When the target does not support these intrinsics they should be converted to a chain of scalar load or store operations.
If the mask is not constant, the scalarizer will build a chain of conditional basic blocks.
I added isLegalMaskedGather() isLegalMaskedScatter() APIs.
Differential Revision: http://reviews.llvm.org/D13722
llvm-svn: 251237
When using the MCU psABI, compiler-generated library calls should pass
some parameters in-register. However, since inreg marking for x86 is currently
done by the front end, it will not be applied to backend-generated calls.
This is a workaround for PR3997, which describes a similar issue for -mregparm.
Differential Revision: http://reviews.llvm.org/D13977
llvm-svn: 251223
We don't need a mask of a rotation result to be a constant splat - any constant scalar/vector can be usefully folded.
Followup to D13851.
llvm-svn: 251197
This patch adds support for lowering to the XOP VPROT / VPROTI vector bit rotation instructions.
This has required changes to the DAGCombiner rotation pattern matching to support vector types - so far I've only changed it to support splat vectors, but generalising this further is feasible in the future.
Differential Revision: http://reviews.llvm.org/D13851
llvm-svn: 251188
Summary:
The logic here isn't straightforward because our support for
TargetOptions::GuaranteedTailCallOpt.
Also fix a bug where we were allowing tail calls to cdecl functions from
fastcall and vectorcall functions. We were special casing thiscall and
stdcall callers rather than checking for any convention that requires
clearing stack arguments before returning.
Reviewers: hans
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D14024
llvm-svn: 251137
Summary:
This ensures that BranchFolding (and similar) won't remove these blocks.
Also allow AsmPrinter::EmitBasicBlockStart to process MBBs which are
address-taken but do not have BBs that are address-taken, since otherwise
its call to getAddrLabelSymbolTableToEmit would fail an assertion on such
blocks. I audited the other callers of getAddrLabelSymbolTableToEmit
(and getAddrLabelSymbol); they all have BBs known to be address-taken
except for the call through getAddrLabelSymbol from
WinException::create32bitRef; that call is actually now unreachable, so
I've removed it and updated the signature of create32bitRef.
This fixes PR25168.
Reviewers: majnemer, andrew.w.kaylor, rnk
Subscribers: pgavlin, llvm-commits
Differential Revision: http://reviews.llvm.org/D13774
llvm-svn: 251113
When we fold "mul ((add x, c1), c1)" -> "add ((mul x, c2), c1*c2)", we bail if (add x, c1) has multiple
users which would result in an extra add instruction.
In such cases, this patch adds a check to see if we can eliminate a multiply instruction in exchange for the extra add.
I also added the capability of doing the existing optimization with non-splatted vectors (splatted also works).
Differential Revision: http://reviews.llvm.org/D13740
llvm-svn: 251028
When we have to convert the masked.load, masked.store to scalar code, we generate a chain of conditional basic blocks.
I added optimization for constant mask vector.
Differential Revision: http://reviews.llvm.org/D13855
llvm-svn: 250893
The mask value type for maskload/maskstore GCC builtins is never a vector of
packed floats/doubles.
This patch fixes the following issues:
1. The mask argument for builtin_ia32_maskloadpd and builtin_ia32_maskstorepd
should be of type llvm_v2i64_ty and not llvm_v2f64_ty.
2. The mask argument for builtin_ia32_maskloadpd256 and
builtin_ia32_maskstorepd256 should be of type llvm_v4i64_ty and not
llvm_v4f64_ty.
3. The mask argument for builtin_ia32_maskloadps and builtin_ia32_maskstoreps
should be of type llvm_v4i32_ty and not llvm_v4f32_ty.
4. The mask argument for builtin_ia32_maskloadps256 and
builtin_ia32_maskstoreps256 should be of type llvm_v8i32_ty and not
llvm_v8f32_ty.
Differential Revision: http://reviews.llvm.org/D13776
llvm-svn: 250817
Currently, in MachineBlockPlacement pass the loop is rotated to let the best exit to be the last BB in the loop chain, to maximize the fall-through from the loop to outside. With profile data, we can determine the cost in terms of missed fall through opportunities when rotating a loop chain and select the best rotation. Basically, there are three kinds of cost to consider for each rotation:
1. The possibly missed fall through edge (if it exists) from BB out of the loop to the loop header.
2. The possibly missed fall through edges (if they exist) from the loop exits to BB out of the loop.
3. The missed fall through edge (if it exists) from the last BB to the first BB in the loop chain.
Therefore, the cost for a given rotation is the sum of costs listed above. We select the best rotation with the smallest cost. This is only for PGO mode when we have more precise edge frequencies.
