In the large code model, we now put __chkstk in %r11 before calling it.
Refactor the code so that we only do this once. Simplify things by using
__chkstk_ms instead of __chkstk on cygming. We already use that symbol
in the prolog emission, and it simplifies our logic.
Second half of PR18582.
llvm-svn: 227519
Reduce integer multiplication by a constant of the form k*2^c, where k is in {3,5,9} into a lea + shl. Previously it was only done for imulq on 64-bit platforms, but it makes sense for imull and 32-bit as well.
Differential Revision: http://reviews.llvm.org/D7196
llvm-svn: 227308
This includes two things:
1) Fix TCRETURNdi and TCRETURN64di patterns to check the right thing (LP64 as opposed to target bitness).
2) Allow LEA64_32 in MatchingStackOffset.
llvm-svn: 227307
By Asaf Badouh and Elena Demikhovsky
Added special nodes for rounding: FMADD_RND, FMSUB_RND..
It will prevent merge between nodes with rounding and other standard nodes.
llvm-svn: 227303
- Added KSHIFTB/D/Q for skx
- Added KORTESTB/D/Q for skx
- Fixed store operation for v8i1 type for KNL
- Store size of v8i1, v4i1 and v2i1 are changed to 8 bits
llvm-svn: 227043
Handle the poor codegen for i64/x86xmm->v2i64 (%mm -> %xmm) moves. Instead of
using stack store/load pair to do the job, use scalar_to_vector directly, which
in the MMX case can use movq2dq. This was the current behavior prior to
improvements for vector legalization of extloads in r213897.
This commit fixes the regression and as a side-effect also remove some
unnecessary shuffles.
In the new attached testcase, we go from:
pshufw $-18, (%rdi), %mm0
movq %mm0, -8(%rsp)
movq -8(%rsp), %xmm0
pshufd $-44, %xmm0, %xmm0
movd %xmm0, %eax
...
To:
pshufw $-18, (%rdi), %mm0
movq2dq %mm0, %xmm0
movd %xmm0, %eax
...
Differential Revision: http://reviews.llvm.org/D7126
rdar://problem/19413324
llvm-svn: 226953
The problem occurs when after vectorization we have type
<2 x i32>. This type is promoted to <2 x i64> and then requires
additional efforts for expanding loads and truncating stores.
I added EXPAND / TRUNCATE attributes to the masked load/store
SDNodes. The code now contains additional shuffles.
I've prepared changes in the cost estimation for masked memory
operations, it will be submitted separately.
llvm-svn: 226808
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
This patch disables target specific combine on X86ISD::INSERTPS dag nodes
if optlevel is CodeGenOpt::None.
The backend currently implements a target specific combine rule that converts
a vector load used by an INSERTPS dag node into a scalar load plus a
scalar_to_vector. This allows ISel to select a single INSERTPSrm instead of
two instructions (i.e. a vector load plus INSERTPSrr).
However, the existing target combine rule on INSERTPS nodes only works under
the assumption that ISel will always be able to match an INSERTPSrm. This is
not true in general at -O0, since the backend only allows folding a load into
the memory operand of an instruction if the optimization level is not
CodeGenOpt::None.
In the example below:
//
__m128 test(__m128 a, __m128 *b) {
__m128 c = _mm_insert_ps(a, *b, 1 << 6);
return c;
}
//
Before this patch, at -O0, the backend would have canonicalized the load to 'b'
into a scalar load plus scalar_to_vector. Later on, ISel would have selected an
INSERTPSrr leaving the insertps mask in an inconsistent state:
movss 4(%rdi), %xmm1
insertps $64, %xmm1, %xmm0 # xmm0 = xmm1[1],xmm0[1,2,3].
With this patch, the backend avoids folding the vector load into the operand of
the INSERTPS. The new codegen at -O0 is:
movaps (%rdi), %xmm1
insertps $64, %xmm1, %xmm0 # %xmm1[1],xmm0[1,2,3].
llvm-svn: 226277
This now handles both 32 and 64-bit element sizes.
In this version, the test are in vector-shuffle-512-v8.ll, canonicalized by
Chandler's update_llc_test_checks.py.
Part of <rdar://problem/17688758>
llvm-svn: 225838
r225551 vector byte shuffle optimization caused an assertion as fully zeroable vectors can be produced under certain circumstances. This fix drops the assert and returns a zero vector where the assert would have failed.
llvm-svn: 225718
This happens in the HINT benchmark, where the SLP-vectorizer created
v2f32 fcmp/select code. The "correct" solution would have been to
teach the vectorizer cost model that v2f32 isn't legal (because really,
it isn't), but if we can vectorize we might as well do so.
We legalize these v2f32 FMIN/FMAX nodes by widening to v4f32 later on.
v3f32 were already widened to v4f32 by the generic unroll-and-build-vector
legalization.
rdar://15763436
Differential Revision: http://reviews.llvm.org/D6557
llvm-svn: 225691
It's possible for the constant pool entry for the shuffle mask to come
from a completely different operation. This occurs when Constants have
the same bit pattern but have different types.
Make DecodePSHUFBMask tolerant of types which, after a bitcast, are
appropriately sized vector types.
This fixes PR22188.
llvm-svn: 225597
Teach the ISelLowering for X86 about the L,M,O target specific constraints.
Although, for the moment, clang performs constraint validation and prevents
passing along inline asm which may have immediate constant constraints violated,
the backend should be able to cope with the invalid inline asm a bit better.
llvm-svn: 225596
In the current code we only attempt to match against insertps if we have exactly one element from the second input vector, irrespective of how much of the shuffle result is zeroable.
This patch checks to see if there is a single non-zeroable element from either input that requires insertion. It also supports matching of cases where only one of the inputs need to be referenced.
We also split insertps shuffle matching off into a new lowerVectorShuffleAsInsertPS function.
Differential Revision: http://reviews.llvm.org/D6879
llvm-svn: 225589
pshufb can shuffle in zero bytes as well as bytes from a source vector - we can use this to avoid having to shuffle 2 vectors and ORing the result when the used inputs from a vector are all zeroable.
Differential Revision: http://reviews.llvm.org/D6878
llvm-svn: 225551
complements the new vector shuffle lowering code path. This flag,
naturally, is *off* because we've not tested or evaluated the results of
this at all. However, the flag will make it much easier to evaluate
whether we can be this aggressive and whether there are missing vector
shuffle lowering optimizations.
llvm-svn: 225491
The call lowering assumes that if the callee is a global, we want to emit a direct call.
This is correct for regular globals, but not for TLS ones.
Differential Revision: http://reviews.llvm.org/D6862
llvm-svn: 225438
type (in addition to the memory type).
The *LoadExt* legalization handling used to only have one type, the
memory type. This forced users to assume that as long as the extload
for the memory type was declared legal, and the result type was legal,
the whole extload was legal.
However, this isn't always the case. For instance, on X86, with AVX,
this is legal:
v4i32 load, zext from v4i8
but this isn't:
v4i64 load, zext from v4i8
Whereas v4i64 is (arguably) legal, even without AVX2.
Note that the same thing was done a while ago for truncstores (r46140),
but I assume no one needed it yet for extloads, so here we go.
Calls to getLoadExtAction were changed to add the value type, found
manually in the surrounding code.
Calls to setLoadExtAction were mechanically changed, by wrapping the
call in a loop, to match previous behavior. The loop iterates over
the MVT subrange corresponding to the memory type (FP vectors, etc...).
I also pulled neighboring setTruncStoreActions into some of the loops;
those shouldn't make a difference, as the additional types are illegal.
(e.g., i128->i1 truncstores on PPC.)
No functional change intended.
Differential Revision: http://reviews.llvm.org/D6532
llvm-svn: 225421
"ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
relocation target a movq or addq instruction.
Prohibit the truncation of such loads to movl or addl.
This fixes PR22083.
Differential Revision: http://reviews.llvm.org/D6839
llvm-svn: 225250
The assembler backend will relax to the long form if necessary. This removes a swap from long form to short form in the MCInstLowering code. Selecting the long form used to be required by the old JIT.
llvm-svn: 225242
If the control flow is modelling an if-statement where the only instruction in
the 'then' basic block (excluding the terminator) is a call to cttz/ctlz,
CodeGenPrepare can try to speculate the cttz/ctlz call and simplify the control
flow graph.
Example:
\code
entry:
%cmp = icmp eq i64 %val, 0
br i1 %cmp, label %end.bb, label %then.bb
then.bb:
%c = tail call i64 @llvm.cttz.i64(i64 %val, i1 true)
br label %end.bb
end.bb:
%cond = phi i64 [ %c, %then.bb ], [ 64, %entry]
\code
In this example, basic block %then.bb is taken if value %val is not zero.
Also, the phi node in %end.bb would propagate the size-of in bits of %val
only if %val is equal to zero.
With this patch, CodeGenPrepare will try to hoist the call to cttz from %then.bb
into basic block %entry only if cttz is cheap to speculate for the target.
Added two new hooks in TargetLowering.h to let targets customize the behavior
(i.e. decide whether it is cheap or not to speculate calls to cttz/ctlz). The
two new methods are 'isCheapToSpeculateCtlz' and 'isCheapToSpeculateCttz'.
By default, both methods return 'false'.
On X86, method 'isCheapToSpeculateCtlz' returns true only if the target has
LZCNT. Method 'isCheapToSpeculateCttz' only returns true if the target has BMI.
Differential Revision: http://reviews.llvm.org/D6728
llvm-svn: 224899
When combining consecutive loads+inserts into a single vector load,
we should keep the alignment of the base load. Doing otherwise can, and does,
lead to using overly aligned instructions. In the included test case, for
example, using a 32-byte vmovaps on a 16-byte aligned value. Oops.
rdar://19190968
llvm-svn: 224746
Previously I tried to plug musttail into the existing vararg lowering
code. That turned out to be a mistake, because non-vararg calls use
significantly different register lowering, even on x86. For example, AVX
vectors are usually passed in registers to normal functions and memory
to vararg functions. Now musttail uses a completely separate lowering.
Hopefully this can be used as the basis for non-x86 perfect forwarding.
Reviewers: majnemer
Differential Revision: http://reviews.llvm.org/D6156
llvm-svn: 224745
Currently, when ctpop is supported for scalar types, the expansion of
@llvm.ctpop.vXiY uses vector element extractions, insertions and individual
calls to @llvm.ctpop.iY. When not, expansion with bit-math operations is used
for the scalar calls.
Local haswell measurements show that we can improve vector @llvm.ctpop.vXiY
expansion in some cases by using a using a vector parallel bit twiddling
approach, based on:
v = v - ((v >> 1) & 0x55555555);
v = (v & 0x33333333) + ((v >> 2) & 0x33333333);
v = ((v + (v >> 4) & 0xF0F0F0F)
v = v + (v >> 8)
v = v + (v >> 16)
v = v & 0x0000003F
(from http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel)
When scalar ctpop isn't supported, the approach above performs better for
v2i64, v4i32, v4i64 and v8i32 (see numbers below). And even when scalar ctpop
is supported, this approach performs ~2x better for v8i32.
Here, x86_64 implies -march=corei7-avx without ctpop and x86_64h includes ctpop
support with -march=core-avx2.
== [x86_64h - new]
v8i32: 0.661685
v4i32: 0.514678
v4i64: 0.652009
v2i64: 0.324289
== [x86_64h - old]
v8i32: 1.29578
v4i32: 0.528807
v4i64: 0.65981
v2i64: 0.330707
== [x86_64 - new]
v8i32: 1.003
v4i32: 0.656273
v4i64: 1.11711
v2i64: 0.754064
== [x86_64 - old]
v8i32: 2.34886
v4i32: 1.72053
v4i64: 1.41086
v2i64: 1.0244
More work for other vector types will come next.
llvm-svn: 224725
Added RegOp2MemOpTable4 to transform 4th operand from register to memory in merge-masked versions of instructions.
Added lowering tests.
llvm-svn: 224516
This handles the case of a BUILD_VECTOR being constructed out of elements extracted from a vector twice the size of the result vector. Previously this was always scalarized. Now, we try to construct a shuffle node that feeds on extract_subvectors.
This fixes PR15872 and provides a partial fix for PR21711.
Differential Revision: http://reviews.llvm.org/D6678
llvm-svn: 224429
The type promotion helper does not support vector type, so when make
such it does not kick in in such cases.
Original commit message:
[CodeGenPrepare] Move sign/zero extensions near loads using type promotion.
This patch extends the optimization in CodeGenPrepare that moves a sign/zero
extension near a load when the target can combine them. The optimization may
promote any operations between the extension and the load to make that possible.
Although this optimization may be beneficial for all targets, in particular
AArch64, this is enabled for X86 only as I have not benchmarked it for other
targets yet.
** Context **
Most targets feature extended loads, i.e., loads that perform a zero or sign
extension for free. In that context it is interesting to expose such pattern in
CodeGenPrepare so that the instruction selection pass can form such loads.
Sometimes, this pattern is blocked because of instructions between the load and
the extension. When those instructions are promotable to the extended type, we
can expose this pattern.
** Motivating Example **
Let us consider an example:
define void @foo(i8* %addr1, i32* %addr2, i8 %a, i32 %b) {
%ld = load i8* %addr1
%zextld = zext i8 %ld to i32
%ld2 = load i32* %addr2
%add = add nsw i32 %ld2, %zextld
%sextadd = sext i32 %add to i64
%zexta = zext i8 %a to i32
%addza = add nsw i32 %zexta, %zextld
%sextaddza = sext i32 %addza to i64
%addb = add nsw i32 %b, %zextld
%sextaddb = sext i32 %addb to i64
call void @dummy(i64 %sextadd, i64 %sextaddza, i64 %sextaddb)
ret void
}
As it is, this IR generates the following assembly on x86_64:
[...]
movzbl (%rdi), %eax # zero-extended load
movl (%rsi), %es # plain load
addl %eax, %esi # 32-bit add
movslq %esi, %rdi # sign extend the result of add
movzbl %dl, %edx # zero extend the first argument
addl %eax, %edx # 32-bit add
movslq %edx, %rsi # sign extend the result of add
addl %eax, %ecx # 32-bit add
movslq %ecx, %rdx # sign extend the result of add
[...]
The throughput of this sequence is 7.45 cycles on Ivy Bridge according to IACA.
