If KMOVB not supported (require AVX512DQ) only KMOVW can be used so store size should be 2 bytes.
Differential Revision: http://reviews.llvm.org/D17138
llvm-svn: 260878
As discussed on PR26491, this patch adds support for lowering v4f32 shuffles to the MOVLHPS/MOVHLPS instructions. It also adds support for memory folding with their MOVLPS/MOVHPS load equivalents.
This first patch only really helps SSE1 targets as SSE2+ targets will widen the shuffle mask and use v2f64 equivalents (although they still combine to MOVLHPS/MOVHLPS for v2f64 splats). This will have to be addressed in a future patch, most likely when we add support for binary target shuffle combines.
Differential Revision: http://reviews.llvm.org/D16956
llvm-svn: 260168
Fix a crash in `getMemOpBaseRegImmOfs` that happens if the base of
`MemOp` is a frame index memory operand. The fix is to have
`getMemOpBaseRegImmOfs` bail out in such cases. We can possibly be more
clever here, if needed.
llvm-svn: 259456
Currently, AnalyzeBranch() fails non-equality comparison between floating points
on X86 (see https://llvm.org/bugs/show_bug.cgi?id=23875). This is because this
function can modify the branch by reversing the conditional jump and removing
unconditional jump if there is a proper fall-through. However, in the case of
non-equality comparison between floating points, this can turn the branch
"unanalyzable". Consider the following case:
jne.BB1
jp.BB1
jmp.BB2
.BB1:
...
.BB2:
...
AnalyzeBranch() will reverse "jp .BB1" to "jnp .BB2" and then "jmp .BB2" will be
removed:
jne.BB1
jnp.BB2
.BB1:
...
.BB2:
...
However, AnalyzeBranch() cannot analyze this branch anymore as there are two
conditional jumps with different targets. This may disable some optimizations
like block-placement: in this case the fall-through behavior is enforced even if
the fall-through block is very cold, which is suboptimal.
Actually this optimization is also done in block-placement pass, which means we
can remove this optimization from AnalyzeBranch(). However, currently
X86::COND_NE_OR_P and X86::COND_NP_OR_E are not reversible: there is no defined
negation conditions for them.
In order to reverse them, this patch defines two new CondCode X86::COND_E_AND_NP
and X86::COND_P_AND_NE. It also defines how to synthesize instructions for them.
Here only the second conditional jump is reversed. This is valid as we only need
them to do this "unconditional jump removal" optimization.
Differential Revision: http://reviews.llvm.org/D11393
llvm-svn: 258847
The red zone consists of 128 bytes beyond the stack pointer so that the
allocation of objects in leaf functions doesn't require decrementing
rsp. In r255656, we introduced an optimization that would cheaply
materialize certain constants via push/pop. Push decrements the stack
pointer and stores it's result at what is now the top of the stack.
However, this means that using push/pop would encroach on the red zone.
PR26023 gives an example where this corrupts an object in the red zone.
llvm-svn: 256808
Unfortunately this fix had the effect of exposing the
-verify-machineinstrs FIXME of X86InstrInfo.cpp in two testcases for
which I disabled it for now.
Two testcases also have additional pushq/popq where the corrected code
cannot prove that %rax is dead any longer. Looking at the examples, this
could potentially be fixed by improving computeRegisterLiveness() to check
the live-in lists of the successors blocks when reaching the end of a
block.
This fixes http://llvm.org/PR25951.
llvm-svn: 256799
We need a frame pointer if there is a push/pop sequence after the
prologue in order to unwind the stack. Scanning the instructions to
figure out if this happened made hasFP not constant-time which is a
violation of expectations. Let's compute this up-front and reuse that
computation when we need it.
llvm-svn: 256730
A frame pointer must be used if stack pointer is modified after the
prologue. LLVM will emit pushf/popf if we need to save/restore the
FLAGS register, requiring us to have a frame pointer for the function.
