8df83c4a24
fix rdar://8785296 - -fcatch-undefined-behavior generates inefficient code
...
The basic issue is that isel (very reasonably!) expects conditional branches
to be folded, so CGP leaving around a bunch dead computation feeding
conditional branches isn't such a good idea. Just fold branches on constants
into unconditional branches.
llvm-svn: 123526
2011-01-15 07:36:13 +00:00
ee588defc6
simplify code, no functionality change.
...
llvm-svn: 123525
2011-01-15 07:29:01 +00:00
1b93be501d
Now that instruction optzns can update the iterator as they go, we can
...
have objectsize folding recursively simplify away their result when it
folds. It is important to catch this here, because otherwise we won't
eliminate the cross-block values at isel and other times.
llvm-svn: 123524
2011-01-15 07:25:29 +00:00
7a2771440f
make the current instruction iterator an ivar, allowing xforms that
...
potentially invalidate it (like inline asm lowering) to be sunk into
their proper place, cleaning up a ton of code.
llvm-svn: 123523
2011-01-15 07:14:54 +00:00
9c10d587f6
implement an instcombine xform that canonicalizes casts outside of and-with-constant operations.
...
This fixes rdar://8808586 which observed that we used to compile:
union xy {
struct x { _Bool b[15]; } x;
__attribute__((packed))
struct y {
__attribute__((packed)) unsigned long b0to7;
__attribute__((packed)) unsigned int b8to11;
__attribute__((packed)) unsigned short b12to13;
__attribute__((packed)) unsigned char b14;
} y;
};
struct x
foo(union xy *xy)
{
return xy->x;
}
into:
_foo: ## @foo
movq (%rdi), %rax
movabsq $1095216660480, %rcx ## imm = 0xFF00000000
andq %rax, %rcx
movabsq $-72057594037927936, %rdx ## imm = 0xFF00000000000000
andq %rax, %rdx
movzbl %al, %esi
orq %rdx, %rsi
movq %rax, %rdx
andq $65280, %rdx ## imm = 0xFF00
orq %rsi, %rdx
movq %rax, %rsi
andq $16711680, %rsi ## imm = 0xFF0000
orq %rdx, %rsi
movl %eax, %edx
andl $-16777216, %edx ## imm = 0xFFFFFFFFFF000000
orq %rsi, %rdx
orq %rcx, %rdx
movabsq $280375465082880, %rcx ## imm = 0xFF0000000000
movq %rax, %rsi
andq %rcx, %rsi
orq %rdx, %rsi
movabsq $71776119061217280, %r8 ## imm = 0xFF000000000000
andq %r8, %rax
orq %rsi, %rax
movzwl 12(%rdi), %edx
movzbl 14(%rdi), %esi
shlq $16, %rsi
orl %edx, %esi
movq %rsi, %r9
shlq $32, %r9
movl 8(%rdi), %edx
orq %r9, %rdx
andq %rdx, %rcx
movzbl %sil, %esi
shlq $32, %rsi
orq %rcx, %rsi
movl %edx, %ecx
andl $-16777216, %ecx ## imm = 0xFFFFFFFFFF000000
orq %rsi, %rcx
movq %rdx, %rsi
andq $16711680, %rsi ## imm = 0xFF0000
orq %rcx, %rsi
movq %rdx, %rcx
andq $65280, %rcx ## imm = 0xFF00
orq %rsi, %rcx
movzbl %dl, %esi
orq %rcx, %rsi
andq %r8, %rdx
orq %rsi, %rdx
ret
We now compile this into:
_foo: ## @foo
## BB#0: ## %entry
movzwl 12(%rdi), %eax
movzbl 14(%rdi), %ecx
shlq $16, %rcx
orl %eax, %ecx
shlq $32, %rcx
movl 8(%rdi), %edx
orq %rcx, %rdx
movq (%rdi), %rax
ret
A small improvement :-)
llvm-svn: 123520
2011-01-15 06:32:33 +00:00
e20dd530d0
one more instcombine variant that is needed to work with future changes,
...
no functionality change currently.
llvm-svn: 123517
2011-01-15 05:50:18 +00:00
497459d5fd
fix typo
...
llvm-svn: 123516
2011-01-15 05:42:47 +00:00
f3c4eefff8
Catch ~x < cst just like ~x < ~y, we currently handle this through
...
means that are about to disappear.
llvm-svn: 123515
2011-01-15 05:41:33 +00:00
311aa63c87
reduce indentation
...
llvm-svn: 123514
2011-01-15 05:40:29 +00:00
b68ec5c339
Generalize LoadAndStorePromoter a bit and switch LICM
...
to use it.
llvm-svn: 123501
2011-01-15 00:12:35 +00:00
3e2f6cf7ae
Fix a false-positive warning.
...
llvm-svn: 123480
2011-01-14 22:31:13 +00:00
9eb7cb48e4
Enhance GlobalOpt to be able evaluate initializers that involve stores through
...
bitcasts, at least in simple cases. This fixes clang's CodeGenCXX/virtual-base-dtor.cpp
llvm-svn: 123477
2011-01-14 22:19:20 +00:00
b498f9aff3
switch SRoA to use LoadAndStorePromoter instead of its own copy of the code.
...
llvm-svn: 123457
2011-01-14 19:50:47 +00:00
95294b8796
Add a new LoadAndStorePromoter class, which implements the general
...
"promote a bunch of load and stores" logic, allowing the code to
be shared and reused.
llvm-svn: 123456
2011-01-14 19:36:13 +00:00
9987a6f49b
split SROA into two passes: one that uses DomFrontiers (-scalarrepl)
...
and one that uses SSAUpdater (-scalarrepl-ssa)
llvm-svn: 123436
2011-01-14 08:13:00 +00:00
543384efb4
Implement full support for promoting allocas to registers using SSAUpdater
...
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
2011-01-14 07:50:47 +00:00
90f3a9a1c7
indentation
...
llvm-svn: 123426
2011-01-14 04:23:53 +00:00
7f60dc1eb0
Move some shift transforms out of instcombine and into InstructionSimplify.
...
While there, I noticed that the transform "undef >>a X -> undef" was wrong.
For example if X is 2 then the top two bits must be equal, so the result can
not be anything. I fixed this in the constant folder as well. Also, I made
the transform for "X << undef" stronger: it now folds to undef always, even
though X might be zero. This is in accordance with the LangRef, but I must
admit that it is fairly aggressive. Also, I added "i32 X << 32 -> undef"
following the LangRef and the constant folder, likewise fairly aggressive.
llvm-svn: 123417
2011-01-14 00:37:45 +00:00
328e91bbe1
Fix whitespace.
...
llvm-svn: 123396
2011-01-13 20:59:44 +00:00
c8056a952e
Check for empty structs, and for consistency, zero-element arrays.
...
llvm-svn: 123383
2011-01-13 18:26:59 +00:00
08713d3c5f
Extend SROA to handle arrays accessed as homogeneous structs and vice versa.
...
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
2011-01-13 17:45:11 +00:00
12eec40c83
Make SROA more aggressive with allocas containing padding.
...
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
2011-01-13 17:45:08 +00:00
30f3ebbc1f
Use SmallVector instead of SmallPtrSet and avoid non-deterministic behavior.
...
llvm-svn: 123318
2011-01-12 19:12:45 +00:00
dd5f60b7a7
revert 123144, reenabling the rest of memset formation.
...
llvm-svn: 123302
2011-01-12 03:25:15 +00:00
654098f411
revert r123146 which disabled code that wasn't the root cause
...
of the bootstrap miscompare issue.
llvm-svn: 123299
2011-01-12 01:52:23 +00:00
fa7c29d255
revert r123149, reenabling an improvement to memcpyopt that wasn't
...
the source of the bootstrap problem.
llvm-svn: 123298
2011-01-12 01:43:46 +00:00
12cc296bd4
Remove the PR8954 workaround.
...
llvm-svn: 123288
2011-01-11 22:56:41 +00:00
f2407aa98b
Fix a non-deterministic loop in llvm::MergeBlockIntoPredecessor.
...
DT->changeImmediateDominator() trivially ignores identity updates, so there is
really no need for the uniqueing provided by SmallPtrSet.