Now, by promoting the additions to form more extended loads we would generate:
[...]
movzbl (%rdi), %eax # zero-extended load
movslq (%rsi), %rdi # sign-extended load
addq %rax, %rdi # 64-bit add
movzbl %dl, %esi # zero extend the first argument
addq %rax, %rsi # 64-bit add
movslq %ecx, %rdx # sign extend the second argument
addq %rax, %rdx # 64-bit add
[...]
The throughput of this sequence is 6.15 cycles on Ivy Bridge according to IACA.
This kind of sequences happen a lot on code using 32-bit indexes on 64-bit
architectures.
Note: The throughput numbers are similar on Sandy Bridge and Haswell.
** Proposed Solution **
To avoid the penalty of all these sign/zero extensions, we merge them in the
loads at the beginning of the chain of computation by promoting all the chain of
computation on the extended type. The promotion is done if and only if we do not
introduce new extensions, i.e., if we do not degrade the code quality.
To achieve this, we extend the existing “move ext to load” optimization with the
promotion mechanism introduced to match larger patterns for addressing mode
(r200947).
The idea of this extension is to perform the following transformation:
ext(promotableInst1(...(promotableInstN(load))))
=>
promotedInst1(...(promotedInstN(ext(load))))
The promotion mechanism in that optimization is enabled by a new TargetLowering
switch, which is off by default. In other words, by default, the optimization
performs the “move ext to load” optimization as it was before this patch.
** Performance **
Configuration: x86_64: Ivy Bridge fixed at 2900MHz running OS X 10.10.
Tested Optimization Levels: O3/Os
Tests: llvm-testsuite + externals.
Results:
- No regression beside noise.
- Improvements:
CINT2006/473.astar: ~2%
Benchmarks/PAQ8p: ~2%
Misc/perlin: ~3%
The results are consistent for both O3 and Os.
<rdar://problem/18310086>
llvm-svn: 224402
This patch extends the optimization in CodeGenPrepare that moves a sign/zero
extension near a load when the target can combine them. The optimization may
promote any operations between the extension and the load to make that possible.
Although this optimization may be beneficial for all targets, in particular
AArch64, this is enabled for X86 only as I have not benchmarked it for other
targets yet.
** Context **
Most targets feature extended loads, i.e., loads that perform a zero or sign
extension for free. In that context it is interesting to expose such pattern in
CodeGenPrepare so that the instruction selection pass can form such loads.
Sometimes, this pattern is blocked because of instructions between the load and
the extension. When those instructions are promotable to the extended type, we
can expose this pattern.
** Motivating Example **
Let us consider an example:
define void @foo(i8* %addr1, i32* %addr2, i8 %a, i32 %b) {
%ld = load i8* %addr1
%zextld = zext i8 %ld to i32
%ld2 = load i32* %addr2
%add = add nsw i32 %ld2, %zextld
%sextadd = sext i32 %add to i64
%zexta = zext i8 %a to i32
%addza = add nsw i32 %zexta, %zextld
%sextaddza = sext i32 %addza to i64
%addb = add nsw i32 %b, %zextld
%sextaddb = sext i32 %addb to i64
call void @dummy(i64 %sextadd, i64 %sextaddza, i64 %sextaddb)
ret void
}
As it is, this IR generates the following assembly on x86_64:
[...]
movzbl (%rdi), %eax # zero-extended load
movl (%rsi), %es # plain load
addl %eax, %esi # 32-bit add
movslq %esi, %rdi # sign extend the result of add
movzbl %dl, %edx # zero extend the first argument
addl %eax, %edx # 32-bit add
movslq %edx, %rsi # sign extend the result of add
addl %eax, %ecx # 32-bit add
movslq %ecx, %rdx # sign extend the result of add
[...]
The throughput of this sequence is 7.45 cycles on Ivy Bridge according to IACA.
Now, by promoting the additions to form more extended loads we would generate:
[...]
movzbl (%rdi), %eax # zero-extended load
movslq (%rsi), %rdi # sign-extended load
addq %rax, %rdi # 64-bit add
movzbl %dl, %esi # zero extend the first argument
addq %rax, %rsi # 64-bit add
movslq %ecx, %rdx # sign extend the second argument
addq %rax, %rdx # 64-bit add
[...]
The throughput of this sequence is 6.15 cycles on Ivy Bridge according to IACA.
This kind of sequences happen a lot on code using 32-bit indexes on 64-bit
architectures.
Note: The throughput numbers are similar on Sandy Bridge and Haswell.
** Proposed Solution **
To avoid the penalty of all these sign/zero extensions, we merge them in the
loads at the beginning of the chain of computation by promoting all the chain of
computation on the extended type. The promotion is done if and only if we do not
introduce new extensions, i.e., if we do not degrade the code quality.
To achieve this, we extend the existing “move ext to load” optimization with the
promotion mechanism introduced to match larger patterns for addressing mode
(r200947).
The idea of this extension is to perform the following transformation:
ext(promotableInst1(...(promotableInstN(load))))
=>
promotedInst1(...(promotedInstN(ext(load))))
The promotion mechanism in that optimization is enabled by a new TargetLowering
switch, which is off by default. In other words, by default, the optimization
performs the “move ext to load” optimization as it was before this patch.
** Performance **
Configuration: x86_64: Ivy Bridge fixed at 2900MHz running OS X 10.10.
Tested Optimization Levels: O3/Os
Tests: llvm-testsuite + externals.
Results:
- No regression beside noise.
- Improvements:
CINT2006/473.astar: ~2%
Benchmarks/PAQ8p: ~2%
Misc/perlin: ~3%
The results are consistent for both O3 and Os.
<rdar://problem/18310086>
llvm-svn: 224351
EltsFromConsecutiveLoads was apparently only ever called for 128-bit vectors, and assumed this implicitly. r223518 started calling it for AVX-sized vectors, causing the code path that had this assumption to crash.
This adds a check to make this path fire only for 128-bit vectors.
Differential Revision: http://reviews.llvm.org/D6579
llvm-svn: 223922
Before this patch, the backend sub-optimally expanded the non-constant shift
count of a v8i16 shift into a sequence of two 'movd' plus 'movzwl'.
With this patch the backend checks if the target features sse4.1. If so, then
it lets the shuffle legalizer deal with the expansion of the shift amount.
Example:
;;
define <8 x i16> @test(<8 x i16> %A, <8 x i16> %B) {
%shamt = shufflevector <8 x i16> %B, <8 x i16> undef, <8 x i32> zeroinitializer
%shl = shl <8 x i16> %A, %shamt
ret <8 x i16> %shl
}
;;
Before (with -mattr=+avx):
vmovd %xmm1, %eax
movzwl %ax, %eax
vmovd %eax, %xmm1
vpsllw %xmm1, %xmm0, %xmm0
retq
Now:
vpxor %xmm2, %xmm2, %xmm2
vpblendw $1, %xmm1, %xmm2, %xmm1
vpsllw %xmm1, %xmm0, %xmm0
retq
llvm-svn: 223660
X86ISelLowering.cpp has a long switch for intrinsics. I moved a part of
this long switch to the new intrinsics table in X86IntrinsicsInfo.h.
No functional changes, just code and compile time optimization.
llvm-svn: 223641
Fix the poor codegen seen in PR21710 ( http://llvm.org/bugs/show_bug.cgi?id=21710 ).
Before we crack 32-byte build vectors into smaller chunks (and then subsequently
glue them back together), we should look for the easy case where we can just load
all elements in a single op.
An example of the codegen change is:
From:
vmovss 16(%rdi), %xmm1
vmovups (%rdi), %xmm0
vinsertps $16, 20(%rdi), %xmm1, %xmm1
vinsertps $32, 24(%rdi), %xmm1, %xmm1
vinsertps $48, 28(%rdi), %xmm1, %xmm1
vinsertf128 $1, %xmm1, %ymm0, %ymm0
retq
To:
vmovups (%rdi), %ymm0
retq
Differential Revision: http://reviews.llvm.org/D6536
llvm-svn: 223518
Summary:
Follow up to [x32] "Use ebp/esp as frame and stack pointer":
http://reviews.llvm.org/D4617
In that earlier patch, NaCl64 was made to always use rbp.
That's needed for most cases because rbp should hold a full
64-bit address within the NaCl sandbox so that load/stores
off of rbp don't require sandbox adjustment (zeroing the top
32-bits, then filling those by adding r15).
However, llvm.frameaddress returns a pointer and pointers
are 32-bit for NaCl64. In this case, use ebp instead, which
will make the register copy type check. A similar mechanism
may be needed for llvm.eh.return, but is not added in this change.
Test Plan: test/CodeGen/X86/frameaddr.ll
Reviewers: dschuff, nadav
Subscribers: jfb, llvm-commits
Differential Revision: http://reviews.llvm.org/D6514
llvm-svn: 223510
SSE2/AVX non-constant packed shift instructions only use the lower 64-bit of
the shift count.
This patch teaches function 'getTargetVShiftNode' how to deal with shifts
where the shift count node is of type MVT::i64.
Before this patch, function 'getTargetVShiftNode' only knew how to deal with
shift count nodes of type MVT::i32. This forced the backend to wrongly
truncate the shift count to MVT::i32, and then zero-extend it back to MVT::i64.
llvm-svn: 223505
When lowering a vector shift node, the backend checks if the shift count is a
shuffle with a splat mask. If so, then it introduces an extra dag node to
extract the splat value from the shuffle. The splat value is then used
to generate a shift count of a target specific shift.
However, if we know that the shift count is a splat shuffle, we can use the
splat index 'I' to extract the I-th element from the first shuffle operand.
The advantage is that the splat shuffle may become dead since we no longer
use it.
Example:
;;
define <4 x i32> @example(<4 x i32> %a, <4 x i32> %b) {
%c = shufflevector <4 x i32> %b, <4 x i32> undef, <4 x i32> zeroinitializer
%shl = shl <4 x i32> %a, %c
ret <4 x i32> %shl
}
;;
Before this patch, llc generated the following code (-mattr=+avx):
vpshufd $0, %xmm1, %xmm1 # xmm1 = xmm1[0,0,0,0]
vpxor %xmm2, %xmm2
vpblendw $3, %xmm1, %xmm2, %xmm1 # xmm1 = xmm1[0,1],xmm2[2,3,4,5,6,7]
vpslld %xmm1, %xmm0, %xmm0
retq
With this patch, the redundant splat operation is removed from the code.
vpxor %xmm2, %xmm2
vpblendw $3, %xmm1, %xmm2, %xmm1 # xmm1 = xmm1[0,1],xmm2[2,3,4,5,6,7]
vpslld %xmm1, %xmm0, %xmm0
retq
llvm-svn: 223461
r32900 introduced custom lowering for fcopysign, with two checks to
change the magnitude value's type if it's larger/smaller than the sign
value's type. r32932 replaced that code for the smaller case.
r43205 did the same for the larger case, but left the old code, now dead.
llvm-svn: 223415
The current DAG combine turns a sequence of extracts from <4 x i32> followed by zexts into a store followed by scalar loads.
According to measurements by Martin Krastev (see PR 21269) for x86-64, a sequence of an extract, movs and shifts gives better performance. However, for 32-bit x86, the previous sequence still seems better.
Differential Revision: http://reviews.llvm.org/D6501
llvm-svn: 223360
Replaced some logic that checked if a build_vector node is doing a splat of a
non-undef value with a call to method BuildVectorSDNode::getSplatValue().
No functional change intended.
llvm-svn: 223354
I'm recommiting the codegen part of the patch.
The vectorizer part will be send to review again.
Masked Vector Load and Store Intrinsics.
Introduced new target-independent intrinsics in order to support masked vector loads and stores. The loop vectorizer optimizes loops containing conditional memory accesses by generating these intrinsics for existing targets AVX2 and AVX-512. The vectorizer asks the target about availability of masked vector loads and stores.
Added SDNodes for masked operations and lowering patterns for X86 code generator.
Examples:
<16 x i32> @llvm.masked.load.v16i32(i8* %addr, <16 x i32> %passthru, i32 4 /* align */, <16 x i1> %mask)
declare void @llvm.masked.store.v8f64(i8* %addr, <8 x double> %value, i32 4, <8 x i1> %mask)
Scalarizer for other targets (not AVX2/AVX-512) will be done in a separate patch.
http://reviews.llvm.org/D6191
llvm-svn: 223348
Commit on
- This patch fixes the bug described in
http://lists.cs.uiuc.edu/pipermail/llvmdev/2013-May/062343.html
The fix allocates an extra slot just below the GPRs and stores the base pointer
there. This is done only for functions containing llvm.eh.sjlj.setjmp that also
need a base pointer. Because code containing llvm.eh.sjlj.setjmp saves all of
the callee-save GPRs in the prologue, the offset to the extra slot can be
computed before prologue generation runs.
Impact at run-time on affected functions is::
- One extra store in the prologue, The store saves the base pointer.
- One extra load after a llvm.eh.sjlj.setjmp. The load restores the base pointer.
Because the extra slot is just above a gap between frame-pointer-relative and
base-pointer-relative chunks of memory, there is no impact on other offset
calculations other than ensuring there is room for the extra slot.
http://reviews.llvm.org/D6388
Patch by Arch Robison <arch.robison@intel.com>
llvm-svn: 223329
This is the second patch in a small series. This patch contains the MachineInstruction and x86-64 backend pieces required to lower Statepoints. It does not include the code to actually generate the STATEPOINT machine instruction and as a result, the entire patch is currently dead code. I will be submitting the SelectionDAG parts within the next 24-48 hours. Since those pieces are by far the most complicated, I wanted to minimize the size of that patch. That patch will include the tests which exercise the functionality in this patch. The entire series can be seen as one combined whole in http://reviews.llvm.org/D5683.
The STATEPOINT psuedo node is generated after all gc values are explicitly spilled to stack slots. The purpose of this node is to wrap an actual call instruction while recording the spill locations of the meta arguments used for garbage collection and other purposes. The STATEPOINT is modeled as modifing all of those locations to prevent backend optimizations from forwarding the value from before the STATEPOINT to after the STATEPOINT. (Doing so would break relocation semantics for collectors which wish to relocate roots.)
The implementation of STATEPOINT is closely modeled on PATCHPOINT. Eventually, much of the code in this patch will be removed. The long term plan is to merge the functionality provided by statepoints and patchpoints. Merging their implementations in the backend is likely to be a good starting point.
