We want to do this for 2 reasons:
1. Value tracking does not recognize the ashr variant, so it would fail to match for cases like D39766.
2. DAGCombiner does better at producing optimal codegen when we have the cmp+sel pattern.
More detail about what happens in the backend:
1. DAGCombiner has a generic transform for all targets to convert the scalar cmp+sel variant of abs
into the shift variant. That is the opposite of this IR canonicalization.
2. DAGCombiner has a generic transform for all targets to convert the vector cmp+sel variant of abs
into either an ABS node or the shift variant. That is again the opposite of this IR canonicalization.
3. DAGCombiner has a generic transform for all targets to convert the exact shift variants produced by #1 or #2
into an ISD::ABS node. Note: It would be an efficiency improvement if we had #1 go directly to an ABS node
when that's legal/custom.
4. The pattern matching above is incomplete, so it is possible to escape the intended/optimal codegen in a
variety of ways.
a. For #2, the vector path is missing the case for setlt with a '1' constant.
b. For #3, we are missing a match for commuted versions of the shift variants.
5. Therefore, this IR canonicalization can only help get us to the optimal codegen. The version of cmp+sel
produced by this patch will be recognized in the DAG and converted to an ABS node when possible or the
shift sequence when not.
6. In the following examples with this patch applied, we may get conditional moves rather than the shift
produced by the generic DAGCombiner transforms. The conditional move is created using a target-specific
decision for any given target. Whether it is optimal or not for a particular subtarget may be up for debate.
define i32 @abs_shifty(i32 %x) {
%signbit = ashr i32 %x, 31
%add = add i32 %signbit, %x
%abs = xor i32 %signbit, %add
ret i32 %abs
}
define i32 @abs_cmpsubsel(i32 %x) {
%cmp = icmp slt i32 %x, zeroinitializer
%sub = sub i32 zeroinitializer, %x
%abs = select i1 %cmp, i32 %sub, i32 %x
ret i32 %abs
}
define <4 x i32> @abs_shifty_vec(<4 x i32> %x) {
%signbit = ashr <4 x i32> %x, <i32 31, i32 31, i32 31, i32 31>
%add = add <4 x i32> %signbit, %x
%abs = xor <4 x i32> %signbit, %add
ret <4 x i32> %abs
}
define <4 x i32> @abs_cmpsubsel_vec(<4 x i32> %x) {
%cmp = icmp slt <4 x i32> %x, zeroinitializer
%sub = sub <4 x i32> zeroinitializer, %x
%abs = select <4 x i1> %cmp, <4 x i32> %sub, <4 x i32> %x
ret <4 x i32> %abs
}
> $ ./opt -instcombine shiftyabs.ll -S | ./llc -o - -mtriple=x86_64 -mattr=avx
> abs_shifty:
> movl %edi, %eax
> negl %eax
> cmovll %edi, %eax
> retq
>
> abs_cmpsubsel:
> movl %edi, %eax
> negl %eax
> cmovll %edi, %eax
> retq
>
> abs_shifty_vec:
> vpabsd %xmm0, %xmm0
> retq
>
> abs_cmpsubsel_vec:
> vpabsd %xmm0, %xmm0
> retq
>
> $ ./opt -instcombine shiftyabs.ll -S | ./llc -o - -mtriple=aarch64
> abs_shifty:
> cmp w0, #0 // =0
> cneg w0, w0, mi
> ret
>
> abs_cmpsubsel:
> cmp w0, #0 // =0
> cneg w0, w0, mi
> ret
>
> abs_shifty_vec:
> abs v0.4s, v0.4s
> ret
>
> abs_cmpsubsel_vec:
> abs v0.4s, v0.4s
> ret
>
> $ ./opt -instcombine shiftyabs.ll -S | ./llc -o - -mtriple=powerpc64le
> abs_shifty:
> srawi 4, 3, 31
> add 3, 3, 4
> xor 3, 3, 4
> blr
>
> abs_cmpsubsel:
> srawi 4, 3, 31
> add 3, 3, 4
> xor 3, 3, 4
> blr
>
> abs_shifty_vec:
> vspltisw 3, -16
> vspltisw 4, 15
> vsubuwm 3, 4, 3
> vsraw 3, 2, 3
> vadduwm 2, 2, 3
> xxlxor 34, 34, 35
> blr
>
> abs_cmpsubsel_vec:
> vspltisw 3, -16
> vspltisw 4, 15
> vsubuwm 3, 4, 3
> vsraw 3, 2, 3
> vadduwm 2, 2, 3
> xxlxor 34, 34, 35
> blr
>
Differential Revision: https://reviews.llvm.org/D40984
llvm-svn: 320921
Changes to the original scalar loop during LV code gen cause the return value
of Legal->isConsecutivePtr() to be inconsistent with the return value during
legal/cost phases (further analysis and information of the bug is in D39346).
