We are combining shuffles to bit shifts before unary permutes, which means we can't fold loads plus the destination register is destructive
llvm-svn: 306978
We are combining shuffles to bit shifts before unary permutes, which means we can't fold loads plus the destination register is destructive
The 32-bit shuffles are a bit tricky and will be dealt with in a later patch
llvm-svn: 306977
[X86][AVX512] Improve lowering of AVX512 compare intrinsics (remove redundant shift left+right instructions).
AVX512 compare instructions return v*i1 types.
In cases where the number of elements in the returned value are less than 8, clang adds zeroes to get a mask of v8i1 type.
Later on it's replaced with CONCAT_VECTORS, which then is lowered to many DAG nodes including insert/extract element and shift right/left nodes.
The fact that AVX512 compare instructions put the result in a k register and zeroes all its upper bits allows us to remove the extra nodes simply by copying the result to the required register class.
When lowering, identify these cases and transform them into an INSERT_SUBVECTOR node (marked legal), then catch this pattern in instructions selection phase and transform it into one avx512 cmp instruction.
Differential Revision: https://reviews.llvm.org/D33188
llvm-svn: 306402
Convert vector increment or decrement to sub/add with an all-ones constant:
add X, <1, 1...> --> sub X, <-1, -1...>
sub X, <1, 1...> --> add X, <-1, -1...>
The all-ones vector constant can be materialized using a pcmpeq instruction that is
commonly recognized as an idiom (has no register dependency), so that's better than
loading a splat 1 constant.
AVX512 uses 'vpternlogd' for 512-bit vectors because there is apparently no better
way to produce 512 one-bits.
The general advantages of this lowering are:
1. pcmpeq has lower latency than a memop on every uarch I looked at in Agner's tables,
so in theory, this could be better for perf, but...
2. That seems unlikely to affect any OOO implementation, and I can't measure any real
perf difference from this transform on Haswell or Jaguar, but...
3. It doesn't look like it from the diffs, but this is an overall size win because we
eliminate 16 - 64 constant bytes in the case of a vector load. If we're broadcasting
a scalar load (which might itself be a bug), then we're replacing a scalar constant
load + broadcast with a single cheap op, so that should always be smaller/better too.
4. This makes the DAG/isel output more consistent - we use pcmpeq already for padd x, -1
and psub x, -1, so we should use that form for +1 too because we can. If there's some
reason to favor a constant load on some CPU, let's make the reverse transform for all
of these cases (either here in the DAG or in a later machine pass).
This should fix:
https://bugs.llvm.org/show_bug.cgi?id=33483
Differential Revision: https://reviews.llvm.org/D34336
llvm-svn: 306289
This is very similar to the transform in:
https://reviews.llvm.org/rL306040
...but in this case, we use cmp X, 1 to set the carry bit as needed.
Again, we can show that all of these are logically equivalent (although
InstCombine currently canonicalizes to a form not seen here), and if
we believe IACA, then this is the smallest/fastest code. Eg, with SNB:
| Num Of | Ports pressure in cycles | |
| Uops | 0 - DV | 1 | 2 - D | 3 - D | 4 | 5 | |
---------------------------------------------------------------------
| 1 | 1.0 | | | | | | | cmp edi, 0x1
| 2 | | 1.0 | | | | 1.0 | CP | sbb eax, eax
The larger motivation is to clean up all select-of-constants combining/lowering
because we're missing some common cases.
llvm-svn: 306072
Our handling of select-of-constants is lumpy in IR (https://reviews.llvm.org/D24480),
lumpy in DAGCombiner, and lumpy in X86ISelLowering. That's why we only had the 'sbb'
codegen in 1 out of the 4 tests. This is a step towards smoothing that out.