I expect this to fix PR8954.
llvm-svn: 123286
2011-01-11 22:54:38 +00:00
cb9c4f85ec
Dial back the speculative fix for PR8954 a bit, so that we only recompute dominators
...
once at the beginning of GVN instead of once per iteration.
llvm-svn: 123278
2011-01-11 22:14:42 +00:00
51eb403907
Attempt to fix the bootstrap buildbot. Rafael says this works for him on x86-64 Linux.
...
llvm-svn: 123270
2011-01-11 20:23:34 +00:00
0022a4b417
Remove dead variable, const-ref-ize an APInt.
...
llvm-svn: 123248
2011-01-11 18:26:37 +00:00
d41db8f9cb
this pass claims to preserve scev, make sure to tell it about deletions.
...
llvm-svn: 123247
2011-01-11 18:14:50 +00:00
8e158495f1
Factor the actual simplification out of SimplifyIndirectBrOnSelect and into a new helper function so it can be reused in e.g. an upcoming SimplifySwitchOnSelect.
...
No functional change.
llvm-svn: 123234
2011-01-11 12:52:11 +00:00
193ce7c4d1
update memdep when an instruction is deleted. This code isn't
...
actually reached in the testcase in PR8954, but it's safe and good
practice.
llvm-svn: 123224
2011-01-11 08:19:16 +00:00
e2523b287c
when MergeBlockIntoPredecessor merges two blocks, update MemDep if it
...
is floating around in the ether.
llvm-svn: 123223
2011-01-11 08:16:49 +00:00
f6ae904e34
Fix FoldSingleEntryPHINodes to update memdep and AA when it deletes
...
phi nodes. It is called from MergeBlockIntoPredecessor which is
called from GVN, which claims to preserve these.
I'm skeptical that this is the actual problem behind PR8954, but
this is a stab in the right direction.
llvm-svn: 123222
2011-01-11 08:13:40 +00:00
dfcfcb49fa
random cleanups
...
llvm-svn: 123221
2011-01-11 08:00:40 +00:00
63fe78de68
remove a bogus assertion: the latch block of a loop is not
...
neccesarily an uncond branch to the header. This fixes
PR8955 (the assertion tripping).
llvm-svn: 123219
2011-01-11 07:47:59 +00:00
d490c2d2ae
Fix a random missed optimization by making InstCombine more aggressive when determining which bits are demanded by
...
a comparison against a constant.
llvm-svn: 123203
2011-01-11 00:36:45 +00:00
cf414cf0a6
Teach instcombine about the rest of the SSE and SSE2 conversion
...
intrinsics element dependencies. Reviewed by Nick.
llvm-svn: 123161
2011-01-10 07:19:37 +00:00
88bc848ab6
another random stab in the dark trying to fix llvm-gcc-i386-linux-selfhost
...
llvm-svn: 123149
2011-01-10 02:34:11 +00:00
4662bd4b13
another (more) aggressive attempt to bring llvm-gcc-i386-linux-selfhost
...
back to life.
llvm-svn: 123146
2011-01-10 00:47:34 +00:00
1017fa6746
temporarily disable memset formation from memsets in an effort to restore buildbot stability.
...
llvm-svn: 123144
2011-01-09 23:52:48 +00:00
caf5c0d037
fix a few old bugs (found by inspection) where we would zap instructions
...
without informing memdep. This could cause nondeterminstic weirdness
based on where instructions happen to get allocated, and will hopefully
breath some life into some broken testers.
llvm-svn: 123124
2011-01-09 19:26:10 +00:00
cc21c4aa98
Instcombine: Fix pattern where the sext did not dominate the icmp using it
...
llvm-svn: 123121
2011-01-09 16:00:11 +00:00
a42e5915bf
LoopInstSimplify preserves LoopSimplify.
...
llvm-svn: 123117
2011-01-09 12:35:16 +00:00
a337f5ec5c
reduce indentation. Print <nuw> and <nsw> when dumping SCEV AddRec's
...
that have the bit set.
llvm-svn: 123104
2011-01-09 02:16:18 +00:00
7d6433ae76
fix a latent bug in memcpyoptimizer that my recent patches exposed: it wasn't
...
updating memdep when fusing stores together. This fixes the crash optimizing
the bullet benchmark.
llvm-svn: 123091
2011-01-08 22:19:21 +00:00
ff6ed2ac5f
tryMergingIntoMemset can only handle constant length memsets.
...
llvm-svn: 123090
2011-01-08 22:11:56 +00:00
9a1d63ba9f
Merge memsets followed by neighboring memsets and other stores into
...
larger memsets. Among other things, this fixes rdar://8760394 and
allows us to handle "Example 2" from http://blog.regehr.org/archives/320 ,
compiling it into a single 4096-byte memset:
_mad_synth_mute: ## @mad_synth_mute
## BB#0: ## %entry
pushq %rax
movl $4096, %esi ## imm = 0x1000
callq ___bzero
popq %rax
ret
llvm-svn: 123089
2011-01-08 21:19:19 +00:00
5120ebf184
fix an issue in IsPointerOffset that prevented us from recognizing that
...
P and P+1 are relative to the same base pointer.
llvm-svn: 123087
2011-01-08 21:07:56 +00:00
4dc1fd938f
enhance memcpyopt to merge a store and a subsequent
...
memset into a single larger memset.
llvm-svn: 123086
2011-01-08 20:54:51 +00:00
c638147e9f
constify TargetData references.
...
Split memset formation logic out into its own
"tryMergingIntoMemset" helper function.
llvm-svn: 123081
2011-01-08 20:24:01 +00:00
59c82f850d
When loop rotation happens, it is *very* common for the duplicated condbr
...
to be foldable into an uncond branch. When this happens, we can make a
much simpler CFG for the loop, which is important for nested loop cases
where we want the outer loop to be aggressively optimized.
Handle this case more aggressively. For example, previously on
phi-duplicate.ll we would get this:
define void @test(i32 %N, double* %G) nounwind ssp {
entry:
%cmp1 = icmp slt i64 1, 1000
br i1 %cmp1, label %bb.nph, label %for.end
bb.nph: ; preds = %entry
br label %for.body
for.body: ; preds = %bb.nph, %for.cond
%j.02 = phi i64 [ 1, %bb.nph ], [ %inc, %for.cond ]
%arrayidx = getelementptr inbounds double* %G, i64 %j.02
%tmp3 = load double* %arrayidx
%sub = sub i64 %j.02, 1
%arrayidx6 = getelementptr inbounds double* %G, i64 %sub
%tmp7 = load double* %arrayidx6
%add = fadd double %tmp3, %tmp7
%arrayidx10 = getelementptr inbounds double* %G, i64 %j.02
store double %add, double* %arrayidx10
%inc = add nsw i64 %j.02, 1
br label %for.cond
for.cond: ; preds = %for.body
%cmp = icmp slt i64 %inc, 1000
br i1 %cmp, label %for.body, label %for.cond.for.end_crit_edge
for.cond.for.end_crit_edge: ; preds = %for.cond
br label %for.end
for.end: ; preds = %for.cond.for.end_crit_edge, %entry
ret void
}
Now we get the much nicer:
define void @test(i32 %N, double* %G) nounwind ssp {
entry:
br label %for.body
for.body: ; preds = %entry, %for.body
%j.01 = phi i64 [ 1, %entry ], [ %inc, %for.body ]
%arrayidx = getelementptr inbounds double* %G, i64 %j.01
%tmp3 = load double* %arrayidx
%sub = sub i64 %j.01, 1
%arrayidx6 = getelementptr inbounds double* %G, i64 %sub
%tmp7 = load double* %arrayidx6
%add = fadd double %tmp3, %tmp7
%arrayidx10 = getelementptr inbounds double* %G, i64 %j.01
store double %add, double* %arrayidx10
%inc = add nsw i64 %j.01, 1
%cmp = icmp slt i64 %inc, 1000
br i1 %cmp, label %for.body, label %for.end
for.end: ; preds = %for.body
ret void
}
With all of these recent changes, we are now able to compile:
void foo(char *X) {
for (int i = 0; i != 100; ++i)
for (int j = 0; j != 100; ++j)
X[j+i*100] = 0;
}
into a single memset of 10000 bytes. This series of changes
should also be helpful for other nested loop scenarios as well.
llvm-svn: 123079
2011-01-08 19:59:06 +00:00
30f318e5d1
split ssa updating code out to its own helper function. Don't bother
...
moving the OrigHeader block anymore: we just merge it away anyway so
its code layout doesn't matter.
llvm-svn: 123077
2011-01-08 19:26:33 +00:00
2615130e1d
Implement a TODO: Enhance loopinfo to merge away the unconditional branch
...
that it was leaving in loops after rotation (between the original latch
block and the original header.