This patch is an alternative fix to PR34965 following the CM_Widen approach
proposed by Ayal and Gil in D39346. It extends InstWidening enum with
CM_Widen_Reverse to properly record the widening decision for consecutive
reverse memory accesses and, consequently, get rid of the
Legal->isConsetuviePtr() call in LV code gen. I think this is a simpler/cleaner
solution to PR34965 than the one in D39346.
Fixes PR34965.
Patch by Diego Caballero, thanks!
Differential Revision: https://reviews.llvm.org/D40742
llvm-svn: 320913
r307148 added an assembly mnemonic spelling correction support and enabled it
on ARM. This enables that support on PowerPC as well.
Patch by Dmitry Venikov, thanks!
Differential Revision: https://reviews.llvm.org/D40552
llvm-svn: 320911
From working on lld I've learned this is generally the
preferred way for several reasons (e.g. more concise, improves
encapsulation).
Differential Revision: https://reviews.llvm.org/D41265
llvm-svn: 320906
Summary:
1. Use stream 0 only for combined module. Previously if combined module was not
processes ThinLTO used the stream for own output. However small changes in input,
could trigger combined module and shuffle outputs making life of llvm::LTO harder.
2. Always process combined module and write output to stream 0. Processing empty
combined module is cheap and allows llvm::LTO users to avoid implementing processing
which is already done in llvm::LTO.
Subscribers: mehdi_amini, inglorion, eraman, hiraditya
Differential Revision: https://reviews.llvm.org/D41267
llvm-svn: 320905
r320895 modified a test so that it needs -enable-import-metadata which
is false by default for NDEBUG, found another place that needs this
added.
llvm-svn: 320903
When unsafe algerbra is allowed calls to cabs(r) can be replaced by:
sqrt(creal(r)*creal(r) + cimag(r)*cimag(r))
Patch by Paul Walker, thanks!
Differential Revision: https://reviews.llvm.org/D40069
llvm-svn: 320901
This is a small step forward to move VPlan stuff to where it should belong (i.e., VPlan.*):
1. VP*Recipe classes in LoopVectorize.cpp are moved to VPlan.h.
2. Many of VP*Recipe::print() and execute() definitions are still left in
LoopVectorize.cpp since they refer to things declared in LoopVectorize.cpp. To
be moved to VPlan.cpp at a later time.
3. InterleaveGroup class is moved from anonymous namespace to llvm namespace.
Referencing it in anonymous namespace from VPlan.h ended up in warning.
Patch by Hideki Saito, thanks!
Differential Revision: https://reviews.llvm.org/D41045
llvm-svn: 320900
Summary:
This implements a missing feature to allow importing of aliases, which
was previously disabled because alias cannot be available_externally.
We instead import an alias as a copy of its aliasee.
Some additional work was required in the IndexBitcodeWriter for the
distributed build case, to ensure that the aliasee has a value id
in the distributed index file (i.e. even when it is not being
imported directly).
This is a performance win in codes that have many aliases, e.g. C++
applications that have many constructor and destructor aliases.
Reviewers: pcc
Subscribers: mehdi_amini, inglorion, eraman, llvm-commits
Differential Revision: https://reviews.llvm.org/D40747
llvm-svn: 320895
Prior to this patch, a predicate wouldn't make sense outside of its
rule. Indeed, it was only during emitting a rule that a predicate would
be made aware of the IDs of the data it is checking. Because of that,
predicates could not be moved around or compared between each other.
