We previously used the default expansion to SELECT_CC, which in turn would
expand to "LHI; BRC; LHI". In most cases it's better to use an IPM-based
sequence instead.
llvm-svn: 192784
Floats are stored in the high 32 bits of an FPR, and the only GPR<->FPR
transfers are full-register transfers. This patch optimizes GPR<->FPR
float transfers when the high word of a GPR is directly accessible.
llvm-svn: 191764
This just adds the basics necessary for allocating the upper words to
virtual registers (move, load and store). The move support is parameterised
in a way that makes it easy to handle zero extensions, but the associated
zero-extend patterns are added by a later patch.
The easiest way of testing this seemed to be add a new "h" register
constraint for high words. I don't expect the constraint to be useful
in real inline asms, but it should work, so I didn't try to hide it
behind an option.
llvm-svn: 191739
Use subreg_hNN and subreg_lNN for the high and low NN bits of a register.
List the low registers first, so that subreg_l32 also means the low 32
bits of a 128-bit register.
Floats are stored in the upper 32 bits of a 64-bit register, so they
should use subreg_h32 rather than subreg_l32.
No behavioral change intended.
llvm-svn: 191659
I'm about to add support for high-word operations, so it seemed better
for the low-word registers to have names like R0L rather than R0W.
No behavioral change intended.
llvm-svn: 191655
The backend previously folded offsets into PC-relative addresses
whereever possible. That's the right thing to do when the address
can be used directly in a PC-relative memory reference (using things
like LRL). But if we have a register-based memory reference and need
to load the PC-relative address separately, it's better to use an anchor
point that could be shared with other accesses to the same area of the
variable.
Fixes a FIXME.
llvm-svn: 191524
Another patch to avoid duplication of encoding information. Things like
NILF, NILL and NILH are used as both 32-bit and 64-bit instructions.
Here the 64-bit versions are defined as aliases of the 32-bit ones.
llvm-svn: 191369
Another patch to reduce the duplication of encoding information.
Rather than define separate patterns for truncating 64-bit stores,
use the 32-bit stores with a subreg. No behavioral changed intended.
llvm-svn: 191365
The main complication here is that TM and TMY (the memory forms) set
CC differently from the register forms. When the tested bits contain
some 0s and some 1s, the register forms set CC to 1 or 2 based on the
value the uppermost bit. The memory forms instead set CC to 1
regardless of the uppermost bit.
Until now, I've tried to make it so that a branch never tests for an
impossible CC value. E.g. NR only sets CC to 0 or 1, so branches on the
result will only test for 0 or 1. Originally I'd tried to do the same
thing for TM and TMY by using custom matching code in ISelDAGToDAG.
That ended up being very ugly though, and would have meant duplicating
some of the chain checks that the common isel code does.
I've therefore gone for the simpler alternative of adding an extra
operand to the TM DAG opcode to say whether a memory form would be OK.
This means that the inverse of a "TM;JE" is "TM;JNE" rather than the
more precise "TM;JNLE", just like the inverse of "TMLL;JE" is "TMLL;JNE".
I suppose that's arguably less confusing though...
llvm-svn: 190400
The architecture has many comparison instructions, including some that
extend one of the operands. The signed comparison instructions use sign
extensions and the unsigned comparison instructions use zero extensions.
In cases where we had a free choice between signed or unsigned comparisons,
we were trying to decide at lowering time which would best fit the available
instructions, taking things like extension type into account. The code
to do that was getting increasingly hairy and was also making some bad
decisions. E.g. when comparing the result of two LLCs, it is better to use
CR rather than CLR, since CR can be fused with a branch while CLR can't.
This patch removes the lowering code and instead adds an operand to
integer comparisons to say whether signed comparison is required,
whether unsigned comparison is required, or whether either is OK.
We can then leave the choice of instruction up to the normal isel code.
llvm-svn: 190138
For now this just handles simple comparisons of an ANDed value with zero.