First, show that all of these IR forms are equivalent:
http://rise4fun.com/Alive/mx
Second, show that the 'sbb' version is faster/smaller. IACA output for SandyBridge
(later Intel and AMD chips are similar based on Agner's tables):
This is the "obvious" x86 codegen (what gcc appears to produce currently):
| Num Of | Ports pressure in cycles | |
| Uops | 0 - DV | 1 | 2 - D | 3 - D | 4 | 5 | |
---------------------------------------------------------------------
| 1* | | | | | | | | xor eax, eax
| 1 | 1.0 | | | | | | CP | test edi, edi
| 1 | | | | | | 1.0 | CP | setnz al
| 1 | | 1.0 | | | | | CP | neg eax
This is the adc version:
| 1* | | | | | | | | xor eax, eax
| 1 | 1.0 | | | | | | CP | cmp edi, 0x1
| 2 | | 1.0 | | | | 1.0 | CP | adc eax, 0xffffffff
And this is sbb:
| 1 | 1.0 | | | | | | | neg edi
| 2 | | 1.0 | | | | 1.0 | CP | sbb eax, eax
If IACA is trustworthy, then sbb became a single uop in Broadwell, so this will be
clearly better than the alternatives going forward.
llvm-svn: 306040
Masked gather for vector length 2 is lowered incorrectly for element type i32.
The type <2 x i32> was automatically extended to <2 x i64> and we generated VPGATHERQQ instead of VPGATHERQD.
The type <2 x float> is extended to <4 x float>, so there is no bug for this type, but the sequence may be more optimal.
In this patch I'm fixing <2 x i32>bug and optimizing <2 x float> sequence for GATHERs only. The same fix should be done for Scatters as well.
Differential revision: https://reviews.llvm.org/D34343
llvm-svn: 305987
There are a couple of potential improvements as seen in the IR and asm:
1. We're unnecessarily extending to a larger type to compare values.
2. The codegen for (select cond, 1, -1) could avoid a cmov.
(or we could change the order of the compares, so we have a select with 0 operand)
llvm-svn: 305802
Target shuffle combining now supports the matching of INSERT_VECTOR_ELT/PINSRW/PINSRB for merging multiple insertions into shuffles/bitmasks.
llvm-svn: 305788
AVX512 compare instructions return v*i1 types.
In cases where the number of elements in the returned value are less than 8, clang adds zeroes to get a mask of v8i1 type.
Later on it's replaced with CONCAT_VECTORS, which then is lowered to many DAG nodes including insert/extract element and shift right/left nodes.
The fact that AVX512 compare instructions put the result in a k register and zeroes all its upper bits allows us to remove the extra nodes simply by copying the result to the required register class.
When lowering, identify these cases and transform them into an INSERT_SUBVECTOR node (marked legal), then catch this pattern in instructions selection phase and transform it into one avx512 cmp instruction.
Differential Revision: https://reviews.llvm.org/D33188
llvm-svn: 305465
We know that shuffle masks are power-of-2 sizes, but there's no way (?) for LLVM to know that,
so hack combineX86ShufflesRecursively() to be much faster by replacing div/rem with shift/mask.
This makes the motivating compile-time bug in PR32037 ( https://bugs.llvm.org/show_bug.cgi?id=32037 )
about 9% faster overall.
Differential Revision: https://reviews.llvm.org/D34174
llvm-svn: 305398
Much of PR32037's compile time regression is due to getTargetConstantBitsFromNode always creating large (>64bit) APInts during the bitcasting from the source data to the destination bitwidth.
This commit avoids this bitcast stage if the data is already the correct bitwidth.
llvm-svn: 305284
This step is just intended to reduce code duplication rather than change any functionality.
A follow-up would be to replace PPCTargetLowering::spliceIntoChain() usage with this new helper.
Differential Revision: https://reviews.llvm.org/D33649
llvm-svn: 305192
I was looking closer at the x86 test diffs in D33866, and the first change seems like it
shouldn't happen in the first place. So this patch will resolve that.
Using Agner's tables and AMD docs, vperm2f128 and vinsertf128 have identical timing for
any given CPU model, so we should be able to interchange those without affecting perf.
But as we can see in some of the diffs here, using vperm2f128 allows load folding, so
we should take that opportunity to reduce code size and register pressure.
