [#uses=1]
%tmp.6 = cast int %tmp.5 to sbyte ; <sbyte> [#uses=1]
%tmp.8 = seteq sbyte %tmp.6, 0 ; <bool> [#uses=1]
br bool %tmp.8, label %then, label %UnifiedReturnBlock
Now it is compiled to:
%tmp.1 = getelementptr %struct.xxx* %P, int 0, uint 0 ; <uint*> [#uses=1]
%tmp.2 = load uint* %tmp.1 ; <uint> [#uses=1]
%tmp.2 = cast uint %tmp.2 to sbyte ; <sbyte> [#uses=1]
%tmp.8 = seteq sbyte %tmp.2, 0 ; <bool> [#uses=1]
br bool %tmp.8, label %then, label %UnifiedReturnBlock
which is the difference between this:
foo:
subl $4, %esp
movl 8(%esp), %eax
movl (%eax), %eax
shll $24, %eax
sarl $24, %eax
testb %al, %al
jne .LBBfoo_2
and this:
foo:
subl $4, %esp
movl 8(%esp), %eax
movl (%eax), %eax
testb %al, %al
jne .LBBfoo_2
This occurs 3243 times total in the External tests, 215x in povray,
6x in each f2c'd program, 1451x in 176.gcc, 7x in crafty, 20x in perl,
25x in gap, 3x in m88ksim, 25x in ijpeg.
Maybe this will cause a little jump on gcc tommorow :)
llvm-svn: 21715
This implements set.ll:test20.
This triggers 2x on povray, 9x on mesa, 11x on gcc, 2x on crafty, 1x on eon,
6x on perlbmk and 11x on m88ksim.
It allows us to compile these two functions into the same code:
struct s { unsigned int bit : 1; };
unsigned foo(struct s *p) {
if (p->bit)
return 1;
else
return 0;
}
unsigned bar(struct s *p) { return p->bit; }
llvm-svn: 21690
library function:
isdigit(chr) -> 0 or 1 if chr is constant
isdigit(chr) -> chr - '0' <= 9 otherwise
Although there are many calls to isdigit in llvm-test, most of them are
compiled away by macros leaving only this:
2 MultiSource/Applications/hexxagon
llvm-svn: 21688
actual spec (int -> uint)
* Add the ability to get/cache the strlen function prototype.
* Make sure generated values are appropriately named for debugging purposes
* Add the SPrintFOptimiation for 4 casts of sprintf optimization:
sprintf(str,cstr) -> llvm.memcpy(str,cstr) (if cstr has no %)
sprintf(str,"") -> store sbyte 0, str
sprintf(str,"%s",src) -> llvm.memcpy(str,src) (if src is constant)
sprintf(str,"%c",chr) -> store chr, str ; store sbyte 0, str+1
The sprintf optimization didn't fire as much as I had hoped:
2 MultiSource/Applications/SPASS
5 MultiSource/Benchmarks/McCat/18-imp
22 MultiSource/Benchmarks/Prolangs-C/TimberWolfMC
1 MultiSource/Benchmarks/Prolangs-C/assembler
6 MultiSource/Benchmarks/Prolangs-C/unix-smail
2 MultiSource/Benchmarks/mediabench/mpeg2/mpeg2dec
llvm-svn: 21679
Neither of these activated as many times as was hoped:
strchr:
9 MultiSource/Applications/siod
1 MultiSource/Applications/d
2 MultiSource/Prolangs-C/archie-client
1 External/SPEC/CINT2000/176.gcc/176.gcc
llvm.memset:
no hits
llvm-svn: 21669
strings passed to Statistic's constructor are not destructable. The stats
are printed during static destruction and the SimplifyLibCalls module was
getting destructed before the statistics.
llvm-svn: 21661
type be obtained from a CallInst we're optimizing.
* Make it possible for getConstantStringLength to return the ConstantArray
that it extracts in case the content is needed by an Optimization.
* Implement the strcmp optimization
* Implement the toascii optimization
This pass is now firing several to many times in the following MultiSource
tests:
Applications/Burg - 7 (strcat,strcpy)
Applications/siod - 13 (strcat,strcpy,strlen)
Applications/spiff - 120 (exit,fputs,strcat,strcpy,strlen)
Applications/treecc - 66 (exit,fputs,strcat,strcpy)
Applications/kimwitu++ - 34 (strcmp,strcpy,strlen)
Applications/SPASS - 588 (exit,fputs,strcat,strcpy,strlen)
llvm-svn: 21626
sinh, cosh, etc.
* Make the name comparisons for the fp libcalls a little more efficient by
switching on the first character of the name before doing comparisons.
llvm-svn: 21611
* Correct stale documentation in a few places
* Re-order the file to better associate things and reduce line count
* Make the pass thread safe by caching the Function* objects needed by the
optimizers in the pass object instead of globally.
* Provide the SimplifyLibCalls pass object to the optimizer classes so they
can access cached Function* objects and TargetData info
* Make sure the pass resets its cache if the Module passed to runOnModule
changes
* Rename CallOptimizer LibCallOptimization. All the classes are named
*Optimization while the objects are *Optimizer.
* Don't cache Function* in the optimizer objects because they could be used
by multiple PassManager's running in multiple threads
* Add an optimization for strcpy which is similar to strcat
* Add a "TODO" list at the end of the file for ideas on additional libcall
optimizations that could be added (get ideas from other compilers).
Sorry for the huge diff. Its mostly reorganization of code. That won't
happen again as I believe the design and infrastructure for this pass is
now done or close to it.
llvm-svn: 21589
call to them into an 'unreachable' instruction.
This triggers a bunch of times, particularly on gcc:
gzip: 36
gcc: 601
eon: 12
bzip: 38
llvm-svn: 21587
* MemCpyOptimization can only be optimized if the 3rd and 4th arguments are
constants and we weren't checking for that.
* The result of llvm.memcpy (and llvm.memmove) is void* not sbyte*, put in
a cast.
llvm-svn: 21570
* Have the SimplifyLibCalls pass acquire the TargetData and pass it down to
the optimization classes so they can use it to make better choices for
the signatures of functions, etc.
* Rearrange the code a little so the utility functions are closer to their
usage and keep the core of the pass near the top of the files.
* Adjust the StrLen pass to get/use the correct prototype depending on the
TargetData::getIntPtrType() result. The result of strlen is size_t which
could be either uint or ulong depending on the platform.
* Clean up some coding nits (cast vs. dyn_cast, remove redundant items from
a switch, etc.)
* Implement the MemMoveOptimization as a twin of MemCpyOptimization (they
only differ in name).
llvm-svn: 21569
named getConstantStringLength. This is the common part of StrCpy and
StrLen optimizations and probably several others, yet to be written. It
performs all the validity checks for looking at constant arrays that are
supposed to be null-terminated strings and then computes the actual
length of the string.
