This is still a hack until we can teach tblgen to generate the
optional CPSR operand rather than an implicit CPSR def. But the
strangeness is now limited to the selection DAG. ADD/SUB MI's no
longer have implicit CPSR defs, nor do we allow flag setting variants
of these opcodes in machine code. There are several corner cases to
consider, and getting one wrong would previously lead to nasty
miscompilation. It's not the first time I've debugged one, so this
time I added enough verification to ensure it won't happen again.
llvm-svn: 140228
It appears that our use of the imp-use and imp-def flags with
sub-registers is not yet robust enough to support this.
The failing test case is complicated, I am working on a reduction.
<rdar://problem/10044201>
llvm-svn: 138861
There is no non-writeback store multiple instruction in Thumb1, so
don't define one. As a result load multiple is the only instantiation of
the multiclass, so refactor that away entirely.
llvm-svn: 138338
This pleases the register scavenger and brings
test/CodeGen/ARM/2011-08-12-vmovqqqq-pseudo.ll a little closer to
working with -verify-machineinstrs.
llvm-svn: 138164
On Cortex-A8, we use the NEON v2f32 instructions for f32 arithmetic. For
better latency, we also send D-register copies down the NEON pipeline by
translating them to vorr instructions.
This patch promotes even S-register copies to D-register copies when
possible so they can also go down the NEON pipeline. Example:
vldr.32 s0, LCPI0_0
loop:
vorr d1, d0, d0
loop2:
...
vadd.f32 d1, d1, d16
The vorr instruction looked like this after regalloc:
%S2<def> = COPY %S0, %D1<imp-def>
Copies involving odd S-registers, and copies that don't define the full
D-register are left alone.
llvm-svn: 137182
The tSpill and tRestore instructions are just copies of the tSTRspi and
tLDRspi instructions, respectively. Just use those directly instead.
llvm-svn: 134092
sink them into MC layer.
- Added MCInstrInfo, which captures the tablegen generated static data. Chang
TargetInstrInfo so it's based off MCInstrInfo.
llvm-svn: 134021
is, it assumes addresses are 64-bit aligned (which should be the more common
case). If the alignment is found not to be aligned, then getOperandLatency()
would adjust the operand latency computation by one to compensate for it.
rdar://9294833
llvm-svn: 129742
entries being compared may not be ARMConstantPoolValue. Without checking
whether they are ARMConstantPoolValue first, and if the stars and moons
are aligned properly, the equality test may return true (when the first few
words of two Constants' values happen to be identical) and very bad things can
happen.
rdar://9125354
llvm-svn: 128203
int tries = INT_MAX;
while (tries > 0) {
tries--;
}
The check should be:
subs r4, #1
cmp r4, #0
bgt LBB0_1
The subs can set the overflow V bit when r4 is INT_MAX+1 (which loop
canonicalization apparently does in this case). cmp #0 would have cleared
it while not changing the N and Z bits. Since BGT is dependent on the V
bit, i.e. (N == V) && !Z, it is not safe to eliminate the cmp #0.
rdar://9172742
llvm-svn: 128179
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
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
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
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
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
DAG scheduling during isel. Most new functionality is currently
guarded by -enable-sched-cycles and -enable-sched-hazard.
Added InstrItineraryData::IssueWidth field, currently derived from
ARM itineraries, but could be initialized differently on other targets.
Added ScheduleHazardRecognizer::MaxLookAhead to indicate whether it is
active, and if so how many cycles of state it holds.
Added SchedulingPriorityQueue::HasReadyFilter to allowing gating entry
into the scheduler's available queue.
ScoreboardHazardRecognizer now accesses the ScheduleDAG in order to
get information about it's SUnits, provides RecedeCycle for bottom-up
scheduling, correctly computes scoreboard depth, tracks IssueCount, and
considers potential stall cycles when checking for hazards.
ScheduleDAGRRList now models machine cycles and hazards (under
flags). It tracks MinAvailableCycle, drives the hazard recognizer and
priority queue's ready filter, manages a new PendingQueue, properly
accounts for stall cycles, etc.
llvm-svn: 122541
Use the same COPY_TO_REGCLASS approach as for the 2-register *_sfp instructions.
