foldMemoryOperand how to "fold" them, by converting them into constant-pool
loads. When they aren't folded, they use xorps/cmpeqd, but for example when
register pressure is high, they may now be folded as memory operands, which
reduces register pressure.
Also, mark V_SET0 isAsCheapAsAMove so that two-address-elimination will
remat it instead of copying zeros around (V_SETALLONES was already marked).
llvm-svn: 60461
delegates to the regular x86-32 convention which handles byval, but only
after it handles a few cases, and it's necessary to handle byval before
handling those cases. This fixes PR3122 (and rdar://6400815), llvm-gcc
miscompiling LLVM.
llvm-svn: 60453
1. ppcf128 select is expanded to f64 select's.
2. f64 select operand 0 is an i1 truncate, it's promoted to i32 zero_extend.
3. f64 select is updated. It's changed back to a "NewNode" and being re-analyzed.
4. f64 select operands are being processed. Operand 0 is a "NewNode". It's being expunged out of ReplacedValues map.
5. ExpungeNode tries to remap f64 select and notice it's a "NewNode" and assert.
Duncan, please take a look. Thanks.
llvm-svn: 60443
- Incorporate Tilmann Scheller's ISD::TRUNCATE custom lowering patch
- Update SPU calling convention info, even if it's not used yet (but can be
at some point or another)
- Ensure that any-extended f32 loads are custom lowered, especially when
they're promoted for use in printf.
llvm-svn: 60438
straight-forward implementation. This does not require any extra
alias analysis queries beyond what we already do for non-local loads.
Some programs really really like load PRE. For example, SPASS triggers
this ~1000 times, ~300 times in 255.vortex, and ~1500 times on 403.gcc.
The biggest limitation to the implementation is that it does not split
critical edges. This is a huge killer on many programs and should be
addressed after the initial patch is enabled by default.
The implementation of this should incidentally speed up rejection of
non-local loads because it avoids creating the repl densemap in cases
when it won't be used for fully redundant loads.
This is currently disabled by default.
Before I turn this on, I need to fix a couple of miscompilations in
the testsuite, look at compile time performance numbers, and look at
perf impact. This is pretty close to ready though.
llvm-svn: 60408
- LowerXADDO lowers [SU]ADDO into an ADD with an implicit EFLAGS define. The
EFLAGS are fed into a SETCC node which has the conditional COND_O or COND_C,
depending on the type of ADDO requested.
- LowerBRCOND now recognizes if it's coming from a SETCC node with COND_O or
COND_C set.
llvm-svn: 60388
figuring out the base of the IV. This produces better
code in the example. (Addresses use (IV) instead of
(BASE,IV) - a significant improvement on low-register
machines like x86).
llvm-svn: 60374
- Fix v2[if]64 vector insertion code before IBM files a bug report.
- Ensure that zero (0) offsets relative to $sp don't trip an assert
(add $sp, 0 gets legalized to $sp alone, tripping an assert)
- Shuffle masks passed to SPUISD::SHUFB are now v16i8 or v4i32
llvm-svn: 60358
multiplies.
Some more cleverness would be nice, though. It would be nice if we
could do this transformation on illegal types. Also, we would
prefer a narrower constant when possible so that we can use a narrower
multiply, which can be cheaper.
llvm-svn: 60283
overflowed on negation. This commit checks to make sure that neithe C nor X
overflows. This requires that the RHS of X (a subtract instruction) be a
constant integer.
llvm-svn: 60275
properly updates the reverse dependency map when it installs updated
dependencies for instructions that depend on the removed instruction.
llvm-svn: 60222
1. Make it fold blocks separated by an unconditional branch. This enables
jump threading to see a broader scope.
2. Make jump threading able to eliminate locally redundant loads when they
feed the branch condition of a block. This frequently occurs due to
reg2mem running.
3. Make jump threading able to eliminate *partially redundant* loads when
they feed the branch condition of a block. This is common in code with
lots of loads and stores like C++ code and 255.vortex.
This implements thread-loads.ll and rdar://6402033.
Per the fixme's, several pieces of this should be moved into Transforms/Utils.
llvm-svn: 60148
performance in most cases on the Grawp tester, but does speed some
things up (like shootout/hash by 15%). This also doesn't impact
compile time in a noticable way on the Grawp tester.
