This was discussed in D48401 as another improvement for:
https://bugs.llvm.org/show_bug.cgi?id=37806
If we have 2 different variable values, then we shuffle (select) those lanes,
shuffle (select) the constants, and then perform the binop. This eliminates a binop.
The new shuffle uses the same shuffle mask as the existing shuffle, so there's no
danger of creating a difficult shuffle.
All of the earlier constraints still apply, but we also check for extra uses to
avoid creating more instructions than we'll remove.
Additionally, we're disallowing the fold for div/rem because that could expose a
UB hole.
Differential Revision: https://reviews.llvm.org/D48678
llvm-svn: 335974
There's no way to expose this difference currently,
but we should use the updated variable because the
original opcodes can go stale if we transform into
something new.
llvm-svn: 335920
This is an enhancement to D48401 that was discussed in:
https://bugs.llvm.org/show_bug.cgi?id=37806
We can convert a shift-left-by-constant into a multiply (we canonicalize IR in the other
direction because that's generally better of course). This allows us to remove the shuffle
as we do in the regular opcodes-are-the-same cases.
This requires a small hack to make sure we don't introduce any extra poison:
https://rise4fun.com/Alive/ZGv
Other examples of opcodes where this would work are add+sub and fadd+fsub, but we already
canonicalize those subs into adds, so there's nothing to do for those cases AFAICT. There
are planned enhancements for opcode transforms such or -> add.
Note that there's a different fold needed if we've already managed to simplify away a binop
as seen in the test based on PR37806, but we manage to get that one case here because this
fold is positioned above the demanded elements fold currently.
Differential Revision: https://reviews.llvm.org/D48485
llvm-svn: 335888
With non-commutative binops, we could be using the same
variable value as operand 0 in 1 binop and operand 1 in
the other, so we have to check for that possibility and
bail out.
llvm-svn: 335312
This is the simplest case from PR37806:
https://bugs.llvm.org/show_bug.cgi?id=37806
If we have a common variable operand used in a pair of binops with vector constants
that are vector selected together, then we can constant shuffle the constant vectors
to eliminate the shuffle instruction.
This has some tricky parts that are hopefully addressed in the tests and their
respective comments:
1. If the shuffle mask contains an undef element, then that lane of the result is
undef:
http://llvm.org/docs/LangRef.html#shufflevector-instruction
Therefore, we can replace the constant in that lane with an undef value except
for div/rem. With div/rem, an undef in the divisor would cause the whole op to
be undef. So I'm using the same hack as in D47686 - replace the undefs with '1'.
2. Intersect the wrapping and FMF of the original binops for the new binop. There
should be no extra poison or fast-math potential in the new binop that wasn't
possible in the original code.
3. Disregard other uses. Given that we're eliminating uses (shortening the
dependency chain), I think that's always the right IR canonicalization. But
I purposely chose the udiv test to demonstrate the scenario where both
intermediate values have other uses because that seems likely worse for
codegen with an expensive math op. This seems like a very rare possibility to
me, so I don't think it requires a backend patch first.
Differential Revision: https://reviews.llvm.org/D48401
llvm-svn: 335283
The 1st attempt at this:
https://reviews.llvm.org/rL314117
was reverted at:
https://reviews.llvm.org/rL314118
because of bot fails for clang tests that were checking optimized IR. That should be fixed with:
https://reviews.llvm.org/rL314144
...so try again.
Original commit message:
The transform to convert an extract-of-a-select-of-vectors was added at:
https://reviews.llvm.org/rL194013
And a question about the validity of this transform was raised in the review:
https://reviews.llvm.org/D1539:
...but not answered AFAICT>
Most of the motivating cases in that patch are now handled by other combines. These are the tests that were added with
the original commit, but they are not regressing even after we remove the transform in this patch.
The diffs we see after removing this transform cause us to avoid increasing the instruction count, so we don't want to do
those transforms as canonicalizations.
The motivation for not turning a vector-select-of-vectors into a scalar operation is shown in PR33301:
https://bugs.llvm.org/show_bug.cgi?id=33301
...in those cases, we'll get vector ops with this patch rather than the vector/scalar mix that we currently see.
Differential Revision: https://reviews.llvm.org/D38006
llvm-svn: 314147
The transform to convert an extract-of-a-select-of-vectors was added at:
rL194013
And a question about the validity of this transform was raised in the review:
https://reviews.llvm.org/D1539:
...but not answered AFAICT>
Most of the motivating cases in that patch are now handled by other combines. These are the tests that were added with
the original commit, but they are not regressing even after we remove the transform in this patch.
The diffs we see after removing this transform cause us to avoid increasing the instruction count, so we don't want to do
those transforms as canonicalizations.