This change made a big difference in the code generated for the
CodeGen/Thumb2/cross-rc-coalescing-2.ll test: The coalescer is still doing
a fine job, but some instructions that were previously moved outside the loop
are not moved now. It's using fewer VFP registers now, which is generally
a good thing, so I think the estimates for register pressure changed and that
affected the LICM behavior. Since that isn't obviously wrong, I've just
changed the test file. This completes the work for Radar 8711675.
llvm-svn: 121730
difficult on current ARM implementations for a few reasons.
1. Even though a single vmla has latency that is one cycle shorter than a pair
of vmul + vadd, a RAW hazard during the first (4? on Cortex-a8) can cause
additional pipeline stall. So it's frequently better to single codegen
vmul + vadd.
2. A vmla folowed by a vmul, vmadd, or vsub causes the second fp instruction to
stall for 4 cycles. We need to schedule them apart.
3. A vmla followed vmla is a special case. Obvious issuing back to back RAW
vmla + vmla is very bad. But this isn't ideal either:
vmul
vadd
vmla
Instead, we want to expand the second vmla:
vmla
vmul
vadd
Even with the 4 cycle vmul stall, the second sequence is still 2 cycles
faster.
Up to now, isel simply avoid codegen'ing fp vmla / vmls. This works well enough
but it isn't the optimial solution. This patch attempts to make it possible to
use vmla / vmls in cases where it is profitable.
A. Add missing isel predicates which cause vmla to be codegen'ed.
B. Make sure the fmul in (fadd (fmul)) has a single use. We don't want to
compute a fmul and a fmla.
C. Add additional isel checks for vmla, avoid cases where vmla is feeding into
fp instructions (except for the #3 exceptional case).
D. Add ARM hazard recognizer to model the vmla / vmls hazards.
E. Add a special pre-regalloc case to expand vmla / vmls when it's likely the
vmla / vmls will trigger one of the special hazards.
Work in progress, only A+B are enabled.
llvm-svn: 120960
Remove movePastCSLoadStoreOps and associated code for simple pointer
increments. Update routines that depended upon other opcodes for save/restore.
Adjust all testcases accordingly.
llvm-svn: 119725
and xor. The 32-bit move immediates can be hoisted out of loops by machine
LICM but the isel hacks were preventing them.
Instead, let peephole optimization pass recognize registers that are defined by
immediates and the ARM target hook will fold the immediates in.
Other changes include 1) do not fold and / xor into cmp to isel TST / TEQ
instructions if there are multiple uses. This happens when the 'and' is live
out, machine sink would have sinked the computation and that ends up pessimizing
code. The peephole pass would recognize situations where the 'and' can be
toggled to define CPSR and eliminate the comparison anyway.
2) Move peephole pass to after machine LICM, sink, and CSE to avoid blocking
important optimizations.
rdar://8663787, rdar://8241368
llvm-svn: 119548
'db', 'ib', 'da') instead of having that mode as a separate field in the
instruction. It's more convenient for the asm parser and much more readable for
humans.
<rdar://problem/8654088>
llvm-svn: 119310
1. Fix pre-ra scheduler so it doesn't try to push instructions above calls to
"optimize for latency". Call instructions don't have the right latency and
this is more likely to use introduce spills.
2. Fix if-converter cost function. For ARM, it should use instruction latencies,
not # of micro-ops since multi-latency instructions is completely executed
even when the predicate is false. Also, some instruction will be "slower"
when they are predicated due to the register def becoming implicit input.
rdar://8598427
llvm-svn: 118135
at more than those which define CPSR. You can have this situation:
(1) subs ...
(2) sub r6, r5, r4
(3) movge ...
(4) cmp r6, 0
(5) movge ...
We cannot convert (2) to "subs" because (3) is using the CPSR set by
(1). There's an analogous situation here:
(1) sub r1, r2, r3
(2) sub r4, r5, r6
(3) cmp r4, ...
(5) movge ...
(6) cmp r1, ...
(7) movge ...
We cannot convert (1) to "subs" because of the intervening use of CPSR.
llvm-svn: 117950
operand and one of them has a single use that is a live out copy, favor the
one that is live out. Otherwise it will be difficult to eliminate the copy
if the instruction is a loop induction variable update. e.g.
