Now that `Metadata` is typeless, reflect that in the assembly. These
are the matching assembly changes for the metadata/value split in
r223802.
- Only use the `metadata` type when referencing metadata from a call
intrinsic -- i.e., only when it's used as a `Value`.
- Stop pretending that `ValueAsMetadata` is wrapped in an `MDNode`
when referencing it from call intrinsics.
So, assembly like this:
define @foo(i32 %v) {
call void @llvm.foo(metadata !{i32 %v}, metadata !0)
call void @llvm.foo(metadata !{i32 7}, metadata !0)
call void @llvm.foo(metadata !1, metadata !0)
call void @llvm.foo(metadata !3, metadata !0)
call void @llvm.foo(metadata !{metadata !3}, metadata !0)
ret void, !bar !2
}
!0 = metadata !{metadata !2}
!1 = metadata !{i32* @global}
!2 = metadata !{metadata !3}
!3 = metadata !{}
turns into this:
define @foo(i32 %v) {
call void @llvm.foo(metadata i32 %v, metadata !0)
call void @llvm.foo(metadata i32 7, metadata !0)
call void @llvm.foo(metadata i32* @global, metadata !0)
call void @llvm.foo(metadata !3, metadata !0)
call void @llvm.foo(metadata !{!3}, metadata !0)
ret void, !bar !2
}
!0 = !{!2}
!1 = !{i32* @global}
!2 = !{!3}
!3 = !{}
I wrote an upgrade script that handled almost all of the tests in llvm
and many of the tests in cfe (even handling many `CHECK` lines). I've
attached it (or will attach it in a moment if you're speedy) to PR21532
to help everyone update their out-of-tree testcases.
This is part of PR21532.
llvm-svn: 224257
This commit adds a weak variant of the cmpxchg operation, as described
in C++11. A cmpxchg instruction with this modifier is permitted to
fail to store, even if the comparison indicated it should.
As a result, cmpxchg instructions must return a flag indicating
success in addition to their original iN value loaded. Thus, for
uniformity *all* cmpxchg instructions now return "{ iN, i1 }". The
second flag is 1 when the store succeeded.
At the DAG level, a new ATOMIC_CMP_SWAP_WITH_SUCCESS node has been
added as the natural representation for the new cmpxchg instructions.
It is a strong cmpxchg.
By default this gets Expanded to the existing ATOMIC_CMP_SWAP during
Legalization, so existing backends should see no change in behaviour.
If they wish to deal with the enhanced node instead, they can call
setOperationAction on it. Beware: as a node with 2 results, it cannot
be selected from TableGen.
Currently, no use is made of the extra information provided in this
patch. Test updates are almost entirely adapting the input IR to the
new scheme.
Summary for out of tree users:
------------------------------
+ Legacy Bitcode files are upgraded during read.
+ Legacy assembly IR files will be invalid.
+ Front-ends must adapt to different type for "cmpxchg".
+ Backends should be unaffected by default.
llvm-svn: 210903
We had disabled use of TBAA during CodeGen (even when otherwise using AA)
because the ptrtoint/inttoptr used by CGP for address sinking caused BasicAA to
miss basic type punning that it should catch (and, thus, we'd fail to override
TBAA when we should).
However, when AA is in use during CodeGen, CGP now uses normal GEPs and
bitcasts, instead of ptrtoint/inttoptr, when doing address sinking. As a
result, BasicAA should be able to make us do the right thing in the face of
type-punning, and it seems safe to enable use of TBAA again. self-hosting seems
fine on PPC64/Linux on the P7, with TBAA enabled and -misched=shuffle.
Note: We still don't update TBAA when merging stack slots, although because
BasicAA should now catch all such cases, this is no longer a blocking issue.