There is a small twist: this sequence might exist in user code via
inline-assembly. For now, conservatively assume that such functions
require a frame pointer. For real world justification, please see
clang's implementation of __readeflags.
This fixes PR25945.
llvm-svn: 256456
The patterns that set a mask register to 0/1
KXOR %kn, %kn, %kn / KXNOR %kn, %kn, %kn
are replaced with
KXOR %k0, %k0, %kn / KXNOR %k0, %k0, %kn - AVX-512 targets optimization.
KNL does not recognize dependency-breaking idioms for mask registers,
so kxnor %k1, %k1, %k2 has a RAW dependence on %k1.
Using %k0 as the undef input register is a performance heuristic based
on the assumption that %k0 is used less frequently than the other mask
registers, since it is not usable as a write mask.
Differential Revision: http://reviews.llvm.org/D15739
llvm-svn: 256365
Use the 3-byte (4 with REX prefix) push-pop sequence for materializing
small constants. This is smaller than using a mov (5, 6 or 7 bytes
depending on size and REX prefix), but it's likely to be slower, so
only used for 'minsize'.
This is a follow-up to r255656.
Differential Revision: http://reviews.llvm.org/D15549
llvm-svn: 255936
"movl $-1, %eax" is 5 bytes, "xorl %eax, %eax; decl %eax" is 3 bytes.
This commit makes LLVM use the latter when optimizing for size.
Differential Revision: http://reviews.llvm.org/D14971
llvm-svn: 255656
computeRegisterLiveness() was broken in that it reported dead for a
register even if a subregister was alive. I assume this was because the
results of analayzePhysRegs() are hard to understand with respect to
subregisters.
This commit: Changes the results of analyzePhysRegs (=struct
PhysRegInfo) to be clearly understandable, also renames the fields to
avoid silent breakage of third-party code (and improve the grammar).
Fix all (two) users of computeRegisterLiveness() in llvm: By reenabling
it and removing workarounds for the bug.
This fixes http://llvm.org/PR24535 and http://llvm.org/PR25033
Differential Revision: http://reviews.llvm.org/D15320
llvm-svn: 255362
These instructions are not supported by all CPUs in 64-bit mode. Emitting them
causes Chromium to crash on start-up for users with such chips.
(GCC puts these instructions behind -msahf on 64-bit for the same reason.)
This patch adds FeatureLAHFSAHF, enables it by default for 32-bit targets
and modern CPUs, and changes X86InstrInfo::copyPhysReg back to the lowering
from before r244503 when the instructions are not available.
Differential Revision: http://reviews.llvm.org/D15240
llvm-svn: 254793
Summary:
computeRegisterLiveness and analyzePhysReg are currently getting
confused about liveness in some cases, breaking copyPhysReg's
calculation of whether AX is dead in some cases. Work around this issue
temporarily by assuming that AX is always live.
See detail in: https://llvm.org/bugs/show_bug.cgi?id=25033#c7
And associated bugs PR24535 PR25033 PR24991 PR24992 PR25201.
This workaround makes the code correct but slightly inefficient, but it
seems to confuse the machine instr verifier which now things EAX was
undefined in some cases where it's being conservatively saved /
restored.
Reviewers: majnemer, sanjoy
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D15198
llvm-svn: 254680
generated for _mm_losd_s{s,d}() intrinsics and used in scalar FMAs generated
for FMA intrinsics _mm_f{madd,msub,nmadd,nmsub}_s{s,d}().