BB:
sub r1, r3, #1
str r0, [r2, r3]
mov r3, r1
cmp
bne BB
=>
BB:
str r0, [r2, r3]
sub r3, r3, #1
cmp
bne BB
This fixed the recent 256.bzip2 regression.
llvm-svn: 117675
- For now, loads of [r, r] addressing mode is the same as the
[r, r lsl/lsr/asr #] variants. ARMBaseInstrInfo::getOperandLatency() should
identify the former case and reduce the output latency by 1.
- Also identify [r, r << 2] case. This special form of shifter addressing mode
is "free".
llvm-svn: 117519
the LDR instructions have. This makes the literal/register forms of the
instructions explicit and allows us to assign scheduling itineraries
appropriately. rdar://8477752
llvm-svn: 117505
explicit about the operands. Split out the different variants into separate
instructions. This gives us the ability to, among other things, assign
different scheduling itineraries to the variants. rdar://8477752.
llvm-svn: 117409
"long latency" enough to hoist even if it may increase spilling. Reloading
a value from spill slot is often cheaper than performing an expensive
computation in the loop. For X86, that means machine LICM will hoist
SQRT, DIV, etc. ARM will be somewhat aggressive with VFP and NEON
instructions.
- Enable register pressure aware machine LICM by default.
llvm-svn: 116781
allow target to correctly compute latency for cases where static scheduling
itineraries isn't sufficient. e.g. variable_ops instructions such as
ARM::ldm.
This also allows target without scheduling itineraries to compute operand
latencies. e.g. X86 can return (approximated) latencies for high latency
instructions such as division.
- Compute operand latencies for those defined by load multiple instructions,
e.g. ldm and those used by store multiple instructions, e.g. stm.
llvm-svn: 115755
stick with a constant estimate of 90% (branch predictors are good!), but we might find that we want to provide
more nuanced estimates in the future.
llvm-svn: 115364
cost modeling for if-conversion. Now if only we had a way to estimate the misprediction probability.
Adjsut CodeGen/ARM/ifcvt10.ll. The pipeline on Cortex-A8 is long enough that it is still profitable
to predicate an ldm, but the shorter pipeline on Cortex-A9 makes it unprofitable.
llvm-svn: 114995
Rather than having arbitrary cutoffs, actually try to cost model the conversion.
For now, the constants are tuned to more or less match our existing behavior, but these will be
changed to reflect realistic values as this work proceeds.
llvm-svn: 114973
I am unable to write a test for this case, help is solicited, though...
What I did is to tickle the code in the debugger and verify that we do the right thing.
llvm-svn: 114430
into OptimizeCompareInstr.
This necessitates the passing of CmpValue around,
so widen the virtual functions to accomodate.
No functionality changes.
llvm-svn: 114428
Recognize VLD1q64Pseudo as a stack slot load.
Reject these if they are loading or storing a subregister. The API (and
VirtRegRewriter) doesn't know how to deal with that.
llvm-svn: 113985
encountered while building llvm-gcc for arm. This is probably the same issue
that the ppc buildbot hit. llvm::prior works on a MachineBasicBlock::iterator,
not a plain MachineInstr.
llvm-svn: 113983
backing out following to get it back to green,
so I can investigate in peace:
svn merge -c -113840 llvm/test/CodeGen/ARM/arm-and-tst-peephole.ll
svn merge -c -113876 -c -113839 llvm/lib/Target/ARM/ARMBaseInstrInfo.cpp
llvm-svn: 113980
by morphing the 'and' to its recording form 'andS'.
This is basically a test commit into this area, to
see whether the bots like me. Several generalizations
can be applied and various avenues of code simplification
are open. I'll introduce those as I go.
I am aware of stylistic input from Bill Wendling, about
where put the analysis complexity, but I am positive
that we can move things around easily and will find a
satisfactory solution.
llvm-svn: 113839
iterator when an optimization took place. This allows us to do more insane
things with the code than just remove an instruction or two.
llvm-svn: 113640
take multiple cycles to decode.
For the current if-converter clients (actually only ARM), the instructions that
are predicated on false are not nops. They would still take machine cycles to
decode. Micro-coded instructions such as LDM / STM can potentially take multiple
cycles to decode. If-converter should take treat them as non-micro-coded
simple instructions.
llvm-svn: 113570
instruction in the class would be decoded to. Or zero if the number of
uOPs must be determined dynamically.