Nevertheless, I plan to commit code to deal with this properly in the near
future.
llvm-svn: 206093
The syntax for "cmpxchg" should now look something like:
cmpxchg i32* %addr, i32 42, i32 3 acquire monotonic
where the second ordering argument gives the required semantics in the case
that no exchange takes place. It should be no stronger than the first ordering
constraint and cannot be either "release" or "acq_rel" (since no store will
have taken place).
rdar://problem/15996804
llvm-svn: 203559
Summary:
AsmPrinter::EmitInlineAsm() will no longer use the EmitRawText() call for
targets with mature MC support. Such targets will always parse the inline
assembly (even when emitting assembly). Targets without mature MC support
continue to use EmitRawText() for assembly output.
The hasRawTextSupport() check in AsmPrinter::EmitInlineAsm() has been replaced
with MCAsmInfo::UseIntegratedAs which when true, causes the integrated assembler
to parse inline assembly (even when emitting assembly output). UseIntegratedAs
is set to true for targets that consider any failure to parse valid assembly
to be a bug. Target specific subclasses generally enable the integrated
assembler in their constructor. The default value can be overridden with
-no-integrated-as.
All tests that rely on inline assembly supporting invalid assembly (for example,
those that use mnemonics such as 'foo' or 'hello world') have been updated to
disable the integrated assembler.
Changes since review (and last commit attempt):
- Fixed test failures that were missed due to configuration of local build.
(fixes crash.ll and a couple others).
- Fixed tests that happened to pass because the local build was on X86
(should fix 2007-12-17-InvokeAsm.ll)
- mature-mc-support.ll's should no longer require all targets to be compiled.
(should fix ARM and PPC buildbots)
- Object output (-filetype=obj and similar) now forces the integrated assembler
to be enabled regardless of default setting or -no-integrated-as.
(should fix SystemZ buildbots)
Reviewers: rafael
Reviewed By: rafael
CC: llvm-commits
Differential Revision: http://llvm-reviews.chandlerc.com/D2686
llvm-svn: 201333
The old mask in f24 wasn't well chosen because the lshr would always be zero.
CodeGen didn't detect this but InstCombine would. The new mask ensures
that both shifts are needed.
f26 is specifically testing for a wrap-around mask. The AND can be applied
to just the shift left, either before or after the shift. Again, CodeGen
kept it in the original form but InstCombine would mask after the shift
instead. The exact choice of NILF isn't important for the test so I just
dropped it and kept the rotate.
llvm-svn: 199115
...into (ashr (shl (anyext X), ...), ...), which requires one fewer
instruction. The (anyext X) can sometimes be simplified too.
I didn't do this in DAGCombiner because widening shifts isn't a win
on all targets.
llvm-svn: 199114
The zext handling added in r197802 wasn't right for RNSBG. This patch
restricts it to ROSBG, RXSBG and RISBG. (The tests for RISBG were added
in r197802 since RISBG was the motivating example.)
llvm-svn: 198862
At the moment we expect rotates to have the form:
(or (shl X, Y), (shr X, Z))
where Y == bitsize(X) - Z or Z == bitsize(X) - Y. This form means that
the (or ...) is undefined for Y == 0 or Z == 0. This undefinedness can
be avoided by using Y == (C * bitsize(X) - Z) & (bitsize(X) - 1) or
Z == (C * bitsize(X) - Y) & (bitsize(X) - 1) for any integer C
(including 0, the most natural choice).
llvm-svn: 198861
InstCombine converts (sub 32, (add X, C)) into (sub 32-C, X),
so a rotate left of a 32-bit Y by X+C could appear as either:
(or (shl Y, (add X, C)), (shr Y, (sub 32, (add X, C))))
without InstCombine or:
(or (shl Y, (add X, C)), (shr Y, (sub 32-C, X)))
with it.
We already matched the first form. This patch handles the second too.
llvm-svn: 198860
...namely LOAD AND ADD, LOAD AND AND, LOAD AND OR and LOAD AND EXCLUSIVE OR.
LOAD AND ADD LOGICAL isn't really separately useful for LLVM.