Differential revision: http://reviews.llvm.org/D10717
llvm-svn: 250754
Add FastISel support for SSE4A scalar float / double non-temporal stores
Follow up to D13698
Differential Revision: http://reviews.llvm.org/D13773
llvm-svn: 250610
Our previous value of "16 + 8 + MaxCallFrameSize" for ParentFrameOffset
is incorrect when CSRs are involved. We were supposed to have a test
case to catch this, but it wasn't very rigorous.
The main effect here is that calling _CxxThrowException inside a
catchpad doesn't immediately crash on MOVAPS when you have an odd number
of CSRs.
llvm-svn: 250583
The motivation for this patch starts with PR20134:
https://llvm.org/bugs/show_bug.cgi?id=20134
void foo(int *a, int i) {
a[i] = a[i+1] + a[i+2];
}
It seems better to produce this (14 bytes):
movslq %esi, %rsi
movl 0x4(%rdi,%rsi,4), %eax
addl 0x8(%rdi,%rsi,4), %eax
movl %eax, (%rdi,%rsi,4)
Rather than this (22 bytes):
leal 0x1(%rsi), %eax
cltq
leal 0x2(%rsi), %ecx
movslq %ecx, %rcx
movl (%rdi,%rcx,4), %ecx
addl (%rdi,%rax,4), %ecx
movslq %esi, %rax
movl %ecx, (%rdi,%rax,4)
The most basic problem (the first test case in the patch combines constants) should also be fixed in InstCombine,
but it gets more complicated after that because we need to consider architecture and micro-architecture. For
example, AArch64 may not see any benefit from the more general transform because the ISA solves the sexting in
hardware. Some x86 chips may not want to replace 2 ADD insts with 1 LEA, and there's an attribute for that:
FeatureSlowLEA. But I suspect that doesn't go far enough or maybe it's not getting used when it should; I'm
also not sure if FeatureSlowLEA should also mean "slow complex addressing mode".
I see no perf differences on test-suite with this change running on AMD Jaguar, but I see small code size
improvements when building clang and the LLVM tools with the patched compiler.
A more general solution to the sext(add nsw(x, C)) problem that works for multiple targets is available
in CodeGenPrepare, but it may take quite a bit more work to get that to fire on all of the test cases that
this patch takes care of.
Differential Revision: http://reviews.llvm.org/D13757
llvm-svn: 250560
Summary:
We now use the block for the catchpad itself, rather than its normal
successor, as the funclet entry.
Putting the normal successor in the map leads downstream funclet
membership computations to erroneous results.
Reviewers: majnemer, rnk
Subscribers: rnk, llvm-commits
Differential Revision: http://reviews.llvm.org/D13798
llvm-svn: 250552
D4796 taught LLVM to fold some atomic integer operations into a single
instruction. The pattern was unaware that the instructions clobbered
flags. I fixed some of this issue in D13680 but had missed INC/DEC.
This patch adds the missing EFLAGS definition.
llvm-svn: 250438
Summary:
x86 codegen is clever about generating good code for relaxed
floating-point operations, but it was being silly when globals and
immediates were involved, forgetting where the global was and
loading/storing from/to the wrong place. The same applied to hard-coded
address immediates.
Don't let it forget about the displacement.
This fixes https://llvm.org/bugs/show_bug.cgi?id=25171
A very similar bug when doing floating-points atomics to the stack is
also fixed by this patch.
This fixes https://llvm.org/bugs/show_bug.cgi?id=25144
Reviewers: pete
Subscribers: llvm-commits, majnemer, rsmith
Differential Revision: http://reviews.llvm.org/D13749
llvm-svn: 250429
AVX-512 bit shuffle fails on 32 bit since we create a vector of 64-bit constants.
I split 8x64-bit const vector to 16x32 on 32-bit mode.
Differential Revision: http://reviews.llvm.org/D13644
llvm-svn: 250390
Recommit r250342: move coal-sections-powerpc.s to subdirectory for powerpc.
Some background on why we don't have to use *coal* sections anymore:
Long ago when C++ was new and "weak" had not been standardized, an attempt was
made in cctools to support C++ inlines that can be coalesced by putting them
into their own section (TEXT/textcoal_nt instead of TEXT/text).
