Simon Pilgrim [Wed, 7 Dec 2016 11:19:00 +0000 (11:19 +0000)]
[X86][XOP] Fix VPERMIL2 non-constant pool shuffle decoding (PR31296)
The non-constant pool version of DecodeVPERMIL2PMask was not offsetting correctly for the second input. I've updated the code to match the implementation in the constant-pool version.
Annoyingly this bug was hidden for so long as it's tricky to combine to useful variable shuffle masks that don't become constant-pool entries.
[InlineFunction] Do not propagate the callsite debug location to instructions inlined from functions with debug info.
When a function F is inlined, InlineFunction extends the debug location of every
instruction inlined from F by adding an InlinedAt.
However, if an instruction has a 'null' debug location, InlineFunction would
propagate the callsite debug location to it. This behavior existed since
revision 210459.
Revision 210459 was originally committed specifically to workaround the lack of
debug information for instructions inlined from intrinsic functions (which are
usually declared with attributes `__always_inline__, __nodebug__`).
The problem with revision 210459 is that it doesn't make any sort of distinction
between instructions inlined from a 'nodebug' function and instructions which
are inlined from a function built with debug info. This issue may lead to
incorrect stepping in the debugger.
This patch works under the assumption that a nodebug function does not have a
DISubprogram. When a function F is inlined into another function G,
InlineFunction checks if F has debug info associated with it.
For nodebug functions, the InlineFunction logic is unchanged (i.e. it would
still propagate the callsite debugloc to the inlined instructions). Otherwise,
InlineFunction no longer propagates the callsite debug location.
Philip Reames [Wed, 7 Dec 2016 04:48:50 +0000 (04:48 +0000)]
Reintroduce a check accidentally removed in 288873 to fix clang bots
I believe this is the cause of the failure, but have not been able to confirm. Note that this is a speculative fix; I'm still waiting for a full build to finish as I synced and ended up doing a clean build which takes 20+ minutes on my machine.
LowerTypeTests: Improve performance by optimising type metadata queries.
Requesting metadata for a global is a relatively expensive operation as it
involves a map lookup, but it's one that we need to do relatively frequently in
this pass to collect the list of type metadata nodes associated with a global.
This change improves the performance of type metadata queries by prebuilding
data structures that keep the global together with its list of type metadata,
and changing the pass to use that data structure wherever we were previously
passing global references around.
This change also eliminates some O(N^2) behavior by collecting the list of
globals associated with each type identifier during the first pass over the
list of globals rather than visiting each global to compute that list every
time we add a new type identifier.
Reduces pass runtime on a module containing Chrome's vtables from over 60s
to 0.9s.
Eli Friedman [Tue, 6 Dec 2016 22:49:36 +0000 (22:49 +0000)]
[CodeGen] Fix result type for SMULO/UMULO legalization
On some platforms (like MSP430) the second element of the result
structure for SMULO/UMULO may have a shorter type than the one
returned by SetCC. We need to truncate it to the right type, or
else some incorrect code may be generated later on.
This fixes issue https://github.com/rust-lang/rust/issues/37829
Sanjay Patel [Tue, 6 Dec 2016 22:09:52 +0000 (22:09 +0000)]
[InstSimplify] fixed (?) to not mutate icmps
As Eli noted in the post-commit thread for r288833, the use of
swapOperands() may not be allowed in InstSimplify, so I'm
removing those calls here pending further review.
The swap mutates the icmp, and there doesn't appear to be precedent
for instruction mutation in InstSimplify.
I didn't actually have any tests for those cases, so I'm adding
a few here.
BDCE has two phases:
1. It asks SimplifyDemandedBits if all the bits of an instruction are dead, and if so,
replaces all its uses with the constant zero.
2. Then, it asks SimplifyDemandedBits again if the instruction is really dead
(no side effects etc..) and if so, eliminates it.
Now, in 1) if all the bits of an instruction are dead, we may end up replacing a dbg use:
%call = tail call i32 (...) @g() #4, !dbg !15
tail call void @llvm.dbg.value(metadata i32 %call, i64 0, metadata !8, metadata !16), !dbg !17
->
%call = tail call i32 (...) @g() #4, !dbg !15
tail call void @llvm.dbg.value(metadata i32 0, i64 0, metadata !8, metadata !16), !dbg !17
but not eliminating the call because it may have arbitrary side effects.
In other words, we lose some debug informations.
This patch fixes the problem making sure that BDCE does nothing with the instruction if
it has side effects and no non-dbg uses.