I'll look at adding reusing the CC results in new year.
llvm-svn: 197985
If the extension of a loaded value is compared against zero and used in
other arithmetic, InstCombine will change the comparison to use the
unextended load. It's also possible that the comparison could be against
the unextended load from the outset.
In DAG form this becomes a truncation of an extending load. We want to
strip the truncation if possible so that we can use load-and-test instructions.
llvm-svn: 197804
The handling of ANY_EXTEND and ZERO_EXTEND was too strict. In this context
we can treat ZERO_EXTEND in much the same way as an AND and then also handle
outermost ZERO_EXTENDs.
I couldn't find a test that benefited from the ANY_EXTEND change, but it's
more obvious to write it this way once SIGN_EXTEND and ZERO_EXTEND are
handled differently.
llvm-svn: 197802
This reverts commit r197466.
The MachineCSE fix that required the -mcpu flag has been disabled
until more work can be done to fix downstream issues. Adding -mcpu
wasn't the right workaround anyway.
llvm-svn: 197624
This originally came about after noticing that InstCombine turns
some of the TMHH (icmp (and...), ...) tests into plain comparisons.
Since there is no instruction to compare with a 64-bit immediate,
TMHH is generally better than an ordered comparison for the cases
that it can handle.
llvm-svn: 197238
This patch makes more use of LPGFR and LNGFR. It builds on top of
the LTGFR selection from r197234. Most of the tests are motivated
by what InstCombine would produce.
llvm-svn: 197236
InstCombine turns (sext (trunc)) into (ashr (shl)), then converts any
comparison of the ashr against zero into a comparison of the shl against zero.
This makes sense in itself, but we want to undo it for z, since the sign-
extension instruction has a CC-setting form.
I've included tests for both the original and InstCombined variants,
but the former already worked. The patch fixes the latter.
llvm-svn: 197234
DAGCombiner could fold (truncate (load)) -> smaller load if the original
load was the width of the truncation result or wider. This patch extends
it to handle cases where the original load was narrower (and so the
extension type stays the same).
llvm-svn: 197030
One unusual feature of the z architecture is that the result of a
previous load can be reused indefinitely for subsequent loads, even if
a cache-coherent store to that location is performed by another CPU.
A special serializing instruction must be used if you want to force
a load to be reattempted.
Since volatile loads are not supposed to be omitted in this way,
we should insert a serializing instruction before each such load.
The same goes for atomic loads.
The patch implements this at the IR->DAG boundary, in a similar way
to atomic fences. It is a no-op for targets other than SystemZ.
llvm-svn: 196906
One unusual feature of the z architecture is that the result of a
previous load can be reused indefinitely for subsequent loads, even if
a cache-coherent store to that location is performed by another CPU.
A special serializing instruction must be used if you want to force
a load to be reattempted.
Since volatile loads are not supposed to be omitted in this way,
we should insert a serializing instruction before each such load.
The same goes for atomic loads.
The patch implements this at the IR->DAG boundary, in a similar way
to atomic fences. It is a no-op for targets other than SystemZ.
llvm-svn: 196905
Since z has no setcc instruction as such, the choice of setBooleanContents
is a bit arbitrary. Currently it's set to ZeroOrOneBooleanContent,
so we produced a branch-free form when selecting between 0 and 1,
but not when selecting between 0 and -1. This patch handles the latter
case too.
At some point I'd like to measure whether it's better to use conditional
moves for constant selects on z196, but that's future work.
llvm-svn: 196578
The backend converts 64-bit ORs into subreg moves if the upper 32 bits
of one operand and the low 32 bits of the other are known to be zero.
It then tries to peel away redundant ANDs from the upper 32 bits.
Since AND masks are canonicalized to exclude known-zero bits,
the test ORs the mask and the known-zero bits together before
checking for redundancy. The problem was that it was using the
wrong node when checking for known-zero bits, so could drop ANDs
that were still needed.
llvm-svn: 196267
We would wrongly transform the testcase into the equivalent of an AND with 1.