With this change, it is possible for rotated loops to have just a single
basic block, which is useful.
llvm-svn: 123075
2011-01-08 19:10:28 +00:00
930b716e1b
various code cleanups, enhance MergeBlockIntoPredecessor to preserve
...
loop info.
llvm-svn: 123074
2011-01-08 19:08:40 +00:00
fee37c5fa3
inline preserveCanonicalLoopForm now that it is simple.
...
llvm-svn: 123073
2011-01-08 18:55:50 +00:00
063dca0f6a
Three major changes:
...
1. Rip out LoopRotate's domfrontier updating code. It isn't
needed now that LICM doesn't use DF and it is super complex
and gross.
2. Make DomTree updating code a lot simpler and faster. The
old loop over all the blocks was just to find a block??
3. Change the code that inserts the new preheader to just use
SplitCriticalEdge instead of doing an overcomplex
reimplementation of it.
No behavior change, except for the name of the inserted preheader.
llvm-svn: 123072
2011-01-08 18:52:51 +00:00
30d95f9f87
reduce nesting.
...
llvm-svn: 123071
2011-01-08 18:47:43 +00:00
7fab23bc1d
LoopRotate requires canonical loop form, so it always has preheaders
...
and latch blocks. Reorder entry conditions to make hte pass faster
and more logical.
llvm-svn: 123069
2011-01-08 18:06:22 +00:00
d62691f4e8
use the LI ivar.
...
llvm-svn: 123068
2011-01-08 17:49:51 +00:00
385f2ec6d8
some cleanups: remove dead arguments and eliminate ivars
...
that are just passed to one function.
llvm-svn: 123067
2011-01-08 17:48:33 +00:00
25ba40a0cc
fix an issue duncan pointed out, which could cause loop rotate
...
to violate LCSSA form
llvm-svn: 123066
2011-01-08 17:38:45 +00:00
b4ab257bcc
Fix coding style issues.
...
llvm-svn: 123065
2011-01-08 17:07:11 +00:00
84986b298a
Make more passes preserve dominators (or state that they preserve dominators if
...
they all ready do). This removes two dominator recomputations prior to isel,
which is a 1% improvement in total llc time for 403.gcc.
The only potentially suspect thing is making GCStrategy recompute dominators if
it used a custom lowering strategy.
llvm-svn: 123064
2011-01-08 17:01:52 +00:00
80bd9af7c5
Contract subloop bodies. However, it is still important to visit the phis at the
...
top of subloop headers, as the phi uses logically occur outside of the subloop.
llvm-svn: 123062
2011-01-08 15:52:22 +00:00
6a1fb8f235
Fix a bug in r123034 (trying to sext/zext non-integers) and clean up a little.
...
llvm-svn: 123061
2011-01-08 10:51:36 +00:00
8c5defd0b0
Have loop-rotate simplify instructions (yay instsimplify!) as it clones
...
them into the loop preheader, eliminating silly instructions like
"icmp i32 0, 100" in fixed tripcount loops. This also better exposes the
bigger problem with loop rotate that I'd like to fix: once this has been
folded, the duplicated conditional branch *often* turns into an uncond branch.
Not aggressively handling this is pessimizing later loop optimizations
somethin' fierce by making "dominates all exit blocks" checks fail.
llvm-svn: 123060
2011-01-08 08:24:46 +00:00
43f8d16482
Revamp the ValueMapper interfaces in a couple ways:
...
1. Take a flags argument instead of a bool. This makes
it more clear to the reader what it is used for.
2. Add a flag that says that "remapping a value not in the
map is ok".
3. Reimplement MapValue to share a bunch of code and be a lot
more efficient. For lookup failures, don't drop null values
into the map.
4. Using the new flag a bunch of code can vaporize in LinkModules
and LoopUnswitch, kill it.
No functionality change.
llvm-svn: 123058
2011-01-08 08:15:20 +00:00
2b3f20e6ec
two minor changes: switch to the standard ValueToValueMapTy
...
map from ValueMapper.h (giving us access to its utilities)
and add a fastpath in the loop rotation code, avoiding expensive
ssa updator manipulation for values with nothing to update.
llvm-svn: 123057
2011-01-08 07:21:31 +00:00
fc3d7f664b
InstCombine: Match min/max hidden by sext/zext
...
X = sext x; x >s c ? X : C+1 --> X = sext x; X <s C+1 ? C+1 : X
X = sext x; x <s c ? X : C-1 --> X = sext x; X >s C-1 ? C-1 : X
X = zext x; x >u c ? X : C+1 --> X = zext x; X <u C+1 ? C+1 : X
X = zext x; x <u c ? X : C-1 --> X = zext x; X >u C-1 ? C-1 : X
X = sext x; x >u c ? X : C+1 --> X = sext x; X <u C+1 ? C+1 : X
X = sext x; x <u c ? X : C-1 --> X = sext x; X >u C-1 ? C-1 : X
Instead of calculating this with mixed types promote all to the
larger type. This enables scalar evolution to analyze this
expression. PR8866
llvm-svn: 123034
2011-01-07 21:33:14 +00:00
411e6eedff
Some whitespace fixes
...
llvm-svn: 123033
2011-01-07 21:33:13 +00:00
134cde912a
Revert 122959, it needs more thought. Add it back to README.txt with additional notes.
...
llvm-svn: 123030
2011-01-07 20:42:20 +00:00
89afb43b1e
Remove all uses of the "ugly" method BranchInst::setUnconditionalDest().
...
llvm-svn: 123025
2011-01-07 20:25:56 +00:00
ae67cc13a9
InstCombine: Turn _chk functions into the "unsafe" variant if length and max langth are equal.
...
This happens when we take the (non-constant) length from a malloc.
llvm-svn: 122961
2011-01-06 14:22:52 +00:00
799b011276
InstCombine: If we call llvm.objectsize on a malloc call we can replace it with the size passed to malloc.
...
llvm-svn: 122959
2011-01-06 13:11:05 +00:00
a76cc117e0
InstCombine: Teach llvm.objectsize folding to look through GEPs.
...
llvm-svn: 122958
2011-01-06 13:07:49 +00:00
9ec19ea06a
Add the CallInst optimizations that don't involve expanding inline assembly to
...
OptimizeInst() so that they can be used on a worklist instruction.
llvm-svn: 122945
2011-01-06 02:56:42 +00:00
d28c78eb4f
Move the GEP handling in CodeGenPrepare to OptimizeInst().
...
llvm-svn: 122944
2011-01-06 02:44:52 +00:00
14ac865ca9
Split the optimizations in CodeGenPrepare that don't manipulate the iterators
...
into a separate function, so that it can be called from a loop using a worklist
rather than a loop traversing a whole basic block.
llvm-svn: 122943
2011-01-06 02:37:26 +00:00
70be93a200
Zap the last two -Wself-assign warnings in llvm.
...
Simplify RALinScan::DowngradeRegister with TRI::getOverlaps while we are there.
llvm-svn: 122940
2011-01-06 01:33:22 +00:00
ce3b930a98
Stop reallocating SunkAddrs for each basic block. When we move to an instruction
...
worklist, the key will need to become std::pair<BasicBlock*, Value*>.
llvm-svn: 122932
2011-01-06 00:42:50 +00:00
b62ccb241b
Add some more statistics to CodeGenPrepare.
...
llvm-svn: 122891
2011-01-05 17:47:38 +00:00
ced753fadf
Add some stats to CodeGenPrepare to make it easier to speed it up without
...
regressing code quality.
llvm-svn: 122887
2011-01-05 17:27:27 +00:00
6a78995369
Use pop_back_val instead of back followed by pop_back.