Reviewed by: atrick, ributzka
llvm-svn: 223085
This reverts commit r222632 (and follow-up r222636), which caused a host
of LNT failures on an internal bot. I'll respond to the commit on the
list with a reproduction of one of the failures.
Conflicts:
lib/Target/X86/X86TargetTransformInfo.cpp
llvm-svn: 222936
including SAE mode and memory operand.
Added AVX512_maskable_scalar template, that should cover all scalar instructions in the future.
The main difference between AVX512_maskable_scalar<> and AVX512_maskable<> is using X86select instead of vselect.
I need it, because I can't create vselect node for MVT::i1 mask for scalar instruction.
http://reviews.llvm.org/D6378
llvm-svn: 222820
Since (v)pslldq / (v)psrldq instructions resolve to a single input argument it is useful to match it much earlier than we currently do - this prevents more complicated shuffles (notably insertion into a zero vector) matching before it.
Differential Revision: http://reviews.llvm.org/D6409
llvm-svn: 222796
This patch teaches function 'transformVSELECTtoBlendVECTOR_SHUFFLE' how to
convert VSELECT dag nodes to shuffles on targets that do not have SSE4.1.
On pre-SSE4.1 targets, we can still perform blend operations using movss/movsd.
Also, removed a target specific combine that performed a premature lowering of
VSELECT nodes to target specific MOVSS/MOVSD nodes.
llvm-svn: 222647
r222375 made some improvements to build_vector lowering of v4x32 and v4xf32 into an insertps, but it missed a case where:
1. A single extracted element is used twice.
2. The lower of the two non-zero indexes should be preserved, and the higher should be used for the dest mask.
This caused a crash, since the source value for the insertps ends-up uninitialized.
Differential Revision: http://reviews.llvm.org/D6377
llvm-svn: 222635
Introduced new target-independent intrinsics in order to support masked vector loads and stores. The loop vectorizer optimizes loops containing conditional memory accesses by generating these intrinsics for existing targets AVX2 and AVX-512. The vectorizer asks the target about availability of masked vector loads and stores.
Added SDNodes for masked operations and lowering patterns for X86 code generator.
Examples:
<16 x i32> @llvm.masked.load.v16i32(i8* %addr, <16 x i32> %passthru, i32 4 /* align */, <16 x i1> %mask)
declare void @llvm.masked.store.v8f64(i8* %addr, <8 x double> %value, i32 4, <8 x i1> %mask)
Scalarizer for other targets (not AVX2/AVX-512) will be done in a separate patch.
http://reviews.llvm.org/D6191
llvm-svn: 222632
No functionality changed yet, but this will prevent subsequent patches
from having to handle permutations of various interleaved shuffle
patterns.
llvm-svn: 222614
This patch adds a feature flag to avoid unaligned 32-byte load/store AVX codegen
for Sandy Bridge and Ivy Bridge. There is no functionality change intended for
those chips. Previously, the absence of AVX2 was being used as a proxy to detect
this feature. But that hindered codegen for AVX-enabled AMD chips such as btver2
that do not have the 32-byte unaligned access slowdown.
Performance measurements are included in PR21541 ( http://llvm.org/bugs/show_bug.cgi?id=21541 ).
Differential Revision: http://reviews.llvm.org/D6355
llvm-svn: 222544
shuffle lowering to allow much better blend matching.
Specifically, with the new structure the code seems clearer to me and we
correctly can hit the cases where merging two 128-bit lanes is a clear
win and can be shuffled cheaply afterward.
llvm-svn: 222539
a bunch more improvements.
Non-lane-crossing is fine, the key is that lane merging only makes sense
for single-input shuffles. Not sure why I got so turned around here. The
code all works, I was just using the wrong model for it.
This only updates v4 and v8 lowering. The v16 and v32 lowering requires
restructuring the entire check sequence.
llvm-svn: 222537
lanes.
By special casing these we can often either reduce the total number of
shuffles significantly or reduce the number of (high latency on Haswell)
AVX2 shuffles that potentially cross 128-bit lanes. Even when these
don't actually cross lanes, they have much higher latency to support
that. Doing two of them and a blend is worse than doing a single insert
across the 128-bit lanes to blend and then doing a single interleaved
shuffle.
While this seems like a narrow case, it kept cropping up on me and the
difference is *huge* as you can see in many of the test cases. I first
hit this trying to perfectly fix the interleaving shuffle patterns used
by Halide for AVX2.
llvm-svn: 222533
Windows itanium targets the MSVCRT, and the stack probe symbol is provided by
MSVCRT. This corrects the emission of stack probes on i686-windows-itanium.
llvm-svn: 222439
This patch improves the lowering of v4f32 and v4i32 build_vector dag nodes
that are known to have at least two non-zero elements.
With this patch, a build_vector that performs a blend with zero is
converted into a shuffle. This is done to let the shuffle legalizer expand
the dag node in a optimal way. For example, if we know that a build_vector
performs a blend with zero, we can try to lower it as a movq/blend instead of
always selecting an insertps.
This patch also improves the logic that lowers a build_vector into a insertps
with zero masking. See for example the extra test cases added to test sse41.ll.
Differential Revision: http://reviews.llvm.org/D6311
llvm-svn: 222375
This patch builds on http://reviews.llvm.org/D5598 to perform byte rotation shuffles (lowerVectorShuffleAsByteRotate) on pre-SSSE3 (palignr) targets - pre-SSSE3 is only enabled on i8 and i16 vector targets where it is a more definite performance gain.
I've also added a separate byte shift shuffle (lowerVectorShuffleAsByteShift) that makes use of the ability of the SLLDQ/SRLDQ instructions to implicitly shift in zero bytes to avoid the need to create a zero register if we had used palignr.
Differential Revision: http://reviews.llvm.org/D5699
llvm-svn: 222340
Updated X86TargetLowering::isShuffleMaskLegal to match SHUFP masks with commuted inputs and PSHUFD masks that reference the second input.
As part of this I've refactored isPSHUFDMask to work in a more general manner and allow it to match against either the first or second input vector.
Differential Revision: http://reviews.llvm.org/D6287
llvm-svn: 222087
getTargetConstant should only be used when you can guarantee the instruction
selected will be able to cope with the raw value. BUILD_VECTOR is rather too
generic for this so we should use getConstant instead. In that case, an
instruction can still consume the constant, but if it doesn't it'll be
materialised through its own round of ISel.
Should fix PR21352.
llvm-svn: 221961
between splitting a vector into 128-bit lanes and recombining them vs.
decomposing things into single-input shuffles and a final blend.
This handles a large number of cases in AVX1 where the cross-lane
shuffles would be much more expensive to represent even though we end up
with a fast blend at the root. Instead, we can do a better job of
shuffling in a single lane and then inserting it into the other lanes.
This fixes the remaining bits of Halide's regression captured in PR21281
for AVX1. However, the bug persists in AVX2 because I've made this
change reasonably conservative. The cases where it makes sense in AVX2
to split into 128-bit lanes are much more rare because we can often do
full permutations across all elements of the 256-bit vector. However,
the particular test case in PR21281 is an example of one of the rare
cases where it is *always* better to work in a single 128-bit lane. I'm
going to try to teach the logic to detect and form the good code even in
AVX2 next, but it will need to use a separate heuristic.
Finally, there is one pesky regression here where we previously would
craftily use vpermilps in AVX1 to shuffle both high and low halves at
the same time. We no longer pull that off, and not for any really good
reason. Ultimately, I think this is just another missing nuance to the
selection heuristic that I'll try to add in afterward, but this change
already seems strictly worth doing considering the magnitude of the
improvements in common matrix math shuffle patterns.
As always, please let me know if this causes a surprising regression for
you.
llvm-svn: 221861
re-combining shuffles because nothing was available in the wider vector
type.
The key observation (which I've put in the comments for future
maintainers) is that at this point, no further combining is really
possible. And so even though these shuffles trivially could be combined,
we need to actually do that as we produce them when producing them this
late in the lowering.
This fixes another (huge) part of the Halide vector shuffle regressions.
As it happens, this was already well covered by the tests, but I hadn't
noticed how bad some of these got. The specific patterns that turn
directly into unpckl/h patterns were occurring *many* times in common
vector processing code.
There are still more problems here sadly, but trying to incrementally
tease them apart and it looks like this is the core of the problem in
the splitting logic.
There is some chance of regression here, you can see it in the test
changes. Specifically, where we stop forming pshufb in some cases, it is
possible that pshufb was in fact faster. Intel "says" that pshufb is
slower than the instruction sequences replacing it.
llvm-svn: 221852
This is a follow-on to r221706 and r221731 and discussed in more detail in PR21385.
This patch also loosens the testcase checking for btver2. We know that the "1.0" will be loaded, but
we can't tell exactly when, so replace the CHECK-NEXT specifiers with plain CHECKs. The CHECK-NEXT
sequence relied on a quirk of post-RA-scheduling that may change independently of anything in these tests.
llvm-svn: 221819
cases from Halide folks. This initial step was extracted from
a prototype change by Clay Wood to try and address regressions found
with Halide and the new vector shuffle lowering.
llvm-svn: 221779
This is a first step for generating SSE rcp instructions for reciprocal
calcs when fast-math allows it. This is very similar to the rsqrt optimization
enabled in D5658 ( http://reviews.llvm.org/rL220570 ).
For now, be conservative and only enable this for AMD btver2 where performance
improves significantly both in terms of latency and throughput.
We may never enable this codegen for Intel Core* chips because the divider circuits
are just too fast. On SandyBridge, divss can be as fast as 10 cycles versus the 21
cycle critical path for the rcp + mul + sub + mul + add estimate.
Follow-on patches may allow configuration of the number of Newton-Raphson refinement
steps, add AVX512 support, and enable the optimization for more chips.
More background here: http://llvm.org/bugs/show_bug.cgi?id=21385
Differential Revision: http://reviews.llvm.org/D6175
llvm-svn: 221706
The ISel lowering for global TLS access in PIC mode was creating a pseudo
instruction that is later expanded to a call, but the code was not
setting the hasCalls flag in the MachineFrameInfo alongside the adjustsStack
flag. This caused some functions to be mistakenly recognized as leaf functions,
and this in turn affected the decision to eliminate the frame pointer.
With the fix, hasCalls is properly set and the leaf frame pointer is correctly
preserved.
llvm-svn: 221695
AVX2 is available.
According to IACA, the new lowering has a throughput of 8 cycles instead of 13
with the previous one.
Althought this lowering kicks in some SPECs benchmarks, the performance
improvement was within the noise.
Correctness testing has been done for the whole range of uint32_t with the
following program:
uint4 v = (uint4) {0,1,2,3};
uint32_t i;
//Check correctness over entire range for uint4 -> float4 conversion
for( i = 0; i < 1U << (32-2); i++ )
{
float4 t = test(v);
float4 c = correct(v);
if( 0xf != _mm_movemask_ps( t == c ))
{
printf( "Error @ %vx: %vf vs. %vf\n", v, c, t);
return -1;
}
v += 4;
}
Where "correct" is the old lowering and "test" the new one.
The patch adds a test case for the two custom lowering instruction.
It also modifies the vector cost model, which is why cast.ll and uitofp.ll are
modified.
2009-02-26-MachineLICMBug.ll is also modified because we now hoist 7
instructions instead of 4 (3 more constant loads).
rdar://problem/18153096>
llvm-svn: 221657
condition to match a blend.
This prevents optimizations that work on VSELECT to perform invalid
transformations. Indeed, the optimized condition does not match the vector
boolean content that is expected and bad things may happen.
This patch yields the exact same code on the whole test-suite + specs (-O3 and
-O3 -march=core-avx2), it improves one test case (vector-blend.ll) and fixes a
bug reduced in vselect-avx.ll.
<rdar://problem/18819506>
llvm-svn: 221429
This patch improves the folding of vector AND nodes into blend operations for
targets that feature SSE4.1. A vector AND node where one of the operands is
a constant build_vector with elements that are either zero or all-ones can be
converted into a blend.
This allows for example to simplify the following code:
define <4 x i32> @test(<4 x i32> %A, <4 x i32> %B) {
%1 = and <4 x i32> %A, <i32 0, i32 0, i32 0, i32 -1>
%2 = and <4 x i32> %B, <i32 -1, i32 -1, i32 -1, i32 0>
%3 = or <4 x i32> %1, %2
ret <4 x i32> %3
}
Before this patch llc (-mcpu=corei7) generated:
andps LCPI1_0(%rip), %xmm0, %xmm0
andps LCPI1_1(%rip), %xmm1, %xmm1
orps %xmm1, %xmm0, %xmm0
retq
With this patch we generate a single 'vpblendw'.
llvm-svn: 221343
For 8-bit divrems where the remainder is used, we used to generate:
divb %sil
shrw $8, %ax
movzbl %al, %eax
That was to avoid an H-reg access, which is problematic mainly because
it isn't possible in REX-prefixed instructions.
This patch optimizes that to:
divb %sil
movzbl %ah, %eax
To do that, we explicitly extend AH, and extract the L-subreg in the
resulting register. The extension is done using the NOREX variants of
MOVZX. To support signed operations, MOVSX_NOREX is also added.
Further, this introduces a new SDNode type, [us]divrem_ext_hreg, which is
then lowered to a sequence containing a single zext (rather than 2).
Differential Revision: http://reviews.llvm.org/D6064
llvm-svn: 221176
"[x86] Simplify vector selection if condition value type matches vselect value type and true value is all ones or false value is all zeros."
llvm-svn: 221028
This transformation worked if selector is produced by SETCC, however SETCC is needed only if we consider to swap operands. So I replaced SETCC check for this case.
Added tests for vselect of <X x i1> values.
llvm-svn: 220777
Ffter commit at rev219046 512-bit broadcasts lowering become non-optimal. Most of tests on broadcasting and embedded broadcasting were changed and they doesn’t produce efficient code.
Example below is from commit changes (it’s the first test from test/CodeGen/X86/avx512-vbroadcast.ll):
define <16 x i32> @_inreg16xi32(i32 %a) {
; CHECK-LABEL: _inreg16xi32:
; CHECK: ## BB#0:
-; CHECK-NEXT: vpbroadcastd %edi, %zmm0
+; CHECK-NEXT: vmovd %edi, %xmm0
+; CHECK-NEXT: vpbroadcastd %xmm0, %ymm0
+; CHECK-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
; CHECK-NEXT: retq
%b = insertelement <16 x i32> undef, i32 %a, i32 0
%c = shufflevector <16 x i32> %b, <16 x i32> undef, <16 x i32> zeroinitializer
ret <16 x i32> %c
}
Here, 256-bit broadcast was generated instead of 512-bit one.