* Implement the MemCpyOptimization class. This just turns memcpy of 1, 2, 4
and 8 byte data blocks that are properly aligned on those boundaries into
a load and a store. Much more could be done here but alignment
restrictions and lack of knowledge of the target instruction set prevent
use from doing significantly more. That will have to be delegated to the
code generators as they lower llvm.memcpy calls.
llvm-svn: 21562
* Change signatures of OptimizeCall and ValidateCalledFunction so they are
non-const, allowing the optimization object to be modified. This is in
support of caching things used across multiple calls.
* Provide two functions for constructing and caching function types
* Modify the StrCatOptimization to cache Function objects for strlen and
llvm.memcpy so it doesn't regenerate them on each call site. Make sure
these are invalidated each time we start the pass.
* Handle both a GEP Instruction and a GEP ConstantExpr
* Add additional checks to make sure we really are dealing with an arary of
sbyte and that all the element initializers are ConstantInt or
ConstantExpr that reduce to ConstantInt.
* Make sure the GlobalVariable is constant!
* Don't use ConstantArray::getString as it can fail and it doesn't give us
the right thing. We must check for null bytes in the middle of the array.
* Use llvm.memcpy instead of memcpy so we can factor alignment into it.
* Don't use void* types in signatures, replace with sbyte* instead.
llvm-svn: 21555
* Don't use std::string for the function names, const char* will suffice
* Allow each CallOptimizer to validate the function signature before
doing anything
* Repeatedly loop over the functions until an iteration produces
no more optimizations. This allows one optimization to insert a
call that is optimized by another optimization.
* Implement the ConstantArray portion of the StrCatOptimization
* Provide a template for the MemCpyOptimization
* Make ExitInMainOptimization split the block, not delete everything
after the return instruction.
(This covers revision 1.3 and 1.4, as the 1.3 comments were botched)
llvm-svn: 21548
* Fix comments at top of file
* Change algorithm for running the call optimizations from n*n to something
closer to n.
* Use a hash_map to store and lookup the optimizations since there will
eventually (or potentially) be a large number of them. This gets lookup
based on the name of the function to O(1). Each CallOptimizer now has a
std::string member named func_name that tracks the name of the function
that it applies to. It is this string that is entered into the hash_map
for fast comparison against the function names encountered in the module.
* Cleanup some style issues pertaining to iterator invalidation
* Don't pass the Function pointer to the OptimizeCall function because if
the optimization needs it, it can get it from the CallInst passed in.
* Add the skeleton for a new CallOptimizer, StrCatOptimizer which will
eventually replace strcat's of constant strings with direct copies.
llvm-svn: 21526
calls. The pass visits all external functions in the module and determines
if such function calls can be optimized. The optimizations are specific to
the library calls involved. This initial version only optimizes calls to
exit(3) when they occur in main(): it changes them to ret instructions.
llvm-svn: 21522
Completely rework the 'setcc (cast x to larger), y' code. This code has
the advantage of implementing setcc.ll:test19 (being more general than
the previous code) and being correct in all cases.
This allows us to unxfail 2004-11-27-SetCCForCastLargerAndConstant.ll,
and close PR454.
llvm-svn: 21491
Make IPSCCP strip off dead constant exprs that are using functions, making
them appear as though their address is taken. This allows us to propagate
some more pool descriptors, lowering the overhead of pool alloc.
llvm-svn: 21363
This pass forward branches through conditions when it can show that the
conditions is either always true or false for a predecessor. This currently
only handles the most simple cases of this, but is successful at threading
across 2489 branches and 65 switch instructions in 176.gcc, which isn't bad.
llvm-svn: 21306
* Loop invariant code does not dominate the loop header, but rather
the end of the loop preheader.
* The base for a reduced GEP isn't a constant unless all of its
operands (preceding the induction variable) are constant.
* Allow induction variable elimination for the simple case after all.
Also made changes recommended by Chris for properly deleting
instructions.
llvm-svn: 20383
This does a simple form of "jump threading", which eliminates CFG edges that
are provably dead. This triggers 90 times in the external tests, and
eliminating CFG edges is always always a good thing! :)
llvm-svn: 20300
and handle incomplete control dependences correctly. This fixes:
Regression/Transforms/ADCE/dead-phi-edge.ll
-> a missed optimization
Regression/Transforms/ADCE/dead-phi-edge.ll
-> a compiler crash distilled from QT4
llvm-svn: 20227
* Properly compile this:
struct a {};
int test() {
struct a b[2];
if (&b[0] != &b[1])
abort ();
return 0;
}
to 'return 0', not abort().
llvm-svn: 19875
The second folds operations into selects, e.g. (select C, (X+Y), (Y+Z))
-> (Y+(select C, X, Z)
This occurs a few times across spec, e.g.
select add/sub
mesa: 83 0
povray: 5 2
gcc 4 2
parser 0 22
perlbmk 13 30
twolf 0 3
llvm-svn: 19706
Disable the xform for < > cases. It turns out that the following is being
miscompiled:
bool %test(sbyte %S) {
%T = cast sbyte %S to uint
%V = setgt uint %T, 255
ret bool %V
}
llvm-svn: 19628
* We can now fold cast instructions into select instructions that
have at least one constant operand.
* We now optimize expressions more aggressively based on bits that are
known to be zero. These optimizations occur a lot in code that uses
bitfields even in simple ways.
* We now turn more cast-cast sequences into AND instructions. Before we
would only do this if it if all types were unsigned. Now only the
middle type needs to be unsigned (guaranteeing a zero extend).
* We transform sign extensions into zero extensions in several cases.
This corresponds to these test/Regression/Transforms/InstCombine testcases:
2004-11-22-Missed-and-fold.ll
and.ll: test28-29
cast.ll: test21-24
and-or-and.ll
cast-cast-to-and.ll
zeroext-and-reduce.ll
llvm-svn: 19220
SimplifyCFG is one of those passes that we use for final cleanup: it should
not rely on other passes to clean up its garbage. This fixes the "why are
trivially dead setcc's in the output of gccas" problem.
llvm-svn: 19212
do not insert a prototype for malloc of: void* malloc(uint): on 64-bit u
targets this is not correct. Instead of prototype it as void *malloc(...),
and pass the correct intptr_t through the "...".
Finally, fix Regression/CodeGen/SparcV9/2004-12-13-MallocCrash.ll, by not
forming constantexpr casts from pointer to uint.
llvm-svn: 18908
in SPEC, the subsequent optimziations that we are after don't play with
with FP values, so disable this xform for them. Really we just don't want
stuff like:
double G; (always 0 or 412312.312)
= G;
turning into:
bool G_b;
= G_b ? 412312.312 : 0;
We'd rather just do the load.
-Chris
llvm-svn: 18819
down to actually BE a bool. This allows simple value range propagation
stuff work harder, deleting comparisons in bzip2 in some hot loops.