...
llvm-svn: 122876
2011-01-05 16:08:47 +00:00
5a2bb998ac
Use a worklist for later iterations just like ordinary instsimplify. The next
...
step is to only process instructions in subloops if they have been modified by
an earlier simplification.
llvm-svn: 122869
2011-01-05 05:47:47 +00:00
4c51d122d5
Change LoopInstSimplify back to a LoopPass. It revisits subloops rather than
...
skipping them, but it should probably use a worklist and only revisit those
instructions in subloops that have actually changed. It should probably also
use a worklist after the first iteration like instsimplify now does. Regardless,
it's only 0.3% of opt -O2 time on 403.gcc if it replaces the instcombine placed
in the middle of the loop passes.
llvm-svn: 122868
2011-01-05 05:15:53 +00:00
7b25ff04bd
Don't bother value numbering instructions with void types in GVN. In theory this should allow us to insert
...
fewer things into the value numbering maps, but any speedup is beneath the noise threshold on my machine
on 403.gcc.
llvm-svn: 122844
2011-01-04 22:15:21 +00:00
e39cb57b09
Complete the NumberTable --> LeaderTable rename.
...
llvm-svn: 122828
2011-01-04 19:29:46 +00:00
d7d06d3aaf
Fix typo in a comment.
...
llvm-svn: 122827
2011-01-04 19:25:18 +00:00
51489b3b28
Prune #include's.
...
llvm-svn: 122826
2011-01-04 19:24:57 +00:00
c7c3bc63f7
Clarify terminology, settling on referring to what was the "number table" as the "leader table", and
...
rename methods to make it much more clear what they're doing.
llvm-svn: 122823
2011-01-04 19:13:25 +00:00
83546f2fe0
When removing a value from GVN's leaders list, don't drop the Next pointer in a corner case.
...
llvm-svn: 122822
2011-01-04 19:10:54 +00:00
a71d2cc88d
Improve the accuracy of the inlining heuristic looking for the
...
case where a static caller is itself inlined everywhere else, and
thus may go away if it doesn't get too big due to inlining other
things into it. If there are references to the caller other than
calls, it will not be removed; account for this.
This results in same-day completion of the case in PR8853.
llvm-svn: 122821
2011-01-04 19:01:54 +00:00
41a1550ef5
Branch instructions don't produce values, so there's no need to generate a value number for them. This
...
avoids adding them to the various value numbering tables, resulting in a minor (~3%) speedup for GVN
on 40.gcc.
llvm-svn: 122819
2011-01-04 18:54:18 +00:00
22c53e277a
Remove commented out code.
...
llvm-svn: 122817
2011-01-04 18:22:08 +00:00
b2a41e9388
Switch to the new style of asterisk placement.
...
llvm-svn: 122815
2011-01-04 18:19:19 +00:00
8643810ede
Teach loop-idiom to turn a loop containing a memset into a larger memset
...
when safe.
The testcase is basically this nested loop:
void foo(char *X) {
for (int i = 0; i != 100; ++i)
for (int j = 0; j != 100; ++j)
X[j+i*100] = 0;
}
which gets turned into a single memset now. clang -O3 doesn't optimize
this yet though due to a phase ordering issue I haven't analyzed yet.
llvm-svn: 122806
2011-01-04 07:46:33 +00:00
a62b01dc37
restructure this a bit. Initialize the WeakVH with "I", the
...
instruction *after* the store. The store will always be deleted
if the transformation kicks in, so we'd do an N^2 scan of every
loop block. Whoops.
llvm-svn: 122805
2011-01-04 07:27:30 +00:00
f4e13699e7
Avoid finding loop back edges when we are not splitting critical edges in
...
CodeGenPrepare (which is the default behavior).
llvm-svn: 122801
2011-01-04 04:43:31 +00:00
e924969380
Address most of Duncan's review comments. Also, make LoopInstSimplify a simple
...
FunctionPass. It probably doesn't have a reason to be a LoopPass, as it will
probably drop the simple fixed point and either use RPO iteration or Duncan's
approach in instsimplify of only revisiting instructions that have changed.
The next step is to preserve LoopSimplify. This looks like it won't be too hard,
although the pass manager doesn't actually seem to respect when non-loop passes
claim to preserve LCSSA or LoopSimplify. This will have to be fixed.
llvm-svn: 122791
2011-01-04 00:12:46 +00:00
0ba473c218
use the very-handy getTruncateOrZeroExtend helper function, and
...
stop setting NSW: signed overflow is possible. Thanks to Dan
for pointing these out.
llvm-svn: 122790
2011-01-04 00:06:55 +00:00
0839d3930a
Fix comment.
...
llvm-svn: 122788
2011-01-03 23:51:56 +00:00
d62d37225a
Use the new addEscapingValue callback to update GlobalsModRef when GVN adds PHIs of GEPs. For the moment,
...
have GlobalsModRef handle this conservatively by simply removing the value from its maps.
llvm-svn: 122787
2011-01-03 23:51:43 +00:00
bde6ec1db6
Duncan deftly points out that readnone functions aren't
...
invalidated by stores, so they can be handled as 'simple'
operations.
llvm-svn: 122785
2011-01-03 23:38:13 +00:00
3a33d0cc4a
Simplify GVN's value expression structure, allowing the elimination of a lot of
...
almost-but-not-quite-identical code. No intended functionality change.
llvm-svn: 122760
2011-01-03 19:00:11 +00:00
16ca19ffc5
stength reduce my previous patch a bit. The only instructions
...
that are allowed to have metadata operands are intrinsic calls,
and the only ones that take metadata currently return void.
Just reject all void instructions, which should not be value
numbered anyway. To future proof things, add an assert to the
getHashValue impl for calls to check that metadata operands
aren't present.
llvm-svn: 122759
2011-01-03 18:43:03 +00:00
142f1cd251
fix PR8895: metadata operands don't have a strong use of their
...
nested values, so they can change and drop to null, which can
change the hash and cause havok.
It turns out that it isn't a good idea to value number stuff
with metadata operands anyway, so... don't.
llvm-svn: 122758
2011-01-03 18:28:15 +00:00
697de77339
Speed up instsimplify by about 10-15% by not bothering to retry
...
InstructionSimplify on instructions that didn't change since the
last time round the loop.
llvm-svn: 122745
2011-01-03 10:50:04 +00:00
43cecb1200
Switch a worklist in CodeGenPrepare to SmallVector and increase the inline
...
capacity on the Visited SmallPtrSet. On 403.gcc, this is about a 4.5% speedup of
CodeGenPrepare time (which itself is 10% of time spent in the backend).
This is progress towards PR8889.
llvm-svn: 122741
2011-01-03 06:33:01 +00:00
9e5e9ed79a
earlycse can do trivial with-a-block dead store
...
elimination as well. This deletes 60 stores in 176.gcc
that largely come from bitfield code.
llvm-svn: 122736
2011-01-03 04:17:24 +00:00
4b9a525742
switch the load table to use a recycling bump pointer allocator,
...
speeding earlycse up by 6%.
llvm-svn: 122733
2011-01-03 03:53:50 +00:00
e0e32a9ef0
now that loads are in their own table, we can implement
...
store->load forwarding. This allows EarlyCSE to zap 600 more
loads from 176.gcc.
llvm-svn: 122732
2011-01-03 03:46:34 +00:00
92bb0f9f9d
split loads and calls into separate tables. Loads are now just indexed
...
by their pointer instead of using MemoryValue to wrap it.
llvm-svn: 122731
2011-01-03 03:41:27 +00:00
4cb365414f
various cleanups, no functionality change.
...
llvm-svn: 122729
2011-01-03 03:28:23 +00:00
b9a8efc960
Teach EarlyCSE to do trivial CSE of loads and read-only calls.
...
On 176.gcc, this catches 13090 loads and calls, and increases the
number of simple instructions CSE'd from 29658 to 36208.
llvm-svn: 122727
2011-01-03 03:18:43 +00:00
79d83067ee
rename InstValue to SimpleValue, add some comments.
...
llvm-svn: 122725
2011-01-03 02:20:48 +00:00
edb5bcdde5
CMake: Add missing source file.
...
llvm-svn: 122724
2011-01-03 02:13:05 +00:00
d815f69b30
Allocate nodes for the scoped hash table from a recyling bump pointer
...
allocator. This speeds up early cse by about 20%
llvm-svn: 122723
2011-01-03 01:42:46 +00:00
02a9776b64
reduce redundancy in the hashing code and other misc cleanups.