NFC.
llvm-svn: 320887
Adds missing support for DW_FORM_data16.
Update of r320852, fixing the unittest to use a hand-coded struct
instead of std::array to guarantee data layout.
Differential Revision: https://reviews.llvm.org/D41090
llvm-svn: 320886
Remove the unused setModule() function; it would be dangerous if someone
actually used it as it wouldn't reset/recompute various other module
related data.
llvm-svn: 320881
This is primarily to reduce stack usage, but ordering the use queue
according to the position in the code (earlier instructions visited
before later ones) reduces the number of unnecessary bottoms due to
visiting instructions out of order, e.g.
%reg1 = copy %reg0
%reg2 = copy %reg0
%reg3 = and %reg1, %reg2
Here, reg3 should be known to be same as reg0-2, but if reg3 is
evaluated after reg1 is updated, but before reg2 is updated, the two
inputs to the and will appear different, causing reg3 to become
bottom.
llvm-svn: 320866
This seemed to work due to a quirk in the X86 MC encoder that didn't emit a REX byte that the AND64ri8 implies when in 32-bit mode. This made the encoding the same as AND32ri8. I tried to add an assert to catch the dropped REX prefix that caught this.
llvm-svn: 320864
The target independent nodes will get legalized to the target specific nodes by their own legalization process. Someday I'd like to stop using a target specific for zero extends and truncates of legal types so the less places we reference the target specific opcode the better.
llvm-svn: 320863
When I wrote it I thought we were missing a potential optimization for KNL. But investigating further shows that for KNL we still do the optimal thing by widening to v4f32 and then using special isel patterns to widen again to zmm a register.
llvm-svn: 320862
This recommits r320823 reverted due to the test failure in sink-foldable.ll and
an unused variable. Added "REQUIRES: aarch64-registered-target" in the test
and removed unused variable.
Original commit message:
Continue trying to sink an instruction if its users in the loop is foldable.
This will allow the instruction to be folded in the loop by decoupling it from
the user outside of the loop.
Reviewers: hfinkel, majnemer, davidxl, efriedma, danielcdh, bmakam, mcrosier
Reviewed By: hfinkel
Subscribers: javed.absar, bmakam, mcrosier, llvm-commits
Differential Revision: https://reviews.llvm.org/D37076
llvm-svn: 320858
Overtime some non-clang formatted code has creeped into llvm-objcopy. This
patch fixes all of that.
Differential Revision: https://reviews.llvm.org/D41262
llvm-svn: 320856
Summary:
Currently we don't handle v32i1/v64i1 insert_vector_elt correctly as we fail to look at the number of elements closely and assume it can only be v16i1 or v8i1.
We also can't type legalize v64i1 insert_vector_elt correctly on KNL due to the type not being byte addressable as required by the legalizing through memory accesses path requires.
For the first issue, the patch now tries to pick a 512-bit register with the correct number of elements and promotes to that.
For the second issue, we now extend the vector to a byte addressable type, do the stores to memory, load the two halves, and then truncate the halves back to the original type. Technically since we changed the type, we may not need two loads, but actually checking that is more work and for the v64i1 case we do need them.
Reviewers: RKSimon, delena, spatel, zvi
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D40942
llvm-svn: 320849
The original memcpy expansion inserted the loop basic block inbetween
the 2 new basic blocks created by splitting the original block the memcpy
call was in. This commit makes the new memcpy expansion do the same to keep the
layout of the IR matching between the old and new implementations.
Differential Review: https://reviews.llvm.org/D41197
llvm-svn: 320848
The asm parser wasn't preventing these from being accepted in 32-bit mode. Instructions that use a GR64 register are protected by the parser rejecting the register in 32-bit mode.
llvm-svn: 320846
This instruction doesn't access memory. It juse use a similar looking memory encoding. Don't require Intel syntax to put "qword ptr" in front of it.
llvm-svn: 320845
This has no effect due to a top level "let Predicates =" around the instructions. But its also not required because the GR64 usage in the instruction guarantees it can never match.
llvm-svn: 320843