The problem was that, when testing whether the shifted-in bits of the right
shift were significant, we used the width of the final zero-extended result
rather than the width of the shifted value.
llvm-svn: 195731
I've no idea why I decided to handle TMxx differently from all the other
high/low logic operations, but it was a stupid thing to do. The high
registers aren't available as separate 32-bit registers on z10,
so subreg_h32 can't be used on a GR64 there.
I've normally been testing with z196 and with -O3 and so hadn't noticed
this until now.
llvm-svn: 195473
As on other hosts, the CPU identification instruction is priveleged,
so we need to look through /proc/cpuinfo. I copied the PowerPC way of
handling "generic".
Several tests were implicitly assuming z10 and so failed on z196.
llvm-svn: 193742
useAA significantly improves the handling of vector code that has TBAA
information attached. It also helps other cases, as shown by the testsuite
changes here. The only real downside I've seen is that it interferes with
MergeConsecutiveStores. The problem is that that optimization works top
down, starting at the first store in the chain, and looks for cases where
the chain result is only used by a single related store. These related
stores don't alias, so useAA will have rewritten all the later stores to
use a different chain input (typically the same one as the first store).
I think the advantages outweigh the disadvantages though, so for now I've
just disabled alias analysis for the unaligned-01.ll test.
llvm-svn: 193521
Making useAA() default to true for SystemZ showed that the combiner alias
analysis wasn't handling volatile accesses. This hit many of the SystemZ
tests, but I arbitrarily picked one for the purpose of this patch.
llvm-svn: 193518
Most SelectionDAG code drops the TBAA info when creating a new form of a
load and store (e.g. during legalization, or when converting a plain
load to an extending one). This patch tries to catch all cases where
the TBAA information can legitimately be carried over.
The patch adds alternative forms of getLoad() and getExtLoad() that take
a MachineMemOperand instead of individual fields. (The corresponding
getTruncStore() already exists.) The idea is to use the MachineMemOperand
forms when all fields are carried over (size, pointer info, isVolatile,
isNonTemporal, alignment and TBAA info). If some adjustment is being
made, e.g. to narrow the load, then we still pass the individual fields
but also pass the TBAA info.
llvm-svn: 193517
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
This is really an extension of the current (shl (shr ...)) -> shl optimization.
The main difference is that certain upper bits must also not be demanded.
The motivating examples are the first two in the testcase, which occur
in llvmpipe output.
llvm-svn: 192783
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
Remove the command line argument "struct-path-tbaa" since we should not depend
on command line argument to decide which format the IR file is using. Instead,
we check the first operand of the tbaa tag node, if it is a MDNode, we treat
it as struct-path aware TBAA format, otherwise, we treat it as scalar TBAA
format.
When clang starts to use struct-path aware TBAA format no matter whether
struct-path-tbaa is no, and we can auto-upgrade existing bc files, the support
for scalar TBAA format can be dropped.
Existing testing cases are updated to use the struct-path aware TBAA format.
llvm-svn: 191538
The backend tries to use block operations like MVC, NC, OC and XC for
simple scalar operations. For correctness reasons, it rejects any case
in which the regions might partially overlap. However, for performance
reasons, it should also reject cases where the regions might be equal,
since the instruction might then not use the fast path.
This fixes a performance regression seen in bzip2. We may want to limit
the optimisation even more in future, or even remove it entirely, but I'll
try with this for now.
llvm-svn: 191525
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
For some reason I never got around to adding these at the same time as
the signed versions. No idea why.
I'm not sure whether this SystemZII::BranchC* stuff is useful, or whether
it should just be replaced with an "is normal" flag. I'll leave that
for later though.
There are some boundary conditions that can be tweaked, such as preferring
unsigned comparisons for equality with [128, 256), and "<= 255" over "< 256",
but again I'll leave those for a separate patch.
llvm-svn: 190930
The port originally had special patterns for extload, mapping them to the
same instructions as sextload. It seemed neater to have patterns that
match "an extension that is allowed to be signed" and "an extension that
is allowed to be unsigned".