...
llvm-svn: 122720
2011-01-03 01:10:08 +00:00
cab9a0abab
Add a new loop-instsimplify pass, with the intention of replacing the instance
...
of instcombine that is currently in the middle of the loop pass pipeline. This
commit only checks in the pass; it will hopefully be enabled by default later.
llvm-svn: 122719
2011-01-03 00:25:16 +00:00
0844c76f9a
fix some pastos
...
llvm-svn: 122718
2011-01-02 23:29:58 +00:00
8fac5db251
add DEBUG and -stats output to earlycse.
...
Teach it to CSE the rest of the non-side-effecting instructions.
llvm-svn: 122716
2011-01-02 23:19:45 +00:00
18ae5436b1
Enhance earlycse to do CSE of casts, instsimplify and die.
...
Add a testcase.
llvm-svn: 122715
2011-01-02 23:04:14 +00:00
bf0aa927cc
split dom frontier handling stuff out to its own DominanceFrontier header,
...
so that Dominators.h is *just* domtree. Also prune #includes a bit.
llvm-svn: 122714
2011-01-02 22:09:33 +00:00
704541bb23
sketch out a new early cse pass. No functionality yet.
...
llvm-svn: 122713
2011-01-02 21:47:05 +00:00
9c69406f2b
fix a miscompilation of tramp3d-v4: when forming a memcpy, we have to make
...
sure that the loop we're promoting into a memcpy doesn't mutate the input
of the memcpy. Before we were just checking that the dest of the memcpy
wasn't mod/ref'd by the loop.
llvm-svn: 122712
2011-01-02 21:14:18 +00:00
5702a43c09
If a loop iterates exactly once (has backedge count = 0) then don't
...
mess with it. We'd rather peel/unroll it than convert all of its
stores into memsets.
llvm-svn: 122711
2011-01-02 20:24:21 +00:00
5361b84184
Also remove functions that use complex constant expressions in terms of
...
another function.
llvm-svn: 122705
2011-01-02 19:16:44 +00:00
8455b6e45e
enhance loop idiom recognition to scan *all* unconditionally executed
...
blocks in a loop, instead of just the header block. This makes it more
aggressive, able to handle Duncan's Ada examples.
llvm-svn: 122704
2011-01-02 19:01:03 +00:00
0cdc6f62a5
make inSubLoop much more efficient.
...
llvm-svn: 122703
2011-01-02 18:53:08 +00:00
27497ece96
rip out isExitBlockDominatedByBlockInLoop, calling DomTree::dominates instead.
...
isExitBlockDominatedByBlockInLoop is a relic of the days when domtree was
*just* a tree and didn't have DFS numbers. Checking DFS numbers is faster
and easier than "limiting the search of the tree".
llvm-svn: 122702
2011-01-02 18:45:39 +00:00
0469e01c02
add a list of opportunities for future improvement.
...
llvm-svn: 122701
2011-01-02 18:32:09 +00:00
64f1c0dcda
Fix PR8702 by not having LoopSimplify claim to preserve LCSSA form. As described
...
in the PR, the pass could break LCSSA form when inserting preheaders. It probably
would be easy enough to fix this, but since currently we always go into LCSSA form
after running this pass, doing so is not urgent.
llvm-svn: 122695
2011-01-02 13:38:21 +00:00
ddf58010bd
Allow loop-idiom to run on multiple BB loops, but still only scan the loop
...
header for now for memset/memcpy opportunities. It turns out that loop-rotate
is successfully rotating loops, but *DOESN'T MERGE THE BLOCKS*, turning "for
loops" into 2 basic block loops that loop-idiom was ignoring.
With this fix, we form many *many* more memcpy and memsets than before, including
on the "history" loops in the viterbi benchmark, which look like this:
for (j=0; j<MAX_history; ++j) {
history_new[i][j+1] = history[2*i][j];
}
Transforming these loops into memcpy's speeds up the viterbi benchmark from
11.98s to 3.55s on my machine. Woo.
llvm-svn: 122685
2011-01-02 07:58:36 +00:00
5b5a043d82
remove debugging code.
...
llvm-svn: 122683
2011-01-02 07:37:13 +00:00
12f91befce
add some -stats output.
...
llvm-svn: 122682
2011-01-02 07:36:44 +00:00
679572e584
improve loop rotation to use CodeMetrics to analyze the
...
size of a loop header instead of its own code size estimator.
This allows it to handle bitcasts etc more precisely.
llvm-svn: 122681
2011-01-02 07:35:53 +00:00
85b6d81d41
teach loop idiom recognition to form memcpy's from simple loops.
...
llvm-svn: 122678
2011-01-02 03:37:56 +00:00
4e250c8245
Remove functions from the FnSet when one of their callee's is being merged. This
...
maintains the guarantee that the DenseSet expects two elements it contains to
not go from inequal to equal under its nose.
As a side-effect, this also lets us switch from iterating to a fixed-point to
actually maintaining a work queue of functions to look at again, and we don't
add thunks to our work queue so we don't need to detect and ignore them.
llvm-svn: 122677
2011-01-02 02:46:33 +00:00
1903c42b97
fix a globalopt crash on two Adobe-C++ testcases that the recent
...
loop idiom pass exposed.
llvm-svn: 122674
2011-01-01 22:31:46 +00:00
a3514441e0
add a validity check that was missed, fixing a crash on the
...
new testcase.
llvm-svn: 122662
2011-01-01 20:12:04 +00:00
91a4435875
improve validity check to handle constant-trip-count loops more
...
aggressively. In practice, this doesn't help anything though,
see the todo.
llvm-svn: 122660
2011-01-01 19:54:22 +00:00
8b3baf6d75
implement the "no aliasing accesses in loop" safety check. This pass
...
should be correct now.
llvm-svn: 122659
2011-01-01 19:39:01 +00:00
2c440fa403
Simplify this pass by using a depth-first iterator to ensure that all
...
operands are visited before the instructions themselves.
llvm-svn: 122647
2010-12-31 17:49:05 +00:00
6cc7126ed9
Zap dead instructions harder.
...
llvm-svn: 122645
2010-12-31 16:17:54 +00:00
570dd787a6
Make a bunch of symbols internal.
...
llvm-svn: 122642
2010-12-30 22:34:44 +00:00
65a699d4d0
simplify this, isBytewiseValue handles the extra check. We still
...
check for "multiple of a byte" in size to make it clear that the
>> 3 below is safe.
llvm-svn: 122604
2010-12-28 18:53:48 +00:00
5cf10e691b
Silence gcc warning about an unused variable when doing a release build.
...
llvm-svn: 122593
2010-12-28 09:41:15 +00:00
cb18bfa3d2
fix some issues Frits noticed, add AliasAnalysis as a dependency
...
llvm-svn: 122585
2010-12-27 18:39:08 +00:00
84bd73c527
BuildLibCalls: Nuke EmitMemCpy, EmitMemMove and EmitMemSet. They are dead and superseded by IRBuilder.
...
llvm-svn: 122576
2010-12-27 00:25:32 +00:00
7cba269dfb
SimplifyLibCalls: Use IRBuilder to simplify code.
...
llvm-svn: 122575
2010-12-27 00:16:46 +00:00
b9fe685b9a
have loop-idiom nuke instructions that feed stores that get removed.
...
llvm-svn: 122574
2010-12-27 00:03:23 +00:00
29e14edc8d
implement enough of the memset inference algorithm to recognize and insert
...
memsets. This is still missing one important validity check, but this is enough
to compile stuff like this:
void test0(std::vector<char> &X) {
for (std::vector<char>::iterator I = X.begin(), E = X.end(); I != E; ++I)
*I = 0;
}
void test1(std::vector<int> &X) {
for (long i = 0, e = X.size(); i != e; ++i)
X[i] = 0x01010101;
}
With:
$ clang t.cpp -S -o - -O2 -emit-llvm | opt -loop-idiom | opt -O3 | llc
to:
__Z5test0RSt6vectorIcSaIcEE: ## @_Z5test0RSt6vectorIcSaIcEE
## BB#0: ## %entry
subq $8, %rsp
movq (%rdi), %rax
movq 8(%rdi), %rsi
cmpq %rsi, %rax
je LBB0_2
## BB#1: ## %bb.nph
subq %rax, %rsi
movq %rax, %rdi
callq ___bzero
LBB0_2: ## %for.end
addq $8, %rsp
ret
...