Tom Stellard [Tue, 6 Dec 2016 21:13:30 +0000 (21:13 +0000)]
AMDGPU/SI: Don't move copies of immediates to the VALU
Summary:
If we write an immediate to a VGPR and then copy the VGPR to an
SGPR, we can replace the copy with a S_MOV_B32 sgpr, imm, rather than
moving the copy to the SALU.
Zvi Rackover [Tue, 6 Dec 2016 19:35:20 +0000 (19:35 +0000)]
[X86] Prefer reduced width multiplication over pmulld on Silvermont
Summary:
Prefer expansions such as: pmullw,pmulhw,unpacklwd,unpackhwd over pmulld.
On Silvermont [source: Optimization Reference Manual]:
PMULLD has a throughput of 1/11 [instruction/cycles].
PMULHUW/PMULHW/PMULLW have a throughput of 1/2 [instruction/cycles].
Handle the case where a sign extension has ended up being split into separate stages (typically to get around vector legal ops) and a zext + sext_in_reg gets inserted.
Sanjay Patel [Tue, 6 Dec 2016 19:05:46 +0000 (19:05 +0000)]
[InstSimplify] add folds for and-of-icmps with same operands
All of these (and a few more) are already handled by InstCombine,
but we shouldn't have to wait until then to simplify these because
they're cheap to deal with here in InstSimplify.
This is the 'and' sibling of the earlier 'or' patch:
https://reviews.llvm.org/rL288833
There were two problems:
+ AArch64 was reusing random data from its binary op tables, which is
complete nonsense for G_SEQUENCE.
+ Even when AArch64 gave up and said it couldn't handle G_SEQUENCE,
the generic code asserted.
Tim Northover [Tue, 6 Dec 2016 18:38:29 +0000 (18:38 +0000)]
GlobalISel: stop the legalizer from trying to handle oddly-sized types.
It'll almost immediately fail because it always tries to half/double the size
until it finds a legal one. Unfortunately, this triggers an assertion
preventing the DAG fallback from being possible.
Sanjay Patel [Tue, 6 Dec 2016 18:09:37 +0000 (18:09 +0000)]
[InstSimplify] add folds for or-of-icmps with same operands
All of these (and a few more) are already handled by InstCombine,
but we shouldn't have to wait until then to simplify these because
they're cheap to deal with here in InstSimplify.
Daniel Sanders [Tue, 6 Dec 2016 14:39:57 +0000 (14:39 +0000)]
[globalisel][aarch64] Fix unintended assumptions about PartialMappingIdx. NFC.
Summary:
This is NFC but prevents assertions when PartialMappingIdx is tablegen-erated.
The assumptions were:
1) FirstGPR is 0
2) FirstGPR is the first of the First* enumerators.
GPR32 is changed to 1 to demonstrate that assumption #1 is fixed. #2 will
be covered by a subsequent patch that tablegen-erates information and swaps
the order of GPR and FPR as a side effect.
Oliver Stannard [Tue, 6 Dec 2016 12:59:08 +0000 (12:59 +0000)]
[ARM] Better error message for invalid flag-preserving Thumb1 insts
When we see a non flag-setting instruction for which only the flag-setting
version is available in Thumb1, we should give a better error message than
"invalid instruction".
Ayman Musa [Tue, 6 Dec 2016 12:24:14 +0000 (12:24 +0000)]
[X86][AVX512] Detect repeated constant patterns in BUILD_VECTOR suitable for broadcasting.
Check if a build_vector node includes a repeated constant pattern and replace it with a broadcast of that pattern.
For example:
"build_vector <0, 1, 2, 3, 0, 1, 2, 3>" would be replaced by "broadcast <0, 1, 2, 3>"
This is the final patch in the series of patches that improves
BUILD_VECTOR handling on PowerPC. This adds a few peephole optimizations
to remove redundant instructions. It also adds a large test case which
encompasses a large set of code patterns that build vectors - this test
case was the motivator for this series of patches.
[LCG] Add some much needed asserts and verify runs to uncover
a hilarious bug and fix it.
We somehow were never verifying the RefSCCs newly formed when
splitting an existing one apart, and when verifying them we weren't
really checking the SCC indices mapping effectively.
If we had been, it would have been blindingly obvious that right after
putting something int `RC.SCCs` we should update `RC.SCCIndices` instead
of `SCCIndices` which we were about to clear and rebuild anyways. =[
Anyways, this is thoroughly covered by existing tests now that we
actually verify things properly.
Craig Topper [Tue, 6 Dec 2016 08:08:12 +0000 (08:08 +0000)]
[X86] Remove another weird scalar sqrt/rcp/rsqrt pattern.
