optimizations to be much more aggressive in the face of
exact/nsw/nuw div and shifts. For example, these (which
are the same except the first is 'exact' sdiv:
define i1 @sdiv_icmp4_exact(i64 %X) nounwind {
%A = sdiv exact i64 %X, -5 ; X/-5 == 0 --> x == 0
%B = icmp eq i64 %A, 0
ret i1 %B
}
define i1 @sdiv_icmp4(i64 %X) nounwind {
%A = sdiv i64 %X, -5 ; X/-5 == 0 --> x == 0
%B = icmp eq i64 %A, 0
ret i1 %B
}
compile down to:
define i1 @sdiv_icmp4_exact(i64 %X) nounwind {
%1 = icmp eq i64 %X, 0
ret i1 %1
}
define i1 @sdiv_icmp4(i64 %X) nounwind {
%X.off = add i64 %X, 4
%1 = icmp ult i64 %X.off, 9
ret i1 %1
}
This happens when you do something like:
(ptr1-ptr2) == 42
where the pointers are pointers to non-unit types.
llvm-svn: 125266
could end up removing a different function than we intended because it was
functionally equivalent, then end up with a comparison of a function against
itself in the next round of comparisons (the one in the function set and the
one on the deferred list). To fix this, I introduce a choice in the form of
comparison for ComparableFunctions, either normal or "pointer only" used to
find exact Function*'s in lookups.
Also add some debugging statements.
llvm-svn: 125180
auto-simplifier). This has a big impact on Ada code, but not much else.
Unfortunately the impact is mostly negative! This is due to PR9004 (aka
SCCP failing to resolve conditional branch conditions in the destination
blocks of the branch), in which simple correlated expressions are not
resolved but complicated ones are, so simplifying has a bad effect!
llvm-svn: 124788
overflow (nsw flag), which was disabled because it breaks 254.gap. I have
informed the GAP authors of the mistake in their code, and arranged for the
testsuite to use -fwrapv when compiling this benchmark.
llvm-svn: 124746
This makes the job of the later optzn passes easier, allowing the vast amount of
icmp transforms to chew on it.
We transform 840 switches in gcc.c, leading to a 16k byte shrink of the resulting
binary on i386-linux.
The testcase from README.txt now compiles into
decl %edi
cmpl $3, %edi
sbbl %eax, %eax
andl $1, %eax
ret
llvm-svn: 124724
to do this and more, but would only do it if X/Y had only one use. Spotted as the
most common missed simplification in SPEC by my auto-simplifier, now that it knows
about nuw/nsw/exact flags. This removes a bunch of multiplications from 447.dealII
and 483.xalancbmk. It also removes a lot from tramp3d-v4, which results in much
more inlining.
llvm-svn: 124560
benchmarks, and that it can be simplified to X/Y. (In general you can only
simplify (Z*Y)/Y to Z if the multiplication did not overflow; if Z has the
form "X/Y" then this is the case). This patch implements that transform and
moves some Div logic out of instcombine and into InstructionSimplify.
Unfortunately instcombine gets in the way somewhat, since it likes to change
(X/Y)*Y into X-(X rem Y), so I had to teach instcombine about this too.
Finally, thanks to the NSW/NUW flags, sometimes we know directly that "Z*Y"
does not overflow, because the flag says so, so I added that logic too. This
eliminates a bunch of divisions and subtractions in 447.dealII, and has good
effects on some other benchmarks too. It seems to have quite an effect on
tramp3d-v4 but it's hard to say if it's good or bad because inlining decisions
changed, resulting in massive changes all over.
llvm-svn: 124487
operand being factorized (and erased) could occur several times in Ops,
resulting in freed memory being used when the next occurrence in Ops was
analyzed.
llvm-svn: 124287
optimized code are:
(non-negative number)+(power-of-two) != 0 -> true
and
(x | 1) != 0 -> true
Instcombine knows about the second one of course, but only does it if X|1
has only one use. These fire thousands of times in the testsuite.
llvm-svn: 124183
occurs because instcombine sinks loads and inserts phis. This kicks in
on such apps as 175.vpr, eon, 403.gcc, xalancbmk and a bunch of times in
spec2006 in some app that uses std::deque.