In this patch
1) I added vector-shuffle lowering through broadcasts
2) Removed asserts and branches likes because this is incorrect
- assert(Subtarget->hasDQI() && "We can only lower v8i64 with AVX-512-DQI");
3) Fixed lowering tests
llvm-svn: 220774
Minor patch to fix an issue in XFormVExtractWithShuffleIntoLoad where a load is unary shuffled, then bitcast (to a type with the same number of elements) before extracting an element.
An undef was created for the second shuffle operand using the original (post-bitcasted) vector type instead of the pre-bitcasted type like the rest of the shuffle node - this was then causing an assertion on the different types later on inside SelectionDAG::getVectorShuffle.
Differential Revision: http://reviews.llvm.org/D5917
llvm-svn: 220592
This is a first step for generating SSE rsqrt instructions for
reciprocal square root calcs when fast-math is allowed.
For now, be conservative and only enable this for AMD btver2
where performance improves significantly - for example, 29%
on llvm/projects/test-suite/SingleSource/Benchmarks/BenchmarkGame/n-body.c
(if we convert the data type to single-precision float).
This patch adds a two constant version of the Newton-Raphson
refinement algorithm to DAGCombiner that can be selected by any target
via a parameter returned by getRsqrtEstimate()..
See PR20900 for more details:
http://llvm.org/bugs/show_bug.cgi?id=20900
Differential Revision: http://reviews.llvm.org/D5658
llvm-svn: 220570
Currently, @llvm.smul.with.overflow.i8 expands to 9 instructions, where
3 are really needed.
This adds X86ISD::UMUL8/SMUL8 SD nodes, and custom lowers them to
MUL8/IMUL8 + SETO.
i8 is a special case because there is no two/three operand variants of
(I)MUL8, so the first operand and return value need to go in AL/AX.
Also, we can't write patterns for these instructions: TableGen refuses
patterns where output operands don't match SDNode results. In this case,
instructions where the output operand is an implicitly defined register.
A related special case (and FIXME) exists for MUL8 (X86InstrArith.td):
// FIXME: Used for 8-bit mul, ignore result upper 8 bits.
// This probably ought to be moved to a def : Pat<> if the
// syntax can be accepted.
[(set AL, (mul AL, GR8:$src)), (implicit EFLAGS)]
Ideally, these go away with UMUL8, but we still need to improve TableGen
support of implicit operands in patterns.
Before this change:
movsbl %sil, %eax
movsbl %dil, %ecx
imull %eax, %ecx
movb %cl, %al
sarb $7, %al
movzbl %al, %eax
movzbl %ch, %esi
cmpl %eax, %esi
setne %al
After:
movb %dil, %al
imulb %sil
seto %al
Also, remove a made-redundant testcase for PR19858, and enable more FastISel
ALU-overflow tests for SelectionDAG too.
Differential Revision: http://reviews.llvm.org/D5809
llvm-svn: 220516
X86 code to lower VSELECT messed a bit with the bits set in the mask of VSELECT
when it knows it can be lowered into BLEND. Indeed, only the high bits need to be
set for those and it optimizes those accordingly.
However, when the mask is a compile time constant, the lowering will be handled
by the generic optimizer and those modifications will generate bad code in the
generic optimizer.
This patch fixes that by preventing the optimization if the VSELECT will be
handled by the generic optimizer.
<rdar://problem/18675020>
llvm-svn: 220242
new vector shuffle lowering.
This is loosely based on a patch by Marius Wachtler to the PR (thanks!).
I refactored it a bi to use std::count_if and a mutable array ref but
the core idea was exactly right. I also added some direct testing of
this case.
I believe PR21137 is now the only remaining regression.
llvm-svn: 219081
shuffles using AVX and AVX2 instructions. This fixes PR21138, one of the
few remaining regressions impacting benchmarks from the new vector
shuffle lowering.
You may note that it "regresses" many of the vperm2x128 test cases --
these were actually "improved" by the naive lowering that the new
shuffle lowering previously did. This regression gave me fits. I had
this patch ready-to-go about an hour after flipping the switch but
wasn't sure how to have the best of both worlds here and thought the
correct solution might be a completely different approach to lowering
these vector shuffles.
I'm now convinced this is the correct lowering and the missed
optimizations shown in vperm2x128 are actually due to missing
target-independent DAG combines. I've even written most of the needed
DAG combine and will submit it shortly, but this part is ready and
should help some real-world benchmarks out.
llvm-svn: 219079
Update the entire regression test suite for the new shuffles. Remove
most of the old testing which was devoted to the old shuffle lowering
path and is no longer relevant really. Also remove a few other random
tests that only really exercised shuffles and only incidently or without
any interesting aspects to them.
Benchmarking that I have done shows a few small regressions with this on
LNT, zero measurable regressions on real, large applications, and for
several benchmarks where the loop vectorizer fires in the hot path it
shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy
Bridge machines. Running on AMD machines shows even more dramatic
improvements.
When using newer ISA vector extensions the gains are much more modest,
but the code is still better on the whole. There are a few regressions
being tracked (PR21137, PR21138, PR21139) but by and large this is
expected to be a win for x86 generated code performance.
It is also more correct than the code it replaces. I have fuzz tested
this extensively with ISA extensions up through AVX2 and found no
crashes or miscompiles (yet...). The old lowering had a few miscompiles
and crashers after a somewhat smaller amount of fuzz testing.
There is one significant area where the new code path lags behind and
that is in AVX-512 support. However, there was *extremely little*
support for that already and so this isn't a significant step backwards
and the new framework will probably make it easier to implement lowering
that uses the full power of AVX-512's table-based shuffle+blend (IMO).
Many thanks to Quentin, Andrea, Robert, and others for benchmarking
assistance. Thanks to Adam and others for help with AVX-512. Thanks to
Hal, Eric, and *many* others for answering my incessant questions about
how the backend actually works. =]
I will leave the old code path in the tree until the 3 PRs above are at
least resolved to folks' satisfaction. Then I will rip it (and 1000s of
lines of code) out. =] I don't expect this flag to stay around for very
long. It may not survive next week.
llvm-svn: 219046
It turns out this combine was always somewhat flawed -- there are cases
where nested VZEXT nodes *can't* be combined: if their types have
a mismatch that can be observed in the result. While none of these show
up in currently, once I switch to the new vector shuffle lowering a few
test cases actually form such nested VZEXT nodes. I've not come up with
any IR pattern that I can sensible write to exercise this, but it will
be covered by tests once I flip the switch.
llvm-svn: 219044
nodes to the DAG combining of them.
This will allow the combine to fire on both old vector shuffle lowering
and the new vector shuffle lowering and generally seems like a cleaner
design. I've trimmed down the code a bit and tried to make it and the
surrounding combine fairly clean while moving it around.
llvm-svn: 219042
the various ways in which blends can be used to do vector element
insertion for lowering with the scalar math instruction forms that
effectively re-blend with the high elements after performing the
operation.
This then allows me to bail on the element insertion lowering path when
we have SSE4.1 and are going to be doing a normal blend, which in turn
restores the last of the blends lost from the new vector shuffle
lowering when I got it to prioritize insertion in other cases (for
example when we don't *have* a blend instruction).
Without the patterns, using blends here would have regressed
sse-scalar-fp-arith.ll *completely* with the new vector shuffle
lowering. For completeness, I've added RUN-lines with the new lowering
here. This is somewhat superfluous as I'm about to flip the default, but
hey, it shows that this actually significantly changed behavior.
The patterns I've added are just ridiculously repetative. Suggestions on
making them better very much welcome. In particular, handling the
commuted form of the v2f64 patterns is somewhat obnoxious.
llvm-svn: 219033
and MOVSD nodes for single element vector inserts.
This is particularly important because a number of patterns in the
backend detect these patterns and leverage them to simplify things. It
also fixes quite a few of the insertion bad code examples. However, it
regresses a specific area: when available, blendps and blendpd are
*dramatically* faster than movss and movsd respectively. But it doesn't
really work to form the blend logic first because the blends *aren't* as
crazy efficient when the data is coming from memory anyways, and thus
will have a movss or movsd regardless. Also, doing that would block
a bunch of the patterns that this is designed to hit.
So my plan is to go into the patterns for lowering MOVSS and MOVSD and
lower them via blends when available. However that's a pretty invasive
restructuring so it will need to be a follow-up patch.
I have already gone into the patterns to lower MOVSS and MOVSD from
memory using MOVLPD, etc. Without that, several of the test cases
I already have regress.
llvm-svn: 218985
lowering to handle the potential mirroring of 2-element vectors (because
we can't reliably sort them one way) in the caller rather than in the
insertion logic.
This will simplify things considerably as more ways to fail to match the
insertion are added because now we have a nice try and retry point.
llvm-svn: 218980
lowering to match VZEXT_MOVL patterns.
I hadn't realized that these had sufficient pattern smarts in the
backend to lower zext-ing from the low element of a vector without it
being a scalar_to_vector node. They do, and this is how to match a bunch
of patterns for movq, movss, etc.
There is a weird propensity to end up using pshufd to place the element
afterward even though it means domain crossing (or rather, to use
xorps+movss to zext the element rather than movq) but that's an
orthogonal problem with VZEXT_MOVL that someone should probably look at.
llvm-svn: 218977
element types to form illegal vector types.
I've added a special SSE1 test case here that makes sure we don't break
this going forward.
llvm-svn: 218974
elements as well as integer elements in order to form simpler shuffle
patterns.
This is the primary reason why we were failing to match some of the
2-and-2 floating point shuffles such as PR21140. Even after fixing this
we need to support some extra patterns in the backend in order to match
the resulting X86ISD::UNPCKL nodes into the correct instructions. This
commit should fix PR21140 and includes more comprehensive testing of
insertion patterns in v4 shuffles.
Not all of the added tests are beautiful. For example, we don't have
clever instructions to insert-via-load in the integer domain. There are
also some places where we aren't sufficiently cunning with our use of
movq and movd, but that's future work.
llvm-svn: 218911
matching and lowering 64-bit insertions.
The first problem was that we weren't looking through bitcasts to
discover that we *could* lower as insertions. Once fixed, we in turn
weren't looking through bitcasts to discover that we could fold a load
into the lowering. Once fixed, we weren't forming a SCALAR_TO_VECTOR
node around the inserted element and instead were passing a scalar to
a DAG node that expected a vector. It turns out there are some patterns
that will "lower" this into the correct asm, but the rest of the X86
backend is very unhappy with such antics.
This should fix a few more edge case regressions I've spotted going
through the regression test suite to enable the new vector shuffle
lowering.
llvm-svn: 218839
Negative FABS of either a scalar or vector should be handled the same way
on x86 with SSE/AVX: a single OR instruction of the FP operand with a
constant to light up the sign bit(s).
http://llvm.org/bugs/show_bug.cgi?id=20578
Differential Revision: http://reviews.llvm.org/D5201
llvm-svn: 218822
that keep cropping up in the regression test suite.
This also addresses one of the issues raised on the mailing list with
failing to form 'movsd' in as many cases as we realistically should.
There will be corresponding patches forthcoming for v4f32 at least. This
was a lot of fuss for a relatively small gain, but all the fuss was on
my end trying different ways of holding the pieces of the x86 fragment
patterns *just right*. Now that it works, the code is reasonably simple.
In the new test cases I'm adding here, v2i64 sticks out as just plain
horrible. I've not come up with any great ideas here other than that it
would be nice to recognize when we're *going* to take a domain crossing
hit and cross earlier to get the decent instructions. At least with AVX
it is slightly less silly....
llvm-svn: 218756
Nothing was relying on this and there are potentially some edge cases
that it would not be correct under. Removing it seems better than trying
to "fix" it as nothing was relying on it.
llvm-svn: 218755
in exposing the scalar value to the broadcast DAG fragment so that we
can catch even reloads and fold them into the broadcast.
This is somewhat magical I'm afraid but seems to work. It is also what
the old lowering did, and I've switched an old test to run both
lowerings demonstrating that we get the same result.
Unlike the old code, I'm not lowering f32 or f64 scalars through this
path when we only have AVX1. The target patterns include pretty heinous
code to re-cast those as shuffles when the scalar happens to not be
spilled because AVX1 provides no broadcast mechanism from registers
what-so-ever. This is terribly brittle. I'd much rather go through our
generic lowering code to get this. If needed, we can add a peephole to
get even more opportunities to broadcast-from-spill-slots that are
exposed post-RA, but my suspicion is this just doesn't matter that much.
llvm-svn: 218734
the same speed as pshufd but we can fold loads into the pmovzx
instructions.
This fixes some regressions that came up in the regression test suite
for the new vector shuffle lowering.
llvm-svn: 218733
VPBROADCAST.
This has the somewhat expected pervasive impact. I don't know why
I forgot about this. Everything seems good with lots of significant
improvements in the tests.
llvm-svn: 218724
Fixed lowering of this intrinsics in case when mask is v2i1 and v4i1.
Now cmp intrinsics lower in the following way:
(i8 (int_x86_avx512_mask_pcmpeq_q_128
(v2i64 %a), (v2i64 %b), (i8 %mask))) ->
(i8 (bitcast
(v8i1 (insert_subvector undef,
(v2i1 (and (PCMPEQM %a, %b),
(extract_subvector
(v8i1 (bitcast %mask)), 0))), 0))))
llvm-svn: 218669
a flawed direction and causing miscompiles. Read on for details.
Fundamentally, the premise of this patch series was to map
VECTOR_SHUFFLE DAG nodes into VSELECT DAG nodes for all blends because
we are going to *have* to lower to VSELECT nodes for some blends to
trigger the instruction selection patterns of variable blend
instructions. This doesn't actually work out so well.
In order to match performance with the existing VECTOR_SHUFFLE
lowering code, we would need to re-slice the blend in order to fit it
into either the integer or floating point blends available on the ISA.
