1. Be careful with cse "cheap" expressions. e.g. constant materialization. Only cse them when the common expression is local or in a direct predecessor. We don't want cse of cheap instruction causing other expressions to be spilled.
2. Watch out for the case where the expression doesn't itself uses a virtual register. e.g. lea of frame object. If the common expression itself is used by copies (common for passing addresses to function calls), don't perform the cse. Since these expressions do not use a register, it creates a live range but doesn't close any, we want to be very careful with increasing register pressure.
Note these are heuristics so machine cse doesn't make register allocator unhappy. Once we have proper live range splitting and re-materialization support in place, these should be evaluated again.
Now machine cse is almost always a win on llvm nightly tests on x86 and x86_64.
llvm-svn: 98121
need to be MCized, but the last debug info thing are LEB and
cygwin specific (which the MC api doesn't support yet) and
one specific form of EmitReference which I'll tackle next.
llvm-svn: 98116
method. With this, comments should end up on the same lines as the .byte
directives (for example) and we now get no output with:
$ llc CodeGen/X86/2009-02-12-DebugInfoVLA.ll -o - -filetype=null -asm-verbose
woot.
llvm-svn: 98105
is preparatory to having PEI's scavenged frame index value reuse logic
properly distinguish types of frame values (e.g., whether the value is
stack-pointer relative or frame-pointer relative).
No functionality change.
llvm-svn: 98086
coalescer) handle sub-register classes.
- Add heuristics to avoid non-profitable cse. Given the current lack of live
range splitting, avoid cse when an expression has PHI use and the would be
new use is in a BB where the expression wasn't already being used.
llvm-svn: 98043
physreg becomes ridiculously high.
std::upper_bound may be log(N), but for sufficiently large live intervals, it
becomes log(N)*cachemiss = a long long time.
This patch improves coalescer time by 4500x for a function with 20000
function calls. The generated code is different, but not significantly worse -
the allocator hints are almost as good as physreg coalescing anyway.
llvm-svn: 98023
available, the only thing this affects is that we produce
.set in one case we didn't before, which shouldn't harm
anything. Make EmitSectionOffset call EmitDifference
instead of duplicating it.
llvm-svn: 98005
is a workaround for <rdar://problem/7672401/> (which I filed).
This let's us build Wine on Darwin, and it gets the Qt build there a little bit
further (so Doug says).
llvm-svn: 97845
CALL ... %RAX<imp-def>
... [not using %RAX]
%EAX = ..., %RAX<imp-use, kill>
RET %EAX<imp-use,kill>
Now we do this:
CALL ... %RAX<imp-def, dead>
... [not using %RAX]
%EAX = ...
RET %EAX<imp-use,kill>
By not artificially keeping %RAX alive, we lower register pressure a bit.
The correct number of instructions for 2008-08-05-SpillerBug.ll is obviously
55, anybody can see that. Sheesh.
llvm-svn: 97838
node which has a flag. That flag in turn was used by an
already-selected adde which turned into an ADC32ri8 which
used a selected load which was chained to the load we
folded. This flag use caused us to form a cycle. Fix
this by not ignoring chains in IsLegalToFold even in
cases where the isel thinks it can.
llvm-svn: 97791
as nounwind are marked with a -1 call-site value. This is necessary to, for
example, correctly process exceptions thrown from within an "unexpected"
execption handler (see SingleSource/Regression/C++/EH/expection_spec_test.cpp).
llvm-svn: 97757
as the very last thing before node emission. This should
dramatically reduce the number of times we do 'MatchAddress'
on X86, speeding up compile time. This also improves comments
in the tables and shrinks the table a bit, now down to
80506 bytes for x86.
llvm-svn: 97703
CSE and recursive RAUW calls delete a node from the use list,
invalidating the use list iterator. There's currently no known
way to reproduce this in an unmodified LLVM, however there's no
fundamental reason why a SelectionDAG couldn't be formed which
would trigger this case.
llvm-svn: 97665
entry we're about to process is obviously going to fail, don't
bother pushing a scope only to have it immediately be popped.