The current macho linker supports the weak-def bit on any symbol to allow it to
be coalesced, but the compiler still puts weak-def functions/data into alternate
section names, which the linker must map back to the base section name.
This patch makes changes that are necessary to prevent the compiler from using
the "coal" sections and have it use the non-coal sections instead when the
target architecture is not powerpc:
TEXT/textcoal_nt instead use TEXT/text
TEXT/const_coal instead use TEXT/const
DATA/datacoal_nt instead use DATA/data
If the target is powerpc, we continue to use the *coal* sections since anyone
targeting powerpc is probably using an old linker that doesn't have support for
the weak-def bits.
Also, have the assembler issue a warning if it encounters a *coal* section in
the assembly file and inform the users to use the non-coal sections instead.
rdar://problem/14265330
Differential Revision: http://reviews.llvm.org/D13188
llvm-svn: 250370
Recommit r250342: add -arch=ppc32 to the RUN lines of powerpc tests.
Some background on why we don't have to use *coal* sections anymore:
Long ago when C++ was new and "weak" had not been standardized, an attempt was
made in cctools to support C++ inlines that can be coalesced by putting them
into their own section (TEXT/textcoal_nt instead of TEXT/text).
The current macho linker supports the weak-def bit on any symbol to allow it to
be coalesced, but the compiler still puts weak-def functions/data into alternate
section names, which the linker must map back to the base section name.
This patch makes changes that are necessary to prevent the compiler from using
the "coal" sections and have it use the non-coal sections instead when the
target architecture is not powerpc:
TEXT/textcoal_nt instead use TEXT/text
TEXT/const_coal instead use TEXT/const
DATA/datacoal_nt instead use DATA/data
If the target is powerpc, we continue to use the *coal* sections since anyone
targeting powerpc is probably using an old linker that doesn't have support for
the weak-def bits.
Also, have the assembler issue a warning if it encounters a *coal* section in
the assembly file and inform the users to use the non-coal sections instead.
rdar://problem/14265330
Differential Revision: http://reviews.llvm.org/D13188
llvm-svn: 250349
Some background on why we don't have to use *coal* sections anymore:
Long ago when C++ was new and "weak" had not been standardized, an attempt was
made in cctools to support C++ inlines that can be coalesced by putting them
into their own section (TEXT/textcoal_nt instead of TEXT/text).
The current macho linker supports the weak-def bit on any symbol to allow it to
be coalesced, but the compiler still puts weak-def functions/data into alternate
section names, which the linker must map back to the base section name.
This patch makes changes that are necessary to prevent the compiler from using
the "coal" sections and have it use the non-coal sections instead when the
target architecture is not powerpc:
TEXT/textcoal_nt instead use TEXT/text
TEXT/const_coal instead use TEXT/const
DATA/datacoal_nt instead use DATA/data
If the target is powerpc, we continue to use the *coal* sections since anyone
targeting powerpc is probably using an old linker that doesn't have support for
the weak-def bits.
Also, have the assembler issue a warning if it encounters a *coal* section in
the assembly file and inform the users to use the non-coal sections instead.
rdar://problem/14265330
Differential Revision: http://reviews.llvm.org/D13188
llvm-svn: 250342
This patch teaches x86 fast-isel how to select nontemporal stores.
On x86, we can use MOVNTI for nontemporal stores of doublewords/quadwords.
Instructions (V)MOVNTPS/PD/DQ can be used for SSE2/AVX aligned nontemporal
vector stores.
Before this patch, fast-isel always selected 'movd/movq' instead of 'movnti'
for doubleword/quadword nontemporal stores. In the case of nontemporal stores
of aligned vectors, fast-isel always selected movaps/movapd/movdqa instead of
movntps/movntpd/movntdq.
With this patch, if we use SSE2/AVX intrinsics for nontemporal stores we now
always get the expected (V)MOVNT instructions.
The lack of fast-isel support for nontemporal stores was spotted when analyzing
the -O0 codegen for nontemporal stores.
Differential Revision: http://reviews.llvm.org/D13698
llvm-svn: 250285
Summary:
D4796 taught LLVM to fold some atomic integer operations into a single
instruction. The pattern was unaware that the instructions clobbered
flags.
This patch adds the missing EFLAGS definition.
Floating point operations don't set flags, the subsequent fadd
optimization is therefore correct. The same applies for surrounding
load/store optimizations.