This will be used to determine the cost-effectiveness of predicating a
micro-coded instruction.
llvm-svn: 113513
all the other LDM/STM instructions. This fixes asm printer crashes when
compiling with -O0. I've changed one of the NEON tests (vst3.ll) to run
with -O0 to check this in the future.
Prior to this change VLDM/VSTM used addressing mode #5, but not really.
The offset field was used to hold a count of the number of registers being
loaded or stored, and the AM5 opcode field was expanded to specify the IA
or DB mode, instead of the standard ADD/SUB specifier. Much of the backend
was not aware of these special cases. The crashes occured when rewriting
a frameindex caused the AM5 offset field to be changed so that it did not
have a valid submode. I don't know exactly what changed to expose this now.
Maybe we've never done much with -O0 and NEON. Regardless, there's no longer
any reason to keep a count of the VLDM/VSTM registers, so we can use
addressing mode #4 and clean things up in a lot of places.
llvm-svn: 112322
relatively expensive comparison analyzer on each instruction. Also rename the
comparison analyzer method to something more in line with what it actually does.
This pass is will eventually be folded into the Machine CSE pass.
llvm-svn: 110539
This pass tries to remove comparison instructions when possible. For instance,
if you have this code:
sub r1, 1
cmp r1, 0
bz L1
and "sub" either sets the same flag as the "cmp" instruction or could be
converted to set the same flag, then we can eliminate the "cmp" instruction all
together. This is a important for ARM where the ALU instructions could set the
CPSR flag, but need a special suffix ('s') to do so.
llvm-svn: 110423
have 4 bits per register in the operand encoding), but have undefined
behavior when the operand value is 13 or 15 (SP and PC, respectively).
The trivial coalescer in linear scan sometimes will merge a copy from
SP into a subsequent instruction which uses the copy, and if that
instruction cannot legally reference SP, we get bad code such as:
mls r0,r9,r0,sp
instead of:
mov r2, sp
mls r0, r9, r0, r2
This patch adds a new register class for use by Thumb2 that excludes
the problematic registers (SP and PC) and is used instead of GPR
for those operands which cannot legally reference PC or SP. The
trivial coalescer explicitly requires that the register class
of the destination for the COPY instruction contain the source
register for the COPY to be considered for coalescing. This prevents
errant instructions like that above.
PR7499
llvm-svn: 109842
The only folding these load/store architectures can do is converting COPY into a
load or store, and the target independent part of foldMemoryOperand already
knows how to do that.
llvm-svn: 108099
of getPhysicalRegisterRegClass with it.
If we want to make a copy (or estimate its cost), it is better to use the
smallest class as more efficient operations might be possible.
llvm-svn: 107140
with the following instructions. This is done via trickery by considering the
instruction preceding the IT to be the hazard. Care must be taken to ensure
it's the first non-debug instruction, or the presence of debug info will
affect codegen.
Part of the continuing work for rdar://7797940, making ARM code-gen unaffected
by the presence of debug information.
llvm-svn: 106871
void t(int *cp0, int *cp1, int *dp, int fmd) {
int c0, c1, d0, d1, d2, d3;
c0 = (*cp0++ & 0xffff) | ((*cp1++ << 16) & 0xffff0000);
c1 = (*cp0++ & 0xffff) | ((*cp1++ << 16) & 0xffff0000);
/* ... */
}
It code gens into something pretty bad. But with this change (analogous to the
X86 back-end), it will use ldm and generate few instructions.
llvm-svn: 106693
- This fixed a number of bugs in if-converter, tail merging, and post-allocation
scheduler. If-converter now runs branch folding / tail merging first to
maximize if-conversion opportunities.
- Also changed the t2IT instruction slightly. It now defines the ITSTATE
register which is read by instructions in the IT block.
- Added Thumb2 specific hazard recognizer to ensure the scheduler doesn't
change the instruction ordering in the IT block (since IT mask has been
finalized). It also ensures no other instructions can be scheduled between
instructions in the IT block.
This is not yet enabled.
llvm-svn: 106344
addresses a longstanding deficiency noted in many FIXMEs scattered
across all the targets.