This implements GlobalOpt/integer-bool.ll, which is the essence of the
loop condition distilled into a testcase.
llvm-svn: 18817
1. Actually increment the Statistic for the GV elim optzn
2. When resolving undef branches, only resolve branches in executable blocks,
avoiding marking a bunch of completely dead blocks live. This has a big
impact on the quality of the generated code.
With this patch, we positively rip up vortex, compiling Ut_MoveBytes to a
single memcpy call. In vortex we get this:
12 ipsccp - Number of globals found to be constant
986 ipsccp - Number of arguments constant propagated
1378 ipsccp - Number of basic blocks unreachable
8919 ipsccp - Number of instructions removed
llvm-svn: 18796
This implements SCCP/ipsccp-basic.ll, rips apart Olden/mst (as described in
PR415), and does other nice things.
There is still more to come with this, but it's a start.
llvm-svn: 18752
successor block. This turns cases like this:
x = a op b
if (c) {
use x
}
into:
if (c) {
x = a op b
use x
}
This triggers 3965 times in spec, and is tested by
Regression/Transforms/InstCombine/sink_instruction.ll
This appears to expose a bug in the X86 backend for 177.mesa, which I'm
looking in to.
llvm-svn: 18677
in scary and unknown ways before we promote it. This fixes the miscompilation
of 188.ammp that has been plauging us since a globalopt patch went in.
Thanks a ton to Tanya for helping me diagnose the problem!
llvm-svn: 18418
if (x) {
code
...
} else {
code
...
}
Turn it into:
code
if (x) {
...
} else {
...
}
This reduces code size and in some common cases allows us to completely
eliminate the conditional. This turns several if/then/else blocks in loops
into straightline code in 179.art, turning the loops into single basic blocks
(good for modsched even!).
Maybe now brg will leave me alone ;-)
llvm-svn: 18366
1. Speedup getValueState by having it not consider Arguments. It's better
to just add them before we start SCCP'ing.
2. SCCP can delete the contents of dead blocks. No really, it's ok! This
reduces the size of the IR for subsequent passes, even though
simplifycfg would do the same job. In practice, simplifycfg does not
run until much later than sccp in gccas
llvm-svn: 17820
class. The only changes are minor:
* Do not try to SCCP instructions that return void in the rewrite loop.
This is silly and fool hardy, wasting a map lookup and adding an entry
to the map which is never used.
* If we decide something has an undefined value, rewrite it to undef,
potentially leading to further simplications.
llvm-svn: 17816
value. This allows us to turn more globals into constants and eliminate them.
This patch implements GlobalOpt/load-store-global.llx.
Note that this patch speeds up 255.vortex from:
Output/255.vortex.out-cbe.time:program 7.640000
Output/255.vortex.out-llc.time:program 9.810000
to:
Output/255.vortex.out-cbe.time:program 7.250000
Output/255.vortex.out-llc.time:program 9.490000
Which isn't bad at all!
llvm-svn: 17746
If this happens, detect it early instead of relying on instcombine to notice
it later. This can be a big speedup, because PHI nodes can have many
incoming values.
llvm-svn: 17741
%X = alloca ...
%Y = alloca ...
X == Y
into false. This allows us to simplify some stuff in eon (and probably
many other C++ programs) where operator= was checking for self assignment.
Folding this allows us to SROA several additional structs.
llvm-svn: 17735
constant value. This makes the return value dead and allows for
simplification in the caller.
This implements IPConstantProp/return-constant.ll
This triggers several dozen times throughout SPEC.
llvm-svn: 17730
of the array is just two. This occurs 8 times in gcc, 6 times in crafty, and
12 times in 099.go.
This implements ScalarRepl/sroa_two.ll
llvm-svn: 17727
argument pointers. This is only valid to do if the function already
unconditionally loaded an argument or if the pointer passed in is known
to be valid. Make sure to do the required checks.
This fixed ArgumentPromotion/control-flow.ll and the Burg program.
llvm-svn: 17718
for (X * C1) + (X * C2) (where * can be mul or shl), allowing us to fold:
Y+Y+Y+Y+Y+Y+Y+Y
into
%tmp.8 = shl long %Y, ubyte 3 ; <long> [#uses=1]
instead of
%tmp.4 = shl long %Y, ubyte 2 ; <long> [#uses=1]
%tmp.12 = shl long %Y, ubyte 2 ; <long> [#uses=1]
%tmp.8 = add long %tmp.4, %tmp.12 ; <long> [#uses=1]
This implements add.ll:test25
Also add support for (X*C1)-(X*C2) -> X*(C1-C2), implementing sub.ll:test18
llvm-svn: 17704
This allows to elimination of a bunch of global pool descriptor args from
programs being pool allocated (and is also generally useful!)
llvm-svn: 17657
loops. This optimization is not turned on by default yet, but may be run
with the opt tool's -loop-reduce flag. There are many FIXMEs listed in the
code that will make it far more applicable to a wide range of code, but you
have to start somewhere :)
This limited version currently triggers on the following tests in the
MultiSource directory:
pcompress2: 7 times
cfrac: 5 times
anagram: 2 times
ks: 6 times
yacr2: 2 times
llvm-svn: 17134
change hacks off 10K of bytecode from perlbmk (.5%) even though the front-end
is not generating them yet and we are not optimizing the resultant code.
This isn't too bad.
llvm-svn: 17111
exercise that I'm not interested in tackling right now. Just punt and treat them
like unwind's.
This 'fixes' test/Regression/Transforms/ADCE/unreachable-function.ll
llvm-svn: 17106
unneccesary. This allows us to delete several hundred phi nodes of the
form PHI(x,x,x,undef) from 253.perlbmk and probably other programs as well.
This implements Mem2Reg/UndefValuesMerge.ll
llvm-svn: 17098
pointer recurrences into expressions from this:
%P_addr.0.i.0 = phi sbyte* [ getelementptr ([8 x sbyte]* %.str_1, int 0, int 0), %entry ], [ %inc.0.i, %no_exit.i ]
%inc.0.i = getelementptr sbyte* %P_addr.0.i.0, int 1 ; <sbyte*> [#uses=2]
into this:
%inc.0.i = getelementptr sbyte* getelementptr ([8 x sbyte]* %.str_1, int 0, int 0), int %inc.0.i.rec
Actually create something nice, like this:
%inc.0.i = getelementptr [8 x sbyte]* %.str_1, int 0, int %inc.0.i.rec
llvm-svn: 16924
First, it allows SRA of globals that have embedded arrays, implementing
GlobalOpt/globalsra-partial.llx. This comes up infrequently, but does allow,
for example, deleting several stores to dead parts of globals in dhrystone.
Second, this implements GlobalOpt/malloc-promote-*.llx, which is the
following nifty transformation:
Basically if a global pointer is initialized with malloc, and we can tell
that the program won't notice, we transform this:
struct foo *FooPtr;
...
FooPtr = malloc(sizeof(struct foo));
...