The motivation for not turning a vector-select-of-vectors into a scalar operation is shown in PR33301:
https://bugs.llvm.org/show_bug.cgi?id=33301
...in those cases, we'll get vector ops with this patch rather than the vector/scalar mix that we currently see.
Differential Revision: https://reviews.llvm.org/D38006
llvm-svn: 314117
Recurse instead of returning on the first found optimization. Also, return early in the caller
instead of continuing because that allows another round of simplification before we might
potentially lose undef information from a shuffle mask by eliminating the shuffle.
As noted in the review, we could probably do better and be more efficient by moving all of
demanded elements into a separate pass, but this is yet another quick fix to instcombine.
Differential Revision: https://reviews.llvm.org/D37236
llvm-svn: 312248
Summary:
If the first insertelement instruction has multiple users and inserts at
position 0, we can re-use this instruction when folding a chain of
insertelement instructions. As we need to generate the first
insertelement instruction anyways, this should be a strict improvement.
We could get rid of the restriction of inserting at position 0 by
creating a different shufflemask, but it is probably worth to keep the
first insertelement instruction with position 0, as this is easier to do
efficiently than at other positions I think.
Reviewers: grosser, mkuper, fpetrogalli, efriedma
Reviewed By: fpetrogalli
Subscribers: gareevroman, llvm-commits
Differential Revision: https://reviews.llvm.org/D37064
llvm-svn: 312110
Previously the InstCombiner class contained a pointer to an IR builder that had been passed to the constructor. Sometimes this would be passed to helper functions as either a pointer or the pointer would be dereferenced to be passed by reference.
This patch makes it a reference everywhere including the InstCombiner class itself so there is more inconsistency. This a large, but mechanical patch. I've done very minimal formatting changes on it despite what clang-format wanted to do.
llvm-svn: 307451
Summary: This matches the behavior we already had for compares and makes us consistent everywhere.
Reviewers: dberlin, hfinkel, spatel
Reviewed By: dberlin
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D33604
llvm-svn: 305049
This fixes a bug that can cause extractelements with operands that
haven't been defined yet to be inserted at a wrong point when
optimising insertelements.
Patch by Karl Hylen.
Differential Revision: https://reviews.llvm.org/D33449
llvm-svn: 304701
insertelement (insertelement X, Y, IdxC1), ScalarC, IdxC2 -->
insertelement (insertelement X, ScalarC, IdxC2), Y, IdxC1
As noted in the code comment and seen in the test changes, the motivation is that by pulling
constant insertion up, we may be able to constant fold some insertelement instructions.
Differential Revision: https://reviews.llvm.org/D31196
llvm-svn: 298520
This adds a combine that canonicalizes a chain of inserts which broadcasts
a value into a single insert + a splat shufflevector.
This fixes PR31286.
Differential Revision: https://reviews.llvm.org/D27992
llvm-svn: 290641
After r289755, the AssumptionCache is no longer needed. Variables affected by
assumptions are now found by using the new operand-bundle-based scheme. This
new scheme is more computationally efficient, and also we need much less
code...
llvm-svn: 289756
Removing the limitation in visitInsertElementInst() causes several regressions
because we're not prepared to fold sequences of shuffles or inserts and extracts
separated by shuffles. Fixing that appears to be a difficult mission because we
are purposely trying to avoid creating shuffles with arbitrary shuffle masks
because some targets may choke on those.
https://llvm.org/bugs/show_bug.cgi?id=30923
llvm-svn: 286423
If inserting more than one constant into a vector:
define <4 x float> @foo(<4 x float> %x) {
%ins1 = insertelement <4 x float> %x, float 1.0, i32 1
%ins2 = insertelement <4 x float> %ins1, float 2.0, i32 2
ret <4 x float> %ins2
}
InstCombine could reduce that to a shufflevector:
define <4 x float> @goo(<4 x float> %x) {
%shuf = shufflevector <4 x float> %x, <4 x float> <float undef, float 1.0, float 2.0, float undef>, <4 x i32><i32 0, i32 5, i32 6, i32 3>
ret <4 x float> %shuf
}
Also, InstCombine tries to convert shuffle instruction to single insertelement, if one of the vectors is a constant vector and only a single element from this constant should be used in shuffle, i.e.
shufflevector <4 x float> %v, <4 x float> <float undef, float 1.0, float
undef, float undef>, <4 x i32> <i32 0, i32 5, i32 undef, i32 undef> ->
insertelement <4 x float> %v, float 1.0, 1
Differential Revision: https://reviews.llvm.org/D24182
llvm-svn: 282237
The motivating case occurs with SSE/AVX scalar intrinsics, so this is a first step towards
shrinking that to a single shufflevector.
Note that the transform is intentionally limited to shuffles that are equivalent to vector
selects to avoid creating arbitrary shuffle masks that may not lower well.