This resolves the last of rdar://7339113.
llvm-svn: 124090
common cases. This triggers a surprising number of times in SPEC2K6
because min/max idioms end up doing this. For example, code from the
STL ends up looking like this to SRoA:
%202 = load i64* %__old_size, align 8, !tbaa !3
%203 = load i64* %__old_size, align 8, !tbaa !3
%204 = load i64* %__n, align 8, !tbaa !3
%205 = icmp ult i64 %203, %204
%storemerge.i = select i1 %205, i64* %__n, i64* %__old_size
%206 = load i64* %storemerge.i, align 8, !tbaa !3
We can now promote both the __n and the __old_size allocas.
This addresses another chunk of rdar://7339113, poor codegen on
stringswitch.
llvm-svn: 124088
that have PHI or select uses of their element pointers. This can often happen
when instcombine sinks two loads into a successor, inserting a phi or select.
With this patch, we can scalarize the alloca, but the pinned elements are not
yet promoted. This is still a win for large aggregates where only one element
is used. This fixes rdar://8904039 and part of rdar://7339113 (poor codegen
on stringswitch).
llvm-svn: 124070
a select. A vector select is pairwise on each element so we'd need a new
condition with the right number of elements to select on. Fixes PR8994.
llvm-svn: 123963
While here, I'd like to complain about how vector is not an aggregate type
according to llvm::Type::isAggregateType(), but they're listed under aggregate
types in the LangRef and zero vectors are stored as ConstantAggregateZero.
llvm-svn: 123956
auto-simplier the transform most missed by early-cse is (zext X) != 0 -> X != 0.
This patch adds this transform and some related logic to InstructionSimplify
and removes some of the logic from instcombine (unfortunately not all because
there are several situations in which instcombine can improve things by making
new instructions, whereas instsimplify is not allowed to do this). At -O2 this
often results in more than 15% more simplifications by early-cse, and results in
hundreds of lines of bitcode being eliminated from the testsuite. I did see some
small negative effects in the testsuite, for example a few additional instructions
in three programs. One program, 483.xalancbmk, got an additional 35 instructions,
which seems to be due to a function getting an additional instruction and then
being inlined all over the place.
llvm-svn: 123911
These were not recommended by my auto-simplifier since they don't fire often enough.
However they do fire from time to time, for example they remove one subtraction from
the final bitcode for 483.xalancbmk.
llvm-svn: 123755
simplification in fully optimized code. It occurs sporadically in the testsuite, and
many times in 403.gcc: the final bitcode has 131 fewer subtractions after this change.
The reason that the multiplies are not eliminated is the same reason that instcombine
did not catch this: they are used by other instructions (instcombine catches this with
a more general transform which in general is only profitable if the operands have only
one use).
llvm-svn: 123754
This fixes the original testcase in PR8927. It also causes a clang
binary built with a patched clang to increase in size by 0.21%.
We can probably get some of the size back by writing a pass that
detects that a global never has its pointer compared and adds
unnamed_addr to it (maybe extend global opt). It is also possible that
there are some other cases clang could add unnamed_addr to.
I will investigate extending globalopt next.
llvm-svn: 123584
then don't try to decimate it into its individual pieces. This will just make a mess of the
IR and is pointless if none of the elements are individually accessed. This was generating
really terrible code for std::bitset (PR8980) because it happens to be lowered by clang
as an {[8 x i8]} structure instead of {i64}.
The testcase now is optimized to:
define i64 @test2(i64 %X) {
br label %L2
L2: ; preds = %0
ret i64 %X
}
before we generated:
define i64 @test2(i64 %X) {
%sroa.store.elt = lshr i64 %X, 56
%1 = trunc i64 %sroa.store.elt to i8
%sroa.store.elt8 = lshr i64 %X, 48
%2 = trunc i64 %sroa.store.elt8 to i8
%sroa.store.elt9 = lshr i64 %X, 40
%3 = trunc i64 %sroa.store.elt9 to i8
%sroa.store.elt10 = lshr i64 %X, 32
%4 = trunc i64 %sroa.store.elt10 to i8
%sroa.store.elt11 = lshr i64 %X, 24
%5 = trunc i64 %sroa.store.elt11 to i8
%sroa.store.elt12 = lshr i64 %X, 16
%6 = trunc i64 %sroa.store.elt12 to i8
%sroa.store.elt13 = lshr i64 %X, 8
%7 = trunc i64 %sroa.store.elt13 to i8
%8 = trunc i64 %X to i8
br label %L2
L2: ; preds = %0
%9 = zext i8 %1 to i64
%10 = shl i64 %9, 56
%11 = zext i8 %2 to i64
%12 = shl i64 %11, 48
%13 = or i64 %12, %10
%14 = zext i8 %3 to i64
%15 = shl i64 %14, 40
%16 = or i64 %15, %13
%17 = zext i8 %4 to i64
%18 = shl i64 %17, 32
%19 = or i64 %18, %16
%20 = zext i8 %5 to i64
%21 = shl i64 %20, 24
%22 = or i64 %21, %19
%23 = zext i8 %6 to i64
%24 = shl i64 %23, 16
%25 = or i64 %24, %22
%26 = zext i8 %7 to i64
%27 = shl i64 %26, 8
%28 = or i64 %27, %25
%29 = zext i8 %8 to i64
%30 = or i64 %29, %28
ret i64 %30
}
In this case, instcombine was able to eliminate the nonsense, but in PR8980 enough
PHIs are in play that instcombine backs off. It's better to not generate this stuff
in the first place.