__Z5test1RSt6vectorIiSaIiEE: ## @_Z5test1RSt6vectorIiSaIiEE
## BB#0: ## %entry
subq $8, %rsp
movq (%rdi), %rax
movq 8(%rdi), %rdx
subq %rax, %rdx
cmpq $4, %rdx
jb LBB1_2
## BB#1: ## %for.body.preheader
andq $-4, %rdx
movl $1, %esi
movq %rax, %rdi
callq _memset
LBB1_2: ## %for.end
addq $8, %rsp
ret
llvm-svn: 122573
2010-12-26 23:42:51 +00:00
6cf8d6cc6e
start using irbuilder to make mem intrinsics in a few passes.
...
llvm-svn: 122572
2010-12-26 22:57:41 +00:00
7c5f9c35d1
sketch more of this out.
...
llvm-svn: 122567
2010-12-26 20:45:45 +00:00
9cb1035f94
move isBytewiseValue out to ValueTracking.h/cpp
...
llvm-svn: 122565
2010-12-26 20:15:01 +00:00
81ae3f299a
actually add the file...
...
llvm-svn: 122563
2010-12-26 19:39:38 +00:00
2ef535a4e4
Start of a pass for recognizing memset and memcpy idioms.
...
No functionality yet.
llvm-svn: 122562
2010-12-26 19:32:44 +00:00
30342fb1fd
Simplify code.
...
llvm-svn: 122561
2010-12-26 15:23:45 +00:00
d729d0dcdb
don't lose TD info
...
llvm-svn: 122556
2010-12-25 20:52:04 +00:00
20fca48341
switch the inliner alignment enforcement stuff to use the
...
getOrEnforceKnownAlignment function, which simplifies the code
and makes it stronger.
llvm-svn: 122555
2010-12-25 20:42:38 +00:00
6fcd32e7d7
Move getOrEnforceKnownAlignment out of instcombine into Transforms/Utils.
...
llvm-svn: 122554
2010-12-25 20:37:57 +00:00
b90b2f0635
Fix a thinko pointed out by Frits van Bommel: looking through global variables in isBytewiseValue is not safe.
...
llvm-svn: 122550
2010-12-24 22:23:59 +00:00
ea9152e551
MemCpyOpt: Turn memcpys from a constant into a memset if possible.
...
This allows us to compile "int cst[] = {-1, -1, -1};" into
movl $-1, 16(%rsp)
movq $-1, 8(%rsp)
instead of
movl _cst+8(%rip), %eax
movl %eax, 16(%rsp)
movq _cst(%rip), %rax
movq %rax, 8(%rsp)
llvm-svn: 122548
2010-12-24 21:17:12 +00:00
226ac14afb
When determining if we can fold (x >> C1) << C2, the bits that we need to verify are zero
...
are not the low bits of x, but the bits that WILL be the low bits after the operation completes.
llvm-svn: 122529
2010-12-23 23:56:24 +00:00
5d690d4168
It is possible for SimplifyCFG to cause PHI nodes to become redundant too late in the optimization
...
pipeline to be caught by instcombine, and it's not feasible to catch them in SimplifyCFG because the
use-lists are in an inconsistent state at the point where it could know that it need to simplify them.
Instead, have CodeGenPrepare look for trivially redundant PHIs as part of its general cleanup effort.
llvm-svn: 122516
2010-12-23 20:57:35 +00:00
18b762a946
Preserve the address space when generating bitcasts for MemTransferInst in ConvertToScalarInfo
...
llvm-svn: 122462
2010-12-23 01:41:32 +00:00
9b43f33620
Change all self assignments X=X to (void)X, so that we can turn on a
...
new gcc warning that complains on self-assignments and
self-initializations.
llvm-svn: 122458
2010-12-23 00:58:24 +00:00
8ef5001b27
InstCombine: creating selects from -1 and 0 is fine, they combine into a sext from i1.
...
llvm-svn: 122453
2010-12-22 23:12:15 +00:00
fbb9ac3cca
Add a generic expansion transform: A op (B op' C) -> (A op B) op' (A op C)
...
if both A op B and A op C simplify. This fires fairly often but doesn't
make that much difference. On gcc-as-one-file it removes two "and"s and
turns one branch into a select.
llvm-svn: 122399
2010-12-22 13:36:08 +00:00
3547d2ebd8
Add some statistics, good for understanding how much more powerful
...
instcombine is compared to instsimplify.
llvm-svn: 122397
2010-12-22 09:40:51 +00:00
5ab8d4b5e5
Give GVN back the ability to perform simple conditional propagation on conditional branch values.
...
I still think that LVI should be handling this, but that capability is some ways off in the future,
and this matters for some significant benchmarks.
llvm-svn: 122378
2010-12-21 23:54:34 +00:00
12470778d7
Remove dead code.
...
llvm-svn: 122371
2010-12-21 22:31:24 +00:00
43493c089f
GVN's Expression is not POD-like (it contains a SmallVector). Simplify code while at it.
...
llvm-svn: 122362
2010-12-21 21:30:19 +00:00
3b8af41a3e
Visit instructions deterministically. Use a FIFO so as to approximately
...
visit instructions before their uses, since InstructionSimplify does a
better job in that case. All this prompted by Frits van Bommel.
llvm-svn: 122343
2010-12-21 17:08:55 +00:00
e7cbb64ec0
If an instruction simplifies, try again to simplify any uses of it. This is
...
not very important since the pass is only used for testing, but it does make
it more realistic. Suggested by Frits van Bommel.
llvm-svn: 122336
2010-12-21 16:12:03 +00:00
d0eb6d39f8
Pull a few more simplifications out of instcombine (there are still
...
plenty left though!), in particular for multiplication.
llvm-svn: 122330
2010-12-21 14:00:22 +00:00
eaff500c7b
Oops, forgot to add the pass itself!
...
llvm-svn: 122265
2010-12-20 21:07:42 +00:00
a436cbe4bf
Add a new convenience pass for testing InstructionSimplify. Previously
...
it could only be tested indirectly, via instcombine, gvn or some other
pass that makes use of InstructionSimplify, which means that testcases
had to be carefully contrived to dance around any other transformations
that that pass did.
llvm-svn: 122264
2010-12-20 20:54:37 +00:00
f7957d0463
Add a check missing from my last commit and avoid a potential overflow situation.
...
llvm-svn: 122258
2010-12-20 20:00:31 +00:00
2bca3a67b3
Reduce indentation.
...
llvm-svn: 122249
2010-12-20 16:21:59 +00:00
68531baea9
Teach InstCombine to merge (icmp ult (X + CA), C1) | (icmp eq X, C2) into (icmp ult (X + CA), C1 + 1) if C2 + CA == C1.
...
InstCombine creates these so now we compile x == 23 || x == 24 || x == 25 to
%x.off = add i32 %x, -23
%1 = icmp ult i32 %x.off, 3
instead of
%x.off = add i32 %x, -23
%1 = icmp ult i32 %x.off, 2
%cmp3 = icmp eq i32 %x, 25
%ret2 = or i1 %1, %cmp3
llvm-svn: 122248
2010-12-20 16:18:51 +00:00
27ca8ebd4b
fix PR8807 by making transformConstExprCastCall aware of byval arguments.
...
llvm-svn: 122238
2010-12-20 08:36:38 +00:00
7398965b67
various cleanups for transformConstExprCastCall
...
llvm-svn: 122237
2010-12-20 08:25:06 +00:00
0f11495289
when eliding a byval copy due to inlining a readonly function, we have
...
to make sure that the reused alloca has sufficient alignment.
llvm-svn: 122236
2010-12-20 08:10:40 +00:00
0099744506
pull byval processing out to its own helper function.