It also, of course, gets the testcase it was designed for right :)
llvm-svn: 60120
-enable-smarter-addr-folding to llc) that gives CGP a better
cost model for when to sink computations into addressing modes.
The basic observation is that sinking increases register
pressure when part of the addr computation has to be available
for other reasons, such as having a use that is a non-memory
operation. In cases where it works, it can substantially reduce
register pressure.
This code is currently an overall win on 403.gcc and 255.vortex
(the two things I've been looking at), but there are several
things I want to do before enabling it by default:
1. This isn't doing any caching of results, so it is much slower
than it could be. It currently slows down release-asserts llc
by 1.7% on 176.gcc: 27.12s -> 27.60s.
2. This doesn't think about inline asm memory operands yet.
3. The cost model botches the case when the needed value is live
across the computation for other reasons.
I'll continue poking at this, and eventually turn it on as llcbeta.
llvm-svn: 60074
optimize addressing modes. This allows us to optimize things like isel-sink2.ll
into:
movl 4(%esp), %eax
cmpb $0, 4(%eax)
jne LBB1_2 ## F
LBB1_1: ## TB
movl $4, %eax
ret
LBB1_2: ## F
movzbl 7(%eax), %eax
ret
instead of:
_test:
movl 4(%esp), %eax
cmpb $0, 4(%eax)
leal 4(%eax), %eax
jne LBB1_2 ## F
LBB1_1: ## TB
movl $4, %eax
ret
LBB1_2: ## F
movzbl 3(%eax), %eax
ret
This shrinks (e.g.) 403.gcc from 1133510 to 1128345 lines of .s.
Note that the 2008-10-16-SpillerBug.ll testcase is dubious at best, I doubt
it is really testing what it thinks it is.
llvm-svn: 60068
(a) Remove conditionally removed code in SelectXAddr. Basically, hope for the
best that the A-form and D-form address predicates catch everything before
the code decides to emit a X-form address.
(b) Expand vector store test cases to include the usual suspects.
llvm-svn: 60034
introduce any new spilling; it just uses unused registers.
Refactor the SUnit topological sort code out of the RRList scheduler and
make use of it to help with the post-pass scheduler.
llvm-svn: 59999
(a) Slight rethink on i64 zero/sign/any extend code - use a shuffle to
directly zero-extend i32 to i64, but use rotates and shifts for
sign extension. Also ensure unified register consistency.
(b) Add new test harness for i64 operations: i64ops.ll
llvm-svn: 59970
(a) Improve the extract element code: there's no need to do gymnastics with
rotates into the preferred slot if a shuffle will do the same thing.
(b) Rename a couple of SPUISD pseudo-instructions for readability and better
semantic correspondence.
(c) Fix i64 sign/any/zero extension lowering.
llvm-svn: 59965
indicate functions that allocate, such as operator new, or list::insert. The
actual definition is slightly less strict (for now).
No changes to the bitcode reader/writer, asm printer or verifier were needed.
llvm-svn: 59934
- When scavenging a register, in addition to the spill, insert a restore before the first use.
- Abort if client is looking to scavenge a register even when a previously scavenged register is still live.
llvm-svn: 59697
Use it to safely handle less-than-or-equals-to exit conditions in loops. These
also occur when the loop exit branch is exit on true because SCEV inverses the
icmp predicate.
Use it again to handle non-zero strides, but only with an unsigned comparison
in the exit condition.
llvm-svn: 59528
If this patch causes a performance regression for anyone, please let me know,
and it can be fixed in a different way with much more effort.
llvm-svn: 59384
to carry a SmallVector of flagged nodes, just calculate the flagged nodes
dynamically when they are needed.
The local-liveness change is due to a trivial scheduling change where
the scheduler arbitrary decision differently.
llvm-svn: 59273
inform the optimizers that the result must be zero/
sign extended from the smaller type. For example,
if a fp to unsigned i16 is promoted to fp to i32,
then we are allowed to assume that the extra 16 bits
are zero (because the result of fp to i16 is undefined
if the result does not fit in an i16). This is
quite aggressive, but should help the optimizers
produce better code. This requires correcting a
test which thought that fp_to_uint is some kind
of truncation, which it is not: in the testcase
(which does fp to i1), either the fp value converts
to 0 or 1 or the result is undefined, which is
quite different to truncation.
llvm-svn: 58991
original code was matching like this:
if (match(A, m_Not(m_Value(B))))
B was already matched as a 'select' instruction. However, this isn't matching
what we think it's matching. It would match B as a 'Value', so basically
anything would match to it. In this case, a Constant matched. B was replaced
with a constant representation. And then the wrong value would be used in the
SelectInst::Create statement, causing a crash.
