This patch removes the old X86ISD::FSRL op - which allowed float vectors to use the byte right shift operations (causing a domain switch....).
Since the refactoring of the shuffle lowering code this no longer has any use.
Differential Revision: http://reviews.llvm.org/D10169
llvm-svn: 238906
in-register LUT technique.
Summary:
A description of this technique can be found here:
http://wm.ite.pl/articles/sse-popcount.html
The core of the idea is to use an in-register lookup table and the
PSHUFB instruction to compute the population count for the low and high
nibbles of each byte, and then to use horizontal sums to aggregate these
into vector population counts with wider element types.
On x86 there is an instruction that will directly compute the horizontal
sum for the low 8 and high 8 bytes, giving vNi64 popcount very easily.
Various tricks are used to get vNi32 and vNi16 from the vNi8 that the
LUT computes.
The base implemantion of this, and most of the work, was done by Bruno
in a follow up to D6531. See Bruno's detailed post there for lots of
timing information about these changes.
I have extended Bruno's patch in the following ways:
0) I committed the new tests with baseline sequences so this shows
a diff, and regenerated the tests using the update scripts.
1) Bruno had noticed and mentioned in IRC a redundant mask that
I removed.
2) I introduced a particular optimization for the i32 vector cases where
we use PSHL + PSADBW to compute the the low i32 popcounts, and PSHUFD
+ PSADBW to compute doubled high i32 popcounts. This takes advantage
of the fact that to line up the high i32 popcounts we have to shift
them anyways, and we can shift them by one fewer bit to effectively
divide the count by two. While the PSHUFD based horizontal add is no
faster, it doesn't require registers or load traffic the way a mask
would, and provides more ILP as it happens on different ports with
high throughput.
3) I did some code cleanups throughout to simplify the implementation
logic.
4) I refactored it to continue to use the parallel bitmath lowering when
SSSE3 is not available to preserve the performance of that version on
SSE2 targets where it is still much better than scalarizing as we'll
still do a bitmath implementation of popcount even in scalar code
there.
With #1 and #2 above, I analyzed the result in IACA for sandybridge,
ivybridge, and haswell. In every case I measured, the throughput is the
same or better using the LUT lowering, even v2i64 and v4i64, and even
compared with using the native popcnt instruction! The latency of the
LUT lowering is often higher than the latency of the scalarized popcnt
instruction sequence, but I think those latency measurements are deeply
misleading. Keeping the operation fully in the vector unit and having
many chances for increased throughput seems much more likely to win.
With this, we can lower every integer vector popcount implementation
using the LUT strategy if we have SSSE3 or better (and thus have
PSHUFB). I've updated the operation lowering to reflect this. This also
fixes an issue where we were scalarizing horribly some AVX lowerings.
Finally, there are some remaining cleanups. There is duplication between
the two techniques in how they perform the horizontal sum once the byte
population count is computed. I'm going to factor and merge those two in
a separate follow-up commit.
Differential Revision: http://reviews.llvm.org/D10084
llvm-svn: 238636
Predicate UseAVX depricates pattern selection on AVX-512.
This predicate is necessary for DAG selection to select EVEX form.
But mapping SSE intrinsics to AVX-512 instructions is not ready yet.
So I replaced UseAVX with HasAVX for intrinsics patterns.
llvm-svn: 237903
This is a follow-on to r236740 where I took Andrea's advice
in D9504 to remove a redundant pattern...except that I removed
the wrong pattern!
AFAICT, there is no change in the final code produced because
subsequent passes would clean up the extra instructions created
by the more complicated pattern.
llvm-svn: 236743
Finish the job that was abandoned in D6958 following the refactoring in
http://reviews.llvm.org/rL230221:
1. Uncomment the intrinsic def for the AVX r_Int instruction.
2. Add missing r_Int entries to the load folding tables; there are already
tests that check these in "test/Codegen/X86/fold-load-unops.ll", so I
haven't added any more in this patch.
3. Add patterns to solve PR21507 ( https://llvm.org/bugs/show_bug.cgi?id=21507 ).
So instead of this:
movaps %xmm0, %xmm1
rcpss %xmm1, %xmm1
movss %xmm1, %xmm0
We should now get:
rcpss %xmm0, %xmm0
And instead of this:
vsqrtss %xmm0, %xmm0, %xmm1
vblendps $1, %xmm1, %xmm0, %xmm0 ## xmm0 = xmm1[0],xmm0[1,2,3]
We should now get:
vsqrtss %xmm0, %xmm0, %xmm0
Differential Revision: http://reviews.llvm.org/D9504
llvm-svn: 236740
We don't need codegen-only intrinsic instructions for the vector forms of these instructions.
This makes the reciprocal estimate instruction lowering identical to how we handle normal
square roots: (V)SQRTPS / (V)SQRTPD.
No existing regression tests fail with this patch.
Differential Revision: http://reviews.llvm.org/D9301
llvm-svn: 236013
This fixes a regression introduced at revision 218263.