This avoids a lot of scope stack traffic in common cases.
Unfortunately, this requires duplicating some of the predicate
dispatch. To avoid duplicating the actual logic I pulled each
predicate out to its own static function which gets used in
both places.
llvm-svn: 97651
SwitchOpcodeMatcher) and have DAGISelMatcherOpt form it. This
speeds up selection, particularly for X86 which has lots of
variants of instructions with only type differences.
llvm-svn: 97645
- Eliminate TargetInstrInfo::isIdentical and replace it with produceSameValue. In the default case, produceSameValue just checks whether two machine instructions are identical (except for virtual register defs). But targets may override it to check for unusual cases (e.g. ARM pic loads from constant pools).
llvm-svn: 97628
long test(long x) { return (x & 123124) | 3; }
Currently compiles to:
_test:
orl $3, %edi
movq %rdi, %rax
andq $123127, %rax
ret
This is because instruction and DAG combiners canonicalize
(or (and x, C), D) -> (and (or, D), (C | D))
However, this is only profitable if (C & D) != 0. It gets in the way of the
3-addressification because the input bits are known to be zero.
llvm-svn: 97616
CopyToReg/CopyFromReg/INLINEASM. These are annoying because
they have the same opcode before an after isel. Fix this by
setting their NodeID to -1 to indicate that they are selected,
just like what automatically happens when selecting things that
end up being machine nodes.
With that done, give IsLegalToFold a new flag that causes it to
ignore chains. This lets the HandleMergeInputChains routine be
the one place that validates chains after a match is successful,
enabling the new hotness in chain processing. This smarter
chain processing eliminates the need for "PreprocessRMW" in the
X86 and MSP430 backends and enables MSP to start matching it's
multiple mem operand instructions more aggressively.
I currently #if out the dead code in the X86 backend and MSP
backend, I'll remove it for real in a follow-on patch.
The testcase changes are:
test/CodeGen/X86/sse3.ll: we generate better code
test/CodeGen/X86/store_op_load_fold2.ll: PreprocessRMW was
miscompiling this before, we now generate correct code
Convert it to filecheck while I'm at it.
test/CodeGen/MSP430/Inst16mm.ll: Add a testcase for mem/mem
folding to make anton happy. :)
llvm-svn: 97596
was that we weren't properly handling the case when interior
nodes of a matched pattern become dead after updating chain
and flag uses. Now we handle this explicitly in
UpdateChainsAndFlags.
llvm-svn: 97561
DoInstructionSelection. Inline "SelectRoot" into it from DAGISelHeader.
Sink some other stuff out of DAGISelHeader into SDISel.
Eliminate the various 'Indent' stuff from various targets, which dates
to when isel was recursive.
17 files changed, 114 insertions(+), 430 deletions(-)
llvm-svn: 97555
stuff now that we don't care about emulating the old broken
behavior of the old isel. This eliminates the
'CheckChainCompatible' check (along with IsChainCompatible) which
did an incorrect and inefficient scan *up* the chain nodes which
happened as the pattern was being formed and does the validation
at the end in HandleMergeInputChains when it forms a structural
pattern. This scans "down" the graph, which means that it is
quickly bounded by nodes already selected. This also handles
token factors that get "trapped" in the dag.
Removing the CheckChainCompatible nodes also shrinks the
generated tables by about 6K for X86 (down to 83K).
There are two pieces remaining before I can nuke PreprocessRMW:
1. I xfailed a test because we're now producing worse code in a
case that has nothing to do with the change: it turns out that
our use of MorphNodeTo will leave dead nodes in the graph
which (depending on how the graph is walked) end up causing
bogus uses of chains and blocking matches. This is really
bad for other reasons, so I'll fix this in a follow-up patch.