FooPtr->A FooPtr->B
Into:
struct foo FooPtrBody;
...
FooPtrBody.A FooPtrBody.B
This comes up occasionally, for example, the 'disp' global in 183.equake (where
the xform speeds the CBE version of the program up from 56.16s to 52.40s (7%)
on apoc), and the 'desired_accept', 'fixLRBT', 'macroArray', & 'key_queue'
globals in 300.twolf (speeding it up from 22.29s to 21.55s (3.4%)).
The nice thing about this xform is that it exposes the resulting global to
global variable optimization and makes alias analysis easier in addition to
eliminating a few loads.
llvm-svn: 16916
still optimize away all of the indirect calls and loads, etc from it.
This turns code like this:
if (G != 0)
G();
into
if (G != 0)
ActualCallee();
This triggers a couple of times in gcc and libstdc++.
llvm-svn: 16901
stored to, but are stored at variable indexes. This occurs at least in
176.gcc, but probably others, and we should handle it for completeness.
llvm-svn: 16876
has a large number of users. Instead, just keep track of whether we're
making changes as we do so.
This patch has no functionlity changes.
llvm-svn: 16874
we know that all uses of the global will trap if the pointer contained is
null. In this case, we forward substitute the stored value to any uses.
This has the effect of devirtualizing trivial globals in trivial cases. For
example, 164.gzip contains this:
gzip.h:extern int (*read_buf) OF((char *buf, unsigned size));
bits.c: read_buf = file_read;
deflate.c: lookahead = read_buf((char*)window,
deflate.c: n = read_buf((char*)window+strstart+lookahead, more);
Since read_buf has to point to file_read at every use, we just replace
the calls through read_buf with a direct call to file_read.
This occurs in several benchmarks, including 176.gcc and 164.gzip. Direct
calls are good and stuff.
llvm-svn: 16871
* Do not lead dangling dead constants prevent optimization
* Iterate global optimization while we're making progress.
These changes allow us to be more aggressive, handling cases like
GlobalOpt/iterate.llx without a problem (turning it into 'ret int 0').
llvm-svn: 16857
optimizations to trigger much more often. This allows the elimination of
several dozen more global variables in Programs/External. Note that we only
do this for non-constant globals: constant globals will already be optimized
out if the accesses to them permit it.
This implements Transforms/GlobalOpt/globalsra.llx
llvm-svn: 16842
* Instead of handling dead functions specially, just nuke them.
* Be more aggressive about cleaning up after constification, in
particular, handle getelementptr instructions and constantexprs.
* Be a little bit more structured about how we process globals.
*** Delete globals that are only stored to, and never read. These are
clearly not useful, so they should go. This implements deadglobal.llx
This last one triggers quite a few times. In particular, 2208 in the
external tests, 1865 of which are in 252.eon. This shrinks eon from
1995094 to 1732341 bytes of bytecode.
llvm-svn: 16802
simplifications of the resultant program to avoid making later passes
do it all.
This allows us to constify globals that just have the same constant that
they are initialized stored into them.
Suprisingly this comes up ALL of the freaking time, dozens of times in
SPEC, 30 times in vortex alone.
For example, on 256.bzip2, it allows us to constify these two globals:
%smallMode = internal global ubyte 0 ; <ubyte*> [#uses=8]
%verbosity = internal global int 0 ; <int*> [#uses=49]
Which (with later optimizations) results in the bytecode file shrinking
from 82286 to 69686 bytes! Lets hear it for IPO :)
For the record, it's nuking lots of "if (verbosity > 2) { do lots of stuff }"
code.
llvm-svn: 16793
an instruction if it can be hoisted to a common dominator of the block.
This implements: test/Regression/Transforms/TailDup/MergeTest.ll
llvm-svn: 16758
* SubOne/AddOne functions always return ConstantInt, declare them as such
* Pull code for handling setcc X, cst, where cst is at the end of the range,
or cc is LE or GE up earlier in visitSetCondInst. This reduces #iterations
in some cases.
* Fold: (div X, C1) op C2 -> range check, implementing div.ll:test6 - test9.
llvm-svn: 16588
This takes something like this:
%A = phi int [ 3, %cond_false.0 ], [ 2, %endif.0.i ], [ 2, %endif.1.i ]
%B = div int %tmp.243, 4
and turns it into:
%A = phi int [ 3/4, %cond_false.0 ], [ 2/4, %endif.0.i ], [ 2/4, %endif.1.i ]
which is later simplified (in this case) into %A = 0.
This triggers thousands of times in spec, for example, 269 times in 176.gcc.
This is tested by InstCombine/add.ll:test23 and set.ll:test18.
llvm-svn: 16582
Instcombine (setcc (truncate X), C1).
This occurs THOUSANDS of times in many benchmarks. Particularlly common
seem to be things like (seteq (cast bool X to int), int 0)
This turns it into (seteq bool %X, false), which then becomes (not %X).
llvm-svn: 16567
This is important for several reasons:
1. Benchmarks have lots of code that looks like this (perlbmk in particular):
%tmp.2.i = setne int %tmp.0.i, 128 ; <bool> [#uses=1]
%tmp.6343 = seteq int %tmp.0.i, 1 ; <bool> [#uses=1]
%tmp.63 = and bool %tmp.2.i, %tmp.6343 ; <bool> [#uses=1]
we now fold away the setne, a clear improvement.
2. In the more important cases, such as (X >= 10) & (X < 20), we now produce
smaller code: (X-10) < 10.
3. Perhaps the nicest effect of this patch is that it really helps out the
code generators. In particular, for a 'range test' like the above,
instead of generating this on X86 (the difference on PPC is even more
pronounced):
cmp %EAX, 50
setge %CL
cmp %EAX, 100
setl %AL
and %CL, %AL
cmp %CL, 0
we now generate this:
add %EAX, -50
cmp %EAX, 50
Furthermore, this causes setcc's to be folded into branches more often.
These combinations trigger dozens of times in the spec benchmarks, particularly
in 176.gcc, 186.crafty, 253.perlbmk, 254.gap, & 099.go.
llvm-svn: 16559
Implement (setcc (shl X, C1), C2) folding.
The second one occurs several dozen times in spec. The first was added
just in case. :)
These are tested by shift.ll:test2[12], and div.ll:test5
llvm-svn: 16549
This latent bug was exposed by recent changes, and is tested as:
llvm/test/Regression/Transforms/InstCombine/2004-09-28-BadShiftAndSetCC.llx
llvm-svn: 16546
triggers often, for example:
6x in povray, 1x in gzip, 279x in gcc, 1x in crafty, 8x in eon, 11x in perlbmk,
362x in gap, 4x in vortex, 14 in m88ksim, 211x in 126.gcc, 1x in compress,
11x in ijpeg, and 4x in 147.vortex.
llvm-svn: 16521
whose addresses where used by trivial phi nodes and select instructions. This
is now performed by the instcombine pass, which is more powerful, is much
simpler, and is faster. This allows the deletion of a bunch of code, two
FIXME's and two gotos.
llvm-svn: 16406
a function being deleted. Due to optimizations done while inlining, there
can be edges from the external call node to a function node that were not
apparent any longer.