When coming from VECTOR_SHUFFLE (or other vNi1 style VSELECT sources)
this works well because the X86 backend ensures that these types of
operands to VSELECT get sign extended into '-1' and '0' for true and
false, allowing us to re-slice the bits in whatever granularity without
changing semantics.
However, if the VSELECT condition comes from some other source, for
example code lowering vector comparisons, it will likely only have the
required bit set -- the high bit. We can't blindly slice up this style
of VSELECT. Reid found some code using Halide that triggers this and I'm
hopeful to eventually get a test case, but I don't need it to understand
why this is A Bad Idea.
There is another aspect that makes this approach flawed. When in
VECTOR_SHUFFLE form, we have very distilled information that represents
the *constant* blend mask. Converting back to a VSELECT form actually
can lose this information, and so I think now that it is better to treat
this as VECTOR_SHUFFLE until the very last moment and only use VSELECT
nodes for instruction selection purposes.
My plan is to:
1) Clean up and formalize the target pre-legalization DAG combine that
converts a VSELECT with a constant condition operand into
a VECTOR_SHUFFLE.
2) Remove any fancy lowering from VSELECT during *legalization* relying
entirely on the DAG combine to catch cases where we can match to an
immediate-controlled blend instruction.
One additional step that I'm not planning on but would be interested in
others' opinions on: we could add an X86ISD::VSELECT or X86ISD::BLENDV
which encodes a fully legalized VSELECT node. Then it would be easy to
write isel patterns only in terms of this to ensure VECTOR_SHUFFLE
legalization only ever forms the fully legalized construct and we can't
cycle between it and VSELECT combining.
llvm-svn: 218658
nodes, and rely exclusively on its logic. This removes a ton of
duplication from the blend lowering and centralizes it in one place.
One downside is that it requires a bunch of hacks to make this work with
the current legalization framework. We have to manually speculate one
aspect of legalizing VSELECT nodes to get everything to work nicely
because the existing legalization framework isn't *actually* bottom-up.
The other grossness is that we somewhat duplicate the analysis of
constant blends. I'm on the fence here. If reviewers thing this would
look better with VSELECT when it has constant operands dumping over tho
VECTOR_SHUFFLE, we could go that way. But it would be a substantial
change because currently all of the actual blend instructions are
matched via patterns in the TD files based around VSELECT nodes (despite
them not being perfect fits for that). Suggestions welcome, but at least
this removes the rampant duplication in the backend.
llvm-svn: 218600
X86 target-specific DAG combining that tried to convert VSELECT nodes
into VECTOR_SHUFFLE nodes that it "knew" would lower into
immediate-controlled blend nodes.
Turns out, we have perfectly good lowering of all these VSELECT nodes,
and indeed that lowering already knows how to handle lowering through
BLENDI to immediate-controlled blend nodes. The code just wasn't getting
used much because this thing forced the world to go through the vector
shuffle lowering. Yuck.
This also exposes that I was too aggressive in avoiding domain crossing
in v218588 with that lowering -- when the other option is to expand into
two 128-bit vectors, it is worth domain crossing. Restore that behavior
now that we have nice tests covering it.
The test updates here fall into two camps. One is where previously we
ended up with an unsigned encoding of the blend operand and now we get
a signed encoding. In most of those places there were elaborate comments
explaining exactly what these operands really mean. Rather than that,
just switch these tests to use the nicely decoded comments that make it
obvious that the final shuffle matches.
The other updates are just removing pointless domain crossing by
blending integers with PBLENDW rather than BLENDPS.
llvm-svn: 218589
crossing and generally work more like the blend emission code in the new
vector shuffle lowering.
My goal is to have the new vector shuffle lowering just produce VSELECT
nodes that are either matched here to BLENDI or are legal and matched in
the .td files to specific blend instructions. That seems much cleaner as
there are other ways to produce a VSELECT anyways. =]
No *observable* functionality changed yet, mostly because this code
appears to be near-dead. The behavior of this lowering routine did
change though. This code being mostly dead and untestable will change
with my next commit which will also point some new tests at it.
llvm-svn: 218588
AVX-512.
There is no interesting logic yet. Everything ends up eventually
delegating to the generic code to split the vector and shuffle the
halves. Interestingly, that logic does a significantly better job of
lowering all of these types than the generic vector expansion code does.
Mostly, it lets most of the cases fall back to nice AVX2 code rather
than all the way back to SSE code paths.
Step 2 of basic AVX-512 support in the new vector shuffle lowering. Next
up will be to incrementally add direct support for the basic instruction
set to each type (adding tests first).
llvm-svn: 218585
assertion, making the name generic, and improving the documentation.
Step 1 in adding very primitive support for AVX-512. No functionality
changed yet.
llvm-svn: 218584
vectors.
Someone will need to build the AVX512 lowering, which should follow
AVX1 and AVX2 *very* closely for AVX512F and AVX512BW resp. I've added
a dummy test which is a port of the v8f32 and v8i32 tests from AVX and
AVX2 to v8f64 and v8i64 tests for AVX512F and AVX512BW. Hopefully this
is enough information for someone to implement proper lowering here. If
not, I'll be happy to help, but right now the AVX-512 support isn't
a priority for me.
llvm-svn: 218583
lowerings.
This was hopelessly broken. First, the x86 backend wants '-1' to be the
element value representing true in a boolean vector, and second the
operand order for VSELECT is backwards from the actual x86 instructions.
To make matters worse, the backend is just using '-1' as the true value
to get the high bit to be set. It doesn't actually symbolically map the
'-1' to anything. But on x86 this isn't quite how it works: there *only*
the high bit is relevant. As a consequence weird non-'-1' values like
0x80 actually "work" once you flip the operands to be backwards.
Anyways, thanks to Hal for helping me sort out what these *should* be.
llvm-svn: 218582
new vector shuffle target DAG combines -- it helps to actually test for
the value you want rather than just using an integer in a boolean
context.
Have I mentioned that I loathe implicit conversions recently? :: sigh ::
llvm-svn: 218576
of widening masks.
We can't widen a zeroing mask unless both elements that would be merged
are either zeroed or undef. This is the only way to widen a mask if it
has a zeroed element.
Also clean up the code here by ordering the checks in a more logical way
and by using the symoblic values for undef and zero. I'm actually torn
on using the symbolic values because the existing code is littered with
the assumption that -1 is undef, and moreover that entries '< 0' are the
special entries. While that works with the values given to these
constants, using the symbolic constants actually makes it a bit more
opaque why this is the case.
llvm-svn: 218575
I spotted this by inspection when debugging something else, so I have no
test case what-so-ever, and am not even sure it is possible to
realistically trigger the bug. But this is what was intended here.
llvm-svn: 218565
and in the target shuffle combining when trying to widen vector
elements.
Previously only one of these was correct, and we didn't correctly
propagate zeroing target shuffle masks (which have a different sentinel
value from undef in non- target shuffle masks now). This isn't just
a missed optimization, this caused us to drop zeroing shuffles on the
floor and miscompile code. The added test case is one example of that.
There are other fixes to the test suite as a consequence of this as well
as restoring the undef elements in some of the masks that were lost when
I brought sanity to the actual *value* of the undef and zero sentinels.
I've also just cleaned up some of the PSHUFD and PSHUFLW and PSHUFHW
combining code, but that code really needs to go. It was a nice initial
attempt, but it isn't very principled and the recursive shuffle combiner
is much more powerful.
llvm-svn: 218562
that managed to elude all of my fuzz testing historically. =/
Something changed to allow this code path to actually be exercised and
it was doing bad things. It is especially heavily exercised by the
patterns that emerge when doing AVX shuffles that end up lowered through
the 128-bit code path.
llvm-svn: 218540
layer of tie-breaking sorting, it really helps to check that you're in
a tie first. =] Otherwise the whole thing cycles infinitely. Test case
added, another one found through fuzz testing.
llvm-svn: 218523
AVX support.
New test cases included. Note that none of the existing test cases
covered these buggy code paths. =/ Also, it is clear from this that
SHUFPS and SHUFPD are the most bug prone shuffle instructions in x86. =[
These were all detected by fuzz-testing. (I <3 fuzz testing.)
llvm-svn: 218522
Summary:
I originally tried doing this specifically for X86 in the backend in D5091,
but it was rather brittle and generally running too late to be general.
Furthermore, other targets may want to implement similar optimizations.
So I reimplemented it at the IR-level, fitting it into AtomicExpandPass
as it interacts with that pass (which could not be cleanly done before
at the backend level).
This optimization relies on a new target hook, which is only used by X86
for now, as the correctness of the optimization on other targets remains
an open question. If it is found correct on other targets, it should be
trivial to enable for them.
Details of the optimization are discussed in D5091.
Test Plan: make check-all + a new test
Reviewers: jfb
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D5422
llvm-svn: 218455
v4f64 and v8f32 shuffles when they are lane-crossing. We have fully
general lane-crossing permutation functions in AVX2 that make this easy.
Part of this also changes exactly when and how these vectors are split
up when we don't have AVX2. This isn't always a win but it usually is
a win, so on the balance I think its better. The primary regressions are
all things that just need to be fixed anyways such as modeling when
a blend can be completely accomplished via VINSERTF128, etc.
Also, this highlights one of the few remaining big features: we do
a really poor job of inserting elements into AVX registers efficiently.
This completes almost all of the big tricks I have in mind for AVX2. The
only things left that I plan to add:
1) element insertion smarts
2) palignr and other fairly specialized lowerings when they happen to
apply
llvm-svn: 218449
256-bit vectors with lane-crossing.
Rather than immediately decomposing to 128-bit vectors, try flipping the
256-bit vector lanes, shuffling them and blending them together. This
reduces our worst case shuffle by a pretty significant margin across the
board.
llvm-svn: 218446
lowering where it only used the mask of the low 128-bit lane rather than
the entire mask.
This allows the new lowering to correctly match the unpack patterns for
v8i32 vectors.
For reference, the reason that we check for the the entire mask rather
than checking the repeated mask is because the repeated masks don't
abide by all of the invariants of normal masks. As a consequence, it is
safer to use the full mask with functions like the generic equivalence
test.
llvm-svn: 218442
reduce the amount of checking we do here.
The first realization is that only non-crossing cases between 128-bit
lanes are handled by almost the entire function. It makes more sense to
handle the crossing cases first.
THe second is that until we actually are going to generate fancy shared
lowering strategies that use the repeated semantics of the v8i16
lowering, we should waste time checking for repeated masks. It is
simplest to directly test for the entire unpck masks anyways, so we
gained nothing from this.
This also matches the structure of v32i8 more closely.
No functionality changed here.
llvm-svn: 218441
lowering.
This completes the basic AVX2 feature support, but there are still some
improvements I'd like to do to really get the last mile of performance
here.
llvm-svn: 218440
for this now.
Should prevent folks from running afoul of this and not knowing why
their code won't instruction select the way I just did...
llvm-svn: 218436
missing test cases for it.
Unsurprisingly, without test cases, there were bugs here. Surprisingly,
this bug wasn't caught at compile time. Yep, there is an X86ISD::BLENDV.
It isn't wired to anything. Oops. I'll fix than next.
llvm-svn: 218434
lowering.
This also implements the fancy blend lowering for v16i16 using AVX2 and
teaches the X86 backend to print shuffle masks for 256-bit PSHUFB
and PBLENDW instructions. It also makes the mask decoding correct for
PBLENDW instructions. The yaks, they are legion.
Tests are updated accordingly. There are some missing tests for the
VBLENDVB lowering, but I'll add those in a follow-up as this commit has
accumulated enough cruft already.
llvm-svn: 218430
into unblended shuffles and a blend.
This is the consistent fallback for the lowering paths that have fast
blend operations available, and its getting quite repetitive.
No functionality changed.
llvm-svn: 218399
the native AVX2 instructions.
Note that the test case is really frustrating here because VPERMD
requires the mask to be in the register input and we don't produce
a comment looking through that to the constant pool. I'm going to
attempt to improve this in a subsequent commit, but not sure if I will
succeed.
llvm-svn: 218347
detection. It was incorrectly handling undef lanes by actually treating
an undef lane in the first 128-bit lane as a *numeric* shuffle value.
Fortunately, this almost always DTRT and disabled detecting repeated
patterns. But not always. =/ This patch introduces a much more
principled approach and fixes the miscompiles I spotted by inspection
previously.
llvm-svn: 218346
shuffles using the AVX2 instructions. This is the first step of cutting
in real AVX2 support.
Note that I have spotted at least one bug in the test cases already, but
I suspect it was already present and just is getting surfaced. Will
investigate next.
llvm-svn: 218338
add VPBLENDD to the InstPrinter's comment generation so we get nice
comments everywhere.
Now that we have the nice comments, I can see the bug introduced by
a silly typo in the commit that enabled VPBLENDD, and have fixed it. Yay
tests that are easy to inspect.
llvm-svn: 218335
Summary:
AtomicExpand already had logic for expanding wide loads and stores on LL/SC
architectures, and for expanding wide stores on CmpXchg architectures, but
not for wide loads on CmpXchg architectures. This patch fills this hole,
and makes use of this new feature in the X86 backend.
Only one functionnal change: we now lose the SynchScope attribute.
It is regrettable, but I have another patch that I will submit soon that will
solve this for all of AtomicExpand (it seemed better to split it apart as it
is a different concern).
Test Plan: make check-all (lots of tests for this functionality already exist)
Reviewers: jfb
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D5404
llvm-svn: 218332
VPBLENDD where appropriate even on 128-bit vectors.
According to Agner's tables, this instruction is significantly higher
throughput (can execute on any port) on Haswell chips so we should
aggressively try to form it when available.
Sadly, this loses our delightful shuffle comments. I'll add those back
for VPBLENDD next.
llvm-svn: 218322
undef in the shuffle mask. This shows up when we're printing comments
during lowering and we still have an IR-level constant hanging around
that models undef.
A nice consequence of this is *much* prettier test cases where the undef
lanes actually show up as undef rather than as a particular set of
values. This also allows us to print shuffle comments in cases that use
undef such as the recently added variable VPERMILPS lowering. Now those
test cases have nice shuffle comments attached with their details.
The shuffle lowering for PSHUFB has been augmented to use undef, and the
shuffle combining has been augmented to comprehend it.
llvm-svn: 218301
trick that I missed.
VPERMILPS has a non-immediate memory operand mode that allows it to do
asymetric shuffles in the two 128-bit lanes. Use this rather than two
shuffles and a blend.