2. CheckFoldableChainNode needs to be improved to handle the TF.
llvm-svn: 97539
ComplexPattern at the root be generated multiple times, once
for each opcode they are part of. This encourages factoring
because the opcode checks get treated just like everything
else in the matcher.
llvm-svn: 97439
to a scope where every child starts with a CheckOpcode, but
executes more efficiently. Enhance DAGISelMatcherOpt to
form it.
This also fixes a bug in CheckOpcode: apparently the SDNodeInfo
objects are not pointer comparable, we have to compare the
enum name.
llvm-svn: 97438
defs or uses. The regular def and use checking below covers them, and
can be more precise. It's safe to hoist an instruction with a dead
implicit def if the register isn't live into the loop header.
llvm-svn: 97352
for alignment into the LSDA. If the TType base offset is emitted, then put the
padding there. Otherwise, put it in the call site table length. There will be no
conflict between the two sites when placing the padding in one place.
llvm-svn: 97277
The PowerPC floating point registers can represent both f32 and f64 via the
two register classes F4RC and F8RC. F8RC is considered a subclass of F4RC to
allow cross-class coalescing. This coalescing only affects whether registers
are spilled as f32 or f64.
Spill slots must be accessed with load/store instructions corresponding to the
class of the spilled register. PPCInstrInfo::foldMemoryOperandImpl was looking
at the instruction opcode which is wrong.
X86 has similar floating point register classes, but doesn't try to fold
memory operands, so there is no problem there.
llvm-svn: 97262
the alignment requirement, if it no longer makes the TType base offset overflow
into extra bytes, then we need to pad to those bytes ourselves.
llvm-svn: 97196
will eliminate the need for padding in the "Call site table length". E.g., if
we have this:
GCC_except_table1:
Lexception1:
.byte 0xff ## @LPStart Encoding = omit
.byte 0x9b ## @TType Encoding = indirect pcrel sdata4
.byte 0x7f ## @TType base offset
.byte 0x03 ## Call site Encoding = udata4
.byte 0x89 ## Call site table length
with padding of 1. We want to emit the padding like this:
GCC_except_table1:
Lexception1:
.byte 0xff ## @LPStart Encoding = omit
.byte 0x9b ## @TType Encoding = indirect pcrel sdata4
.byte 0xff ## @TType base offset
.space 1,0 ## Padding
.byte 0x03 ## Call site Encoding = udata4
.byte 0x89 ## Call site table length
and not with padding on the "Call site table length" entry.
llvm-svn: 97183
terms of store and load, which means bitcasting between scalar
integer and vector has endian-specific results, which undermines
this whole approach.
llvm-svn: 97137
GCC_except_table label but before the Lexception, which the FDE references.
This causes problems as the FDE does not point to the start of an LSDA chunk.
Use an unnormalized uleb128 for the call-site table length that includes the
padding.
llvm-svn: 97078
the number of value bits, not the number of bits of allocation for in-memory
storage.
Make getTypeStoreSize and getTypeAllocSize work consistently for arrays and
vectors.
Fix several places in CodeGen which compute offsets into in-memory vectors
to use TargetData information.
This fixes PR1784.
llvm-svn: 97064
necessary to swap the operands to handle NaN and negative zero properly.
Also, reintroduce logic for checking for NaN conditions when forming
SSE min and max instructions, fixed to take into consideration NaNs and
negative zeros. This allows forming min and max instructions in more
cases.
llvm-svn: 97025
to adding them in a determinstic order (bottom up from
the root) based on the structure of the graph itself.
This updates tests for some random changes, interesting
bits: CodeGen/Blackfin/promote-logic.ll no longer crashes.
I have no idea why, but that's good right?
CodeGen/X86/2009-07-16-LoadFoldingBug.ll also fails, but
now compiles to have one fewer constant pool entry, making
the expected load that was being folded disappear. Since it
is an unreduced mass of gnast, I just removed it.