This fixes the compiler crash while compiling 175.vpr
llvm-svn: 16399
Move include/Config and include/Support into include/llvm/Config,
include/llvm/ADT and include/llvm/Support. From here on out, all LLVM
public header files must be under include/llvm/.
llvm-svn: 16137
block (common in a switch), make sure to remove extra edges in successor
blocks. This fixes CodeExtractor/2004-08-12-BlockExtractPHI.ll and should
be pulled into LLVM 1.3 (though the regression test need not be, as that
would require pulling in the LoopExtract.cpp changes).
llvm-svn: 15717
instructions in the body of the function (not the entry block). This fixes
test/Programs/SingleSource/Regression/C/2004-08-12-InlinerAndAllocas.c
and test/Programs/External/SPEC/CINT2000/176.gcc on zion.
This should obviously be pulled into 1.3.
llvm-svn: 15684
dangling constant users were removed from a function, causing it to be dead,
we never removed the call graph edge from the external node to the function.
In most cases, this didn't cause a problem (by luck). This should definitely
go into 1.3
llvm-svn: 15570
1. Fix a REALLY nasty cyclic replacement issue that Anshu discovered, causing
nondeterminstic crashes and memory corruption.
2. For performance, don't go inserting constantexpr casts of GV pointers.
This should definitely go into 1.3
llvm-svn: 15568
assumed that a constant on the RHS of a multiplication was either an
IntConstant or an FPConstant. It checked for an IntConstant and then,
if it did not find one, did a hard cast to an FPConstant. That code
would crash if the RHS were a ConstantExpr that was neither an
IntConstant nor an FPConstant. This version replaces the hard cast
with a dyn_cast. It performs the same way for IntConstants and
FPConstants but does nothing, instead of crashing, for constant
expressions.
The regression test for this change is 2004-07-27-ConstantExprMul.ll.
llvm-svn: 15291
a bug in DSE).
* Delete dead operand uses iteratively instead of recursively, using a
SetVector.
* Defer deletion of dead operand uses until the end of processing, which means
we don't have to bother with updating the AliasSetTracker. This speeds up
DSE substantially.
llvm-svn: 15204
* Test for whether bits are shifted out during the optzn.
If so, the fold is illegal, though it can be handled explicitly for setne/seteq
This fixes the miscompilation of 254.gap last night, which was a latent bug
exposed by other optimizer improvements.
llvm-svn: 15085
actually care about. Someday when the cast instruction is gone, we can do
better here, but this will do for now. This implements
instcombine/cast.ll:test17/18 as well.
llvm-svn: 15018
This eliminates an N*N*logN algorithm from the loop simplify pass, replacing
it with a much simpler and faster alternative. In a debug build, this reduces
gccas time on eon from 85s to 42s.
llvm-svn: 14851
"load (cast foo)". This allows us to compile C++ code like this:
class Bclass {
public: virtual int operator()() { return 666; }
};
class Dclass: public Bclass {
public: virtual int operator()() { return 667; }
} ;
int main(int argc, char** argv) {
Dclass x;
return x();
}
Into this:
int %main(int %argc, sbyte** %argv) {
entry:
call void %__main( )
ret int 667
}
Instead of this:
int %main(int %argc, sbyte** %argv) {
entry:
%x = alloca "struct.std::bad_typeid" ; <"struct.std::bad_typeid"*> [#uses=3]
call void %__main( )
%tmp.1.i.i = getelementptr "struct.std::bad_typeid"* %x, uint 0, uint 0, uint 0 ; <int (...)***> [#uses=1]
store int (...)** getelementptr ([3 x int (...)*]* %vtable for Bclass, int 0, long 2), int (...)*** %tmp.1.i.i
%tmp.3.i = getelementptr "struct.std::bad_typeid"* %x, int 0, uint 0, uint 0 ; <int (...)***> [#uses=1]
store int (...)** getelementptr ([3 x int (...)*]* %vtable for Dclass, int 0, long 2), int (...)*** %tmp.3.i
%tmp.5 = load int ("struct.std::bad_typeid"*)** cast (int (...)** getelementptr ([3 x int (...)*]* %vtable for Dclass, int 0, long 2) to int
("struct.std::bad_typeid"*)**) ; <int ("struct.std::bad_typeid"*)*> [#uses=1]
%tmp.6 = call int %tmp.5( "struct.std::bad_typeid"* %x ) ; <int> [#uses=1]
ret int %tmp.6
ret int 0
}
In order words, we now resolve the virtual function call.
llvm-svn: 14783
Don't touch GEPs for which DecomposeArrayRef is not going to do anything
special (e.g., < 2 indices, or 2 indices and the last one is a constant.)
llvm-svn: 14647
since May 1st. In this code, the pred iterator was being invalidated sometimes
causing the wrong entries to be added to PHI nodes.
The fix for this is to defererence and safe the *PI value before we hack on
branch instructions, which changes use/def chains, which SOMETIMES invalidates
the iterator.
llvm-svn: 14278
non-deterministic things like the ordering of blocks in the dominance
frontier of a BB. Unfortunately, I don't know of a better way to solve
this problem than to explicitly sort the BB's in function-order before
processing them. This is guaranteed to slow the pass down a bit, but
is absolutely necessary to get usable diffs between two different tools
executing the mem2reg or scalarrepl pass.
Before this, bazillions of spurious diff failures occurred all over the
place due to the different order of processing PHIs:
- %tmp.111 = getelementptr %struct.Connector_struct* %upcon.0.0, uint 0, uint 0
+ %tmp.111 = getelementptr %struct.Connector_struct* %upcon.0.1, uint 0, uint 0
Now, the diffs match.
llvm-svn: 14244
nondeterministic results that depend on where these objects land in memory.
Instead, sort by the value of the constant, which is stable.
Before this patch, the -simplifycfg pass run from two different compilers
could cause different code to be generated, though it was semantically the
same:
@@ -12258,8 +12258,8 @@
%s_addr.1 = phi sbyte* [ %s, %entry ], [ %inc.0, %no_exit ] ; <sbyte*> [#uses=5]
%tmp.1 = load sbyte* %s_addr.1 ; <sbyte> [#uses=1]
switch sbyte %tmp.1, label %no_exit [
- sbyte 0, label %loopexit
sbyte 46, label %loopexit
+ sbyte 0, label %loopexit
]
We need to stomp all of this stuff out.
llvm-svn: 14243
is write an autoconf macro that checks whether __isnan or isnan actually works
**using the C++ compiler after #include <cmath>**, instead of doing it the easy
way with AC_CHECK_FUNCS().
llvm-svn: 14171
186.crafty, fhourstones and 132.ijpeg.