After thinking on this for a moment, and after Nick L. told me how the pattern
matching stuff was supposed to work, the solution was to match NOT an m_Value,
but an m_Select.
llvm-svn: 58946
to generate signed ICMP instructions to replace the FCMP. This would violate
the following:
define i1 @test1(i32 %val) {
%1 = uitofp i32 %val to double
%2 = fcmp ole double %1, 0.000000e+00
ret i1 %2
}
would be transformed into:
define i1 @test1(i32 %val) {
%1 = icmp slt i33 %val, 1
ret i1 %1
}
which is obviously wrong. This patch modifes InstCombiner::FoldFCmp_IntToFP_Cst
to handle when the LHS comes from UIToFP.
llvm-svn: 58929
is noticeably worse than previous PPC-specific code.
Since the latter was also wrong in some cases and
correctness is more important than efficiency, I'm
disabling this test temporarily while I fix it.
llvm-svn: 58876
This allows SCEV users to effectively calculate trip count.
LSR later on transforms back integer IVs to floating point IVs
later on to avoid int-to-float casts inside the loop.
llvm-svn: 58625
bits, use a union of a SimpleValueType enum and a regular Type*.
This increases the size of MVT on 64-bit hosts from 32 bits to 64 bits.
In most cases, this doesn't add significant overhead. There are places
in codegen that use arrays of MVTs, so these are now larger, but
they're small in common cases.
This eliminates restrictions on the size of integer types and vector
types that can be represented in codegen. As the included testcase
demonstrates, it's now possible to codegen very large add operations.
There are still some complications with using very large types. PR2880
is still open so they can't be used as return values on normal targets,
there are no libcalls defined for very large integers so operations
like multiply and divide aren't supported.
This also introduces a minimal tablgen Type library, capable of
handling IntegerType and VectorType. This will allow parts of
TableGen that don't depend on using SimpleValueType values to handle
arbitrary integer and vector types.
llvm-svn: 58623
* merge two weak functions by making them both alias a third non-weak fn
* don't reimplement CallSite::hasArgument
* whitelist the safe linkage types
llvm-svn: 58568
This triggers only 60 times in llvm-test (look at .llvm.bc, not .linked.rbc)
and so it probably wont be turned on by default. Also, may of those are likely
to go away when PR2973 is fixed.
llvm-svn: 58557
function.
- This explicitly models the costs for functions which should
"always" or "never" be inlined. This fixes bugs where such costs
were not previously respected.
llvm-svn: 58450
so that va_start/va_arg/et.al. will walk arguments correctly for Cell SPU.
N.B.: Because neither clang nor llvm-gcc-4.2 can be built for CellSPU, this is
still unexorcised code.
llvm-svn: 58415
ppcf128 to i32 conversion and expand it into a code
sequence like in LegalizeDAG. This needs custom
ppc lowering of FP_ROUND_INREG, so turn that on and
make it work with LegalizeTypes. Probably PPC should
simply custom lower the original conversion.
llvm-svn: 58329
a memset using 16-byte XMM stores, but where the stack realignment code
didn't work. Until it does (PR2962) disable use of xmm regs in memcpy
and memset formation for linux and other targets with insufficiently
aligned stacks.
This is part of PR2888
llvm-svn: 58317
LHS is a foldable load, then LHS and RHS are swapped
and SetCCOpcode is changed to SETUGT. But the later
code is expecting operands to be the wrong way round
for SETUGT, but they are not in this case, resulting
in an inverted compare. The solution is to move the
load normalization before the correction for SETUGT.
This bug was tickled by LegalizeTypes which happened
to legalize the testcase slightly differently to
LegalizeDAG.
llvm-svn: 58092
in the 32-bit signed offset field of addresses. Even though this
may be intended, some linkers refuse to relocate code where the
relocated address computation overflows.