llvm-svn: 123571
multiple uses. In some cases, all the uses are the same operation,
so instcombine can go ahead and promote the phi. In the testcase
this pushes an add out of the loop.
llvm-svn: 123568
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
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
simplification present in fully optimized code (I think instcombine fails to
transform some of these when "X-Y" has more than one use). Fires here and
there all over the test-suite, for example it eliminates 8 subtractions in
the final IR for 445.gobmk, 2 subs in 447.dealII, 2 in paq8p etc.
llvm-svn: 123442
threading of shifts over selects and phis while there. This fires here and
there in the testsuite, to not much effect. For example when compiling spirit
it fires 5 times, during early-cse, resulting in 6 more cse simplifications,
and 3 more terminators being folded by jump threading, but the final bitcode
doesn't change in any interesting way: other optimizations would have caught
the opportunity anyway, only later.
llvm-svn: 123441
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
This is a minor extension of SROA to handle a special case that is
important for some ARM NEON operations. Some of the NEON intrinsics
return multiple values, which are handled as struct types containing
multiple elements of the same vector type. The corresponding return
types declared in the arm_neon.h header have equivalent arrays. We
need SROA to recognize that it can split up those arrays and structs
into separate vectors, even though they are not always accessed with
the same type. SROA already handles loads and stores of an entire
alloca by using insertvalue/extractvalue to access the individual
pieces, and that code works the same regardless of whether the type
is a struct or an array. So, all that needs to be done is to check
for compatible arrays and homogeneous structs.
llvm-svn: 123381
SROA only split up structs and arrays one level at a time, so padding can
only cause trouble if it is located in between the struct or array elements.
llvm-svn: 123380
is "X != 0 -> X" when X is a boolean. This occurs a lot because of the way
llvm-gcc converts gcc's conditional expressions. Add this, and a few other
similar transforms for completeness.
llvm-svn: 123372
point values to their integer representation through the SSE intrinsic
calls. This is the last part of a README.txt entry for which I have real
world examples.
llvm-svn: 123206
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
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
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
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
ret i64 ptrtoint (i8* getelementptr ([1000 x i8]* @X, i64 1, i64 sub (i64 0, i64 ptrtoint ([1000 x i8]* @X to i64))) to i64)
to "ret i64 1000". This allows us to correctly compute the trip count
on a loop in PR8883, which occurs with std::fill on a char array. This
allows us to transform it into a memset with a constant size.
llvm-svn: 122950
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
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
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
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
numbering, in which it considers (for example) "%a = add i32 %x, %y" and
"%b = add i32 %x, %y" to be equal because the operands are equal and the
result of the instructions only depends on the values of the operands.
This has almost no effect (it removes 4 instructions from gcc-as-one-file),
and perhaps slows down compilation: I measured a 0.4% slowdown on the large
gcc-as-one-file testcase, but it wasn't statistically significant.
llvm-svn: 122654
the original instruction, half the cases were missed (making it not
wrong but suboptimal). Also correct a typo (A <-> B) in the second
chunk.
llvm-svn: 122414
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
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
a couple of existing transforms. This fires surprisingly often, for
example when compiling gcc "(X+(-1))+1->X" fires quite a lot as well
as various "and" simplifications (usually with a phi node operand).
Most of the time this doesn't make a real difference since the same
thing would have been done elsewhere anyway, eg: by instcombine, but
there are a few places where this results in simplifications that we
were not doing before.
llvm-svn: 122326
(they had just been forgotten before). Adding Xor causes "main" in the
existing testcase 2010-11-01-lshr-mask.ll to be hugely more simplified.
llvm-svn: 122245
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
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