This was originally meant to be a clean-up, but it does improve the handling
of promoted integers a little, as shown by args-06.ll.
llvm-svn: 190777
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
If we had a store of an integer to memory, and the integer and store size
were suitable for a form of MV..., we used MV... no matter what. We could
then have sequences like:
lay %r2, 0(%r3,%r4)
mvi 0(%r2), 4
In these cases it seems better to force the constant into a register
and use a normal store:
lhi %r2, 4
stc %r2, 0(%r3, %r4)
since %r2 is more likely to be hoisted and is easier to rematerialize.
llvm-svn: 189098
...so that it can be used for z too. Most of the code is the same.
The only real change is to use TargetTransformInfo to test when a sqrt
instruction is available.
The pass is opt-in because at the moment it only handles sqrt.
llvm-svn: 189097
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
Post-RA LICM keeps three sets of registers: PhysRegDefs, PhysRegClobbers
and TermRegs. When it sees a definition of R it adds all aliases of R
to the corresponding set, so that when it needs to test for membership
it only needs to test a single register, rather than worrying about
aliases there too. E.g. the final candidate loop just has:
unsigned Def = Candidates[i].Def;
if (!PhysRegClobbers.test(Def) && ...) {
to test whether register Def is multiply defined.
However, there was also a shortcut in ProcessMI to make sure we didn't
add candidates if we already knew that they would fail the final test.
This shortcut was more pessimistic than the final one because it
checked whether _any alias_ of the defined register was multiply defined.
This is too conservative for targets that define register pairs.
E.g. on z, R0 and R1 are sometimes used as a pair, so there is a
128-bit register that aliases both R0 and R1. If a loop used
R0 and R1 independently, and the definition of R0 came first,
we would be able to hoist the R0 assignment (because that used
the final test quoted above) but not the R1 assignment (because
that meant we had two definitions of the paired R0/R1 register
and would fail the shortcut in ProcessMI).
This patch just uses the same check for the ProcessMI shortcut as
we use in the final candidate loop.
llvm-svn: 188774
For now this matches the equivalent of (neg (abs ...)), which did hit a few
times in projects/test-suite. We should probably also match cases where
absolute-like selects are used with reversed arguments.
llvm-svn: 188671
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
Generalize r188163 to cope with return types other than MVT::i32, just
as the existing visitMemCmpCall code did. I've split this out into a
subroutine so that it can be used for other upcoming patches.
I also noticed that I'd used the wrong API to record the out chain.
It's a load that uses DAG.getRoot() rather than getRoot(), so the out
chain should go on PendingLoads. I don't have a testcase for that because
we don't do any interesting scheduling on z yet.
llvm-svn: 188540
r188163 used CLC to implement memcmp. Code that compares the result
directly against zero can test the CC value produced by CLC, but code
that needs an integer result must use IPM. The sequence I'd used was:
ipm <reg>
sll <reg>, 2
sra <reg>, 30
but I'd forgotten that this inverts the order, so that CC==1 ("less")
becomes an integer greater than zero, and CC==2 ("greater") becomes
an integer less than zero. This sequence should only be used if the
CLC arguments are reversed to compensate. The problem then is that
the branch condition must also be reversed when testing the CLC
result directly.
Rather than do that, I went for a different sequence that works with
the natural CLC order:
ipm <reg>
srl <reg>, 28
rll <reg>, <reg>, 31
One advantage of this is that it doesn't clobber CC. A disadvantage
is that any sign extension to 64 bits must be done separately,
rather than being folded into the shifts.
llvm-svn: 188538
- Instead of setting the suffixes in a bunch of places, just set one master
list in the top-level config. We now only modify the suffix list in a few
suites that have one particular unique suffix (.ml, .mc, .yaml, .td, .py).