On AVX, if we optimize for size, a splat build_vector of a load
is lowered into a VBROADCAST node. This is done even if the value type of the
splat build_vector node is v2i64.
Since AVX doesn't support v2f64/v2i64 broadcasts, revision 218263 added two
extra tablegen patterns to allow selecting a VMOVDDUPrm from an X86VBroadcast
where the scalar element comes from a loadi64/loadf64.
However, revision 218263 forgot to add an extra fallback pattern for the case
where we have a X86VBroadcast of a loadi64 with multiple uses.
This patch adds the missing tablegen pattern in X86InstrSSE.td.
This patch also adds an extra test to 'splat-for-size.ll' to verify that ISel
doesn't crash with a 'fatal error in the backend' due to a missing AVX pattern
to select v2i64 X86ISD::BROADCAST nodes.
llvm-svn: 235509
This is an updated version of Chandler's patch D7402 that got accepted but never committed, and has bit-rotted a bit since.
I've updated the execution domain declarations to match the approach of the packed templates and also added some extra scalar unary tests.
Differential Revision: http://reviews.llvm.org/D9095
llvm-svn: 235372
For code like this:
define <8 x i32> @load_v8i32() {
ret <8 x i32> <i32 7, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
}
We produce this AVX code:
_load_v8i32: ## @load_v8i32
movl $7, %eax
vmovd %eax, %xmm0
vxorps %ymm1, %ymm1, %ymm1
vblendps $1, %ymm0, %ymm1, %ymm0 ## ymm0 = ymm0[0],ymm1[1,2,3,4,5,6,7]
retq
There are at least 2 bugs in play here:
We're generating a blend when a move scalar does the same job using 2 less instruction bytes (see FIXMEs).
We're not matching an existing pattern that would eliminate the xor and blend entirely. The zero bytes are free with vmovd.
The 2nd fix involves an adjustment of "AddedComplexity" [1] and mostly masks the 1st problem.
[1] AddedComplexity has close to no documentation in the source.
The best we have is this comment: "roughly corresponds to the number of nodes that are covered".
It appears that x86 has bastardized this definition by inflating its values for some other
undocumented reason. For example, we have a pattern with "AddedComplexity = 400" (!).
I searched my way to this page:
https://groups.google.com/forum/#!topic/llvm-dev/5UX-Og9M0xQ
Differential Revision: http://reviews.llvm.org/D8794
llvm-svn: 233931
We used to miss non-Q YMM integer vectors, and, non-Q/D XMM integer
vectors.
While there, change the v4i32 patterns to prefer MOVNTDQ.
llvm-svn: 233668
This should complete the job started in r231794 and continued in r232045:
We want to replace as much custom x86 shuffling via intrinsics
as possible because pushing the code down the generic shuffle
optimization path allows for better codegen and less complexity
in LLVM.
AVX2 introduced proper integer variants of the hacked integer insert/extract
C intrinsics that were created for this same functionality with AVX1.
This should complete the removal of insert/extract128 intrinsics.
The Clang precursor patch for this change was checked in at r232109.
llvm-svn: 232120
This is a follow-up to r231182. This adds the "vbroadcasti128" instruction
back, but without the intrinsic mapping. Also add a test to check the
instriction encoding.
This is related to rdar://problem/18742778.
llvm-svn: 231945
The intrinsic is no longer generated by the front-end. Remove the intrinsic and
auto-upgrade it to a vector shuffle.
Reviewed by Nadav
This is related to rdar://problem/18742778.
llvm-svn: 231182
I made the templates general, no need to define pattern separately for each instruction/intrinsic.
Now only need to add r_Int pattern for AVX.
llvm-svn: 230221
This canonicalization step saves us 3 pattern matching possibilities * 4 math ops
for scalar FP math that uses xmm regs. The backend can re-commute the operands
post-instruction-selection if that makes register allocation better.
The tests in llvm/test/CodeGen/X86/sse-scalar-fp-arith.ll cover this scenario already,
so there are no new tests with this patch.
Differential Revision: http://reviews.llvm.org/D7777
llvm-svn: 230024
Change the memory operands in sse12_fp_packed_scalar_logical_alias from scalars to vectors.
That's what the hardware packed logical FP instructions define: 128-bit memory operands.
There are no scalar versions of these instructions...because this is x86.
Generating the wrong code (folding a scalar load into a 128-bit load) is still possible
using the peephole optimization pass and the load folding tables. We won't completely
solve this bug until we either fix the lowering in fabs/fneg/fcopysign and any other
places where scalar FP logic is created or fix the load folding in foldMemoryOperandImpl()
to make sure it isn't changing the size of the load.
Differential Revision: http://reviews.llvm.org/D7474
llvm-svn: 229531
Patch to explicitly add the SSE MOVQ (rr,mr,rm) instructions to SSEPackedInt domain - prevents a number of costly domain switches.
Differential Revision: http://reviews.llvm.org/D7600
llvm-svn: 229439
This patch refactors the existing lowerVectorShuffleAsByteShift function to add support for 256-bit vectors on AVX2 targets.