The CC value provides enough information to do any comparison for a
2-bit mask, and some nonzero comparisons with more populated masks,
but that's all future work.
llvm-svn: 189819
For now just handles simple comparisons of an ANDed value with zero.
The CC value provides enough information to do any comparison for a
2-bit mask, and some nonzero comparisons with more populated masks,
but that's all future work.
llvm-svn: 189469
Lengths up to a certain threshold (currently 6 * 256) use a series of MVCs.
Lengths above that threshold use a loop to handle X*256 bytes followed
by a single MVC to handle the excess (if any). This loop will also be
needed in future when support for variable lengths is added.
Because the same tablegen classes are used to define MVC and CLC,
the patch also has the side-effect of defining a pseudo loop instruction
for CLC. That instruction isn't used yet (and wouldn't be handled correctly
if it were). I'm planning to use it soon though.
llvm-svn: 189331
The initial port used MLG(R) for i64 UMUL_LOHI but left the other three
combinations as not-legal-or-custom. Although 32x32->{32,32}
multiplications exist, they're not as quick as doing a normal 64-bit
multiplication, so it didn't seem like i32 SMUL_LOHI and UMUL_LOHI
would be useful. There's also no direct instruction for i64 SMUL_LOHI,
so it needs to be implemented in terms of UMUL_LOHI.
However, not defining these patterns means that we don't convert
division by a constant into multiplication, so this patch fills
in the other cases. The new i64 SMUL_LOHI sequence is simpler
than the one that we used previously for 64x64->128 multiplication,
so int-mul-08.ll now tests the full sequence.
llvm-svn: 188898
SystemZTargetLowering::emitStringWrapper() previously loaded the character
into R0 before the loop and made R0 live on entry. I'd forgotten that
allocatable registers weren't allowed to be live across blocks at this stage,
and it confused LiveVariables enough to cause a miscompilation of f3 in
memchr-02.ll.
This patch instead loads R0 in the loop and leaves LICM to hoist it
after RA. This is actually what I'd tried originally, but I went for
the manual optimisation after noticing that R0 often wasn't being hoisted.
This bug forced me to go back and look at why, now fixed as r188774.
We should also try to optimize null checks so that they test the CC result
of the SRST directly. The select between null and the SRST GPR result could
then usually be deleted as dead.
llvm-svn: 188779
This first cut is pretty conservative. The final argument register (R6)
is call-saved, so we would need to make sure that the R6 argument to a
sibling call is the same as the R6 argument to the calling function,
which seems worth keeping as a separate patch.
Saying that integer truncations are free means that we no longer
use the extending instructions LGF and LLGF for spills in int-conv-09.ll
and int-conv-10.ll. Instead we treat the registers as 64 bits wide and
truncate them to 32-bits where necessary. I think it's unlikely we'd
use LGF and LLGF for spills in other situations for the same reason,
so I'm removing the tests rather than replacing them. The associated
code is generic and applies to many more instructions than just
LGF and LLGF, so there is no corresponding code removal.
llvm-svn: 188669
This also fixes a bug in the predication of LR to LOCR: I'd forgotten
that with these in-place instruction builds, the implicit operands need
to be added manually. I think this was latent until now, but is tested
by int-cmp-45.c. It also adds a CC valid mask to STOC, again tested by
int-cmp-45.c.
llvm-svn: 187573
Convert >= 1 to > 0, etc. Using comparison with zero isn't a win on its own,
but it exposes more opportunities for CC reuse (the next patch).
llvm-svn: 187571
The loop optimizers were assuming that scales > 1 were OK. I think this
is actually a bug in TargetLoweringBase::isLegalAddressingMode(),
since it seems to be trying to reject anything that isn't r+i or r+r,
but it has no default case for scales other than 0, 1 or 2. Implementing
the hook for z means that z can no longer test any change there though.
llvm-svn: 187497
System z branches have a mask to select which of the 4 CC values should
cause the branch to be taken. We can invert a branch by inverting the mask.