However, it turns out the variable shuffle path to VPERMILPS (and
VPERMILPD, although that one offers no functional differenc from the
immediate operand other than variability) wasn't even plumbed through
codegen. Do such plumbing so that we can reasonably emit
a variable-masked VPERMILP instruction. Also plumb basic comment parsing
and printing through so that the tests are reasonable.
There are still a few tests which don't show the shuffle pattern. These
are tests with undef lanes. I'll teach the shuffle decoding and printing
to handle undef mask entries in a follow-up. I've looked at the masks
and they seem reasonable.
llvm-svn: 218300
td pattern). Currently we only model the immediate operand variation of
VPERMILPS and VPERMILPD, we should make that clear in the pseudos used.
Will be adding support for the variable mask variant in my next commit.
llvm-svn: 218282
We generate broadcast instructions on CPUs with AVX2 to load some constant splat vectors.
This patch should preserve all existing behavior with regular optimization levels,
but also use splats whenever possible when optimizing for *size* on any CPU with AVX or AVX2.
The tradeoff is up to 5 extra instruction bytes for the broadcast instruction to save
at least 8 bytes (up to 31 bytes) of constant pool data.
Differential Revision: http://reviews.llvm.org/D5347
llvm-svn: 218263
Summary:
Update segmented-stacks*.ll tests with x32 target case and make
corresponding changes to make them pass.
Test Plan: tests updated with x32 target
Reviewers: nadav, rafael, dschuff
Subscribers: llvm-commits, zinovy.nis
Differential Revision: http://reviews.llvm.org/D5245
llvm-svn: 218247
a more sane approach to AVX2 support.
Fundamentally, there is no useful way to lower integer vectors in AVX.
None. We always end up with a VINSERTF128 in the end, so we might as
well eagerly switch to the floating point domain and do everything
there. This cleans up lots of weird and unlikely to be correct
differences between integer and floating point shuffles when we only
have AVX1.
The other nice consequence is that by doing things this way we will make
it much easier to write the integer lowering routines as we won't need
to duplicate the logic to check for AVX vs. AVX2 in each one -- if we
actually try to lower a 256-bit vector as an integer vector, we have
AVX2 and can rely on it. I think this will make the code much simpler
and more comprehensible.
Currently, I've disabled *all* support for AVX2 so that we always fall
back to AVX. This keeps everything working rather than asserting. That
will go away with the subsequent series of patches that provide
a baseline AVX2 implementation.
Please note, I'm going to implement AVX2 *without access to hardware*.
That means I cannot correctness test this path. I will be relying on
those with access to AVX2 hardware to do correctness testing and fix
bugs here, but as a courtesy I'm trying to sketch out the framework for
the new-style vector shuffle lowering in the context of the AVX2 ISA.
llvm-svn: 218228
input v8f32 shuffles which are not 128-bit lane crossing but have
different shuffle patterns in the low and high lanes. This removes most
of the extract/insert traffic that was unnecessary and is particularly
good at lowering cases where only one of the two lanes is shuffled at
all.
I've also added a collection of test cases with undef lanes because this
lowering is somewhat more sensitive to undef lanes than others.
llvm-svn: 218226
lowering when it can use a symmetric SHUFPS across both 128-bit lanes.
This required making the SHUFPS lowering tolerant of other vector types,
and adjusting our canonicalization to canonicalize harder.
This is the last of the clever uses of symmetry I've thought of for
v8f32. The rest of the tricks I'm aware of here are to work around
assymetry in the mask.
llvm-svn: 218216
a generic vector shuffle mask into a helper that isn't specific to the
other things that influence which choice is made or the specific types
used with the instruction.
No functionality changed.
llvm-svn: 218215
of a single element into a zero vector for v4f64 and v4i64 in AVX.
Ironically, there is less to see here because xor+blend is so crazy fast
that we can't really beat that to zero the high 128-bit lane.
llvm-svn: 218214
UNPCKHPS with AVX vectors by recognizing those patterns when they are
repeated for both 128-bit lanes.
With this, we now generate the exact same (really nice) code for
Quentin's avx_test_case.ll which was the most significant regression
reported for the new shuffle lowering. In fact, I'm out of specific test
cases for AVX lowering, the rest were AVX2 I think. However, there are
a bunch of pretty obvious remaining things to improve with AVX...
llvm-svn: 218213
important bits of cleverness: to detect and lower repeated shuffle
patterns between the two 128-bit lanes with a single instruction.
This patch just teaches it how to lower single-input shuffles that fit
this model using VPERMILPS. =] There is more that needs to happen here.
llvm-svn: 218211
v8f32 shuffles in the new vector shuffle lowering code.
This is very cheap to do and makes it much more clear that anything more
expensive but overlapping with this lowering should be selected
afterward (for example using AVX2's VPERMPS). However, no functionality
changed here as without this code we would fall through to create no-op
shuffles of each input and a blend. =]
llvm-svn: 218209
VBLENDPD over using VSHUFPD. While the 256-bit variant of VBLENDPD slows
down to the same speed as VSHUFPD on Sandy Bridge CPUs, it has twice the
reciprocal throughput on Ivy Bridge CPUs much like it does everywhere
for 128-bits. There isn't a downside, so just eagerly use this
instruction when it suffices.
llvm-svn: 218208
awkward conditions. The readability improvement of this will be even
more important as I generalize it to handle more types.
No functionality changed.
llvm-svn: 218205
128-bit lane crossings, not 'half' crossings. This came up in code
review ages ago, but I hadn't really addresesd it. Also added some
documentation for the helper.
No functionality changed.
llvm-svn: 218203
actual support for complex AVX shuffling tricks. We can do independent
blends of the low and high 128-bit lanes of an avx vector, so shuffle
the inputs into place and then do the blend at 256 bits. This will in
many cases remove one blend instruction.
The next step is to permute the low and high halves in-place rather than
extracting them and re-inserting them.
llvm-svn: 218202
single-input shuffles with doubles. This allows them to fold memory
operands into the shuffle, etc. This is just the analog to the v4f32
case in my prior commit.
llvm-svn: 218193
instruction for single-vector floating point shuffles. This in turn
allows the shuffles to fold a load into the instruction which is one of
the common regressions hit with the new shuffle lowering.
llvm-svn: 218190
tricky case of single-element insertion into the zero lane of a zero
vector.
We can't just use the same pattern here as we do in every other vector
type because the general insertion logic can handle insertion into the
non-zero lane of the vector. However, in SSE4.1 with v4f32 vectors we
have INSERTPS that is a much better choice than the generic one for such
lowerings. But INSERTPS can do lots of other lowerings as well so
factoring its logic into the general insertion logic doesn't work very
well. We also can't just extract the core common part of the general
insertion logic that is faster (forming VZEXT_MOVL synthetic nodes that
lower to MOVSS when they can) because VZEXT_MOVL is often *faster* than
a blend while INSERTPS is slower! So instead we do a restrictive
condition on attempting to use the generic insertion logic to narrow it
to those cases where VZEXT_MOVL won't need a shuffle afterward and thus
will do better than INSERTPS. Then we try blending. Then we go back to
INSERTPS.
This still doesn't generate perfect code for some silly reasons that can
be fixed by tweaking the td files for lowering VZEXT_MOVL to use
XORPS+BLENDPS when available rather than XORPS+MOVSS when the input ends
up in a register rather than a load from memory -- BLENDPSrr has twice
the reciprocal throughput of MOVSSrr. Don't you love this ISA?
llvm-svn: 218177
analysis used elsewhere. This removes the last duplicate of this logic.
Also simplify the code here quite a bit. No functionality changed.
llvm-svn: 218176
floating point types and use it for both v2f64 and v2i64 single-element
insertion lowering.
This fixes the last non-AVX performance regression test case I've gotten
of for the new vector shuffle lowering. There is obvious analogous
lowering for v4f32 that I'll add in a follow-up patch (because with
INSERTPS, v4f32 requires special treatment). After that, its AVX stuff.
llvm-svn: 218175
vector lanes can be modeled as zero with a call to the new function that
computes a bit-vector representing that information.
No functionality changed here, but will allow doing more clever things
with the zero-test.
llvm-svn: 218174
lowering to support both anyext and zext and to custom lower for many
different microarchitectures.
Using this allows us to get *exactly* the right code for zext and anyext
shuffles in all the vector sizes. For v16i8, the improvement is *huge*.
The new SSE2 test case added I refused to add before this because it was
sooooo muny instructions.
llvm-svn: 218143
to undef lanes as well as defined widenable lanes. This dramatically
improves the lowering we use for undef-shuffles in a zext-ish pattern
for SSE2.
llvm-svn: 218115
shuffles that are zext-ing.
Not a lot to see here; the undef lane variant is better handled with
pshufd, but this improves the actual zext pattern.
llvm-svn: 218112
to the new vector shuffle lowering code.
This allows us to emit PMOVZX variants consistently for patterns where
it is a viable lowering. This instruction is both fast and allows us to
fold loads into it. This only hooks the new lowering up for i16 and i8
element widths, mostly so I could manage the change to the tests. I'll
add the i32 one next, although it is significantly less interesting.
One thing to note is that we already had some tests for these patterns
but those tests had far less horrible instructions. The problem is that
those tests weren't checking the strict start and end of the instruction
sequence. =[ As a consequence something changed in the lowering making
us generate *TERRIBLE* code for these patterns in SSE2 through SSSE3.
I've consolidated all of the tests and spelled out the madness that we
currently emit for these shuffles. I'm going to try to figure out what
has gone wrong here.
llvm-svn: 218102
There is no purpose in using it for single-input shuffles as
pshufd is just as fast and doesn't tie the two operands. This removes
a substantial amount of wrong-domain blend operations in SSSE3 mode. It
also completes the usage of PALIGNR for integer shuffles and addresses
one of the test cases Quentin hit with the new vector shuffle lowering.
There is still the question of whether and when to use this for floating
point shuffles. It is faster than shufps or shufpd but in the integer
domain. I don't yet really have a good heuristic here for when to use
this instruction for floating point vectors.
llvm-svn: 218038
PALIGNR. This just adds it to the v8i16 and v16i8 lowering steps where
it is completely unmatched. It also introduces the logic for detecting
rotation shuffle masks even in the presence of single input or blend
masks and arbitrarily undef lanes.
I've added fairly comprehensive tests for the matching logic in v8i16
because the tests at that size are much easier to write and manage.
I've not checked the SSE2 code generated for these tests because the
code is *horrible*. It is absolute madness. Testing it will just make
the test brittle without giving any interesting improvements in the
correctness confidence.
llvm-svn: 218013
This required a new hook called hasLoadLinkedStoreConditional to know whether
to expand atomics to LL/SC (ARM, AArch64, in a future patch Power) or to
CmpXchg (X86).
Apart from that, the new code in AtomicExpandPass is mostly moved from
X86AtomicExpandPass. The main result of this patch is to get rid of that
pass, which had lots of code duplicated with AtomicExpandPass.
llvm-svn: 217928
the blend that is matched by this are "used" in any sense, and so any
build_vector or other nodes feeding these will already drop other lanes.
llvm-svn: 217855
matching. This design just fundamentally didn't work because ADDSUB is
available prior to any legal lowerings of BLENDI nodes. Instead, we have
a dedicated ADDSUB synthetic ISD node which is pattern matched trivially
into the instructions. These nodes are then recognized by both the
existing and a trivial new lowering combine in the backend. Removing
these patterns required adding 2 missing shuffle masks to the DAG
combine, without which tests would have failed. Added the masks and
a helpful assert as well to catch if anything ever goes wrong here.
llvm-svn: 217851
that we don't use VSELECT and directly emit an addsub synthetic node.
Also remove a stale comment referencing VSELECT.
The test case is updated to use 'core2' which only has SSE3, not SSE4.1,
and it still passes. Previously it would not because we lacked
sufficient blend support to legalize the VSELECT.
llvm-svn: 217849
ADDSUBPD nodes out of blends of adds and subs.
This allows us to actually form these instructions with SSE3 rather than
only forming them when we had both SSE3 for the ADDSUB instructions and
SSE4.1 for the blend instructions. ;] Kind-of important.
I've adjusted the CPU requirements on one of the tests to demonstrate
this kicking in nicely for an SSE3 cpu configuration.
llvm-svn: 217848
introducing a synthetic X86 ISD node representing this generic
operation.
The relevant patterns for mapping these nodes into the concrete
instructions are also added, and a gnarly bit of C++ code in the
target-specific DAG combiner is replaced with simple code emitting this
primitive.
The next step is to generically combine blends of adds and subs into
this node so that we can drop the reliance on an SSE4.1 ISD node
(BLENDI) when matching an SSE3 feature (ADDSUB).
llvm-svn: 217819
when SSE4.1 is available.
This removes a ton of domain crossing from blend code paths that were
ending up in the floating point code path.
This is just the tip of the iceberg though. The real switch is for
integer blend lowering to more actively rely on this instruction being
available so we don't hit shufps at all any longer. =] That will come in
a follow-up patch.
Another place where we need better support is for using PBLENDVB when
doing so avoids the need to have two complementary PSHUFB masks.
llvm-svn: 217767
instructions from the relevant shuffle patterns.
This is the last tweak I'm aware of to generate essentially perfect
v4f32 and v2f64 shuffles with the new vector shuffle lowering up through
SSE4.1. I'm sure I've missed some and it'd be nice to check since v4f32
is amenable to exhaustive exploration, but this is all of the tricks I'm
aware of.
With AVX there is a new trick to use the VPERMILPS instruction, that's
coming up in a subsequent patch.
llvm-svn: 217761
instructions when it finds an appropriate pattern.
These are lovely instructions, and its a shame to not use them. =] They
are fast, and can hand loads folded into their operands, etc.
I've also plumbed the comment shuffle decoding through the various
layers so that the test cases are printed nicely.
llvm-svn: 217758
AVX is available, and generally tidy up things surrounding UNPCK
formation.
Originally, I was thinking that the only advantage of PSHUFD over UNPCK
instruction variants was its free copy, and otherwise we should use the
shorter encoding UNPCK instructions. This isn't right though, there is
a larger advantage of being able to fold a load into the operand of
a PSHUFD. For UNPCK, the operand *must* be in a register so it can be
the second input.