A secondary advantage is making AVX1 and AVX2 codegen more similar. Given that vperm2f128
was introduced with AVX1, we should be selecting it in all of the same situations that we
would with AVX2. If there's some reason that an AVX1 CPU would not want to use this
instruction, that should be fixed up in a later pass.
Differential Revision: https://reviews.llvm.org/D33938
llvm-svn: 305171
If we know that both operands of an unsigned integer vector comparison are non-negative,
then it's safe to directly use a signed-compare-greater-than instruction (the only non-equality
integer vector compare predicate provided by SSE/AVX).
We're intentionally not changing the condition code to signed in order to preserve the
existing transforms that use min/max/psubus below here.
This should solve PR33276:
https://bugs.llvm.org/show_bug.cgi?id=33276
Differential Revision: https://reviews.llvm.org/D33862
llvm-svn: 304909
We currently generate BUILD_VECTOR as a tree of UNPCKL shuffles of the same type:
e.g. for v4f32:
Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
: unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
The issue is because we are not placing sequential vector elements together early enough, we fail to recognise many combinable patterns - consecutive scalar loads, extractions etc.
Instead, this patch unpacks progressively larger sequential vector elements together:
e.g. for v4f32:
Step 1: unpcklps 0, 2 ==> X: <?, ?, 1, 0>
: unpcklps 1, 3 ==> Y: <?, ?, 3, 2>
Step 2: unpcklpd X, Y ==> <3, 2, 1, 0>
This does mean that we are creating UNPCKL shuffle of different value types, but the relevant combines that benefit from this are quite capable of handling the additional BITCASTs that are now included in the shuffle tree.
Differential Revision: https://reviews.llvm.org/D33864
llvm-svn: 304688
Since r288804, we try to lower build_vectors on AVX using broadcasts of
float/double. However, when we broadcast integer values that happen to
have a NaN float bitpattern, we lose the NaN payload, thereby changing
the integer value being broadcast.
This is caused by ConstantFP::get, to which we pass the splat i32 as
a float (by bitcasting it using bitsToFloat). ConstantFP::get takes
a double parameter, so we end up lossily converting a single-precision
NaN to double-precision.
Instead, avoid any kinds of conversions by directly building an APFloat
from the splatted APInt.
Note that this also fixes another piece of code (broadcast of
subvectors), that currently isn't susceptible to the same problem.
Also note that we could really just use APInt and ConstantInt
throughout: the constant pool type doesn't matter much. Still, for
consistency, use the appropriate type.
llvm-svn: 304590
This might give a few better opportunities to optimize these to memcpy
rather than loops - also a few minor cleanups (StringRef-izing,
templating (to avoid std::function indirection), etc).
The SmallVector::assign(iter, iter) could be improved with the use of
SFINAE, but the (iter, iter) ctor and append(iter, iter) need it to and
don't have it - so, workaround it for now rather than bothering with the
added complexity.
(also, as noted in the added FIXME, these assign ops could potentially
be optimized better at least for non-trivially-copyable types)
llvm-svn: 304566
Summary:
Add an early combine to match patterns such as:
(i16 bitcast (v16i1 x))
->
(i16 movmsk (v16i8 sext (v16i1 x)))
This combine needs to happen early enough before
type-legalization scalarizes the result of the setcc.
Reviewers: igorb, craig.topper, RKSimon
Subscribers: delena, llvm-commits
Differential Revision: https://reviews.llvm.org/D33311
llvm-svn: 304406
Summary:
This is a continuation of the work started in D29872 . Passing the carry down as a value rather than as a glue allows for further optimizations. Introducing setcccarry makes the use of addc/subc unecessary and we can start the removal process.
This patch only introduce the optimization strictly required to get the same level of optimization as was available before nothing more.
Reviewers: jyknight, nemanjai, mkuper, spatel, RKSimon, zvi, bkramer
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D33374
llvm-svn: 304404
Simon Pilgrim