This effectively moves the problem up one level, replacing eleven
FIXMEs in the targets with eight FIXMEs in CodeGen, plus one path
through FastISel where we actually supply a DebugLoc, fixing Radar
7421831.
llvm-svn: 106243
call must not be callee-saved; following x86, add a new
regclass to represent this. Also fixes a couple of bugs.
Still disabled by default; Thumb doesn't work yet.
llvm-svn: 106053
In file included from X86InstrInfo.cpp:16:
X86GenInstrInfo.inc:2789: error: integer constant is too large for 'long' type
X86GenInstrInfo.inc:2790: error: integer constant is too large for 'long' type
X86GenInstrInfo.inc:2792: error: integer constant is too large for 'long' type
X86GenInstrInfo.inc:2793: error: integer constant is too large for 'long' type
X86GenInstrInfo.inc:2808: error: integer constant is too large for 'long' type
X86GenInstrInfo.inc:2809: error: integer constant is too large for 'long' type
X86GenInstrInfo.inc:2816: error: integer constant is too large for 'long' type
X86GenInstrInfo.inc:2817: error: integer constant is too large for 'long' type
llvm-svn: 105524
instruction defines subregisters.
Any existing subreg indices on the original instruction are preserved or
composed with the new subreg index.
Also substitute multiple operands mentioning the original register by using the
new MachineInstr::substituteRegister() function. This is necessary because there
will soon be <imp-def> operands added to non read-modify-write partial
definitions. This instruction:
%reg1234:foo = FLAP %reg1234<imp-def>
will reMaterialize(%reg3333, bar) like this:
%reg3333:bar-foo = FLAP %reg333:bar<imp-def>
Finally, replace the TargetRegisterInfo pointer argument with a reference to
indicate that it cannot be NULL.
llvm-svn: 105358
Reverse-merging r103156 into '.':
U lib/Target/ARM/ARMInstrNEON.td
U lib/Target/ARM/ARMRegisterInfo.h
U lib/Target/ARM/ARMBaseRegisterInfo.cpp
U lib/Target/ARM/ARMBaseInstrInfo.cpp
U lib/Target/ARM/ARMRegisterInfo.td
llvm-svn: 103159
These instructions are only needed for codegen, so I've removed all the
explicit encoding bits for now; they should be set in the same way as the for
VLDMD and VSTMD whenever we add encodings for VFP. The use of addrmode5
requires that the instructions be custom-selected so that the number of
registers can be set in the AM5Opc value.
llvm-svn: 99309
with changes to add a separate optional register update argument. Change all
the NEON instructions with address register writeback to use it.
llvm-svn: 99095
- 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
branch in ARM v4 code, since it gets clobbered by the return address before
it is used. Instead of adding a new register class containing all the GPRs
except LR, just use the existing tGPR class.
llvm-svn: 96360
into TargetOpcodes.h. #include the new TargetOpcodes.h
into MachineInstr. Add new inline accessors (like isPHI())
to MachineInstr, and start using them throughout the
codebase.
llvm-svn: 95687
function can support dynamic stack realignment. That's a much easier question
to answer at instruction selection stage than whether the function actually
will have dynamic alignment prologue. This allows the removal of the
stack alignment heuristic pass, and improves code quality for cases where
the heuristic would result in dynamic alignment code being generated when
it was not strictly necessary.
llvm-svn: 93885
for all the processors where I have tried it, and even when it might not help
performance, the cost is quite low. The opportunities for duplicating
indirect branches are limited by other factors so code size does not change
much due to tail duplicating indirect branches aggressively.
llvm-svn: 90144
Make tail duplication of indirect branches much more aggressive (for targets
that indicate that it is profitable), based on further experience with
this transformation. I compiled 3 large applications with and without
this more aggressive tail duplication and measured minimal changes in code
size. ("size" on Darwin seems to round the text size up to the nearest
page boundary, so I can only say that any code size increase was less than
one 4k page.) Radar 7421267.
llvm-svn: 89814
contents of the block to be duplicated. Use this for ARM Cortex A8/9 to
be more aggressive tail duplicating indirect branches, since it makes it
much more likely that they will be predicted in the branch target buffer.
Testcase coming soon.
llvm-svn: 89187
Andrew Trick