Reviewers: rsmith, rtrieu
Subscribers: llvm-commits, reames, morisset
Differential Revision: http://reviews.llvm.org/D13680
llvm-svn: 250135
When lowering invoke statement, all unwind destinations are directly added as successors of call site block, and the weight of those new edges are not assigned properly. Actually, default weight 16 are used for those edges. This patch calculates the proper edge weights for those edges when collecting all unwind destinations.
Differential revision: http://reviews.llvm.org/D13354
llvm-svn: 250119
We made them SP relative back in March (r233137) because that's the
value the runtime passes to EH functions. With the new cleanuppad IR,
funclets adjust their frame argument from SP to FP, so our offsets
should now be FP-relative.
llvm-svn: 250088
Function LowerVSETCC (in X86ISelLowering.cpp) worked under the wrong
assumption that for non-AVX512 targets, the source type and destination type
of a type-legalized setcc node were always the same type.
This assumption was unfortunately incorrect; the type legalizer is not always
able to promote the return type of a setcc to the same type as the first
operand of a setcc.
In the case of a vsetcc node, the legalizer firstly checks if the first input
operand has a legal type. If so, then it promotes the return type of the vsetcc
to that same type. Otherwise, the return type is promoted to the 'next legal
type', which, for vectors of MVT::i1 is always a 128-bit integer vector type.
Example (-mattr=+avx):
%0 = trunc <8 x i32> %a to <8 x i23>
%1 = icmp eq <8 x i23> %0, zeroinitializer
The initial selection dag for the code above is:
v8i1 = setcc t5, t7, seteq:ch
t5: v8i23 = truncate t2
t2: v8i32,ch = CopyFromReg t0, Register:v8i32 %vreg1
t7: v8i32 = build_vector of all zeroes.
The type legalizer would firstly check if 't5' has a legal type. If so, then it
would reuse that same type to promote the return type of the setcc node.
Unfortunately 't5' is of illegal type v8i23, and therefore it cannot be used to
promote the return type of the setcc node. Consequently, the setcc return type
is promoted to v8i16. Later on, 't5' is promoted to v8i32 thus leading to the
following dag node:
v8i16 = setcc t32, t25, seteq:ch
where t32 and t25 are now values of type v8i32.
Before this patch, function LowerVSETCC would have wrongly expanded the setcc
to a single X86ISD::PCMPEQ. Surprisingly, ISel was still able to match an
instruction. In our case, ISel would have matched a VPCMPEQWrr:
t37: v8i16 = X86ISD::VPCMPEQWrr t36, t25
However, t36 and t25 are both VR256, while the result type is instead of class
VR128. This inconsistency ended up causing the insertion of COPY instructions
like this:
%vreg7<def> = COPY %vreg3; VR128:%vreg7 VR256:%vreg3
Which is an invalid full copy (not a sub register copy).
Eventually, the backend would have hit an UNREACHABLE "Cannot emit physreg copy
instruction" in the attempt to expand the malformed pseudo COPY instructions.
This patch fixes the problem adding the missing logic in LowerVSETCC to handle
the corner case of a setcc with 128-bit return type and 256-bit operand type.
This problem was originally reported by Dimitry as PR25080. It has been latent
for a very long time. I have added the minimal reproducible from that bugzilla
as test setcc-lowering.ll.
Differential Revision: http://reviews.llvm.org/D13660
llvm-svn: 250085
This patch fixes a problem in function 'combineX86ShuffleChain' that causes a
chain of shuffles to be wrongly folded away when the combined shuffle mask has
only one element.
We may end up with a combined shuffle mask of one element as a result of
multiple calls to function 'canWidenShuffleElements()'.
Function canWidenShuffleElements attempts to simplify a shuffle mask by widening
the size of the elements being shuffled.
For every pair of shuffle indices, function canWidenShuffleElements checks if
indices refer to adjacent elements. If all pairs refer to "adjacent" elements
then the shuffle mask is safely widened. As a consequence of widening, we end up
with a new shuffle mask which is half the size of the original shuffle mask.
The byte shuffle (pshufb) from test pr24562.ll has a mask of all SM_SentinelZero
indices. Function canWidenShuffleElements would combine each pair of
SM_SentinelZero indices into a single SM_SentinelZero index. So, in a
logarithmic number of steps (4 in this case), the pshufb mask is simplified to
a mask with only one index which is equal to SM_SentinelZero.
Before this patch, function combineX86ShuffleChain wrongly assumed that a mask
of size one is always equivalent to an identity mask. So, the entire shuffle
chain was just folded away as the combined shuffle mask was treated as a no-op
mask.