Reviewer: David Kreitzer
Differential Revision: http://reviews.llvm.org/D14762
llvm-svn: 254140
We had duplicated definitions for the same hardware '[v]movq' instructions. For example with SSE:
def MOVZQI2PQIrr : RS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
"mov{d|q}\t{$src, $dst|$dst, $src}", // X86-64 only
[(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 (scalar_to_vector GR64:$src)))))],
IIC_SSE_MOVDQ>;
def MOV64toPQIrr : RS2I<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
"mov{d|q}\t{$src, $dst|$dst, $src}",
[(set VR128:$dst, (v2i64 (scalar_to_vector GR64:$src)))],
IIC_SSE_MOVDQ>, Sched<[WriteMove]>;
As shown in the test case and PR25554:
https://llvm.org/bugs/show_bug.cgi?id=25554
This causes us to miss reusing an operand because later passes don't know these 'movq' are the same instruction.
This patch deletes one pair of these defs.
Sadly, this won't fix the original test case in the bug report. Something else is still broken.
Differential Revision: http://reviews.llvm.org/D14941
llvm-svn: 253988
Copying one mask register to another under BW should be done with kmovq instruction, otherwise we can loose some bits.
Copying 8 bits under DQ may be done with kmovb.
Differential Revision: http://reviews.llvm.org/D14812
llvm-svn: 253563
It made it possible to apply the memory folding optimization for the 2nd
operand of FMA*_Int instructions.
Reviewer: Quentin Colombet
Differential Revision: http://reviews.llvm.org/D14550
llvm-svn: 252973
All 3 operands of FMA3 instructions are commutable now.
Patch by Slava Klochkov
Reviewers: Quentin Colombet(qcolombet), Ahmed Bougacha(ab).
Differential Revision: http://reviews.llvm.org/D13269
llvm-svn: 252335
This patch improves the memory folding of the inserted float element for the (V)INSERTPS instruction.
The existing implementation occurs in the DAGCombiner and relies on the narrowing of a whole vector load into a scalar load (and then converted into a vector) to (hopefully) allow folding to occur later on. Not only has this proven problematic for debug builds, it also prevents other memory folds (notably stack reloads) from happening.
This patch removes the old implementation and moves the folding code to the X86 foldMemoryOperand handler. A new private 'special case' function - foldMemoryOperandCustom - has been added to deal with memory folding of instructions that can't just use the lookup tables - (V)INSERTPS is the first of several that could be done.
It also tweaks the memory operand folding code with an additional pointer offset that allows existing memory addresses to be modified, in this case to convert the vector address to the explicit address of the scalar element that will be inserted.
Unlike the previous implementation we now set the insertion source index to zero, although this is ignored for the (V)INSERTPSrm version, anything that relied on shuffle decodes (such as unfolding of insertps loads) was incorrectly calculating the source address - I've added a test for this at insertps-unfold-load-bug.ll
Differential Revision: http://reviews.llvm.org/D13988
llvm-svn: 252074
Patch by Slava Klochkov
The key difference between FMA* and FMA*_Int opcodes is that FMA*_Int opcodes are handled more conservatively. It is illegal to commute the 1st operand of FMA*_Int instructions as the upper bits of scalar FMA intrinsic result must be taken from the 1st operand, but such commute transformation would change those upper bits and invalidate the intrinsic's result.
Reviewers: Quentin Colombet, Elena Demikhovsky
Differential Revision: http://reviews.llvm.org/D13710
llvm-svn: 252060
Catchret transfers control from a catch funclet to an earlier funclet.
However, it is not completely clear which funclet the catchret target is
part of. Make this clear by stapling the catchret target's funclet
membership onto the CATCHRET SDAG node.
llvm-svn: 249052
AVX-512 does not provide an instruction that shuffles mask register. So I do the following way:
mask-2-simd , shuffle simd , simd-2-mask
Differential Revision: http://reviews.llvm.org/D12727
llvm-svn: 247876
The changes in:
test/CodeGen/X86/machine-cp.ll
are just due to scheduling differences after some logic instructions were reassociated.
llvm-svn: 247516
This takes the existing static function hasLiveCondCodeDef and makes it a member function of the X86InstrInfo class. This is a useful utility function that an upcoming change would like to use. NFC.