This removes the UNPCK formation in the target-specific DAG combine for
v4i32 shuffles. It also lifts the v8 and v16 cases out of the
AVX-specific check as they are potentially replacing multiple
instructions with a single instruction and so should always be valuable.
The floating point checks are simplified accordingly.
This also adjusts the formation of PSHUFD instructions to attempt to
match the shuffle mask to one which would fit an UNPCK instruction
variant. This was originally motivated to allow it to match the UNPCK
instructions in the combiner, but clearly won't now.
Eventually, we should add a MachineCombiner pass that can form UNPCK
instructions post-RA when the operand is known to be in a register and
thus there is no loss.
llvm-svn: 217755
'punpckhwd' instructions when suitable rather than falling back to the
generic algorithm.
While we could canonicalize to these patterns late in the process, that
wouldn't help when the freedom to use them is only visible during
initial lowering when undef lanes are well understood. This, it turns
out, is very important for matching the shuffle patterns that are used
to lower sign extension. Fixes a small but relevant regression in
gcc-loops with the new lowering.
When I changed this I noticed that several 'pshufd' lowerings became
unpck variants. This is bad because it removes the ability to freely
copy in the same instruction. I've adjusted the widening test to handle
undef lanes correctly and now those will correctly continue to use
'pshufd' to lower. However, this caused a bunch of churn in the test
cases. No functional change, just churn.
Both of these changes are part of addressing a general weakness in the
new lowering -- it doesn't sufficiently leverage undef lanes. I've at
least a couple of patches that will help there at least in an academic
sense.
llvm-svn: 217752
These are super simple. They even take precedence over crazy
instructions like INSERTPS because they have very high throughput on
modern x86 chips.
I still have to teach the integer shuffle variants about this to avoid
so many domain crossings. However, due to the particular instructions
available, that's a touch more complex and so a separate patch.
Also, the backend doesn't seem to realize it can commute blend
instructions by negating the mask. That would help remove a number of
copies here. Suggestions on how to do this welcome, it's an area I'm
less familiar with.
llvm-svn: 217744
support transforming the forms from the new vector shuffle lowering to
use 'movddup' when appropriate.
A bunch of the cases where we actually form 'movddup' don't actually
show up in the test results because something even later than DAG
legalization maps them back to 'unpcklpd'. If this shows back up as
a performance problem, I'll probably chase it down, but it is at least
an encoded size loss. =/
To make this work, also always do this canonicalizing step for floating
point vectors where the baseline shuffle instructions don't provide any
free copies of their inputs. This also causes us to canonicalize
unpck[hl]pd into mov{hl,lh}ps (resp.) which is a nice encoding space
win.
There is one test which is "regressed" by this: extractelement-load.
There, the test case where the optimization it is testing *fails*, the
exact instruction pattern which results is slightly different. This
should probably be fixed by having the appropriate extract formed
earlier in the DAG, but that would defeat the purpose of the test.... If
this test case is critically important for anyone, please let me know
and I'll try to work on it. The prior behavior was actually contrary to
the comment in the test case and seems likely to have been an accident.
llvm-svn: 217738
r189189 implemented AVX512 unpack by essentially performing a 256-bit unpack
between the low and the high 256 bits of src1 into the low part of the
destination and another unpack of the low and high 256 bits of src2 into the
high part of the destination.
I don't think that's how unpack works. AVX512 unpack simply has more 128-bit
lanes but other than it works the same way as AVX. So in each 128-bit lane,
we're always interleaving certain parts of both operands rather different
parts of one of the operands.
E.g. for this:
__v16sf a = { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 };
__v16sf b = { 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31 };
__v16sf c = __builtin_shufflevector(a, b, 0, 8, 1, 9, 4, 12, 5, 13, 16,
24, 17, 25, 20, 28, 21, 29);
we generated punpcklps (notice how the elements of a and b are not interleaved
in the shuffle). In turn, c was set to this:
0 16 1 17 4 20 5 21 8 24 9 25 12 28 13 29
Obviously this should have just returned the mask vector of the shuffle
vector.
I mostly reverted this change and made sure the original AVX code worked
for 512-bit vectors as well.
Also updated the tests because they matched the logic from the code.
llvm-svn: 217602
When compiling without SSE2, isTruncStoreLegal(F64, F32) would return Legal, whereas with SSE2 it would return Expand. And since the Target doesn't seem to actually handle a truncstore for double -> float, it would just output a store of a full double in the space for a float hence overwriting other bits on the stack.
Patch by Luqman Aden!
llvm-svn: 217410
support for MOVDDUP which is really important for matrix multiply style
operations that do lots of non-vector-aligned load and splats.
The original motivation was to add support for MOVDDUP as the lack of it
regresses matmul_f64_4x4 by 5% or so. However, all of the rules here
were somewhat suspicious.
First, we should always be using the floating point domain shuffles,
regardless of how many copies we have to make as a movapd is *crazy*
faster than the domain switching cost on some chips. (Mostly because
movapd is crazy cheap.) Because SHUFPD can't do the copy-for-free trick
of the PSHUF instructions, there is no need to avoid canonicalizing on
UNPCK variants, so do that canonicalizing. This also ensures we have the
chance to form MOVDDUP. =]
Second, we assume SSE2 support when doing any vector lowering, and given
that we should just use UNPCKLPD and UNPCKHPD as they can operate on
registers or memory. If vectors get spilled or come from memory at all
this is going to allow the load to be folded into the operation. If we
want to optimize for encoding size (the only difference, and only
a 2 byte difference) it should be done *much* later, likely after RA.
llvm-svn: 217332
computation was totally wrong, but somehow it didn't really show up with
llc.
I've added an assert that triggers on multiple existing test cases and
updated one of them to show the correct value.
There appear to still be more bugs lurking around insertps's mask. =/
However, note that this only really impacts the new vector shuffle
lowering.
llvm-svn: 217289
shuffle lowering for integer vectors and share it from v4i32, v8i16, and
v16i8 code paths.
Ironically, the SSE2 v16i8 code for this is now better than the SSSE3!
=] Will have to fix the SSSE3 code next to just using a single pshufb.
llvm-svn: 217240
vzext patterns and insert-element patterns that for SSE4 have dedicated
instructions.
With this we can enable the experimental mode in a regression test that
happens to cover some of the past set of issues. You can see that the
new logic does significantly better here on the floating point cases.
A follow-up to this change and the previous ones will hoist the logic
into helpers so it can be shared across element type sizes as in this
particular case it generalizes cleanly.
llvm-svn: 217136
abilities of INSERTPS which are really powerful and come up in very
important contexts such as forming diagonal matrices, etc.
With this I ended up being able to remove the somewhat weird helper
I added for INSERTPS because we can collapse the entire state to a no-op
mask. Added a bunch of tests for inserting into a zero-ish vector.
llvm-svn: 217117
'insertps' patterns.
This replaces two shuffles with a single insertps in very common cases.
My next patch will extend this to leverage the zeroing capabilities of
insertps which will allow it to be used in a much wider set of cases.
llvm-svn: 217100
We duplicate ~30 lines of code to lower FABS and FNEG for x86, so this patch combines them into one function.
No functional change intended, so no additional test cases. Test-suite behavior is unchanged.
Differential Revision: http://reviews.llvm.org/D5064
llvm-svn: 216942
This change will ease refactoring LowerFABS() and LowerFNEG()
since they have a lot of overlap.
Remove the creation of a floating point constant from an integer
because it's going to be used for a bitwise integer op anyway.
No change to codegen expected, but the verbose comment string
for asm output may change from float values to hex (integer),
depending on whether the constant already exists or not.
Differential Revision: http://reviews.llvm.org/D5052
llvm-svn: 216889
Summary:
If a variadic function body contains a musttail call, then we copy all
of the remaining register parameters into virtual registers in the
function prologue. We track the virtual registers through the function
body, and add them as additional registers to pass to the call. Because
this is all done in virtual registers, the register allocator usually
gives us good code. If the function does a call, however, it will have
to spill and reload all argument registers (ew).
Forwarding regparms on x86_32 is not implemented because most compilers
don't support varargs in 32-bit with regparms.
Reviewers: majnemer
Subscribers: aemerson, llvm-commits
Differential Revision: http://reviews.llvm.org/D5060
llvm-svn: 216780
We've rejected these kinds of functions since r28405 in 2006 because
it's impossible to lower the return of a callee cleanup varargs
function. However there are lots of legal ways to leave such a function
without returning, such as aborting. Today we can leave a function with
a musttail call to another function with the correct prototype, and
everything works out.
I'm removing the verifier check declaring that a normal return from such
a function is UB.
Reviewed By: nlewycky
Differential Revision: http://reviews.llvm.org/D5059
llvm-svn: 216779
functionality changed.
Separating this into two functions wasn't helping. There was a decent
amount of boilerplate duplicated, and some subsequent refactorings here
will pull even more common code out.
llvm-svn: 216644
we stopped efficiently lowering sextload using the SSE41 instructions
for that operation.
This is a consequence of a bad predicate I used thinking of the memory
access needs. The code actually handles the cases where the predicate
doesn't apply, and handles them much better. =] Simple fix and a test
case added. Fixes PR20767.
llvm-svn: 216538
This combine is essentially combining target-specific nodes back into target
independent nodes that it "knows" will be combined yet again by a target
independent DAG combine into a different set of target-independent nodes that
are legal (not custom though!) and thus "ok". This seems... deeply flawed. The
crux of the problem is that we don't combine un-legalized shuffles that are
introduced by legalizing other operations, and thus we don't see a very
profitable combine opportunity. So the backend just forces the input to that
combine to re-appear.
However, for this to work, the conditions detected to re-form the unlegalized
nodes must be *exactly* right. Previously, failing this would have caused poor
code (if you're lucky) or a crasher when we failed to select instructions.
After r215611 we would fall back into the legalizer. In some cases, this just
"fixed" the crasher by produces bad code. But in the test case added it caused
the legalizer and the dag combiner to iterate forever.
The fix is to make the alignment checking in the x86 side of things match the
alignment checking in the generic DAG combine exactly. This isn't really a
satisfying or principled fix, but it at least make the code work as intended.
It also highlights that it would be nice to detect the availability of under
aligned loads for a given type rather than bailing on this optimization. I've
left a FIXME to document this.
Original commit message for r215611 which covers the rest of the chang:
[SDAG] Fix a case where we would iteratively legalize a node during
combining by replacing it with something else but not re-process the
node afterward to remove it.
In a truly remarkable stroke of bad luck, this would (in the test case
attached) end up getting some other node combined into it without ever
getting re-processed. By adding it back on to the worklist, in addition
to deleting the dead nodes more quickly we also ensure that if it
*stops* being dead for any reason it makes it back through the
legalizer. Without this, the test case will end up failing during
instruction selection due to an and node with a type we don't have an
instruction pattern for.
It took many million runs of the shuffle fuzz tester to find this.
llvm-svn: 216537
This actually was caught by existing tests but those tests were disabled
with an XFAIL because of PR20736. While working on fixing that,
I noticed the test failure, and tracked it down to this.
We even have a really nice Clang warning that would have caught this but
it isn't enabled in LLVM! =[ I may look at enabling it.
llvm-svn: 216391
these DAG combines.
The DAG auto-CSE thing is truly terrible. Due to it, when RAUW-ing
a node with its operand, you can cause its uses to CSE to itself, which
then causes their uses to become your uses which causes them to be
picked up by the RAUW. For nodes that are determined to be "no-ops",
this is "fine". But if the RAUW is one of several steps to enact
a transformation, this causes the DAG to really silently eat an discard
nodes that you would never expect. It took days for me to actually
pinpoint a test case triggering this and a really frustrating amount of
time to even comprehend the bug because I never even thought about the
ability of RAUW to iteratively consume nodes due to CSE-ing them into
itself.
To fix this, we have to build up a brand-new chain of operations any
time we are combining across (potentially) intervening nodes. But once
the logic is added to do this, another issue surfaces: CombineTo eagerly
deletes the one node combined, *but no others*. This is... really
frustrating. If deleting it makes its operands become dead, those
operand nodes often won't go onto the worklist in the
order you would want -- they're already on it and not near the top. That
means things higher on the worklist will get combined prior to these
dead nodes being GCed out of the worklist, and if the chain is long, the
immediate users won't be enough to re-detect where the root of the chain
is that became single-use again after deleting the dead nodes. The
better way to do this is to never immediately delete nodes, and instead
to just enqueue them so we can recursively delete them. The
combined-from node is typically not on the worklist anyways by virtue of
having been popped off.... But that in turn breaks other tests that
*require* CombineTo to delete unused nodes. :: sigh ::
Fortunately, there is a better way. This whole routine should have been
returning the replacement rather than using CombineTo which is quite
hacky. Switch to that, and all the pieces fall together.
I suspect the same kind of miscompile is possible in the half-shuffle
folding code, and potentially the recursive folding code. I'll be
switching those over to a pattern more like this one for safety's sake
even though I don't immediately have any test cases for them. Note that
the only way I got a test case for this instance was with *heavily* DAG
combined 256-bit shuffle sequences generated by my fuzzer. ;]
llvm-svn: 216319
There's no need to do this if the user doesn't call va_start. In the
future, we're going to have thunks that forward these register
parameters with musttail calls, and they won't need these spills for
handling va_start.
Most of the test suite changes are adding va_start calls to existing
tests to keep things working.
llvm-svn: 216294
This (mostly) reverts commit r216119.
Somewhere during the review Reid committed r214980 which fixed this
another way, and I neglected to check that the testcase still failed
before committing.
I've left test/CodeGen/X86/aligned-variadic.ll around in case it adds
extra coverage.
llvm-svn: 216246
Fix for PR20648 - http://llvm.org/bugs/show_bug.cgi?id=20648
This patch checks the operands of a vselect to see if all values are constants.
If yes, bail out of any further attempts to create a blend or shuffle because
SelectionDAGLegalize knows how to turn this kind of vselect into a single load.
This already happens for machines without SSE4.1, so the added checks just send
more targets down that path.
Differential Revision: http://reviews.llvm.org/D4934
llvm-svn: 216121
The goal of the patch is to implement section 3.2.3 of the AMD64 ABI
correctly. The controlling sentence is, "The size of each argument gets
rounded up to eightbytes. Therefore the stack will always be eightbyte
aligned." The equivalent sentence in the i386 ABI page 37 says, "At all
times, the stack pointer should point to a word-aligned area." For both
architectures, the stack pointer is not being rounded up to the nearest
eightbyte or word between the last normal argument and the first
variadic argument.