It also fixes a tablegen issue that prevented the lowering of vpslldq/vpsrldq vec256 instructions.
Differential Revision: http://reviews.llvm.org/D7596
llvm-svn: 229311
This takes the preposterous number of patterns in this section
that were last added to in r219033 down to just plain obnoxious.
With a little more work, we might get this down to just comical.
I've added more test cases to the existing file that checks these
patterns, but it seems that some of these patterns simply don't
exist with today's shuffle lowering.
llvm-svn: 229158
This associates movss and movsd with the packed single and packed double
execution domains (resp.). While this is largely cosmetic, as we now
don't have weird ping-pong-ing between single and double precision, it
is also useful because it avoids the domain fixing algorithm from seeing
domain breaks that don't actually exist. It will also be much more
important if we have an execution domain default other than packed
single, as that would cause us to mix movss and movsd with integer
vector code on a regular basis, a very bad mixture.
llvm-svn: 228135
Specifically, the existing patterns were scalar-only. These cover the
packed vector bitwise operations when specifically requested with pseudo
instructions. This is particularly important in SSE1 where we can't
actually emit a logical operation on a v2i64 as that isn't a legal type.
This will be tested in subsequent patches which form the floating point
and patterns in more places.
llvm-svn: 228123
This patch detects consecutive vector loads using the existing
EltsFromConsecutiveLoads() logic. This fixes:
http://llvm.org/bugs/show_bug.cgi?id=22329
This patch effectively reverts the tablegen additions of D6492 /
http://reviews.llvm.org/rL224344 ...which in hindsight were a horrible hack.
The test cases that were added with that patch are simply modified to load
from varying offsets of a base pointer. These loads did not match the existing
tablegen patterns.
A happy side effect of doing this optimization earlier is that we can now fold
the load into a math op where possible; this is shown in some of the updated
checks in the test file.
Differential Revision: http://reviews.llvm.org/D7303
llvm-svn: 228006
For ordered, unordered, equal and not-equal tests, packed float and double comparison instructions can be safely commuted without affecting the results. This patch checks the comparison mode of the (v)cmpps + (v)cmppd instructions and commutes the result if it can.
Differential Revision: http://reviews.llvm.org/D7178
llvm-svn: 227145
Patch to allow (v)pclmulqdq to be commuted - swaps the src registers and inverts the immediate (low/high) src mask.
Differential Revision: http://reviews.llvm.org/D7180
llvm-svn: 227141
This patch fixes the following miscompile:
define void @sqrtsd(<2 x double> %a) nounwind uwtable ssp {
%0 = tail call <2 x double> @llvm.x86.sse2.sqrt.sd(<2 x double> %a) nounwind
%a0 = extractelement <2 x double> %0, i32 0
%conv = fptrunc double %a0 to float
%a1 = extractelement <2 x double> %0, i32 1
%conv3 = fptrunc double %a1 to float
tail call void @callee2(float %conv, float %conv3) nounwind
ret void
}
Current codegen:
sqrtsd %xmm0, %xmm1 ## high element of %xmm1 is undef here
xorps %xmm0, %xmm0
cvtsd2ss %xmm1, %xmm0
shufpd $1, %xmm1, %xmm1
cvtsd2ss %xmm1, %xmm1 ## operating on undef value
jmp _callee
This is a continuation of http://llvm.org/viewvc/llvm-project?view=revision&revision=224624 ( http://reviews.llvm.org/D6330 )
which was itself a continuation of r167064 ( http://llvm.org/viewvc/llvm-project?view=revision&revision=167064 ).
All of these patches are partial fixes for PR14221 ( http://llvm.org/bugs/show_bug.cgi?id=14221 );
this should be the final patch needed to resolve that bug.
Differential Revision: http://reviews.llvm.org/D6885
llvm-svn: 227111
This patch adds shuffle matching for the SSE3 MOVDDUP, MOVSLDUP and MOVSHDUP instructions. The big use of these being that they avoid many single source shuffles from needing to use (pre-AVX) dual source instructions such as SHUFPD/SHUFPS: causing extra moves and preventing load folds.
Adding these instructions uncovered an issue in XFormVExtractWithShuffleIntoLoad which crashed on single operand shuffle instructions (now fixed). It also involved fixing getTargetShuffleMask to correctly identify theses instructions as unary shuffles.
Also adds a missing tablegen pattern for MOVDDUP.
Differential Revision: http://reviews.llvm.org/D7042
llvm-svn: 226716
Now that we can fully specify extload legality, we can declare them
legal for the PMOVSX/PMOVZX instructions. This for instance enables
a DAGCombine to fire on code such as
(and (<zextload-equivalent> ...), <redundant mask>)
to turn it into:
(zextload ...)
as seen in the testcase changes.
There is one regression, in widen_load-2.ll: we're no longer able
to do store-to-load forwarding with illegal extload memory types.
This will be addressed separately.
Differential Revision: http://reviews.llvm.org/D6533
llvm-svn: 226676