With this patch we know check if the only element of a combined shuffle mask is
SM_SentinelZero. In case, we propagate a zero vector.
Differential Revision: http://reviews.llvm.org/D13364
llvm-svn: 250027
The XOP vector integer comparisons can deal with all signed/unsigned comparison cases directly and can be easily commuted as well (D7646).
llvm-svn: 249976
Also Fix a buglet where SEH tables had ranges that spanned funclets.
The remaining tests using the old landingpad IR are preparation tests,
and will be deleted along with the old preparation.
llvm-svn: 249917
x64 catchpads use rax to inform the unwinder where control should go
next. However, we must initialize rax before the epilogue sequence so
as to not perturb the unwinder.
llvm-svn: 249910
This wasn't very observable in execution tests, because usually there is
an invoke in the catchpad that unwinds the the catchendpad but never
actually throws.
llvm-svn: 249898
its own variable.
This is needed so that we can explicitly turn off MMX without turning
off SSE and also so that we can diagnose feature set incompatibilities
that involve MMX without SSE.
Rationale:
// sse3
__m128d test_mm_addsub_pd(__m128d A, __m128d B) {
return _mm_addsub_pd(A, B);
}
// mmx
void shift(__m64 a, __m64 b, int c) {
_mm_slli_pi16(a, c);
_mm_slli_pi32(a, c);
_mm_slli_si64(a, c);
_mm_srli_pi16(a, c);
_mm_srli_pi32(a, c);
_mm_srli_si64(a, c);
_mm_srai_pi16(a, c);
_mm_srai_pi32(a, c);
}
clang -msse3 -mno-mmx file.c -c
For this code we should be able to explicitly turn off MMX
without affecting the compilation of the SSE3 function and then
diagnose and error on compiling the MMX function.
This matches the existing gcc behavior and follows the spirit of
the SSE/MMX separation in llvm where we can (and do) turn off
MMX code generation except in the presence of intrinsics.
Updated a couple of tests, but primarily tested with a couple of tests
for turning on only mmx and only sse.
This is paired with a patch to clang to take advantage of this behavior.
llvm-svn: 249731
We emit 1 compact unwind encoding per function, and this can’t represent
the varying stack pointer that will be generated by X86CallFrameOptimization.
Disable the optimization on Darwin.
(It might be possible to split the function into multiple ranges
and emit 1 compact unwind info per range. The compact unwind emission
code isn’t ready for that and this kind of info certainly isn’t
tested/used anywhere. It might be worth exploring this path if we want
to get the space savings at some point though)
llvm-svn: 249694
This instructions doesn't have intrincis.
Added tests for lowering and encoding.
Differential Revision: http://reviews.llvm.org/D12317
llvm-svn: 249688
This fixes two separate bugs:
1) The mask for the high lane was not set correctly. That fixes PR24532.
2) The transformation should bail out if it believes it involves more than
2 lanes, as it does not currently do anything sensible in this case.
Differential Revision: http://reviews.llvm.org/D13505
llvm-svn: 249669
In particular, passing non-trivially copyable objects by value on win32
uses a dynamic alloca (inalloca). We would clobber ESP in the epilogue
and end up returning to outer space.
llvm-svn: 249637
The relocation for the filter funclet will be against a symbol table
entry for a function instead of the section, making it easier to
understand what is going on.
llvm-svn: 249621
The __CxxFrameHandler3 tables for 32-bit are supposed to hold stack
offsets relative to EBP, not ESP. I blindly updated the win-catchpad.ll
test case, and immediately noticed that 32-bit catching stopped working.
While I'm at it, move the frame index to frame offset WinEH table logic
out of PEI. PEI shouldn't have to know about WinEHFuncInfo. I realized
we can calculate frame index offsets just fine from the table printer.
llvm-svn: 249618
We remove unreachable blocks because it is pointless to consider them
for coloring. However, we still had stale pointers to these blocks in
some data structures after we removed them from the function.
Instead, remove the unreachable blocks before attempting to do anything
with the function.
This fixes PR25099.
llvm-svn: 249617
There was an off-by-one bug in ip2state tables which manifested when one
call immediately preceded the try-range of the next. The return address
of the previous call would appear to be within the try range of the next
scope, resulting in extra destructors or catches running.
We also computed the wrong offset for catch parameter stack objects. The
offset should be from RSP, not from RBP.
llvm-svn: 249578
When outgoing function arguments are passed using push instructions, and EH
is enabled, we may need to indicate to the stack unwinder that the stack
pointer was adjusted before the call.