...
llvm-svn: 122235
2010-12-20 07:57:41 +00:00
7394680a00
fix PR8769, a miscompilation by inliner when inlining a function with a byval
...
argument. The generated alloca has to have at least the alignment of the
byval, if not, the client may be making assumptions that the new alloca won't
satisfy.
llvm-svn: 122234
2010-12-20 07:45:28 +00:00
1991c47ec1
Avoid dropping the address space when InstCombine optimizes memset
...
llvm-svn: 122215
2010-12-20 01:05:30 +00:00
4fb9dd4c74
fix an oversight caught by Frits!
...
llvm-svn: 122204
2010-12-19 23:24:04 +00:00
b6252a376a
tidy up
...
llvm-svn: 122190
2010-12-19 20:24:28 +00:00
3e635d2e99
move a transformation to a more logical place, simplifying it.
...
llvm-svn: 122183
2010-12-19 19:43:52 +00:00
5e0c0c72e9
recognize an unsigned add with overflow idiom into uadd.
...
This resolves a README entry and technically resolves PR4916,
but we still get poor code for the testcase in that PR because
GVN isn't CSE'ing uadd with add, filed as PR8817.
Previously we got:
_test7: ## @test7
addq %rsi, %rdi
cmpq %rdi, %rsi
movl $42, %eax
cmovaq %rsi, %rax
ret
Now we get:
_test7: ## @test7
addq %rsi, %rdi
movl $42, %eax
cmovbq %rsi, %rax
ret
llvm-svn: 122182
2010-12-19 19:37:52 +00:00
33dc3f0cfa
optimize uadd(x, cst) into a comparison when the normal
...
result is dead. This is required for my next patch to not
regress the testsuite.
llvm-svn: 122181
2010-12-19 19:35:32 +00:00
ce2995ae58
use IC.ReplaceInstUsesWith instead of a raw RAUW so that uses of
...
the old thing end up on the instcombine worklist. Not doing this
can cause an extra top-level iteration of instcombine, burning
compile time.
llvm-svn: 122179
2010-12-19 18:38:44 +00:00
79874566ce
generalize the sadd creation code to not require that the
...
sadd formed is half the size of the original type. We can
now compile this into a sadd.i8:
unsigned char X(char a, char b) {
int res = a+b;
if ((unsigned )(res+128) > 255U)
abort();
return res;
}
llvm-svn: 122178
2010-12-19 18:35:09 +00:00
c56c845377
fix another miscompile in the llvm.sadd formation logic: it wasn't
...
checking to see if the high bits of the original add result were dead.
Inserting a smaller add and zexting back to that size is not good enough.
This is likely to be the fix for 8816.
llvm-svn: 122177
2010-12-19 18:22:06 +00:00
f29562db25
fix a bug (possibly 8816) in the sadd forming xform: it isn't
...
profitable (or safe) to promote code when the add-with-constant
has other uses.
llvm-svn: 122175
2010-12-19 17:59:02 +00:00
ee61c1d820
rework the code added in r122072 to pull it out to its own
...
helper function, clean up comments, and reduce indentation.
No functionality change.
llvm-svn: 122174
2010-12-19 17:52:50 +00:00
408a684d29
Enhance LICM to promote alias sets whose pointers themselves are stored,
...
which doesn't affect the memory address being promoted.
llvm-svn: 122172
2010-12-19 05:57:25 +00:00
3337a81450
fix PR8602, a bug in an assertion: a volatile store *of* a pointer
...
does not make the alias set for that pointer volatile, just stores
*to* the pointer.
llvm-svn: 122171
2010-12-19 05:51:54 +00:00
fb888622c3
revert r122164, I'm going to go with a different approach.
...
llvm-svn: 122168
2010-12-19 04:23:03 +00:00
583ec6fa44
first step to fixing PR8642: don't fold away empty basic blocks
...
which have trapping constant exprs in them due to PHI nodes.
Eliminating them can cause the constant expr to be evalutated
on new paths if the input edges are critical.
llvm-svn: 122164
2010-12-19 03:02:34 +00:00
6b8b4855ff
simplify this a bit.
...
llvm-svn: 122156
2010-12-18 20:22:49 +00:00
5e3605552e
Whitespace fixes. No functionality change.
...
llvm-svn: 122110
2010-12-17 23:27:41 +00:00
7aa18bf46a
Add vector versions of some existing scalar transforms to aid codegen in matching psign & pblend operations to the IR produced by clang/gcc for their C idioms.
...
llvm-svn: 122105
2010-12-17 23:12:19 +00:00
1294ea7d53
Reapply r121905 (automatic synthesis of @llvm.sadd.with.overflow) with a fix for a bug that manifested itself
...
on the DragonEgg self-host bot. Unfortunately, the testcase is pretty messy and doesn't reduce well due to
interactions with other parts of InstCombine.
llvm-svn: 122072
2010-12-17 18:08:00 +00:00
e5f49c4ff2
SimplifyCFG: Ranges can be larger than 64 bits. Fixes Release-selfhost build.
...
llvm-svn: 122054
2010-12-17 10:48:14 +00:00
d14b0f1db7
improve switch formation to handle small range
...
comparisons formed by comparisons. For example,
this:
void foo(unsigned x) {
if (x == 0 || x == 1 || x == 3 || x == 4 || x == 6)
bar();
}
compiles into:
_foo: ## @foo
## BB#0: ## %entry
cmpl $6, %edi
ja LBB0_2
## BB#1: ## %entry
movl %edi, %eax
movl $91, %ecx
btq %rax, %rcx
jb LBB0_3
instead of:
_foo: ## @foo
## BB#0: ## %entry
cmpl $2, %edi
jb LBB0_4
## BB#1: ## %switch.early.test
cmpl $6, %edi
ja LBB0_3
## BB#2: ## %switch.early.test
movl %edi, %eax
movl $88, %ecx
btq %rax, %rcx
jb LBB0_4
This catches a bunch of cases in GCC, which look like this:
%804 = load i32* @which_alternative, align 4, !tbaa !0
%805 = icmp ult i32 %804, 2
%806 = icmp eq i32 %804, 3
%or.cond121 = or i1 %805, %806
%807 = icmp eq i32 %804, 4
%or.cond124 = or i1 %or.cond121, %807
br i1 %or.cond124, label %.thread, label %808
turning this into a range comparison.
llvm-svn: 122045
2010-12-17 06:20:15 +00:00
93dc2b808f
Revert r64460. strtol and friends cannot be marked readonly, even with
...
a null endptr argument, because they may write to errno.
This fixes a seflhost miscompile observed on Linux targets when TBAA
was enabled.
llvm-svn: 122014
2010-12-17 01:09:43 +00:00
9bbe849fc3
Fix a bug in the loop in JumpThreading::ProcessThreadableEdges() where it could falsely produce a MultipleDestSentinel value if the first predecessor ended with an 'indirectbr'. If that happened, it caused an unnecessary FindMostPopularDest() call.
...
This wasn't a correctness problem, but it broke the fast path for single-predecessor blocks.
llvm-svn: 121966
2010-12-16 12:16:00 +00:00
8d1ab6f6e1
Speculatively revert commit 121905 since it looks like it might have broken the
...
dragonegg self-host buildbot. Original commit message:
Add an InstCombine transform to recognize instances of manual overflow-safe addition
(performing the addition in a wider type and explicitly checking for overflow), and
fold them down to intrinsics. This currently only supports signed-addition, but could
be generalized if someone works out the magic constant formulas for other operations.
llvm-svn: 121965
2010-12-16 09:40:54 +00:00
e1a17a3473
Make memcpyopt TBAA-aware.
...
llvm-svn: 121944
2010-12-16 02:51:19 +00:00
4467aa5294
Preserve TBAA tags when doing load PRE.
...
llvm-svn: 121921
2010-12-15 23:53:55 +00:00
1cf8881299
Add an InstCombine transform to recognize instances of manual overflow-safe addition
...
(performing the addition in a wider type and explicitly checking for overflow), and
fold them down to intrinsics. This currently only supports signed-addition, but could
be generalized if someone works out the magic constant formulas for other operations.
Fixes <rdar://problem/8558713>.
llvm-svn: 121905
2010-12-15 22:32:38 +00:00
a4fcd2418d
Move Value::getUnderlyingObject to be a standalone
...
function so that it can live in Analysis instead of
VMCore.
llvm-svn: 121885
2010-12-15 20:02:24 +00:00
0a2c416894
Move Sub simplifications and additional Add simplifications out of
...
instcombine and into InstructionSimplify.
llvm-svn: 121861
2010-12-15 14:07:39 +00:00
3d1803495e
Teach jump threading to "look through" a select when the branch direction of a terminator depends on it.