Bugpoint makes really nasty miscompilations embarassingly easy to find. It
narrowed it down to the instcombiner and this testcase (from fhourstones):
bool %l7153_l4706_htstat_loopentry_2E_4_no_exit_2E_4(int* %i, [32 x int]* %works, int* %tmp.98.out) {
newFuncRoot:
%tmp.96 = load int* %i ; <int> [#uses=1]
%tmp.97 = getelementptr [32 x int]* %works, long 0, int %tmp.96 ; <int*> [#uses=1]
%tmp.98 = load int* %tmp.97 ; <int> [#uses=2]
%tmp.99 = load int* %i ; <int> [#uses=1]
%tmp.100 = and int %tmp.99, 7 ; <int> [#uses=1]
%tmp.101 = seteq int %tmp.100, 7 ; <bool> [#uses=2]
%tmp.102 = cast bool %tmp.101 to int ; <int> [#uses=0]
br bool %tmp.101, label %codeRepl4.exitStub, label %codeRepl3.exitStub
codeRepl4.exitStub: ; preds = %newFuncRoot
store int %tmp.98, int* %tmp.98.out
ret bool true
codeRepl3.exitStub: ; preds = %newFuncRoot
store int %tmp.98, int* %tmp.98.out
ret bool false
}
... which only has one combination performed on it:
$ llvm-as < t.ll | opt -instcombine -debug | llvm-dis
IC: Old = %tmp.101 = seteq int %tmp.100, 7 ; <bool> [#uses=1]
New = setne int %tmp.100, 0 ; <bool>:<badref> [#uses=0]
IC: MOD = br bool %tmp.101, label %codeRepl3.exitStub, label %codeRepl4.exitStub
IC: MOD = %tmp.97 = getelementptr [32 x int]* %works, uint 0, int %tmp.96 ; <int*> [#uses=1]
It doesn't get much better than this. :)
llvm-svn: 14109
collapse this:
bool %le(int %A, int %B) {
%c1 = setgt int %A, %B
%tmp = select bool %c1, int 1, int 0
%c2 = setlt int %A, %B
%result = select bool %c2, int -1, int %tmp
%c3 = setle int %result, 0
ret bool %c3
}
into:
bool %le(int %A, int %B) {
%c3 = setle int %A, %B ; <bool> [#uses=1]
ret bool %c3
}
which is handy, because the Java FE makes these sequences all over the place.
This is tested as: test/Regression/Transforms/InstCombine/JavaCompare.ll
llvm-svn: 14086
This code hadn't been updated after the "structs with more than 256 elements"
related changes to the GEP instruction. Also it was not handling the
ConstantAggregateZero class.
Now it does!
llvm-svn: 13834
Add support for acos/asin/atan. 188.ammp contains three calls to acos with
constant arguments. Constant folding it allows elimination of those 3 calls
and three FP divisions of the results.
llvm-svn: 13821
into (X & (C2 << C1)) != (C3 << C1), where the shift may be either left or
right and the compare may be any one.
This triggers 1546 times in 176.gcc alone, as it is a common pattern that
occurs for bitfield accesses.
llvm-svn: 13740
CloneTrace, and because it is primarily an operation on ValueMaps. It
is now a global (non-static) function which can be pulled in using
ValueMapper.h.
llvm-svn: 13600
Add better comments, including a better head-of-file comment.
Prune #includes.
Fix a FIXME that Chris put here by using doInitialization().
Use DEBUG() to print out debug msgs.
Give names to basic blocks inserted by this pass.
Expand tabs.
Use InsertProfilingInitCall() from ProfilingUtils to insert the initialize call.
llvm-svn: 13581
in the size calculation.
This is not something you want to see:
Loop Unroll: F[main] Loop %no_exit Loop Size = 2 Trip Count = 2147483648 - UNROLLING!
The problem was that 2*2147483648 == 0.
Now we get:
Loop Unroll: F[main] Loop %no_exit Loop Size = 2 Trip Count = 2147483648 - TOO LARGE: 4294967296>100
Thanks to some anonymous person playing with the demo page that repeatedly
caused zion to go into swapping land. That's one way to ensure you'll get
a quick bugfix. :)
Testcase here: Transforms/LoopUnroll/2004-05-13-DontUnrollTooMuch.ll
llvm-svn: 13564
PHI node entries from multiple outside-the-region blocks. This also fixes
extraction of the entry block in a function. Yaay.
This has successfully block extracted all (but one) block from the score_move
function in obsequi (out of 33). Hrm, I wonder which block the bug is in. :)
llvm-svn: 13489
* Add a stub for the severSplitPHINodes which will allow us to bbextract
bb's with PHI nodes in them soon.
* Remove unused arguments from findInputsOutputs
* Dramatically simplify the code in findInputsOutputs. In particular,
nothing really cares whether or not a PHI node is using something.
* Move moveCodeToFunction to after emitCallAndSwitchStatement as that's the
order they get called.
* Fix a bug where we would code extract a region that included a call to
vastart. Like 'alloca', calls to vastart must stay in the function that
they are defined in.
* Add some comments.
llvm-svn: 13482
from the extracted region. If the return has 0 or 1 exit blocks, the new
function returns void. If it has 2 exits, it returns bool, otherwise it
returns a ushort as before.
This allows us to use a conditional branch instruction when there are two
exit blocks, as often happens during block extraction.
llvm-svn: 13481
1. Get rid of the silly abort block. When doing bb extraction, we get one
abort block for every block extracted, which is kinda annoying.
2. If the switch ends up having a single destination, turn it into an
unconditional branch.
I would like to add support for conditional branches, but to do this we will
want to have the function return a bool instead of a ushort.
llvm-svn: 13478
%tmp.0 = getelementptr [50 x sbyte]* %ar, uint 0, int 5 ; <sbyte*> [#uses=2]
%tmp.7 = getelementptr sbyte* %tmp.0, int 8 ; <sbyte*> [#uses=1]
together. This patch actually allows us to simplify and generalize the code.
llvm-svn: 13415
is only used by a cast, and the casted type is the same size as the original
allocation, it would eliminate the cast by folding it into the allocation.
Unfortunately, it was placing the new allocation instruction right before
the cast, which could pull (for example) alloca instructions into the body
of a function. This turns statically allocatable allocas into expensive
dynamically allocated allocas, which is bad bad bad.
This fixes the problem by placing the new allocation instruction at the same
place the old one was, duh. :)
llvm-svn: 13289
loop. This eliminates the extra add from the previous case, but it's
not clear that this will be a performance win overall. Tommorows test
results will tell. :)
llvm-svn: 13103
Eventually it would be nice if CallGraph maintained an ilist of CallGraphNode's instead
of a vector of pointers to them, but today is not that day.
llvm-svn: 13100
block. The primary motivation for doing this is that we can now unroll nested loops.
