1. Fixed ARM pc adjustment.
2. Fixed dynamic-no-pic codegen
3. CSE of pc-relative load of global addresses.
It's now enabled by default for Darwin.
llvm-svn: 123991
qadd and qdadd uses "rd, rm, rn", the same applies to the 'sub' variants. This
is described in ARM manuals and matches the encoding used by the gnu assembler.
llvm-svn: 123975
DAG. Disable using "-disable-sched-cycles".
For ARM, this enables a framework for modeling the cpu pipeline and
counting stalls. It also activates several heuristics to drive
scheduling based on the model. Scheduling is inherently imprecise at
this stage, and until spilling is improved it may defeat attempts to
schedule. However, this framework provides greater control over
tuning codegen.
Although the flag is not target-specific, it should have very little
affect on the default scheduler used by x86. The only two changes that
affect x86 are:
- scheduling a high-latency operation bumps the current cycle so independent
operations can have their latency covered. i.e. two independent 4
cycle operations can produce results in 4 cycles, not 8 cycles.
- Two operations with equal register pressure impact and no
latency-based stalls on their uses will be prioritized by depth before height
(height is irrelevant if no stalls occur in the schedule below this point).
llvm-svn: 123971
flags. They are still not enable in this revision.
Added TargetInstrInfo::isZeroCost() to fix a fundamental problem with
the scheduler's model of operand latency in the selection DAG.
Generalized unit tests to work with sched-cycles.
llvm-svn: 123969
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
value, the "add pc" must be CSE'ed at the same time. We could follow the same
approach as T2 by adding pseudo instructions that combine the ldr + "add pc".
But the better approach is to use movw + movt (which I will enable soon), so
I'll leave this as a TODO.
llvm-svn: 123949
in cdp/cdp2 instructions. Also increase the hack with cdp/cdp2 instructions.
- Fix the encoding of cdp/cdp2 instructions for ARM (no thumb and thumb2 yet) and add testcases for t
hem.
llvm-svn: 123927
The value mapping gets confused about which original values have multiple new
definitions so they may need phi insertions.
This could probably be simplified by letting enterIntvBefore() take a live range
to be added following the instruction. As long as the range stays inside the
same basic block, value mapping shouldn't be a problem.
llvm-svn: 123926
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
to add/sub by doing the normal operation and then checking for overflow
afterwards. This generally relies on the DAG handling the later invalid
operations as well.
Fixes the 64-bit part of rdar://8622122 and rdar://8774702.
llvm-svn: 123908
TargetInstrInfo:
Change produceSameValue() to take MachineRegisterInfo as an optional argument.
When in SSA form, targets can use it to make more aggressive equality analysis.
Machine LICM:
1. Eliminate isLoadFromConstantMemory, use MI.isInvariantLoad instead.
2. Fix a bug which prevent CSE of instructions which are not re-materializable.
3. Use improved form of produceSameValue.
ARM:
1. Teach ARM produceSameValue to look pass some PIC labels.
2. Look for operands from different loads of different constant pool entries
which have same values.
3. Re-implement PIC GA materialization using movw + movt. Combine the pair with
a "add pc" or "ldr [pc]" to form pseudo instructions. This makes it possible
to re-materialize the instruction, allow machine LICM to hoist the set of
instructions out of the loop and make it possible to CSE them. It's a bit
hacky, but it significantly improve code quality.
4. Some minor bug fixes as well.
With the fixes, using movw + movt to materialize GAs significantly outperform the
load from constantpool method. 186.crafty and 255.vortex improved > 20%, 254.gap
and 176.gcc ~10%.
llvm-svn: 123905
Added a check for already live regs before claiming HighRegPressure.
Fixed a few cases of checking the wrong number of successors.
Added some tracing until these heuristics are better understood.
llvm-svn: 123892
with an invalid type then split the result and perform the overflow check
normally.
Fixes the 32-bit parts of rdar://8622122 and rdar://8774702.
llvm-svn: 123864
interval after an instruction. The leaveIntvAfter() method only adds liveness
from the instruction's boundary index to the inserted copy.