...
When it sees a promising select it now tries to figure out whether the condition of the select is known in any of the predecessors and if so it maps the operands appropriately.
llvm-svn: 121859
2010-12-15 09:51:20 +00:00
e893e2601e
make qsort predicate more conformant by returning 0 for equal values.
...
llvm-svn: 121838
2010-12-15 04:52:41 +00:00
35609d97ae
Fix PR8790, another instance where unreachable code can cause instruction simplification to fail,
...
this case involve a select that simplifies to itself.
llvm-svn: 121817
2010-12-15 00:55:35 +00:00
15c85c916f
Cleanup trailing whitespace.
...
llvm-svn: 121816
2010-12-15 00:52:44 +00:00
7499b452c1
- Insert new instructions before DomBlock's terminator,
...
which is simpler than finding a place to insert in BB.
- Don't perform the 'if condition hoisting' xform on certain
i1 PHIs, as it interferes with switch formation.
This re-fixes "example 7", without breaking the world hopefully.
llvm-svn: 121764
2010-12-14 08:46:09 +00:00
335f0e4ad4
fix two significant issues with FoldTwoEntryPHINode:
...
first, it can kick in on blocks whose conditions have been
folded to a constant, even though one of the edges will be
trivially folded.
second, it doesn't clean up the "if diamond" that it just
eliminated away. This is a problem because other simplifycfg
xforms kick in depending on the order of block visitation,
causing pointless work.
llvm-svn: 121762
2010-12-14 08:01:53 +00:00
dc20a7d38c
remove the instsimplify logic I added in r121754. It is apparently
...
breaking the selfhost builds, though I can't fathom how.
llvm-svn: 121761
2010-12-14 07:53:03 +00:00
9ac168d0ab
clean up logic, convert std::set to SmallPtrSet, handle the case
...
when all 2-entry phis are simplified away.
llvm-svn: 121760
2010-12-14 07:41:39 +00:00
9fd838d31b
tidy up a bit, move DEBUG down to when we commit to doing the transform so we
...
don't print it unless the xform happens.
llvm-svn: 121758
2010-12-14 07:23:10 +00:00
b42d293faa
use SimplifyInstruction instead of reimplementing part of it.
...
llvm-svn: 121757
2010-12-14 07:20:29 +00:00
fb73de482c
simplify GetIfCondition by using getSinglePredecessor.
...
llvm-svn: 121756
2010-12-14 07:15:21 +00:00
0f4d67bd88
use AddPredecessorToBlock in 3 places instead of a manual loop.
...
llvm-svn: 121755
2010-12-14 07:09:42 +00:00
a07cc6f4fd
make FoldTwoEntryPHINode use instsimplify a bit, make
...
GetIfCondition faster by avoiding pred_iterator. No
really interesting change.
llvm-svn: 121754
2010-12-14 07:00:00 +00:00
afd2a8cfbb
remove the dead (and terrible) llvm::RemoveSuccessor function.
...
llvm-svn: 121753
2010-12-14 06:51:55 +00:00
d7beca3782
improve DEBUG's a bit, switch to eraseFromParent() to simplify
...
code a bit, switch from constant folding to instsimplify.
llvm-svn: 121751
2010-12-14 06:17:25 +00:00
5a9d59d918
reapply my recent change that disables a piece of the switch formation
...
work, but fixes 400.perlbmk.
llvm-svn: 121749
2010-12-14 05:57:30 +00:00
3e5648896e
Fix recent buildbot breakage by pulling SimplifyCFG back to its state as of r121694, the most recent state
...
where I'm confident there were no crashes or miscompilations. XFAIL the test added since then for now.
llvm-svn: 121733
2010-12-13 23:49:28 +00:00
a6e5d5694a
temporarily disable part of my previous patch, which causes an iterator invalidation issue, causing a crash on some versions of perlbmk.
...
llvm-svn: 121728
2010-12-13 23:02:19 +00:00
2d434e594e
add some DEBUG's.
...
llvm-svn: 121711
2010-12-13 19:55:30 +00:00
1e155ab7e1
Fix sort predicate. qsort(3)'s predicate semantics differ from std::sort's. Fixes PR 8780.
...
llvm-svn: 121705
2010-12-13 18:20:38 +00:00
fb836f8c1a
reinstate my patch: the miscompile was caused by an inverted branch in the
...
'and' case.
llvm-svn: 121695
2010-12-13 08:12:19 +00:00
79db357d80
Completely disable the optimization I added in r121680 until
...
I can track down a miscompile. This should bring the buildbots
back to life
llvm-svn: 121693
2010-12-13 07:41:29 +00:00
fbeb55844b
Make simplifycfg reprocess newly formed "br (cond1 | cond2)" conditions
...
when simplifying, allowing them to be eagerly turned into switches. This
is the last step required to get "Example 7" from this blog post:
http://blog.regehr.org/archives/320
On X86, we now generate this machine code, which (to my eye) seems better
than the ICC generated code:
_crud: ## @crud
## BB#0: ## %entry
cmpb $33, %dil
jb LBB0_4
## BB#1: ## %switch.early.test
addb $-34, %dil
cmpb $58, %dil
ja LBB0_3
## BB#2: ## %switch.early.test
movzbl %dil, %eax
movabsq $288230376537592865, %rcx ## imm = 0x400000017001421
btq %rax, %rcx
jb LBB0_4
LBB0_3: ## %lor.rhs
xorl %eax, %eax
ret
LBB0_4: ## %lor.end
movl $1, %eax
ret
llvm-svn: 121690
2010-12-13 07:00:06 +00:00
1d05761df4
make this logic a bit simpler.
...
llvm-svn: 121689
2010-12-13 06:36:51 +00:00
25c3af35d8
split all the guts of SimplifyCFGOpt::run out into one function
...
per terminator kind.
llvm-svn: 121688
2010-12-13 06:25:44 +00:00
cb570f87e5
fix a bug in r121680 that upset the various buildbots.
...
llvm-svn: 121687
2010-12-13 05:34:18 +00:00
a6db741f3d
refactor the speculative execution logic to be factored into the cond branch code instead of
...
doing a cfg search for every block simplified.
llvm-svn: 121686
2010-12-13 05:26:52 +00:00
466f54ffcf
simplify a bunch of code.
...
llvm-svn: 121685
2010-12-13 05:20:28 +00:00
6df7bdd810
move HoistThenElseCodeToIf up to a more logical and efficient-to-handle place.
...
llvm-svn: 121684
2010-12-13 05:15:29 +00:00
2e3832d9a0
move 'MergeBlocksIntoPredecessor' call earlier. Use
...
getSinglePredecessor to simplify code.
llvm-svn: 121683
2010-12-13 05:10:48 +00:00
a69c443459
factor new code out to a SimplifyBranchOnICmpChain helper function.
...
llvm-svn: 121681
2010-12-13 05:03:41 +00:00
a442f24a36
enhance the "change or icmp's into switch" xform to handle one value in an
...
'or sequence' that it doesn't understand. This allows us to optimize
something insane like this:
int crud (unsigned char c, unsigned x)
{
if(((((((((( (int) c <= 32 ||
(int) c == 46) || (int) c == 44)
|| (int) c == 58) || (int) c == 59) || (int) c == 60)
|| (int) c == 62) || (int) c == 34) || (int) c == 92)
|| (int) c == 39) != 0)
foo();
}
into:
define i32 @crud(i8 zeroext %c, i32 %x) nounwind ssp noredzone {
entry:
%cmp = icmp ult i8 %c, 33
br i1 %cmp, label %if.then, label %switch.early.test
switch.early.test: ; preds = %entry
switch i8 %c, label %if.end [
i8 39, label %if.then
i8 44, label %if.then
i8 58, label %if.then
i8 59, label %if.then
i8 60, label %if.then
i8 62, label %if.then
i8 46, label %if.then
i8 92, label %if.then
i8 34, label %if.then
]
by pulling the < comparison out ahead of the newly formed switch.
llvm-svn: 121680
2010-12-13 04:50:38 +00:00
Chris Lattner