This makes a pretty big difference in some cases. For example, in 183.equake,
we are now beating the native compiler with the CBE, and we are a lot closer
with LLC.
I'm now going to play around a bit with the unroll factor and see what effect
it really has.
llvm-svn: 13034
limited. Even in it's extremely simple state (it can only *fully* unroll single
basic block loops that execute a constant number of times), it already helps improve
performance a LOT on some benchmarks, particularly with the native code generators.
llvm-svn: 13028
Basically we were using SimplifyCFG as a huge sledgehammer for a simple
optimization. Because simplifycfg does so many things, we can't use it
for this purpose.
llvm-svn: 12977
Instead of producing code like this:
Loop:
X = phi 0, X2
...
X2 = X + 1
if (X != N-1) goto Loop
We now generate code that looks like this:
Loop:
X = phi 0, X2
...
X2 = X + 1
if (X2 != N) goto Loop
This has two big advantages:
1. The trip count of the loop is now explicit in the code, allowing
the direct implementation of Loop::getTripCount()
2. This reduces register pressure in the loop, and allows X and X2 to be
put into the same register.
As a consequence of the second point, the code we generate for loops went
from:
.LBB2: # no_exit.1
...
mov %EDI, %ESI
inc %EDI
cmp %ESI, 2
mov %ESI, %EDI
jne .LBB2 # PC rel: no_exit.1
To:
.LBB2: # no_exit.1
...
inc %ESI
cmp %ESI, 3
jne .LBB2 # PC rel: no_exit.1
... which has two fewer moves, and uses one less register.
llvm-svn: 12961
This transforms code like this:
%C = or %A, %B
%D = select %cond, %C, %A
into:
%C = select %cond, %B, 0
%D = or %A, %C
Since B is often a constant, the select can often be eliminated. In any case,
this reduces the usage count of A, allowing subsequent optimizations to happen.
This xform applies when the operator is any of:
add, sub, mul, or, xor, and, shl, shr
llvm-svn: 12800
Now we collect all of the call sites we are interested in inlining, then inline
them. This entirely avoids issues with trying to inline a call site we got by
inlining another call site. This also eliminates iterator invalidation issues.
llvm-svn: 12770
This also implements some new features for the indvars pass, including
linear function test replacement, exit value substitution, and it works with
a much more general class of induction variables and loops.
llvm-svn: 12620
#1 is to unconditionally strip constantpointerrefs out of
instruction operands where they are absolutely pointless and inhibit
optimization. GRRR!
#2 is to implement InstCombine/getelementptr_const.ll
llvm-svn: 12519
1. Names were not put on the new arguments created (ok, this just helps sanity :)
2. Fix outgoing pointer values
3. Do not insert stores for values that had not been computed
4. Fix some wierd problems with the outset calculation
This fixes CodeExtractor/2004-03-14-DominanceProblem.ll, making the extractor
work on at least one simple case!
llvm-svn: 12484
as it is making effectively arbitrary modifications to the CFG and we don't
have a domset/domfrontier implementations that can handle the dynamic updates.
Instead of having a bunch of code that doesn't actually work in practice,
just demote any potentially tricky values to the stack (causing the problem
to go away entirely). Later invocations of mem2reg will rebuild SSA for us.
This fixes all of the major performance regressions with tail duplication
from LLVM 1.1. For example, this loop:
---
int popcount(int x) {
int result = 0;
while (x != 0) {
result = result + (x & 0x1);
x = x >> 1;
}
return result;
}
---
Used to be compiled into:
int %popcount(int %X) {
entry:
br label %loopentry
loopentry: ; preds = %entry, %no_exit
%x.0 = phi int [ %X, %entry ], [ %tmp.9, %no_exit ] ; <int> [#uses=3]
%result.1.0 = phi int [ 0, %entry ], [ %tmp.6, %no_exit ] ; <int> [#uses=2]
%tmp.1 = seteq int %x.0, 0 ; <bool> [#uses=1]
br bool %tmp.1, label %loopexit, label %no_exit
no_exit: ; preds = %loopentry
%tmp.4 = and int %x.0, 1 ; <int> [#uses=1]
%tmp.6 = add int %tmp.4, %result.1.0 ; <int> [#uses=1]
%tmp.9 = shr int %x.0, ubyte 1 ; <int> [#uses=1]
br label %loopentry
loopexit: ; preds = %loopentry
ret int %result.1.0
}
And is now compiled into:
int %popcount(int %X) {
entry:
br label %no_exit
no_exit: ; preds = %entry, %no_exit
%x.0.0 = phi int [ %X, %entry ], [ %tmp.9, %no_exit ] ; <int> [#uses=2]
%result.1.0.0 = phi int [ 0, %entry ], [ %tmp.6, %no_exit ] ; <int> [#uses=1]
%tmp.4 = and int %x.0.0, 1 ; <int> [#uses=1]
%tmp.6 = add int %tmp.4, %result.1.0.0 ; <int> [#uses=2]
%tmp.9 = shr int %x.0.0, ubyte 1 ; <int> [#uses=2]
%tmp.1 = seteq int %tmp.9, 0 ; <bool> [#uses=1]
br bool %tmp.1, label %loopexit, label %no_exit
loopexit: ; preds = %no_exit
ret int %tmp.6
}
llvm-svn: 12457
time from 615s to 1.49s on a large testcase that has a gigantic switch statement
that all of the blocks in the function go to (an intepreter).
llvm-svn: 12442
http://mail.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20040308/013095.html
Basically, this patch only updated the immediate dominatees of the header node
to tell them that the preheader also dominated them. In practice, ALL
dominatees of the header node are also dominated by the preheader.
This fixes: LoopSimplify/2004-03-15-IncorrectDomUpdate.
and PR293
llvm-svn: 12434
Simplify the input/output finder. All elements of a basic block are
instructions. Any used arguments are also inputs. An instruction can only
be used by another instruction.
llvm-svn: 12405
extracted, and a function that contained a single top-level loop never had
the loop extracted, regardless of how much non-loop code there was.
llvm-svn: 12403
* Don't insert a branch to the switch instruction after the call, just
make it a single block.
* Insert the new alloca instructions in the entry block of the original
function instead of having them execute dynamically
* Don't make the default edge of the switch instruction go back to the switch.
The loop extractor shouldn't create new loops!
* Give meaningful names to the alloca slots and the reload instructions
* Some minor code simplifications
llvm-svn: 12402
This also implements a two minor improvements:
* Don't insert live-out stores IN the region, insert them on the code path
that exits the region
* If the region is exited to the same block from multiple paths, share the
switch statement entry, live-out store code, and the basic block.
llvm-svn: 12401
a member of the class. While we're at it, turn the collection into a set
instead of a vector to improve efficiency and make queries simpler.
llvm-svn: 12400
loop information won't see it, and we could have unreachable blocks pointing to
the non-header node of blocks in a natural loop. This isn't tidy, so have the
loopsimplify pass clean it up.
llvm-svn: 12380
* Be a lot more accurate about what the effects will be when inlining a call
to a function when an argument is an alloca.