Patch by: Kevin B. Smith
Differential Revision: http://reviews.llvm.org/D12371
llvm-svn: 246073
This is a 'no functional change intended' patch. It removes one FIXME, but adds several more.
Motivation: the FeatureFastUAMem attribute may be too general. It is used to determine if any
sized misaligned memory access under 32-bytes is 'fast'. From the added FIXME comments, however,
you can see that we're not consistent about this. Changing the name of the attribute makes it
clearer to see the logic holes.
Changing this to a 'slow' attribute also means we don't have to add an explicit 'fast' attribute
to new chips; fast unaligned accesses have been standard for several generations of CPUs now.
Differential Revision: http://reviews.llvm.org/D12154
llvm-svn: 245729
This commit removes the global manager variable which is responsible for
storing and allocating pseudo source values and instead it introduces a new
manager class named 'PseudoSourceValueManager'. Machine functions now own an
instance of the pseudo source value manager class.
This commit also modifies the 'get...' methods in the 'MachinePointerInfo'
class to construct pseudo source values using the instance of the pseudo
source value manager object from the machine function.
This commit updates calls to the 'get...' methods from the 'MachinePointerInfo'
class in a lot of different files because those calls now need to pass in a
reference to a machine function to those methods.
This change will make it easier to serialize pseudo source values as it will
enable me to transform the mips specific MipsCallEntry PseudoSourceValue
subclass into two target independent subclasses.
Reviewers: Akira Hatanaka
llvm-svn: 244693
NaCl's sandbox doesn't allow PUSHF/POPF out of security concerns (priviledged emulators have forgotten to mask system bits in the past, and EFLAGS's DF bit is a constant source of hilarity). Commit r220529 fixed PR20376 by saving cmpxchg's flags result using EFLAGS, this commit now generated LAHF/SAHF instead, for all of x86 (not just NaCl) because it leads to an overall performance gain over PUSHF/POPF.
As with the previous patch this code generation is pretty bad because it occurs very later, after register allocation, and in many cases it rematerializes flags which were already available (e.g. already in a register through SETE). Fortunately it's somewhat rare that this code needs to fire.
I did [[ https://github.com/jfbastien/benchmark-x86-flags | a bit of benchmarking ]], the results on an Intel Haswell E5-2690 CPU at 2.9GHz are:
| Time per call (ms) | Runtime (ms) | Benchmark |
| 0.000012514 | 6257 | sete.i386 |
| 0.000012810 | 6405 | sete.i386-fast |
| 0.000010456 | 5228 | sete.x86-64 |
| 0.000010496 | 5248 | sete.x86-64-fast |
| 0.000012906 | 6453 | lahf-sahf.i386 |
| 0.000013236 | 6618 | lahf-sahf.i386-fast |
| 0.000010580 | 5290 | lahf-sahf.x86-64 |
| 0.000010304 | 5152 | lahf-sahf.x86-64-fast |
| 0.000028056 | 14028 | pushf-popf.i386 |
| 0.000027160 | 13580 | pushf-popf.i386-fast |
| 0.000023810 | 11905 | pushf-popf.x86-64 |
| 0.000026468 | 13234 | pushf-popf.x86-64-fast |
Clearly `PUSHF`/`POPF` are suboptimal. It doesn't really seems to be worth teaching LLVM about individual flags, at least not for this purpose.
Reviewers: rnk, jvoung, t.p.northover
Subscribers: llvm-commits
Differential revision: http://reviews.llvm.org/D6629
llvm-svn: 244503
This commit implements the initial serialization of the machine operand target
flags. It extends the 'TargetInstrInfo' class to add two new methods that help
to provide text based serialization for the target flags.
This commit can serialize only the X86 target flags, and the target flags for
the other targets will be serialized in the follow-up commits.
Reviewers: Duncan P. N. Exon Smith
llvm-svn: 244185
Create wrapper methods in the Function class for the OptimizeForSize and MinSize
attributes. We want to hide the logic of "or'ing" them together when optimizing
just for size (-Os).
