Splitting a basic block will create a new ALU clause, so we need to make
sure we aren't moving uses of registers that are local to their
current clause into a new one.
I had a test case for this, but unfortunately unrelated schedule changes
invalidated it, and I wasn't been able to come up with another one.
NOTE: This is a candidate for the 3.4 branch.
llvm-svn: 195399
The LDS output queue is accessed via the OQAP register. The OQAP
register cannot be live across clauses, so if value is written to the
output queue, it must be retrieved before the end of the clause.
With the machine scheduler, we cannot statisfy this constraint, because
it lacks proper alias analysis and it will mark some LDS accesses as
having a chain dependency on vertex fetches. Since vertex fetches
require a new clauses, the dependency may end up spiltting OQAP uses and
defs so the end up in different clauses. See the lds-output-queue.ll
test for a more detailed explanation.
To work around this issue, we now combine the LDS read and the OQAP
copy into one instruction and expand it after register allocation.
This patch also adds some checks to the EmitClauseMarker pass, so that
it doesn't end a clause with a value still in the output queue and
removes AR.X and OQAP handling from the scheduler (AR.X uses and defs
were already being expanded post-RA, so the scheduler will never see
them).
Reviewed-by: Vincent Lejeune <vljn at ovi.com>
llvm-svn: 194755
The AMDGPUIndirectAddressing pass was previously responsible for
lowering private loads and stores to indirect addressing instructions.
However, this pass was buggy and way too complicated. The only
advantage it had over the new simplified code was that it saved one
instruction per direct write to private memory. This optimization
likely has a minimal impact on performance, and we may be able
to duplicate it using some other transformation.
For the private address space, we now:
1. Lower private loads/store to Register(Load|Store) instructions
2. Reserve part of the register file as 'private memory'
3. After regalloc lower the Register(Load|Store) instructions to
MOV instructions that use indirect addressing.
llvm-svn: 193179
For targets that have instruction itineraries this means no change. Targets
that move over to the new schedule model will use be able the new schedule
module for instruction latencies in the if-converter (the logic is such that if
there is no itineary we will use the new sched model for the latencies).
Before, we queried "TTI->getInstructionLatency()" for the instruction latency
and the extra prediction cost. Now, we query the TargetSchedule abstraction for
the instruction latency and TargetInstrInfo for the extra predictation cost. The
TargetSchedule abstraction will internally call "TTI->getInstructionLatency" if
an itinerary exists, otherwise it will use the new schedule model.
ATTENTION: Out of tree targets!
(I will also send out an email later to LLVMDev)
This means, if your target implements
unsigned getInstrLatency(const InstrItineraryData *ItinData,
const MachineInstr *MI,
unsigned *PredCost);
and returns a value for "PredCost", you now also need to implement
unsigned getPredictationCost(const MachineInstr *MI);
(if your target uses the IfConversion.cpp pass)
radar://15077010
llvm-svn: 191671
This fixes some regressions in the piglit local memory store tests
introduced by recent commits which made the scheduler aware of the trans
slot.
It's not possible to test this using lit, because there is no way to
determine from the assembly dumps whether or not an instruction is in
the trans slot.
Even if this were possible, the test would be highly sensitive to
changes in the scheduler and might generate confusing false negatives.
Reviewed-by: Vincent Lejeune<vljn at ovi.com>
llvm-svn: 190574
This increases the number of opportunites we have for folding. With the
previous implementation we were unable to fold into any instructions
other than the first when multiple instructions were selected from a
single SDNode.
Reviewed-by: Vincent Lejeune <vljn at ovi.com>
llvm-svn: 186919
Dot4 now uses 8 scalar operands instead of 2 vectors one which allows register
coalescer to remove some unneeded COPY.
This patch also defines some structures/functions that can be used to handle
every vector instructions (CUBE, Cayman special instructions...) in a similar
fashion.
llvm-svn: 182126
Only implemented for R600 so far. SI is missing implementations of a
few callbacks used by the Indirect Addressing pass and needs code to
handle frame indices.
At the moment R600 only supports array sizes of 16 dwords or less.
Register packing of vector types is currently disabled, which means that a
vec4 is stored in T0_X, T1_X, T2_X, T3_X, rather than T0_XYZW. In order
to correctly pack registers in all cases, we will need to implement an
analysis pass for R600 that determines the correct vector width for each
array.
v2:
- Add support for i8 zext load from stack.
- Coding style fixes
v3:
- Don't reserve registers for indirect addressing when it isn't
being used.
- Fix bug caused by LLVM limiting the number of SubRegIndex
declarations.
v4:
- Fix 64-bit defines
llvm-svn: 174525
utils/sort_includes.py script.
Most of these are updating the new R600 target and fixing up a few
regressions that have creeped in since the last time I sorted the
includes.
llvm-svn: 171362
Alp Toker