This pattern turned a vector sqrt/rcp/rsqrt operation of sse_load_f32/f64 into the the scalar instruction for the operation and put undef into the upper bits. For correctness, the resulting code should still perform the sqrt/rcp/rsqrt on the upper bits after the load is extended since that's what the operation asked for. Particularly in the case where the upper bits are 0, in that case we need calculate the sqrt/rcp/rsqrt of the zeroes and keep the result in the upper-bits. This implies we should be using the packed instruction still.
The only test case for this pattern is one I just added so there was no coverage of this.
Craig Topper [Tue, 6 Dec 2016 08:08:09 +0000 (08:08 +0000)]
[X86] Add test case demonstrating a case where a vector sqrt being passed (scalar_to_vector loadf64) uses a scalar sqrt instruction.
This occurs due to a pattern that uses sse_load_f32/f64 with vector sqrt/rcp/rsqrt operations and turns them into scalar instructions. Perhaps for the case were the upper bits come from undef this is ok. I believe a (vzmovl load64) would do the same thing but those seems to become vzload instead and selectScalarSSELoad doesn't handle that today. In that case we should be performing the vector operation on the zeros in the upper bits which is not equivalent to using a scalar instruction.
I will remove this pattern in a follow up patch. There appears to be no other test content for it.
Craig Topper [Tue, 6 Dec 2016 08:08:04 +0000 (08:08 +0000)]
[X86] Remove bad pattern that caused 128-bit loads being used by scalar sqrt/rcp/rsqrt intrinsics to select the memory form of the corresponding instruction and violate the semantics of the intrinsic.
The intrinsics are supposed to pass the upper bits straight through to their output register. This means we need to make sure we still perform the 128-bit load to get those upper bits to pass to give to the instruction since the memory form of the instruction only reads 32 or 64 bits.
Craig Topper [Tue, 6 Dec 2016 08:08:01 +0000 (08:08 +0000)]
[X86] Add test case that shows a scalar sqrtsd intrinsic of a 128-bit vector load using the load form of the sqrtsd instruction which violates the intrinsic semantics.
The sqrtsd instruction only loads 64-bits and writes bits 63:0 with the sqrt result. Bits 127:64 are preserved in the destination register. The semantics of the intrinsic indicate bits 127:64 should come from the intrinsic argument which in this case is a 128-bit load. So the generated code should have a 128-bit load and use a register form of sqrtsd.
Craig Topper [Tue, 6 Dec 2016 08:07:58 +0000 (08:07 +0000)]
[X86] Correct pattern for VSQRTSSr_Int, VSQRTSDr_Int, VRCPSSr_Int, and VRSQRTSSr_Int to not have an IMPLICIT_DEF on the first input. The semantics of the intrinsic are clear and not undefined.
The intrinsic takes one argument, the lower bits are affected by the operation and the upper bits should be passed through. The instruction itself takes two operands, the high bits of the first operand are passed through and the low bits of the second operand are modified by the operation. To match this to the intrinsic we should pass the single intrinsic input to both operands.
I had to remove the stack folding test for these instructions since they depended on the incorrect behavior. The same register is now used for both inputs so the load can't be folded.
Chris Bieneman [Tue, 6 Dec 2016 06:00:49 +0000 (06:00 +0000)]
[ObjectYAML] First bit of support for encoding DWARF in MachO
This patch adds the starting support for encoding data from the MachO __DWARF segment. The first section supported is the __debug_str section because it is the simplest.
Craig Topper [Tue, 6 Dec 2016 04:58:39 +0000 (04:58 +0000)]
[X86] Remove scalar logical op alias instructions. Just use COPY_FROM/TO_REGCLASS and the normal packed instructions instead
Summary:
This patch removes the scalar logical operation alias instructions. We can just use reg class copies and use the normal packed instructions instead. This removes the need for putting these instructions in the execution domain fixing tables as was done recently.
I removed the loadf64_128 and loadf32_128 patterns as DAG combine creates a narrower load for (extractelt (loadv4f32)) before we ever get to isel.
I plan to add similar patterns for AVX512DQ in a future commit to allow use of the larger register class when available.
Chris Bieneman [Tue, 6 Dec 2016 04:45:11 +0000 (04:45 +0000)]
[CMake] Cleanup TableGen include flags
It is kinda crazy to have llvm/include and llvm/lib/Target in the include path for every tablegen invocation for every tablegen-like tool.
This patch removes those flags from the tablgen function that is called everywhere by instead creating a variable LLVM_TABLEGEN_FLAGS which is setup in the LLVM source directories.
