Just like llvm.assume, there are a lot of cases where we can just ignore llvm.experimental.noalias.scope.decl.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D93042
Noticed while looking at D92701 - we only really handle TCK_RecipThroughput gather/scatter costs - for now drop back to the default implementation for non-legal gathers/scatters.
Without FMF, we lower these intrinsics into something like this:
vmaxsd %xmm0, %xmm1, %xmm2
vcmpunordsd %xmm0, %xmm0, %xmm0
vblendvpd %xmm0, %xmm1, %xmm2, %xmm0
But if we can ignore NANs, the single min/max instruction is enough
because there is no need to fix up the x86 logic that corresponds to
X > Y ? X : Y.
We probably want to make other adjustments for FP intrinsics with FMF
to account for specialized codegen (for example, FSQRT).
Differential Revision: https://reviews.llvm.org/D92337
Update costs now that D92095 and D92102 have tweaked the SSE2 implementation
The SSE42 BLENDVPD cost can actually be used on SSE41 as we don't attempt to generate PCMPGT anymore
Add scalar i16/i32/i64 costs as we can do this cheaply with CMOV
This might be a regression for some ARM targets, but that should
be changed in the target-specific overrides.
There is apparently still no default lowering for these nodes,
so I am assuming these intrinsics are not in common use.
X86, PowerPC, and RISC-V for example, just crash given the most
basic IR.
This is re-applying a combination of f7eac51b9b3f and 8ec7ea3ddce7 as one patch
to avoid regressions now that we have better testing in place.
Those were reverted with 32dd5870ee31 because of crashing in experimental intrinsics.
That bug should be fixed with 7ae346434.
Paraphrased original commit messages:
This is the last step in removing cost-kind as a consideration in the
basic class model for intrinsics.
See D89461 for the start of that.
Subsequent commits dealt with each of the special-case intrinsics that
had customization here in the basic class. This should remove a barrier
to retrying D87188 (canonicalization to the abs intrinsic).
The ARM and x86 cost diffs seen here may be wrong because the
target-specific overrides have their own bugs, but we hope this is
less wrong - if something has a significant throughput cost, then it
should have a significant size / blended cost too by default.
The only behavioral diff in current regression tests is shown in the
x86 scatter-gather test (which is misplaced or broken because it runs
the entire -O3 pipeline) - we unrolled less, and we assume that is
a improvement.
Exception: in general, we want the *size* cost for a scalar call to be
cheap even if the other costs are expensive - we expect it to just be
a branch with some optional stack manipulation.
It is likely that we will want to carve out some
exceptions/overrides to this rule as follow-up patches for
calls that have some general and/or target-specific difference
to the expected lowering.
This was noticed as a regression in unrolling, so we have a test
for that now along with a couple of direct cost model tests.
If the assumed scalarization costs for the oversized vector
calls are not realistic, that would be another follow-up
refinement of the cost models.
Differential Revision: https://reviews.llvm.org/D90554
The constrained intrinsics have metadata arguments, so the
tests here were crashing as noted in D90554 (and that was
reverted even though this bug exists independently of that
change).
This is a partial un-revert of 32dd5870ee31 (originally df09f82599 ).
I'm adding back the baseline tests first, so we don't have
to back-track as much in case there are still problems.
as it's causing crashes in the optimizer. A reduced testcase has been posted as a follow-up.
This reverts commit f7eac51b9b3f780c96ca41913293851c5acb465b.
Temporarily Revert "[CostModel] make default size cost for libcalls small (again)" as it depends upon the primary revert.
This reverts commit 8ec7ea3ddce7379e13e8dfb4a5260a6d2004aa1c.
Temporarily Revert "[CostModel] add tests for math library calls; NFC" as it depends upon the primary revert.
This reverts commit df09f825995b10da03f148133c119f52c94fd6e4.
Temporarily Revert "[LoopUnroll] add test for full unroll that is sensitive to cost-model; NFC" as it depends upon the primary revert.
This reverts commit 618d555e8d926a83161774df2035519c387269db.
We can use GF2P8AFFINEQB to reverse bits in a byte. Shuffles are needed to reverse the bytes in elements larger than i8. LegalizeVectorOps takes care of inserting the shuffle for the larger element size.
We already have Custom lowering for v16i8 with SSSE3, v32i8 with AVX, and v64i8 with AVX512BW.
I think we might be able to use this for scalars too by moving into a vector and back. But I'll save that for a follow up as its a little more involved.
Reviewed By: RKSimon, pengfei
Differential Revision: https://reviews.llvm.org/D91515
This was changed recently with D90554 / f7eac51b9b3f
...because we had a regression testing blindspot for intrinsics
that are expected to be lowered to libcalls.
