- There is a minor semantic change here (evidenced by the test change) for
Darwin triples that have no version component. I debated changing the default
behavior of isOSVersionLT, but decided it made more sense for triples to be
explicit.
llvm-svn: 129802
Making use of VFP / NEON floating point multiply-accumulate / subtraction is
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.
Enable these fp vmlx codegen changes for Cortex-A9.
llvm-svn: 129775
Add a avoidWriteAfterWrite() target hook to identify register classes that
suffer from write-after-write hazards. For those register classes, try to avoid
writing the same register in two consecutive instructions.
This is currently disabled by default. We should not spill to avoid hazards!
The command line flag -avoid-waw-hazard can be used to enable waw avoidance.
llvm-svn: 129772
when they are a truncate from something else. This eliminates fully half of all the
fastisel rejections on a test c++ file I'm working with, which should make a substantial
improvement for -O0 compile of c++ code.
This fixed rdar://9297003 - fast isel bails out on all functions taking bools
llvm-svn: 129752
Before we would bail out on i1 arguments all together, now we just bail on
non-constant ones. Also, we used to emit extraneous code. e.g. test12 was:
movb $0, %al
movzbl %al, %edi
callq _test12
and test13 was:
movb $0, %al
xorl %edi, %edi
movb %al, 7(%rsp)
callq _test13f
Now we get:
movl $0, %edi
callq _test12
and:
movl $0, %edi
callq _test13f
llvm-svn: 129751
registers for fast allocation a different way. This has us updating
used registers only when we're using that exact register.
Fixes rdar://9207598
llvm-svn: 129711
value constraints on them (when defined as ImmLeaf's). This is particularly important
for X86-64, where almost all reg/imm instructions take a i64immSExt32 immediate operand,
which has a value constraint. Before this patch we ended up iseling the examples into
such amazing code as:
movabsq $7, %rax
imulq %rax, %rdi
movq %rdi, %rax
ret
now we produce:
imulq $7, %rdi, %rax
ret
This dramatically shrinks the generated code at -O0 on x86-64.
llvm-svn: 129691
2. implement rdar://9289501 - fast isel should fold trivial multiplies to shifts
3. teach tblgen to handle shift immediates that are different sizes than the
shifted operands, eliminating some code from the X86 fast isel backend.
4. Have FastISel::SelectBinaryOp use (the poorly named) FastEmit_ri_ function
instead of FastEmit_ri to simplify code.
llvm-svn: 129666
when we have a global variable base an an index. Instead, just give up on
folding the global variable.
Before we'd geenrate:
_test: ## @test
## BB#0:
movq _rtx_length@GOTPCREL(%rip), %rax
leaq (%rax), %rax
addq %rdi, %rax
movzbl (%rax), %eax
ret
now we generate:
_test: ## @test
## BB#0:
movq _rtx_length@GOTPCREL(%rip), %rax
movzbl (%rax,%rdi), %eax
ret
The difference is even more significant when there is a scale
involved.
This fixes rdar://9289558 - total fail with addr mode formation at -O0/x86-64
llvm-svn: 129664
less trivial things) into a dummy lea. Before we generated:
_test: ## @test
movq _G@GOTPCREL(%rip), %rax
leaq (%rax), %rax
ret
now we produce:
_test: ## @test
movq _G@GOTPCREL(%rip), %rax
ret
This is part of rdar://9289558
llvm-svn: 129662
The basic issue here is that bottom-up isel is matching the branch
and compare, and was failing to fold the load into the branch/compare
combo. Fixing this (by allowing folding into any instruction of a
sequence that is selected) allows us to produce things like:
cmpb $0, 52(%rax)
je LBB4_2
instead of:
movb 52(%rax), %cl
cmpb $0, %cl
je LBB4_2
This makes the generated -O0 code run a bit faster, but also speeds up
compile time by putting less pressure on the register allocator and
generating less code.
This was one of the biggest classes of missing load folding. Implementing
this shrinks 176.gcc's c-decl.s (as a random example) by about 4% in (verbose-asm)
line count.
llvm-svn: 129656
Change ELF systems to use CFI for producing the EH tables. This reduces the
size of the clang binary in Debug builds from 690MB to 679MB.
llvm-svn: 129571
This is done by pushing physical register definitions close to their
use, which happens to handle flag definitions if they're not glued to
the branch. This seems to be generally a good thing though, so I
didn't need to add a target hook yet.