This removes TableGen.cmake's dependency on LLVM_MAIN_SRC_DIR, and LLVM_MAIN_INCLUDE_DIR.
Philip Reames [Tue, 6 Dec 2016 03:34:33 +0000 (03:34 +0000)]
[LVI] Remove dead code in mergeIn
Integers are expressed in the lattice via constant ranges. They can never be represented by constants or not-constants; those are reserved for non-integer types. This code has been dead for literaly years.
Philip Reames [Tue, 6 Dec 2016 03:01:08 +0000 (03:01 +0000)]
[LVI] Hide the last markX function on LVILatticeVal
This completes a small series of patches to hide the stateful updates of LVILatticeVal from the consuming code. The only remaining stateful API is mergeIn.
Summary:
We recently introduced a feature that enforce at link-time that the
LLVM headers used by a clients are matching the ABI setting of the
LLVM library linked to.
However for clients that are using only headers from ADT and promise
they won't call into LLVM, this is forcing to link libSupport. This
new flag is intended to provide a way to configure LLVM with this
promise for such client.
Matt Arsenault [Tue, 6 Dec 2016 01:02:51 +0000 (01:02 +0000)]
AMDGPU: Don't required structured CFG
The structured CFG is just an aid to inserting exec
mask modification instructions, once that is done
we don't really need it anymore. We also
do not analyze blocks with terminators that
modify exec, so this should only be impacting
true branches.
[libFuzzer] refactor the code to allow collecting features in different ways. Also initialize a couple of Fuzzer:: members that might have been used uninitialized :(
Tim Northover [Mon, 5 Dec 2016 23:10:19 +0000 (23:10 +0000)]
GlobalISel: avoid looking too closely at PHIs when we bail.
The function used to finish off PHIs by adding the relevant basic blocks can
fail if we're aborting and still don't actually have the needed
MachineBasicBlocks. So avoid trying in that case.
Davide Italiano [Mon, 5 Dec 2016 23:04:21 +0000 (23:04 +0000)]
[SCCP] Remove manual folding of terminator instructions.
There are two cases handled here:
1) a branch on undef
2) a switch with an undef condition.
Both cases are currently handled by ResolvedUndefsIn. If we have
a branch on undef, we force its value to false (which is trivially
foldable). If we have a switch on undef, we force to the first
constant (which is also foldable).
Bob Haarman [Mon, 5 Dec 2016 22:44:00 +0000 (22:44 +0000)]
[pdb] handle missing pdb streams more gracefully
Summary: The code we use to read PDBs assumed that streams we ask it to read exist, and would read memory outside a vector and crash if this wasn't the case. This would, for example, cause llvm-pdbdump to crash on PDBs generated by lld. This patch handles such cases more gracefully: the PDB reading code in LLVM now reports errors when asked to get a stream that is not present, and llvm-pdbdump will report missing streams and continue processing streams that are present.
Tim Northover [Mon, 5 Dec 2016 22:40:13 +0000 (22:40 +0000)]
GlobalISel: place constants correctly in the entry block.
When the entry block was empty after arg lowering, we were always placing
constants at the end. This is probably hamrless while translating the same
block, but horribly wrong once its terminator has been translated. So switch to
inserting at the beginning.
Tim Northover [Mon, 5 Dec 2016 21:47:07 +0000 (21:47 +0000)]
GlobalISel: make G_CONSTANT take a ConstantInt rather than int64_t.
This makes it more similar to the floating-point constant, and also allows for
larger constants to be translated later. There's no real functional change in
this patch though, just syntax updates.
Tim Northover [Mon, 5 Dec 2016 21:40:33 +0000 (21:40 +0000)]
GlobalISel: improve translation fallback for constants.
Returning 0 (NoReg) from getOrCreateVReg leads to unexpected situations later
in the translation. It's better to return a valid (if undefined) register and
let the rest of the instruction carry on as planned.
Keno Fischer [Mon, 5 Dec 2016 21:25:03 +0000 (21:25 +0000)]
[LAA] Prevent invalid IR for loop-invariant bound in loop body
Summary:
If LAA expands a bound that is loop invariant, but not hoisted out
of the loop body, it used to use that value anyway, causing a
non-domination error, because the memcheck block is of course not
dominated by the scalar loop body. Detect this situation and expand
the SCEV expression instead.
[X86] Fix non-intrinsic roundss/roundsd to not read the destination register
This changes the scalar non-intrinsic non-avx roundss/sd instruction
definitions not to read their destination register - allowing partial dependency
breaking.
projects / llvm / log