Currently, we are not consistent about this and rely on a front-end to always set
OptimizeForSize (-Os) if MinSize (-Oz) is on. Thus, there are 18 FIXME changes here
that should be added as follow-on patches with regression tests.
This patch is NFC-intended: it just replaces existing direct accesses of the attributes
by the equivalent wrapper call.
Differential Revision: http://reviews.llvm.org/D11734
llvm-svn: 243994
In the commentary for D11660, I wasn't sure if it was alright to create new
integer machine instructions without also creating the implicit EFLAGS operand.
From what I can see, the implicit operand is always created by the MachineInstrBuilder
based on the instruction type, so we don't have to do that explicitly. However, in
reviewing the debug output, I noticed that the operand was not marked as 'dead'.
The machine combiner should do that to preserve future optimization opportunities
that may be checking for that dead EFLAGS operand themselves.
Differential Revision: http://reviews.llvm.org/D11696
llvm-svn: 243990
Add i16, i32, i64 imul machine instructions to the list of reassociation
candidates.
A new bit of logic is needed to handle integer instructions: they have an
implicit EFLAGS operand, so we have to make sure it's dead in order to do
any reassociation with integer ops.
Differential Revision: http://reviews.llvm.org/D11660
llvm-svn: 243756
This is a follow-up to the FIXME that was added with D7474 ( http://reviews.llvm.org/rL229531 ).
I thought this load folding bug had been made hard-to-hit, but it turns out to be very easy
when targeting 32-bit x86 and causes a miscompile/crash in Wine:
https://bugs.winehq.org/show_bug.cgi?id=38826https://llvm.org/bugs/show_bug.cgi?id=22371#c25
The quick fix is to simply remove the scalar FP logical instructions from the load folding table
in X86InstrInfo, but that causes us to miss load folds that should be possible when lowering fabs,
fneg, fcopysign. So the majority of this patch is altering those lowerings to use *vector* FP
logical instructions (because that's all x86 gives us anyway). That lets us do the load folding
legally.
Differential Revision: http://reviews.llvm.org/D11477
llvm-svn: 243361
Adds pushes to the folding tables.
This also required a fix to the TD definition, since the memory forms of
the push instructions did not have the right mayLoad/mayStore flags.
Differential Revision: http://reviews.llvm.org/D11340
llvm-svn: 243010
canFoldMemoryOperand is not actually used anywhere in the codebase - all existing users instead call foldMemoryOperand directly when they wish to fold and can correctly deduce what they need from the return value.
This patch removes the canFoldMemoryOperand base function and the target implementations; only x86 had a real (bit-rotted) implementation, although AMDGPU had a preparatory stub that had never needed to be completed.
Differential Revision: http://reviews.llvm.org/D11331
llvm-svn: 242638
MOVSDto64rr and MOV64toSDrr are defined to convert between FR64 (%xmm)
<-> GR64 registers, not VR64 (%mm) <-> GR64. This is wrong.
I found this by inspection and could not find a suitable testcase for it
since (1) we don't handle MMX bitcasts in Peephole optimizer as to
generate COPYs that (2) could be expanded back to the appropriate x86
instruction in ExpandPostRA.
Switch to use the appropriate instructions: MMX_MOVD64from64rr and
MMX_MOVD64to64rr here.
llvm-svn: 242191
Extend the reassociation optimization of http://reviews.llvm.org/rL240361 (D10460)
to SSE scalar FP SP adds in addition to AVX scalar FP SP adds.
With the 'switch' in place, we can trivially add other opcodes and test cases in
future patches.
Differential Revision: http://reviews.llvm.org/D10975
llvm-svn: 241515
Currently ( D10321, http://reviews.llvm.org/rL239486 ), we can use the machine combiner pass
to reassociate the following sequence to reduce the critical path:
A = ? op ?
B = A op X
C = B op Y
-->
A = ? op ?