This should fix the exception handling issues in PR24792.
Differential Revision: http://reviews.llvm.org/D13132
llvm-svn: 249522
Our current emission strategy is to emit the funclet prologue in the
CatchPad's normal destination. This is problematic because
intra-funclet control flow to the normal destination is not erroneous
and results in us reevaluating the prologue if said control flow is
taken.
Instead, use the CatchPad's location for the funclet prologue. This
correctly models our desire to have unwind edges evaluate the prologue
but edges to the normal destination result in typical control flow.
Differential Revision: http://reviews.llvm.org/D13424
llvm-svn: 249483
Most importantly, this keeps constant hoisting from preventing instruction selections ability to turn an AND with 0xffffffff into a move into a 32-bit subregister.
llvm-svn: 249370
Track which basic blocks belong to which funclets. Permit branch
folding to fire but only if it can prove that doing so will not cause
code in one funclet to be reused in another.
llvm-svn: 249257
This patch teaches FastIsel the following two things:
1) On SSE2, no instructions are needed for bitcasts between 128-bit vector types;
2) On AVX, no instructions are needed for bitcasts between 256-bit vector types.
Example:
%1 = bitcast <4 x i31> %V to <2 x i64>
Before (-fast-isel -fast-isel-abort=1):
FastIsel miss: %1 = bitcast <4 x i31> %V to <2 x i64>
Now we don't fall back to SelectionDAG and we correctly fold that computation
propagating the register associated to %V.
Originally reviewed here: http://reviews.llvm.org/D13347
llvm-svn: 249147
This patch teaches FastIsel the following two things:
1) On SSE2, no instructions are needed for bitcasts between 128-bit vector types;
2) On AVX, no instructions are needed for bitcasts between 256-bit vector types.
Example:
%1 = bitcast <4 x i31> %V to <2 x i64>
Before (-fast-isel -fast-isel-abort=1):
FastIsel miss: %1 = bitcast <4 x i31> %V to <2 x i64>
Now we don't fall back to SelectionDAG and we correctly fold that computation
propagating the register associated to %V.
Differential Revision: http://reviews.llvm.org/D13347
llvm-svn: 249121
We emit denormalized tables, where every range of invokes in the same
state gets a complete list of EH action entries. This is significantly
simpler than trying to infer the correct nested scoping structure from
the MI. Fortunately, for SEH, the nesting structure is really just a
size optimization.
With this, some basic __try / __except examples work.
llvm-svn: 249078
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
The custom code produces incorrect results if later reassociated.
Since r221657, on x86, vNi32 uitofp is lowered using an optimized
sequence:
movdqa LCPI0_0(%rip), %xmm1 ## xmm1 = [65535, ...]
pand %xmm0, %xmm1
por LCPI0_1(%rip), %xmm1 ## [0x4b000000, ...]
psrld $16, %xmm0
por LCPI0_2(%rip), %xmm0 ## [0x53000000, ...]
addps LCPI0_3(%rip), %xmm0 ## [float -5.497642e+11, ...]
addps %xmm1, %xmm0
Since r240361, the machine combiner opportunistically reassociates
2-instruction sequences (with -ffast-math). In the new code sequence,
the ADDPS' are eligible. In isolation, for simple examples (without
reassociable users), this makes no performance difference (the goal
being to enable reassociation of longer chains).
In the trivial example (just one uitofp), the reassociation doesn't
happen, because (I think) it would require the emission of a separate
movaps for a constantpool load (instead of folding it into addps).
However, when we have multiple uitofp sequences, and the constantpool
loads are CSE'd earlier, the machine combiner can do the reassociation.
When the ADDPS' are reassociated, the resulting sequence isn't correct
anymore, as we'd be adding large (2**39) constants with comparatively
smaller values (~2**23). Given that two of the three inputs are powers
of 2 larger than 2**16, and that ulp(2**39) == 2**(39-24) == 2**15,
the reassociated chain will produce 0 for any input in [0, 2**14[.
In my testing, it also produces wrong results for 99.5% of [0, 2**32[.
Avoid this by disabling the new lowering when -ffast-math. It does
mean that we'll get slower code than without it, but at least we
won't get egregiously incorrect code.
One might argue that, considering -ffast-math is all but meaningless,
uitofp producing wrong results isn't a compiler bug. But it really is.
Fixes PR24512.
...though this is really more of a workaround.