- Aside from removing the need for a bunch of lit.local.cfg files, this enables
4 tests that were inadvertently being skipped (one in
Transforms/BranchFolding, a .s file each in DebugInfo/AArch64 and
CodeGen/PowerPC, and one in CodeGen/SI which is now failing and has been
XFAILED).
- This commit also fixes a bunch of config files to use config.root instead of
older copy-pasted code.
llvm-svn: 188513
This follows the same lines as the integer code. In the end it seemed
easier to have a second 4-bit mask in TSFlags to specify the compare-like
CC values. That eats one more TSFlags bit than adding a CCHasUnordered
would have done, but it feels more concise.
llvm-svn: 187883
This patch just uses a peephole test for "add; compare; branch" sequences
within a single block. The IR optimizers already convert loops to
decrement-and-branch-on-nonzero form in some cases, so even this
simplistic test triggers many times during a clang bootstrap and
projects/test-suite run. It looks like there are still cases where we
need to more strongly prefer branches on nonzero though. E.g. I saw a
case where a loop that started out with a check for 0 ended up with a
check for -1. I'll try to look at that sometime.
I ended up adding the Reference class because MachineInstr::readsRegister()
doesn't check for subregisters (by design, as far as I could tell).
llvm-svn: 187723
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
Extend r187495 to conditional loads. I split this out because the
easiest way seemed to be to force a particular operand order in
SystemZISelDAGToDAG.cpp.
llvm-svn: 187496
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
r186399 aggressively used the RISBG instruction for immediate ANDs,
both because it can handle some values that AND IMMEDIATE can't,
and because it allows the destination register to be different from
the source. I realized later while implementing the distinct-ops
support that it would be better to leave the choice up to
convertToThreeAddress() instead. The AND IMMEDIATE form is shorter
and is less likely to be cracked.
This is a problem for 32-bit ANDs because we assume that all 32-bit
operations will leave the high word untouched, whereas RISBG used in
this way will either clear the high word or copy it from the source
register. The patch uses the z196 instruction RISBLG for this instead.
This means that z10 will be restricted to NILL, NILH and NILF for
32-bit ANDs, but I think that should be OK for now. Although we're
using z10 as the base architecture, the optimization work is going
to be focused more on z196 and zEC12.
llvm-svn: 187492
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
The atomic tests assume the two-operand forms, so I've restricted them to z10.
Running and-01.ll, or-01.ll and xor-01.ll for z196 as well as z10 shows why
using convertToThreeAddress() is better than exposing the three-operand forms
first and then converting back to two operands where possible (which is what
I'd originally tried). Using the three-operand form first stops us from
taking advantage of NG, OG and XG for spills.
llvm-svn: 186683
The original code only folded SRA into ROTATE ... SELECTED BITS
if there was no outer shift. This patch splits out that check
and generalises it slightly. The extra cases aren't really that
interesting, but this is paving the way for RNSBG support.
llvm-svn: 186571
Another patch in the series to make more use of R.SBG. This one extends
r186072 and r186073 to handle cases where the AND is inside the shift.
llvm-svn: 186399
This update was done with the following bash script:
find test/CodeGen -name "*.ll" | \
while read NAME; do
echo "$NAME"
if ! grep -q "^; *RUN: *llc.*debug" $NAME; then
TEMP=`mktemp -t temp`
cp $NAME $TEMP
sed -n "s/^define [^@]*@\([A-Za-z0-9_]*\)(.*$/\1/p" < $NAME | \
while read FUNC; do
sed -i '' "s/;\(.*\)\([A-Za-z0-9_-]*\):\( *\)$FUNC: *\$/;\1\2-LABEL:\3$FUNC:/g" $TEMP
done
sed -i '' "s/;\(.*\)-LABEL-LABEL:/;\1-LABEL:/" $TEMP
sed -i '' "s/;\(.*\)-NEXT-LABEL:/;\1-NEXT:/" $TEMP
sed -i '' "s/;\(.*\)-NOT-LABEL:/;\1-NOT:/" $TEMP
sed -i '' "s/;\(.*\)-DAG-LABEL:/;\1-DAG:/" $TEMP
mv $TEMP $NAME
fi
done
llvm-svn: 186280
Normal (sext (setcc ...)) sequences are optimised into
(select_cc ..., -1, 0) by DAGCombiner::visitSIGN_EXTEND.