Also, fix the sign-extension of constant offsets to use the
actual pointer size, rather than the size of the GlobalAddress
node, which may be different, for example on x86-64 where MVT::i32
is used when the address is being fit into the 32-bit displacement
field.
llvm-svn: 57885
Where previously LLVM might emit code like this:
ucomisd %xmm1, %xmm0
setne %al
setp %cl
orb %al, %cl
jne .LBB4_2
it now emits this:
ucomisd %xmm1, %xmm0
jne .LBB4_2
jp .LBB4_2
It has fewer instructions and uses fewer registers, but it does
have more branches. And in the case that this code is followed by
a non-fallthrough edge, it may be followed by a jmp instruction,
resulting in three branch instructions in sequence. Some effort
is made to avoid this situation.
To achieve this, X86ISelLowering.cpp now recognizes FCMP_OEQ and
FCMP_UNE in lowered form, and replace them with code that emits
two branches, except in the case where it would require converting
a fall-through edge to an explicit branch.
Also, X86InstrInfo.cpp's branch analysis and transform code now
knows now to handle blocks with multiple conditional branches. It
uses loops instead of having fixed checks for up to two
instructions. It can now analyze and transform code generated
from FCMP_OEQ and FCMP_UNE.
llvm-svn: 57873
the copy instruction from the instruction list before asking the
target to create the new instruction. This gets the old instruction
out of the way so that it doesn't interfere with the target's
rematerialization code. In the case of x86, this helps it find
more cases where EFLAGS is not live.
Also, in the X86InstrInfo.cpp, teach isSafeToClobberEFLAGS to check
to see if it reached the end of the block after scanning each
instruction, instead of just before. This lets it notice when the
end of the block is only two instructions away, without doing any
additional scanning.
These changes allow rematerialization to clobber EFLAGS in more
cases, for example using xor instead of mov to set the return value
to zero in the included testcase.
llvm-svn: 57872
for strange asm conditions earlier. In this case, we have a
double being passed in an integer reg class. Convert to like
sized integer register so that we allocate the right number
for the class (two i32's for the f64 in this case).
llvm-svn: 57862
the previous patch this one actually passes make check.
"Fix PR2356 on PowerPC: if we have an input and output that are tied together
that have different sizes (e.g. i32 and i64) make sure to reserve registers for
the bigger operand."
llvm-svn: 57771
and add a TargetLowering hook for it to use to determine when this
is legal (i.e. not in PIC mode, etc.)
This allows instruction selection to emit folded constant offsets
in more cases, such as the included testcase, eliminating the need
for explicit arithmetic instructions.
This eliminates the need for the C++ code in X86ISelDAGToDAG.cpp
that attempted to achieve the same effect, but wasn't as effective.
Also, fix handling of offsets in GlobalAddressSDNodes in several
places, including changing GlobalAddressSDNode's offset from
int to int64_t.
The Mips, Alpha, Sparc, and CellSPU targets appear to be
unaware of GlobalAddress offsets currently, so set the hook to
false on those targets.
llvm-svn: 57748
have an unreachable block in a function. This was triggering the assert. This is
a horrid hack to cover this up.
Oh! for a good debug info architecture!
llvm-svn: 57714
in 32-bit mode instead of assigning a register pair. This has nothing to
do with PR2356, but I happened to notice it while working on it.
llvm-svn: 57704
use a SUB instruction instead of an ADD, because -128 can be
encoded in an 8-bit signed immediate field, while +128 can't be.
This avoids the need for a 32-bit immediate field in this case.
A similar optimization applies to 64-bit adds with 0x80000000,
with the 32-bit signed immediate field.
To support this, teach tablegen how to handle 64-bit constants.
llvm-svn: 57663
shift counts, and patterns that match dynamic shift counts
when the subtract is obscured by a truncate node.
Add DAGCombiner support for recognizing rotate patterns
when the shift counts are defined by truncate nodes.
Fix and simplify the code for commuting shld and shrd
instructions to work even when the given instruction doesn't
have a parent, and when the caller needs a new instruction.
These changes allow LLVM to use the shld, shrd, rol, and ror
instructions on x86 to replace equivalent code using two
shifts and an or in many more cases.
llvm-svn: 57662