Ideally, we'd have some sort of Machine FMF, but that's a problem
that's not worth tackling until we do more with machine IR.
llvm-svn: 248965
The Win64 unwinder disassembles forwards from each PC to try to
determine if this PC is in an epilogue. If so, it skips calling the EH
personality function for that frame. Typically, this means you cannot
catch an exception in the same frame that you threw it, because 'throw'
calls a noreturn runtime function.
Previously we avoided this problem with the TrapUnreachable
TargetOption, but that's a much bigger hammer than we need. All we need
is a 1 byte non-epilogue instruction right after the call. Instead,
what we got was an unconditional branch to a shared block containing the
ud2, potentially 7 bytes instead of 1. So, this reverts r206684, which
added TrapUnreachable, and replaces it with something better.
The new code pattern matches for invoke/call followed by unreachable and
inserts an int3 into the DAG. To be 100% watertight, we would need to
insert SEH_Epilogue instructions into all basic blocks ending in a call
with no terminators or successors, but in practice this is unlikely to
come up.
llvm-svn: 248959
The XOP shifts just have logical/arithmetic versions and the left/right shifts are controlled by whether the value is positive/negative. Because of this I've added new X86ISD nodes instead of trying to force them to use the existing shift nodes.
Additionally Excavator cores (bdver4) support XOP and AVX2 - meaning that it should use the AVX2 shifts when it can and fall back to XOP in other cases.
Differential Revision: http://reviews.llvm.org/D8690
llvm-svn: 248878
The x64 ABI requires that epilogues do not contain code other than stack
adjustments and some limited control flow. However, we'd insert code to
initialize the return address after stack adjustments. Instead, insert
EAX/RAX with the current value before we create the stack adjustments in
the epilogue.
llvm-svn: 248839
HHVM calling convention, hhvmcc, is used by HHVM JIT for
functions in translated cache. We currently support LLVM back end to
generate code for X86-64 and may support other architectures in the
future.
In HHVM calling convention any GP register could be used to pass and
return values, with the exception of R12 which is reserved for
thread-local area and is callee-saved. Other than R12, we always
pass RBX and RBP as args, which are our virtual machine's stack pointer
and frame pointer respectively.
When we enter translation cache via hhvmcc function, we expect
the stack to be aligned at 16 bytes, i.e. skewed by 8 bytes as opposed
to standard ABI alignment. This affects stack object alignment and stack
adjustments for function calls.
One extra calling convention, hhvm_ccc, is used to call C++ helpers from
HHVM's translation cache. It is almost identical to standard C calling
convention with an exception of first argument which is passed in RBP
(before we use RDI, RSI, etc.)
Differential Revision: http://reviews.llvm.org/D12681
llvm-svn: 248832
Summary:
Funclets have been turned into functions by the time they hit the object
file. Make sure that they have decent names for the symbol table and
CFI directives explaining how to reason about their prologues.
Differential Revision: http://reviews.llvm.org/D13261
llvm-svn: 248824
alignment requirements, for example in the case of vectors.
These requirements are exploited by the code generator by using
move instructions that have similar alignment requirements, e.g.,
movaps on x86.
Although the code generator properly aligns the arguments with
respect to the displacement of the stack pointer it computes,
the displacement itself may cause misalignment. For example if
we have
%3 = load <16 x float>, <16 x float>* %1, align 64
call void @bar(<16 x float> %3, i32 0)
the x86 back-end emits:
movaps 32(%ecx), %xmm2
movaps (%ecx), %xmm0
movaps 16(%ecx), %xmm1
movaps 48(%ecx), %xmm3
subl $20, %esp <-- if %esp was 16-byte aligned before this instruction, it no longer will be afterwards
movaps %xmm3, (%esp) <-- movaps requires 16-byte alignment, while %esp is not aligned as such.
movl $0, 16(%esp)
calll __bar
To solve this, we need to make sure that the computed value with which
the stack pointer is changed is a multiple af the maximal alignment seen
during its computation. With this change we get proper alignment:
subl $32, %esp
movaps %xmm3, (%esp)
Differential Revision: http://reviews.llvm.org/D12337
llvm-svn: 248786
This is a redo of D7208 ( r227242 - http://llvm.org/viewvc/llvm-project?view=revision&revision=227242 ).
The patch was reverted because an AArch64 target could infinite loop after the change in DAGCombiner
to merge vector stores. That happened because AArch64's allowsMisalignedMemoryAccesses() wasn't telling
the truth. It reported all unaligned memory accesses as fast, but then split some 128-bit unaligned
accesses up in performSTORECombine() because they are slow.