However, this is deliberately not done for vectors, and after
vector type legalization we have (sext_inreg (setcc ...)) instead.
I wondered about trying to extend DAGCombiner to handle this case too,
but it seemed to be a loss on some other targets I tried, even those for
which SETCC isn't "legal" and SELECT_CC is.
llvm-svn: 186149
If the source of these instructions is spilled we should load the destination.
If the destination is spilled we should store the source.
llvm-svn: 186147
RISBG can handle some ANDs for which no AND IMMEDIATE exists.
It also acts as a three-operand AND for some cases where an
AND IMMEDIATE could be used instead.
It might be worth adding a pass to replace RISBG with AND IMMEDIATE
in cases where the register operands end up being the same and where
AND IMMEDIATE is smaller.
llvm-svn: 186072
Look for patterns of the form (store (load ...), ...) in which the two
locations are known not to partially overlap. (Identical locations are OK.)
These sequences are better implemented by MVC unless either the load or
the store could use RELATIVE LONG instructions.
The testcase showed that we weren't using LHRL and LGHRL for extload16,
only sextloadi16. The patch fixes that too.
llvm-svn: 185919
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
The stack coloring pass has code to delete stores and loads that become
trivially dead after coloring. Extend it to cope with single instructions
that copy from one frame index to another.
The testcase happens to show an example of this kicking in at the moment.
It did occur in Real Code too though.
llvm-svn: 185705
The stack coloring pass renumbered frame indexes with a loop of the form:
for each frame index FI
for each instruction I that uses FI
for each use of FI in I
rename FI to FI'
This caused problems if an instruction used two frame indexes F0 and F1
and if F0 was renamed to F1 and F1 to F2. The first time we visited the
instruction we changed F0 to F1, then we changed both F1s to F2.
In other words, the problem was that SSRefs recorded which instructions
used an FI, but not which MachineOperands and MachineMemOperands within
that instruction used it.
This is easily fixed for MachineOperands by walking the instructions
once and processing each operand in turn. There's already a loop to
do that for dead store elimination, so it seemed more efficient to
fuse the two at the block level.
MachineMemOperands are more tricky because they can be shared between
instructions. The patch handles them by making SSRefs an array of
MachineMemOperands rather than an array of MachineInstrs. We might end
up processing the same MachineMemOperand twice, but that's OK because
we always know from the SSRefs index what the original frame index was.
llvm-svn: 185703
...now that the problem that prompted the restriction has been fixed.
The original spill-02.py was a compromise because at the time I couldn't
find an example that actually failed without the two scavenging slots.
The version included here did.
llvm-svn: 185701
This is another prerequisite for frame-to-frame MVC copies.
I'll commit the patch that makes use of the slot separately.
The downside of trying to test many corner cases with each of the
available addressing modes is that a fair few tests need to account
for the new frame layout. I do still think it's useful to have all
these tests though, since it's something that wouldn't get much coverage
otherwise.
llvm-svn: 185698
Add a mapping from register-based <INSN>R instructions to the corresponding
memory-based <INSN>. Use it to cut down on the number of spill loads.
Some instructions extend their operands from smaller fields, so this
required a new TSFlags field to say how big the unextended operand is.
This optimisation doesn't trigger for C(G)R and CL(G)R because in practice
we always combine those instructions with a branch. Adding a test for every
other case probably seems excessive, but it did catch a missed optimisation
for DSGF (fixed in r185435).
llvm-svn: 185529
Fixes some cases where we were using full 64-bit division for (sdiv i32, i32)
and (sdiv i64, i32).