B = X op Y
C = A op B
'op' is currently limited to x86 AVX scalar FP adds (with fast-math on), but in theory, it could
be any associative math/logic op (see TODO in code comment).
This patch generalizes the pattern match to ignore the instruction that defines 'A'. So instead of
a sequence of 3 adds, we now only need to find 2 dependent adds and decide if it's worth
reassociating them.
This generalization has a compile-time cost because we can now match more instruction sequences
and we rely more heavily on the machine combiner to discard sequences where reassociation doesn't
improve the critical path.
For example, in the new test case:
A = M div N
B = A add X
C = B add Y
We'll match 2 reassociation patterns, but this transform doesn't reduce the critical path:
A = M div N
B = A add Y
C = B add X
We need the combiner to reject that pattern but select this:
A = M div N
B = X add Y
C = B add A
Differential Revision: http://reviews.llvm.org/D10460
llvm-svn: 240361
The _Int instructions are special, in that they operate on the full
VR128 instead of FR32. The load folding then looks at MOVSS, at the
user, and bails out when it sees a size mismatch.
What we really know is that the rm_Int instructions don't load the
higher lanes, so folding is fine.
This happens for the straightforward intrinsic code, e.g.:
_mm_add_ss(a, _mm_load_ss(p));
Fixes PR23349.
Differential Revision: http://reviews.llvm.org/D10554
llvm-svn: 240326
The patch is generated using this command:
tools/clang/tools/extra/clang-tidy/tool/run-clang-tidy.py -fix \
-checks=-*,llvm-namespace-comment -header-filter='llvm/.*|clang/.*' \
llvm/lib/
Thanks to Eugene Kosov for the original patch!
llvm-svn: 240137
Summary:
NFC: no one uses AnalyzeBranchPredicate yet.
Add TargetInstrInfo::AnalyzeBranchPredicate and implement for x86. A
later change adding support for page-fault based implicit null checks
depends on this.
Reviewers: reames, ab, atrick
Reviewed By: atrick
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D10200
llvm-svn: 239742
Summary:
TargetInstrInfo::getLdStBaseRegImmOfs to
TargetInstrInfo::getMemOpBaseRegImmOfs and implement for x86. The
implementation only handles a few easy cases now and will be made more
sophisticated in the future.
This is NFCI: the only user of `getLdStBaseRegImmOfs` (now
`getmemOpBaseRegImmOfs`) is `LoadClusterMotion` and `LoadClusterMotion`
is disabled for x86.
Reviewers: reames, ab, MatzeB, atrick
Reviewed By: MatzeB, atrick
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D10199
llvm-svn: 239741
This will use Itinieraries if available, but will also work if just a
MCSchedModel is available.
Differential Revision: http://reviews.llvm.org/D10428
llvm-svn: 239658
This is a reimplementation of D9780 at the machine instruction level rather than the DAG.
Use the MachineCombiner pass to reassociate scalar single-precision AVX additions (just a
starting point; see the TODO comments) to increase ILP when it's safe to do so.
The code is closely based on the existing MachineCombiner optimization that is implemented
for AArch64.
This patch should not cause the kind of spilling tragedy that led to the reversion of r236031.
Differential Revision: http://reviews.llvm.org/D10321
llvm-svn: 239486
Summary:
This was a longstanding FIXME and is a necessary precursor to cases
where foldOperandImpl may have to create more than one instruction
(e.g. to constrain a register class). This is the split out NFC changes from
D6262.
Reviewers: pete, ributzka, uweigand, mcrosier
Reviewed By: mcrosier
Subscribers: mcrosier, ted, llvm-commits
Differential Revision: http://reviews.llvm.org/D10174
llvm-svn: 239336
Implemented DAG lowering for all these forms.
Added tests for DAG lowering and encoding.
Differential Revision: http://reviews.llvm.org/D10310
llvm-svn: 239300
Ahmed Bougacha