This patch attempts to fix the problem in AArch's allowsMisalignedMemoryAccesses() while preserving
existing (perhaps questionable) lowering behavior.
The x86 test shows that store merging is working as intended for a target with fast 32-byte unaligned
stores.
Differential Revision: http://reviews.llvm.org/D12635
llvm-svn: 248622
Fix for D12561 - we weren't correctly ensuring that the base element for extension was moved to start on a boundary suitable for UNPCKL/H
llvm-svn: 248536
Fixes the overflow case of llvm.*absdiff intrinsic also updats the tests and LangRef.rst accordingly.
Differential Revision: http://reviews.llvm.org/D11678
llvm-svn: 248483
If the stores are storing values from loads which partially
alias the stores, we could end up placing the merged loads
and stores on the same chain which has the potential to break.
Each store may have a different chain dependency on only some
of the original loads. Create a new TokenFactor to capture all
of the required dependencies of the stores rather than assuming
all stores can use the same chain.
The testcase is a situation where this happens, although
it does not have an observable change from this. The DAG nodes
just happened to not be reordered before despite this missing
chain dependency.
This is based on an off-list report for an out of tree target
which regressed due to r246307 and I haven't managed to find a case
where the nodes do end up reordered with an in tree target.
llvm-svn: 248468
This patches removes the x86.sse41.pmovsx* intrinsics, provides a suitable upgrade path and updates relevant tests to sign extend a subvector instead.
LLVM counterpart to D12835
Differential Revision: http://reviews.llvm.org/D13002
llvm-svn: 248368
The C standard has historically not specified whether or not these functions should raise the inexact flag. Traditionally on Darwin, these functions *did* raise inexact, and the llvm lowerings followed that conventions. n1778 (C bindings for IEEE-754 (2008)) clarifies that these functions should not set inexact. This patch brings the lowerings for arm64 and x86 in line with the newly specified behavior. This also lets us fold some logic into TD patterns, which is nice.
Differential Revision: http://reviews.llvm.org/D12969
llvm-svn: 248266
This patch generalizes the lowering of shuffles as zero extensions to allow extensions that don't start from the first element. It now recognises extensions starting anywhere in the lower 128-bits or at the start of any higher 128-bit lane.
The motivation was to reduce the number of high cost pshufb calls, but it also improves the SSE2 case as well.
Differential Revision: http://reviews.llvm.org/D12561
llvm-svn: 248250
This patch adds support for combining patterns such as (FMUL(FADD(1.0, x), y)) and (FMUL(FSUB(x, 1.0), y)) to their FMA equivalents.
This is useful in particular for linear interpolation cases such as (FADD(FMUL(x, t), FMUL(y, FSUB(1.0, t))))
Differential Revision: http://reviews.llvm.org/D13003
llvm-svn: 248210
If storing multiple FP constants, some subset of the stores
would be replaced with integers due to visit order, so
MergeConsecutiveStores would only partially merge
these.
llvm-svn: 248169
Now that we have fast vector CTPOP implementations we can use this to speed up vector CTTZ using the pattern (cttz(x) = ctpop((x & -x) - 1))
Additionally, for AVX512CD that provides lzcnt instructions we can use the pattern (cttz_undef(x) = (width - 1) - ctlz(x & -x))
Differential Revision: http://reviews.llvm.org/D12663
llvm-svn: 248091
- Strenghten the logic to be sure we hoist the restore point out of the current
loop. (The fixes a bug with infinite loop, added as part of the patch.)
- Walk over the exit blocks of the current loop to conver to the desired restore
point in one iteration of the update loop.
llvm-svn: 247958
Windows EH funclets need to be contiguous. The FuncletLayout pass will
ensure that the funclets are together and begin with a funclet entry MBB.
Differential Revision: http://reviews.llvm.org/D12943
llvm-svn: 247937
This makes catchret look more like a branch, and less like a weird use
of BlockAddress. It also lets us get away from
llvm.x86.seh.restoreframe, which relies on the old parentfpoffset label
arithmetic.
llvm-svn: 247936
getLandingPadSuccessor assumes that each invoke can have at most one EH
pad successor, but WinEH invokes can have more than one. Two out of
three callers of getLandingPadSuccessor don't use the returned
landingpad, so we can make them use this simple predicate instead.
Eventually we'll have to circle back and fix SplitKit.cpp so that
register allocation works. Baby steps.
llvm-svn: 247904
Sanjay Patel