However, not all instructions can produce all 4 CC values, so inverting
the branch like this can lead to some oddities. For example, integer
comparisons only produce a CC of 0 (equal), 1 (less) or 2 (greater).
If an integer EQ is reversed to NE before instruction selection,
the branch will test for 1 or 2. If instead the branch is reversed
after instruction selection (by inverting the mask), it will test for
1, 2 or 3. Both are correct, but the second isn't really canonical.
This patch therefore keeps track of which CC values are possible
and uses this when inverting a mask.
Although this is mostly cosmestic, it fixes undefined behavior
for the CIJNLH in branch-08.ll. Another fix would have been
to mask out bit 0 when generating the fused compare and branch,
but the point of this patch is that we shouldn't need to do that
in the first place.
The patch also makes it easier to reuse CC results from other instructions.
llvm-svn: 187495
r187116 moved compare-and-branch generation from the instruction-selection
pass to the peephole optimizer (via optimizeCompare). It turns out that even
this is a bit too early. Fused compare-and-branch instructions don't
interact well with predication, where a CC result is needed. They also
make it harder to reuse the CC side-effects of earlier instructions
(not yet implemented, but the subject of a later patch).
Another problem was that the AnalyzeBranch family of routines weren't
handling compares and branches, so we weren't able to reverse the fused
form in cases where we would reverse a separate branch. This could have
been fixed by extending AnalyzeBranch, but given the other problems,
I've instead moved the fusing to the long-branch pass, which is also
responsible for the opposite transformation: splitting out-of-range
compares and branches into separate compares and long branches.
I've added a test for the AnalyzeBranch problem. A test for the
predication problem is included in the next patch, which fixes a bug
in the choice of CC mask.
llvm-svn: 187494
Before the patch we took advantage of the fact that the compare and
branch are glued together in the selection DAG and fused them together
(where possible) while emitting them. This seemed to work well in practice.
However, fusing the compare so early makes it harder to remove redundant
compares in cases where CC already has a suitable value. This patch
therefore uses the peephole analyzeCompare/optimizeCompareInstr pair of
functions instead.
No behavioral change intended, but it paves the way for a later patch.
llvm-svn: 187116
These instructions are allowed to trap even if the condition is false,
so for now they are only used for "*ptr = (cond ? x : *ptr)"-style
constructs.
llvm-svn: 187111
GPR and FPR constraints like "{r2}" and "{f2}" weren't handled correctly
because the name-to-regno mapping depends on the value type and
(because of that) the internal names in RegStrings are not the
same as the AsmName.
CC constraints like "{cc}" didn't work either because there was no
associated register class.
llvm-svn: 186148
in-tree implementations of TargetLoweringBase::isFMAFasterThanMulAndAdd in
order to resolve the following issues with fmuladd (i.e. optional FMA)
intrinsics:
1. On X86(-64) targets, ISD::FMA nodes are formed when lowering fmuladd
intrinsics even if the subtarget does not support FMA instructions, leading
to laughably bad code generation in some situations.
2. On AArch64 targets, ISD::FMA nodes are formed for operations on fp128,
resulting in a call to a software fp128 FMA implementation.
3. On PowerPC targets, FMAs are not generated from fmuladd intrinsics on types
like v2f32, v8f32, v4f64, etc., even though they promote, split, scalarize,
etc. to types that support hardware FMAs.
The function has also been slightly renamed for consistency and to force a
merge/build conflict for any out-of-tree target implementing it. To resolve,
see comments and fixed in-tree examples.
llvm-svn: 185956
Use "STC;MVC" for memsets that are too big for two STCs or MV...Is yet
small enough for a single MVC. As with memcpy, I'm leaving longer cases
till later.
The number of tests might seem excessive, but f33 & f34 from memset-04.ll
failed the first cut because I'd not added the "?:" on the calculation
of Size1.
llvm-svn: 185918
I was originally going to use MVC for memmove too, but that's less of
a clear win. Remove some accidental left-overs in the previous commit.
llvm-svn: 185804
Richard Sandiford