This fixes PR6370
llvm-svn: 97023
Previously, LiveIntervalAnalysis would infer phi joins by looking for multiply
defined registers. That doesn't work if the phi join is implicitly defined in
all but one of the predecessors.
llvm-svn: 96994
no id's would cause early exit allowing IsLegalToFold to return true
instead of false, producing a cyclic dag.
This was striking the new isel because it isn't using SelectNodeTo yet,
which theoretically is just an optimization.
llvm-svn: 96972
126.gcc nightly tests. These failures uncovered latent bugs that machine DCE
could remove one half of a stack adjust down/up pair, causing PEI to assert.
This update fixes that, and the tests now pass.
llvm-svn: 96822
dragonegg self-host build. I reverted 96640 in order to revert
96556 (96640 goes on top of 96556), but it also looks like with
both of them applied the breakage happens even earlier. The
symptom of the 96556 miscompile is the following crash:
llvm[3]: Compiling AlphaISelLowering.cpp for Release build
cc1plus: /home/duncan/tmp/tmp/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp:4982: void llvm::SelectionDAG::ReplaceAllUsesWith(llvm::SDNode*, llvm::SDNode*, llvm::SelectionDAG::DAGUpdateListener*): Assertion `(!From->hasAnyUseOfValue(i) || From->getValueType(i) == To->getValueType(i)) && "Cannot use this version of ReplaceAllUsesWith!"' failed.
Stack dump:
0. Running pass 'X86 DAG->DAG Instruction Selection' on function '@_ZN4llvm19AlphaTargetLowering14LowerOperationENS_7SDValueERNS_12SelectionDAGE'
g++: Internal error: Aborted (program cc1plus)
This occurs when building LLVM using LLVM built by LLVM (via
dragonegg). Probably LLVM has miscompiled itself, though it
may have miscompiled GCC and/or dragonegg itself: at this point
of the self-host build, all of GCC, LLVM and dragonegg were built
using LLVM. Unfortunately this kind of thing is extremely hard
to debug, and while I did rummage around a bit I didn't find any
smoking guns, aka obviously miscompiled code.
Found by bisection.
r96556 | evancheng | 2010-02-18 03:13:50 +0100 (Thu, 18 Feb 2010) | 5 lines
Some dag combiner goodness:
Transform br (xor (x, y)) -> br (x != y)
Transform br (xor (xor (x,y), 1)) -> br (x == y)
Also normalize (and (X, 1) == / != 1 -> (and (X, 1)) != / == 0 to match to "test on x86" and "tst on arm"
r96640 | evancheng | 2010-02-19 01:34:39 +0100 (Fri, 19 Feb 2010) | 16 lines
Transform (xor (setcc), (setcc)) == / != 1 to
(xor (setcc), (setcc)) != / == 1.
e.g. On x86_64
%0 = icmp eq i32 %x, 0
%1 = icmp eq i32 %y, 0
%2 = xor i1 %1, %0
br i1 %2, label %bb, label %return
=>
testl %edi, %edi
sete %al
testl %esi, %esi
sete %cl
cmpb %al, %cl
je LBB1_2
llvm-svn: 96672
for ARM to just check if a function has a FP to determine if it's safe
to simplify the stack adjustment pseudo ops prior to eliminating frame
indices. Allow targets to override the default behavior and does so for ARM
and Thumb2.
llvm-svn: 96634
Moderate the weight given to very small intervals.
The spill weight given to new intervals created when spilling was not
normalized in the same way as the original spill weights calculated by
CalcSpillWeights. That meant that restored registers would tend to hang around
because they had a much higher spill weight that unspilled registers.
This improves the runtime of a few tests by up to 10%, and there are no
significant regressions.
llvm-svn: 96613
checking whether AnalyzeBranch disagrees with the CFG
directly, rather than looking for EH_LABEL instructions.
EH_LABEL instructions aren't always at the end of the
block, due to FP_REG_KILL and other things. This fixes
an infinite loop compiling MultiSource/Benchmarks/Bullet.
llvm-svn: 96611
Chris Lattner