Ideally, SplitKit should be smarter about this, perhaps by combining useIntv()
and leaveIntvAfter() into one method that guarantees continuity.
llvm-svn: 123858
Region splitting includes loop splitting as a subset, and it is more generic.
The splitting heuristics for variables that are live in more than one block are
now:
1. Try to create a region that covers multiple basic blocks.
2. Try to create a new live range for each block with multiple uses.
3. Spill.
Steps 2 and 3 are similar to what the standard spiller is doing.
llvm-svn: 123853
by indvars through the scev expander.
trunc(add x, y) --> add(trunc x, y). Currently SCEV largely folds the other way
which is probably wrong, but preserved to minimize churn. Instcombine doesn't
do this fold either, demonstrating a missed optz'n opportunity on code doing
add+trunc+add.
llvm-svn: 123838
of the floating point types less than 64-bits. It's somewhat of a temporary
hack but forces more accurate modeling of register pressure and results
in fewer spills.
llvm-svn: 123811
Unfortunately, while this is the "right" thing to do, it breaks some ARM
asm parsing tests because MemMode5 and ThumbMemModeReg are ambiguous. This
is tricky to resolve since neither is a subset of the other.
XFAIL the test for now. The old way was broken in other ways, just ways
we didn't happen to be testing, and our ARM asm parsing is going to require
significant revisiting at a later point anyways.
llvm-svn: 123786
are pointing to the same object, one pointer is accessing the entire
object, and the other is access has a non-zero size. This prevents
TBAA from kicking in and saying NoAlias in such cases.
llvm-svn: 123775
Analyze the live range's behavior entering and leaving basic blocks. Compute an
interference pattern for each allocation candidate, and use SpillPlacement to
find an optimal region where that register can be live.
This code is still not enabled.
llvm-svn: 123774
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
without whatever this was trying to do. When/if someone has the time to do some empirical
evaluations, it might be worth it to figure out what this code was trying to do and see if
it's worth resurrecting/fixing.
llvm-svn: 123684
checks enabled:
1) Use '<' to compare integers in a comparison function rather than '<='.
2) Use the uniqued set DefBlocks rather than Info.DefiningBlocks to initialize
the priority queue.
The speedup of scalarrepl on test-suite + SPEC2000 + SPEC2006 is a bit less, at
just under 16% rather than 17%.
llvm-svn: 123662
This shaves off 4 popcounts from the hacked 186.crafty source.
This is enabled even when a native popcount instruction is available. The
combined code is one operation longer but it should be faster nevertheless.
llvm-svn: 123621
movw r0, :lower16:(L_foo$non_lazy_ptr-(LPC0_0+4))
movt r0, :upper16:(L_foo$non_lazy_ptr-(LPC0_0+4))
LPC0_0:
add r0, pc, r0
It's not yet enabled by default as some tests are failing. I suspect bugs in
down stream tools.
llvm-svn: 123619
eliminating a potentially quadratic data structure, this also gives a 17%
speedup when running -scalarrepl on test-suite + SPEC2000 + SPEC2006. My initial
experiment gave a greater speedup around 25%, but I moved the dominator tree
level computation from dominator tree construction to PromoteMemToReg.
Since this approach to computing IDFs has a much lower overhead than the old
code using precomputed DFs, it is worth looking at using this new code for the
second scalarrepl pass as well.
llvm-svn: 123609
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
into and/shift would cause nodes to move around and a dangling pointer
to happen. The code tried to avoid this with a HandleSDNode, but
got the details wrong.
llvm-svn: 123578
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
http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
In a silly microbenchmark on a 65 nm core2 this is 1.5x faster than the old
code in 32 bit mode and about 2x faster in 64 bit mode. It's also a lot shorter,
especially when counting 64 bit population on a 32 bit target.
I hope this is fast enough to replace Kernighan-style counting loops even when
the input is rather sparse.
llvm-svn: 123547
half a million non-local queries, each of which would otherwise have triggered a
linear scan over a basic block.
Also fix a fixme for memory intrinsics which dereference pointers. With this,
we prove that a pointer is non-null because it was dereferenced by an intrinsic
112 times in llvm-test.
llvm-svn: 123533