The "32" in "SDIVREM32" just refers to the second operand. The first operand
of all *DIVREM*s is a GR128.
llvm-svn: 185435
Try to use MVC when spilling the destination of a simple load or the source
of a simple store. As explained in the comment, this doesn't yet handle
the case where the load or store location is also a frame index, since
that could lead to two simultaneous scavenger spills, something the
backend can't handle yet. spill-02.py tests that this restriction kicks in,
but unfortunately I've not yet found a case that would fail without it.
The volatile trick I used for other scavenger tests doesn't work here
because we can't use MVC for volatile accesses anyway.
I'm planning on relaxing the restriction later, hopefully with a test
that does trigger the problem...
Tests @f8 and @f9 also showed that L(G)RL and ST(G)RL were wrongly
classified as SimpleBDX{Load,Store}. It wouldn't be easy to test for
that bug separately, which is why I didn't split out the fix as a
separate patch.
llvm-svn: 185434
Add pseudo conditional store instructions, so that we use:
branch foo:
store
foo:
instead of:
load
branch foo:
move
foo:
store
z196 has real 32-bit and 64-bit conditional stores, but we don't use
any z196 instructions yet.
llvm-svn: 185065
Unlike most -- hopefully "all other", but I'm still checking -- memory
instructions we support, LOAD REVERSED and STORE REVERSED may access
the memory location several times. This means that they are not suitable
for volatile loads and stores.
This patch is a prerequisite for better atomic load and store support.
The same principle applies there: almost all memory instructions we
support are inherently atomic ("block concurrent"), but LOAD REVERSED
and STORE REVERSED are exceptions.
Other instructions continue to allow volatile operands. I will add
positive "allows volatile" tests at the same time as the "allows atomic
load or store" tests.
llvm-svn: 183002
The code to distinguish between unaligned and aligned addresses was
already there, so this is mostly just a switch-on-and-test process.
llvm-svn: 182920
This patch adds support for the CRJ and CGRJ instructions. Support for
the immediate forms will be a separate patch.
The architecture has a large number of comparison instructions. I think
it's generally better to concentrate on using the "best" comparison
instruction first and foremost, then only use something like CRJ if
CR really was the natual choice of comparison instruction. The patch
therefore opportunistically converts separate CR and BRC instructions
into a single CRJ while emitting instructions in ISelLowering.
llvm-svn: 182764
Before this change, the SystemZ backend would use BRCL for all branches
and only consider shortening them to BRC when generating an object file.
E.g. a branch on equal would use the JGE alias of BRCL in assembly output,
but might be shortened to the JE alias of BRC in ELF output. This was
a useful first step, but it had two problems:
(1) The z assembler isn't traditionally supposed to perform branch shortening
or branch relaxation. We followed this rule by not relaxing branches
in assembler input, but that meant that generating assembly code and
then assembling it would not produce the same result as going directly
to object code; the former would give long branches everywhere, whereas
the latter would use short branches where possible.
(2) Other useful branches, like COMPARE AND BRANCH, do not have long forms.
We would need to do something else before supporting them.
(Although COMPARE AND BRANCH does not change the condition codes,
the plan is to model COMPARE AND BRANCH as a CC-clobbering instruction
during codegen, so that we can safely lower it to a separate compare
and long branch where necessary. This is not a valid transformation
for the assembler proper to make.)
This patch therefore moves branch relaxation to a pre-emit pass.
For now, calls are still shortened from BRASL to BRAS by the assembler,
although this too is not really the traditional behaviour.
The first test takes about 1.5s to run, and there are likely to be
more tests in this vein once further branch types are added. The feeling
on IRC was that 1.5s is a bit much for a single test, so I've restricted
it to SystemZ hosts for now.
The patch exposes (and fixes) some typos in the main CodeGen/SystemZ tests.
A later patch will remove the {{g}}s from that directory.
llvm-svn: 182274
Duncan P. N. Exon Smith