This should solve PR29126:
https://llvm.org/bugs/show_bug.cgi?id=29126
Differential Revision: https://reviews.llvm.org/D23886
llvm-svn: 280504
scalarizePHI only looked for phis that have exactly two uses - the "latch"
use, and an extract. Unfortunately, we can not assume all equivalent extracts
are CSE'd, since InstCombine itself may create an extract which is a duplicate
of an existing one. This extends it to handle several distinct extracts from
the same index.
This should fix at least some of the performance regressions from PR27988.
Differential Revision: http://reviews.llvm.org/D20983
llvm-svn: 271961
We would infinite loop because we created a shufflevector that was wider than
needed and then failed to combine that with the insertelement. When subsequently
visiting the extractelement from that shuffle, we see that it's unnecessary,
delete it, and trigger another visit to the insertelement.
llvm-svn: 259236
Limit this transform to a basic block and guard against PHIs.
Hopefully, this fixes the remaining failures in PR25999:
https://llvm.org/bugs/show_bug.cgi?id=25999
llvm-svn: 257133
This is an extension of the shuffle combining from r203229:
http://reviews.llvm.org/rL203229
The idea is to widen a short input vector with undef elements so the
existing shuffle transform for extract/insert can kick in.
The motivation is to finally solve PR2109:
https://llvm.org/bugs/show_bug.cgi?id=2109
For that example, the IR becomes:
%1 = bitcast <2 x i32>* %P to <2 x float>*
%ld1 = load <2 x float>, <2 x float>* %1, align 8
%2 = shufflevector <2 x float> %ld1, <2 x float> undef, <4 x i32> <i32 0, i32 1, i32 undef, i32 undef>
%i2 = shufflevector <4 x float> %A, <4 x float> %2, <4 x i32> <i32 0, i32 1, i32 4, i32 5>
ret <4 x float> %i2
And x86 SSE output improves from:
movq (%rdi), %xmm1 ## xmm1 = mem[0],zero
movdqa %xmm1, %xmm2
shufps $229, %xmm2, %xmm2 ## xmm2 = xmm2[1,1,2,3]
shufps $48, %xmm0, %xmm1 ## xmm1 = xmm1[0,0],xmm0[3,0]
shufps $132, %xmm1, %xmm0 ## xmm0 = xmm0[0,1],xmm1[0,2]
shufps $32, %xmm0, %xmm2 ## xmm2 = xmm2[0,0],xmm0[2,0]
shufps $36, %xmm2, %xmm0 ## xmm0 = xmm0[0,1],xmm2[2,0]
retq
To the almost optimal:
movhpd (%rdi), %xmm0
Note: There's a tension in the existing transform related to generating
arbitrary shufflevector masks. We avoid that in other places in InstCombine
because we're scared that codegen can't handle strange masks, but it looks
like we're ok with producing those here. I purposely chose weird insert/extract
indexes for the regression tests to see the effect in these cases.
For PowerPC+Altivec, AArch64, and X86+SSE/AVX, I think the codegen is equal or
better for these examples.
Differential Revision: http://reviews.llvm.org/D15096
llvm-svn: 256394
InstCombine didn't realize that it needs to use DataLayout to determine
how wide pointers are. This lead to assertion failures.
This fixes PR23113.
llvm-svn: 234046
Adding nullptr to all the IRBuilder stuff because it's the first thing
that fails to build when testing without the back-compat functions, so
I'll keep having to re-add these locally for each chunk of migration I
do. Might as well check them in to save me the churn. Eventually I'll
have to migrate these too, but I'm going breadth-first.
llvm-svn: 232270
Summary:
Now that the DataLayout is a mandatory part of the module, let's start
cleaning the codebase. This patch is a first attempt at doing that.
This patch is not exactly NFC as for instance some places were passing
a nullptr instead of the DataLayout, possibly just because there was a
default value on the DataLayout argument to many functions in the API.
Even though it is not purely NFC, there is no change in the
validation.
I turned as many pointer to DataLayout to references, this helped
figuring out all the places where a nullptr could come up.
I had initially a local version of this patch broken into over 30
independant, commits but some later commit were cleaning the API and
touching part of the code modified in the previous commits, so it
seemed cleaner without the intermediate state.
Test Plan:
Reviewers: echristo
Subscribers: llvm-commits
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 231740
Summary: SROA generates code that isn't quite as easy to optimize and contains unusual-sized shuffles, but that code is generally correct. As discussed in D7487 the right place to clean things up is InstCombine, which will pick up the type-punning pattern and transform it into a more obvious bitcast+extractelement, while leaving the other patterns SROA encounters as-is.
Test Plan: make check
Reviewers: jvoung, chandlerc
Subscribers: llvm-commits
llvm-svn: 230560