In general, we want the *size* cost for a scalar call to be cheap
even if the other costs are expensive - we expect it to just be
a branch with some optional stack manipulation.
It is likely that we will want to carve out some
exceptions/overrides to this rule as follow-up patches for
calls that have some general and/or target-specific difference
to the expected lowering.
This was noticed as a regression in unrolling, so we have a test
for that now along with a couple of direct cost model tests.
If the assumed scalarization costs for the oversized vector
calls are not realistic, that would be another follow-up
refinement of the cost models.
This is the last step in removing cost-kind as a consideration in the basic class model for intrinsics.
See D89461 for the start of that.
Subsequent commits dealt with each of the special-case intrinsics that had customization here in the
basic class. This should remove a barrier to retrying
D87188 (canonicalization to the abs intrinsic).
The ARM and x86 cost diffs seen here may be wrong because the target-specific overrides have their own
bugs, but we hope this is less wrong - if something has a significant throughput cost, then it should
have a significant size / blended cost too by default.
The only behavioral diff in current regression tests is shown in the x86 scatter-gather test (which is
misplaced or broken because it runs the entire -O3 pipeline) - we unrolled less, and we assume that is
a improvement.
Differential Revision: https://reviews.llvm.org/D90554
As noted in D90554, there's an opcode typo in using an easily
misused cost model API: getCmpSelInstrCost(). Beyond that, the
assumed sequence of ops is questionable, but that would be
another patch.
My guess is that the x86 test diffs show that we are probably
wrong both before and after this change, so there will be no
practical difference.
As an example, I tried this test which shows a cost of '7'
either way:
define <4 x i32> @sadd(<4 x i32> %va, <4 x i32> %vb) {
%V4I32 = call {<4 x i32>, <4 x i1>} @llvm.sadd.with.overflow.v4i32(<4 x i32> %va, <4 x i32> %vb)
%ov = extractvalue {<4 x i32>, <4 x i1>} %V4I32, 1
%r = extractvalue {<4 x i32>, <4 x i1>} %V4I32, 0
%z = select <4 x i1> %ov, <4 x i32> <i32 42, i32 42, i32 42, i32 42>, <4 x i32> %r
ret <4 x i32> %z
}
$ llc -o - sadd.ll -mattr=avx
vpaddd %xmm1, %xmm0, %xmm2
vpcmpgtd %xmm2, %xmm0, %xmm0
vpxor %xmm0, %xmm1, %xmm0
vblendvps %xmm0, LCPI0_0(%rip), %xmm2, %xmm0a
Differential Revision: https://reviews.llvm.org/D90681
As noticed in D90554 ,
the AVX2 costs for 256-bit vectors did not include FMAXNUM entries,
so we fell back to AVX1 which assumes those ops will be split into
128-bit halves or something close to that.
Differential Revision: https://reviews.llvm.org/D90613
I'm assuming the standard size integer instructions for this end up as something like:
mulq %rsi
seto %al
And the 'mul' generally has reciprocal throughput of 1 on typical implementations
(higher latency, but that's not handled here).
The default costs may end up much higher than that, and that's what we see in the test diffs.
Vector types are left as a 'TODO'.
Differential Revision: https://reviews.llvm.org/D90431
Completing the series of FIXME removals for special-case intrinsics:
50dfa19cc799
f2c25c70791d
c963bde0152a
01ea93d85d6e
This one looks quite different than the others. The size/blended
cost is still potentially very far off from the throughput cost,
but this is hopefully not worse on the whole. It looks like the
underlying costs for the expanded shift/logic have their own
cost-kind limitations. Also, we are not asking the target if
it has a legal funnel shift op, so we just assume that the
intrinsic gets expanded.
This was originally part of:
f2c25c70791d
but that was reverted because there was an underlying bug in
processing the vector type of these intrinsics. That was
fixed with:
74ffc823ed21
This is similar in spirit to 01ea93d85d6e (memcpy) except that
here the underlying caller assumptions were created for vectorizer
use (throughput) rather than other passes.
That meant targets could have an enormous throughput cost with no
corresponding size, latency, or blended cost increase.
Paraphrasing from the previous commits:
This may not make sense for some callers, but at least now the
costs will be consistently wrong instead of mysteriously wrong.
Targets should provide better overrides if the current modeling
is not accurate.
In each 128-lane, if there is at least one index is demanded and not all
indices are demanded and this 128-lane is not the first 128-lane of the
legalized-vector, then this 128-lane needs a extracti128;
If in each 128-lane, there is at least one index is demanded, this 128-lane
needs a inserti128.
The following cases will help you build a better understanding:
Assume we insert several elements into a v8i32 vector in avx2,
Case#1: inserting into 1th index needs vpinsrd + inserti128
Case#2: inserting into 5th index needs extracti128 + vpinsrd +
inserti128
Case#3: inserting into 4,5,6,7 index needs 4*vpinsrd + inserti128.