* Dramatically reduce the penalty for inlining a call in a large function.
This heuristic made it almost impossible to inline a function into a large
function, no matter how small the callee is.
llvm-svn: 12363
On the testcase from GCC PR12440, which has a LOT of loops (1392 of which require
preheaders to be inserted), this speeds up the loopsimplify pass from 1.931s to
0.1875s. The loop in question goes from 1.65s -> 0.0097s, which isn't bad. All of
these times are a debug build.
This adds a dependency on DominatorTree analysis that was not there before, but
we always had dominatortree available anyway, because LICM requires both loop
simplify and DT, so this doesn't add any extra analysis in practice.
llvm-svn: 12362
This allows pointers to aggregate objects, whose elements are only read, to
be promoted and passed in by element instead of by reference. This can
enable a LOT of subsequent optimizations in the caller function.
It's worth pointing out that this stuff happens a LOT of C++ programs, because
objects in templates are generally passed around by reference. When these
templates are instantiated on small aggregate or scalar types, however, it is
more efficient to pass them in by value than by reference.
This transformation triggers most on C++ codes (e.g. 334 times on eon), but
does happen on C codes as well. For example, on mesa it triggers 72 times,
and on gcc it triggers 35 times. this is amazingly good considering that
we are using 'basicaa' so far.
llvm-svn: 12202
Note that this is a band-aid put over a band-aid. This just undisables
tail duplication in on very specific case that it seems to work in.
llvm-svn: 11989
assume that if they don't intend to write to a global variable, that they
would mark it as constant. However, there are people that don't understand
that the compiler can do nice things for them if they give it the information
it needs.
This pass looks for blatently obvious globals that are only ever read from.
Though it uses a trivially simple "alias analysis" of sorts, it is still able
to do amazing things to important benchmarks. 253.perlbmk, for example,
contains several ***GIANT*** function pointer tables that are not marked
constant and should be. Marking them constant allows the optimizer to turn
a whole bunch of indirect calls into direct calls. Note that only a link-time
optimizer can do this transformation, but perlbmk does have several strings
and other minor globals that can be marked constant by this pass when run
from GCCAS.
176.gcc has a ton of strings and large tables that are marked constant, both
at compile time (38 of them) and at link time (48 more). Other benchmarks
give similar results, though it seems like big ones have disproportionally
more than small ones.
This pass is extremely quick and does good things. I'm going to enable it
in gccas & gccld. Not bad for 50 SLOC.
llvm-svn: 11836
This case occurs many times in various benchmarks, especially when combined
with the previous patch. This allows it to get stuff like:
if (X == 4 || X == 3)
if (X == 5 || X == 8)
and
switch (X) {
case 4: case 5: case 6:
if (X == 4 || X == 5)
llvm-svn: 11797
allowed in invoke instructions. Thus, if we are inlining a call to an intrinsic
function into an invoke site, we don't need to turn the call into an invoke!
llvm-svn: 11384
more of a testcase for profiling information than anything that should reasonably
be used, but it's a starting point. When I have more time I will whip this into
better shape.
llvm-svn: 11311
Having a proper 'select' instruction would allow the elimination of a lot
of the special case cruft in this patch, but we don't have one yet.
llvm-svn: 11307
The problem is that the dominator update code didn't "realize" that it's
possible for the newly inserted basic block to dominate anything. Because
it IS possible, stuff was getting updated wrong.
llvm-svn: 11137
1. Don't scan to the end of alloca instructions in the caller function to
insert inlined allocas, just insert at the top. This saves a lot of
time inlining into functions with a lot of allocas.
2. Use splice to move the alloca instructions over, instead of remove/insert.
This allows us to transfer a block at a time, and eliminates a bunch of
silly symbol table manipulations.
This speeds up the inliner on the testcase in PR209 from 1.73s -> 1.04s (67%)
llvm-svn: 11118
and that basic block ends with a return instruction. In this case, we can just splice
the cloned "body" of the function directly into the source basic block, avoiding a lot
of rearrangement and splitBasicBlock's linear scan over the split block. This speeds up
the inliner on the testcase in PR209 from 2.3s to 1.7s, a 35% reduction.
llvm-svn: 11116
process. The only optimization we did so far is to avoid creating a
PHI node, then immediately destroying it in the common case where the
callee has one return statement. Instead, we just don't create the return
value. This has no noticable performance impact, but paves the way for
future improvements.
llvm-svn: 11108
to add the cloned block to. This allows the block to be added to the function
immediately, and all of the instructions to be immediately added to the function
symbol table, which speeds up the inliner from 3.7 -> 3.38s on the PR209.
llvm-svn: 11107
constants as being "true" when evaluating branches. This was introduced
because we now create constantexprs for the constants instead of failing the
fold.
llvm-svn: 10778
* Implement SCCP of load instructions, implementing Transforms/SCCP/loadtest.ll
This allows us to fold expressions like "foo"[2], even if the pointer is only
a conditional constant.
llvm-svn: 10767
Fix iterator invalidation problems which was causing -mstrip to miss some
entries, and read free'd memory. This shrinks the symbol table of 254.gap
from 333 to 284 bytes! :)
llvm-svn: 10751
that it needs. This is pretty scary code! This fixes
Regression.Transforms.LevelRaise.2002-07-16-SourceAndDestCrash
Regression.Transforms.LevelRaise.2002-07-31-AssertionFailure
llvm-svn: 10406
We now insert accumulator variables as necessary to eliminate tail recursion
more aggressively. This is still fairly limited, but allows us to transform
fib/factorial, and other functions into nice happy loops. :)
llvm-svn: 10332
The only functionality change is that we now implement:
Regression/Transforms/TailCallElim/intervening-inst.ll
Which is really kinda pointless, because it means that trivially dead code
does not interfere with -tce, but trivially dead code probably wouldn't be
around anytime when this pass is run anyway.
The point of including this change it to support other more aggressive
transformations when we have the analysis capabilities to do so.
llvm-svn: 10312
* Make the cost metric for passing constants in as arguments to functions MUCH
more accurate, by actually estimating the amount of code that will be constant
propagated away.
llvm-svn: 10136
* Implement FuncResolve/2003-11-20-BogusResolveWarning.ll
... which eliminates a large number of annoying warnings. I know misha
will miss them though!
llvm-svn: 10123
pool allocator no end of trouble, and doesn't make a lot of sense anyway. This
does not solve the problem with mutually recursive functions, but they are much less common.
llvm-svn: 9828
Added assert() to ensure symbol table is well formed.
Added code to remember the value that was found; resolving types can change
the symbol table and invalidate the value of the iterator.
Added comments to the ResolveTypes() function (mainly for my own benefit).
Please feel free to correct the comments if they are not accurate.
llvm-svn: 9693
Chris Lattner