The primary motivation is to generate code closer to what people
expect and rule out missed opportunity from enabling macro-op
fusion. As a side benefit, we get several 2-5% gains on x86
benchmarks. There is one regression:
SingleSource/Benchmarks/Shootout/lists slows down be -10%. But this is
an independent scheduler bug that will be tracked separately.
See rdar://problem/9283108.
Incidentally, pre-RA scheduling is only half the solution. Fixing the
later passes is tracked by:
<rdar://problem/8932804> [pre-RA-sched] on x86, attempt to schedule CMP/TEST adjacent with condition jump
Fixes:
<rdar://problem/9262453> Scheduler unnecessary break of cmp/jump fusion
llvm-svn: 129508
ignored. There was a test to catch this, but it was just blindly updated in
a large change. This fixes another part of <rdar://problem/9275290>.
llvm-svn: 129466
the max itself, so it is not easy to write a test case for this, but I added a
test case that would fail if the code in AsmPrinter were removed.
llvm-svn: 129432
Additional fixes:
Do something reasonable for subtargets with generic
itineraries by handle node latency the same as for an empty
itinerary. Now nodes default to unit latency unless an itinerary
explicitly specifies a zero cycle stage or it is a TokenFactor chain.
Original fixes:
UnitsSharePred was a source of randomness in the scheduler: node
priority depended on the queue data structure. I rewrote the recent
VRegCycle heuristics to completely replace the old heuristic without
any randomness. To make the ndoe latency adjustments work, I also
needed to do something a little more reasonable with TokenFactor. I
gave it zero latency to its consumers and always schedule it as low as
possible.
llvm-svn: 129421
Now that we have a first-class way to represent unaligned loads, the unaligned
load intrinsics are superfluous.
First part of <rdar://problem/8460511>.
llvm-svn: 129401
UnitsSharePred was a source of randomness in the scheduler: node
priority depended on the queue data structure. I rewrote the recent
VRegCycle heuristics to completely replace the old heuristic without
any randomness. To make these heuristic adjustments to node latency work,
I also needed to do something a little more reasonable with TokenFactor. I
gave it zero latency to its consumers and always schedule it as low as
possible.
llvm-svn: 129383
induction variable. The preRA scheduler is unaware of induction vars,
so we look for potential "virtual register cycles" instead.
Fixes <rdar://problem/8946719> Bad scheduling prevents coalescing
llvm-svn: 129100
There can be multiple defs for a single virtual register when they are defining
sub-registers.
The missing <dead> flag was stopping the inline spiller from eliminating dead
code after rematerialization.
llvm-svn: 128888
When a virtual register has a single value that is defined as a copy of a
reserved register, permit that copy to be joined. These virtual register are
usually copies of the stack pointer:
%vreg75<def> = COPY %ESP; GR32:%vreg75
MOV32mr %vreg75, 1, %noreg, 0, %noreg, %vreg74<kill>
MOV32mi %vreg75, 1, %noreg, 8, %noreg, 0
MOV32mi %vreg75<kill>, 1, %noreg, 4, %noreg, 0
CALLpcrel32 ...
Coalescing these virtual registers early decreases register pressure.
Previously, they were coalesced by RALinScan::attemptTrivialCoalescing after
register allocation was completed.
The lower register pressure causes the mcinst-lowering-cmp0.ll test case to fail
because it depends on linear scan spilling a particular register.
I am deleting 2008-08-05-SpillerBug.ll because it is counting the number of
instructions emitted, and its revision history shows the 'correct' count being
edited many times.
llvm-svn: 128845
The code inserted by PPCTargetLowering::EmitInstrWithCustomInserter for ppc64 is
wrong, and I don't know how to fix it. It seems to be using the correct register
classes for pointers, but it inserts all 32-bit instructions.
llvm-svn: 128835
registers that arise from argument shuffling with the soft float ABI. These
instructions are particularly slow on Cortex A8. This fixes one half of
<rdar://problem/8674845>.
llvm-svn: 128759
This way, shrinkToUses() will ignore the instruction that is about to be
deleted, and we avoid leaving invalid live ranges that SplitKit doesn't like.
Fix a misunderstanding in MachineVerifier about <def,undef> operands. The
<undef> flag is valid on def operands where it has the same meaning as <undef>
on a use operand. It only applies to sub-register defines which also read the
full register.
llvm-svn: 128642
The rematerialized instruction may require a more constrained register class
than the register being spilled. In the test case, the spilled register has been
inflated to the DPR register class, but we are rematerializing a load of the
ssub_0 sub-register which only exists for DPR_VFP2 registers.
The register class is reinflated after spilling, so the conservative choice is
only temporary.
llvm-svn: 128610
Daniel Dunbar