Reviewed By: pengfei, RKSimon
Differential Revision: https://reviews.llvm.org/D89767
This allows using annotation in a much more contexts than it currently has.
especially when annotation with template or constexpr.
Reviewed By: aaron.ballman
Differential Revision: https://reviews.llvm.org/D88645
This is a modified 2nd try of 22d10b8ab44f
(reverted by 1c8371692d because it managed
to expose an existing crashing bug that should be fixed by
74ffc823 ).
Original commit message:
This is similar in spirit to 01ea93d85d6e (memcpy) except that
here the underlying caller assumptions were created for vectorizer
use (throughput) rather than other passes.
That meant targets could have an enormous throughput cost with no
corresponding size, latency, or blended cost increase.
The ARM costs show a small difference between throughput and
size because there's an underlying difference in cmp/sel
costs that is also predicated on cost-kind.
Paraphrasing from the previous commits:
This may not make sense for some callers, but at least now the
costs will be consistently wrong instead of mysteriously wrong.
Targets should provide better overrides if the current modeling
is not accurate.
I'm not sure if/how this ever worked, but it must not be tested
currently because the basic tests added here were crashing as
noted in the post-review comments for 1c83716 (which reverted
another cost-model fix in 22d10b8ab44f).
This is similar in spirit to 01ea93d85d6e (memcpy) except that
here the underlying caller assumptions were created for vectorizer
use (throughput) rather than other passes.
That meant targets could have an enormous throughput cost with no
corresponding size, latency, or blended cost increase.
The ARM costs show a small difference between throughput and
size because there's an underlying difference in cmp/sel
costs that is also predicated on cost-kind.
Paraphrasing from the previous commits:
This may not make sense for some callers, but at least now the
costs will be consistently wrong instead of mysteriously wrong.
Targets should provide better overrides if the current modeling
is not accurate.
The default implementation base returns TCC_Expensive (currently
set to '4'), so that explains the test diff. This probably does
not make sense for most callers, but at least now the costs will
be consistently wrong instead of mysteriously wrong.
The ARM target has an override that tries to model codegen expansion,
and that should likely be adapted for general usage.
This probably does not affect anything because the vectorizers are
the primary users of the throughput cost, but memcpy is not listed
as a trivially vectorizable intrinsic.
The cost modeling for intrinsics is a patchwork based on different
expectations from the callers, so it's a mess. I'm hoping to untangle
this to allow canonicalization to the new min/max intrinsics in IR.
The general goal is to remove the cost-kind restriction here in the
basic implementation class. Ie, if some intrinsic has throughput cost
of 104, assume that it has the same size, latency, and blended costs.
Effectively, an intrinsic with cost N is composed of N simple
instructions. If that's not correct, the target should provide a more
accurate override.
The x86-64 SSE2 subtarget cost diffs require explanation:
1. The scalar ctlz/cttz are assuming "BSR+XOR+CMOV" or
"TEST+BSF+CMOV/BRANCH", so not cheap.
2. The 128-bit SSE vector width versions assume cost of 18 or 26
(no explanation provided in the tables, but this corresponds to a
bunch of shift/logic/compare).
3. The 512-bit vectors in the test file are scaled up by a factor of
4 from the legal vector width costs.
4. The plain latency cost-kind is not affected in this patch because
that calc is diverted before we get to getIntrinsicInstrCost().
Differential Revision: https://reviews.llvm.org/D89461
This is my first LLVM patch, so please tell me if there are any process issues.
The main observation for this patch is that we can lower UMIN/UMAX with v8i16 by using unsigned saturated subtractions in a clever way. Previously this operation was lowered by turning the signbit of both inputs and the output which turns the unsigned minimum/maximum into a signed one.
We could use this trick in reverse for lowering SMIN/SMAX with v16i8 instead. In terms of latency/throughput this is the needs one large move instruction. It's just that the sign bit turning has an increased chance of being optimized further. This is particularly apparent in the "reduce" test cases. However due to the slight regression in the single use case, this patch no longer proposes this.
Unfortunately this argument also applies in reverse to the new lowering of UMIN/UMAX with v8i16 which regresses the "horizontal-reduce-umax", "horizontal-reduce-umin", "vector-reduce-umin" and "vector-reduce-umax" test cases a bit with this patch. Maybe some extra casework would be possible to avoid this. However independent of that I believe that the benefits in the common case of just 1 to 3 chained min/max instructions outweighs the downsides in that specific case.
Patch By: @TomHender (Tom Hender) ActuallyaDeviloper
Differential Revision: https://reviews.llvm.org/D87236
Other types can be handled in future patches but their uniform / non-uniform costs are more similar and don't appear to cause many vectorization issues.