Patch by Thomas Jablin!
llvm-svn: 216119
Summary: This fixes http://llvm.org/bugs/show_bug.cgi?id=19530.
The problem is that X86ISelLowering erroneously thought the third call
was eligible for tail call elimination.
It would have been if it's return value was actually the one returned
by the calling function, but here that is not the case and
additional values are being returned.
Test Plan: Test case from the original bug report is included.
Reviewers: rafael
Reviewed By: rafael
Subscribers: rafael, llvm-commits
Differential Revision: http://reviews.llvm.org/D4968
llvm-svn: 216117
It should remove dosens of lines in handling instrinsics (in a huge switch) and give an easy way to add new intrinsics.
I did not completed to move al intrnsics to the table, I'll do this in the upcomming commits.
llvm-svn: 215826
MSVC gives this awesome diagnostic:
..\lib\Target\X86\X86ISelLowering.cpp(7085) : error C2971: 'llvm::VariadicFunction1' : template parameter 'Func' : 'isShuffleEquivalentImpl' : a local variable cannot be used as a non-type argument
..\include\llvm/ADT/VariadicFunction.h(153) : see declaration of 'llvm::VariadicFunction1'
..\lib\Target\X86\X86ISelLowering.cpp(7061) : see declaration of 'isShuffleEquivalentImpl'
Using an anonymous namespace makes the problem go away.
llvm-svn: 215744
the new shuffle lowering and an implementation for v4 shuffles.
This allows us to handle non-half-crossing shuffles directly for v4
shuffles, both integer and floating point. This currently misses places
where we could perform the blend via UNPCK instructions, but otherwise
generates equally good or better code for the test cases included to the
existing vector shuffle lowering. There are a few cases that are
entertainingly better. ;]
llvm-svn: 215702
target-specific shuffl DAG combines.
We were recognizing the paired shuffles backwards. This code needs to be
replaced anyways as we have the same functionality elsewhere, but I'll
do the refactoring in a follow-up, this is the minimal fix to the
behavior.
In addition to fixing miscompiles with the new vector shuffle lowering,
it also causes the canonicalization to kick in much better, selecting
the smaller encoding variants in lots of places in the new AVX path.
This still isn't quite ideal as we don't need both the shufpd and the
punpck instructions, but that'll get fixed in a follow-up patch.
llvm-svn: 215690
broken logic for merging shuffle masks in the face of SM_SentinelZero
mask operands.
While these are '-1' they don't mean 'undef' the way '-1' means in the
pre-legalized shuffle masks. Instead, they mean that the shuffle
operation is forcibly zeroing that lane. Reflect this and explicitly
handle it in a bunch of places. In one place the effect is equivalent
but much more clear. In the rest it was really weirdly broken.
Also, rewrite the entire merging thing to be a more directy operation
with a single loop and just doing math to map the indices through the
various masks.
Also add a bunch of asserts to try to make in extremely clear what the
different masks can possibly look like.
Finally, add some comments to clarify that we're merging shuffle masks
*up* here rather than *down* as we do everywhere else, and thus the
logic is quite confusing.
Thanks to several different people for sending test cases, and for
Robert Khasanov for an initial attempt at fixing.
llvm-svn: 215687
lowering scheme.
Currently, this just directly bails to the fallback path of splitting
the 256-bit vector into two 128-bit vectors, operating there, and then
joining the results back together. While the results are far from
perfect, they are *shockingly* good for what we're doing here. I'll be
layering the rest of the functionality on top of this piece by piece and
updating tests as I go.
Note that 256-bit vectors in this mode are still somewhat WIP. While
I think the code paths that I'm adding here are clean and good-to-go,
there are still a lot of 128-bit assumptions that I'll need to stomp out
as I march through the functional spread here.
llvm-svn: 215637
one pesky test case correctly.
This test case caused the old code to infloop occilating between solving
the low-half and the high-half. The 'side balancing' part of
single-input v8 shuffle lowering didn't handle the one pattern which can
cause it to occilate. Fortunately the fuzz testing found this case.
Unfortuately it was *terrible* to handle. I'm really sorry for the
amount and density of the code here, I'd love suggestions on how to
simplify it. I feel like there *must* be a simpler form here, but after
a lot of days I've not found it. This is the only one I've found that
even works. I've added the one pesky test case along with some nice
comments explaining the core problem that we have to solve here.
So far this has survived approximately 32k test cases. More strenuous
fuzzing commencing.
llvm-svn: 215519
I think that this will scale better in most cases than adding a Pat<> for each
mapping from the intrinsic DAG to the intruction (i.e. rri, rrik, rrikz). We
can just lower to the SDNode and have the resulting DAG be matches by the DAG
patterns.
Alternatively (long term), we could keep the Pat<>s but generate them via the
new AVX512_masking multiclass. The difficulty is that in order to formulate
that we would have to concatenate DAGs. Currently this is only supported if
the operators of the input DAGs are identical.
llvm-svn: 215473
shuffle lowering.
This is closely related to the previous one. Here we failed to use the
source offset when swapping in the other case -- where we end up
swapping the *final* shuffle. The cause of this bug is a bit different:
I simply wasn't thinking about the fact that this mask is actually
a slice of a wide mask and thus has numbers that need SourceOffset
applied. Simple fix. Would be even more simple with an algorithm-y thing
to use here, but correctness first. =]
llvm-svn: 215095
via the fuzz tester.
Here I missed an offset when round-tripping a value through a shuffle
mask. I got it right 2 lines below. See a problem? I do. ;] I'll
probably be adding a little "swap" algorithm which accepts a range and
two values and swaps those values where they occur in the range. Don't
really have a name for it, let me know if you do.
llvm-svn: 215094
through the new fuzzer.
This one is great: bad operator precedence led the modulus to happen at
the wrong point. All the asserts didn't fire because there were usually
the right values past the end of the 4 element region we were looking
at. Probably could have gotten a crash here with ASan + fuzzing, but the
correctness tests pinpointed this really nicely.
llvm-svn: 215092
Summary:
Since pointers are 32-bit on x32 we can use ebp and esp as frame and stack
pointer. Some operations like PUSH/POP and CFI_INSTRUCTION still
require 64-bit register, so using 64-bit MachineFramePtr where required.
X86_64 NaCl uses 64-bit frame/stack pointers, however it's been found that
both isTarget64BitLP64 and isTarget64BitILP32 are true for NaCl. Addressing
this issue here as well by making isTarget64BitLP64 false.
Also mark hasReservedSpillSlot unreachable on X86. See inlined comments.
Test Plan: Add one new simple test and upgrade 2 existing with x32 target case.
Reviewers: nadav, dschuff
Subscribers: llvm-commits, zinovy.nis
Differential Revision: http://reviews.llvm.org/D4617
llvm-svn: 215091
fuzz testing.
The function which tested for adjacency did what it said on the tin, but
when I called it, I wanted it to do something more thorough: I wanted to
know if the *pairs* of shuffle elements were adjacent and started at
0 mod 2. In one place I had the decency to try to test for this, but in
the other it was completely skipped, miscompiling this test case. Fix
this by making the helper actually do what I wanted it to do everywhere
I called it (and removing the now redundant code in one place).
I *really* dislike the name "canWidenShuffleElements" for this
predicate. If anyone can come up with a better name, please let me know.
The other name I thought about was "canWidenShuffleMask" but is it
really widening the mask to reduce the number of lanes shuffled? I don't
know. Naming things is hard.
llvm-svn: 215089
to get the subtarget and that's accessible from the MachineFunction
now. This helps clear the way for smaller changes where we getting
a subtarget will require passing in a MachineFunction/Function as
well.
llvm-svn: 214988
test case to actually generate correct code.
The primary miscompile fixed here is that we weren't correctly handling
in-place elements in one half of a single-input v8i16 shuffle when
moving a dword of elements from that half to the other half. Some times,
we would clobber the in-place elements in forming the dword to move
across halves.
The fix to this involves forcibly marking the in-place inputs even when
there is no need to gather them into a dword, and to much more carefully
re-arrange the elements when grouping them into a dword to move across
halves. With these two changes we would generate correct shuffles for
the test case, but found another miscompile. There are also some random
perturbations of the generated shuffle pattern in SSE2. It looks like
a wash; more instructions in some cases fewer in others.
The second miscompile would corrupt the results into nonsense. This is
a buggy pattern in one of the added DAG combines. Mapping elements
through a PSHUFD when pairing redundant half-shuffles is *much* harder
than this code makes it out to be -- it requires reasoning about *all*
of where the input is used in the PSHUFD, not just one part of where it
is used. Plus, we can't combine a half shuffle *into* a PSHUFD but the
code didn't guard against it. I think this was just a bad idea and I've
just removed that aspect of the combine. No tests regress as
a consequence so seems OK.
llvm-svn: 214954
not corrupting the mask by mutating it more times than intended. No
functionality changed (the results were non-overlapping so the old
version "worked" but was non-obvious).
llvm-svn: 214953
a test case.
We also miscompile this test case which is showing a serious flaw in the
single-input v8i16 shuffle code. I've left the specific instruction
checks FIXME-ed out until I can address the bug in the single-input
code, but I wanted to separate out a significant functionality change to
produce correct code from a very simple and targeted crasher fix.
The miscompile problem stems from keeping track of inputs by value
rather than by index. As a consequence of doing this, we can't reliably
update those inputs because they might swap and we can't detect this
without copying the mask.
The blend code now uses indices for the input lists and this seems
strictly better. It also should make it easier to sort things and do
other cleanups. I think the time has come to simplify The Great Lambda
here.
llvm-svn: 214914
This was currently part of lowering to PALIGNR with some special-casing to
make interlane shifting work. Since AVX512F has interlane alignr (valignd/q)
and AVX512BW has vpalignr we need to support both of these *at the same time*,
e.g. for SKX.
This patch breaks out the common code and then add support to check both of
these lowering options from LowerVECTOR_SHUFFLE.
I also added some FIXMEs where I think the AVX512BW and AVX512VL additions
should probably go.
llvm-svn: 214888
They have different semantics (valign is interlane while palingr is intralane)
and palingr is still needed even in the AVX512 context. According to the
latest spec AVX512BW provides these.
llvm-svn: 214887
found by a single test reduced out of a failure on llvm-stress.
The start of the problem (and the crash) came when we tried to use
a find of a non-used slot in the move-to half of the move-mask as the
target for two bad-half inputs. While if lucky this will be the first of
a pair of slots which we can place the bad-half inputs into, it isn't
actually guaranteed. This really isn't surprising, not sure what I was
thinking. The correct way to find the two unused slots is to look for
one of the *used* slots. We know it isn't that pair, and we can use some
modular arithmetic to find the other pair by masking off the odd bit and
adding 2 modulo 4. With this, we reliably found a viable pair of slots
for the bad-half inputs.
Sadly, that wasn't enough. We also had a wrong code bug that surfaced
when I reduced the test case for this where we would use the same slot
twice for the two bad inputs. This is because both of the bad inputs
could be in odd slots originally and thus the mod-2 mapping would
actually be the same. The whole point of the weird indexing into the
pair of empty slots was to try to leverage when the end result needed
the two bad-half inputs to be paired in a dword and pre-pair them in the
correct orrientation. This is less important with the powerful combining
we're now doing, and also easier and more reliable to achieve be noting
that we add the bad-half inputs in order. Thus, if they are in a dword
pair, the low part of that will be the first input in the sequence.
Always putting that in the low element will just do the right thing in
addition to computing the correct result.
Test case added. =]
llvm-svn: 214849
shorter/easier and have the DAG use that to do the same lookup. This
can be used in the future for TargetMachine based caching lookups from
the MachineFunction easily.
Update the MIPS subtarget switching machinery to update this pointer
at the same time it runs.
llvm-svn: 214838
use of PACKUS. It's cleaner that way.
I looked at implementing clever combine-based folding of PACKUS chains
into PSHUFB but it is quite hard and doesn't seem likely to be worth it.
The most annoying part would be detecting that the correct masking had
been done to use PACKUS-style instructions as a blend operation rather
than there being any saturating as is indicated by its name. We generate
really nice code for what few test cases I've come up with that aren't
completely contrived for this by just directly prefering PSHUFB and so
let's go with that strategy for now. =]
llvm-svn: 214707
patterns of v16i8 shuffles.
This implements one of the more important FIXMEs for the SSE2 support in
the new shuffle lowering. We now generate the optimal shuffle sequence
for truncate-derived shuffles which show up essentially everywhere.
Unfortunately, this exposes a weakness in other parts of the shuffle
logic -- we can no longer form PSHUFB here. I'll add the necessary
support for that and other things in a subsequent commit.
llvm-svn: 214702
I spent some time looking into a better or more principled way to handle
this. For example, by detecting arbitrary "unneeded" ORs... But really,
there wasn't any point. We just shouldn't build blatantly wrong code so
late in the pipeline rather than adding more stages and logic later on
to fix it. Avoiding this is just too simple.
llvm-svn: 214680
lowering with a small addition to it and adding PSHUFB combining.
There is one obvious place in the new vector shuffle lowering where we
should form PSHUFBs directly: when without them we will unpack a vector
of i8s across two different registers and do a potentially 4-way blend
as i16s only to re-pack them into i8s afterward. This is the crazy
expensive fallback path for i8 shuffles and we can just directly use
pshufb here as it will always be cheaper (the unpack and pack are
two instructions so even a single shuffle between them hits our
three instruction limit for forming PSHUFB).
However, this doesn't generate very good code in many cases, and it
leaves a bunch of common patterns not using PSHUFB. So this patch also
adds support for extracting a shuffle mask from PSHUFB in the X86
lowering code, and uses it to handle PSHUFBs in the recursive shuffle
combining. This allows us to combine through them, combine multiple ones
together, and generally produce sufficiently high quality code.
Extracting the PSHUFB mask is annoyingly complex because it could be
either pre-legalization or post-legalization. At least this doesn't have
to deal with re-materialized constants. =] I've added decode routines to
handle the different patterns that show up at this level and we dispatch
through them as appropriate.
The two primary test cases are updated. For the v16 test case there is
still a lot of room for improvement. Since I was going through it
systematically I left behind a bunch of FIXME lines that I'm hoping to
turn into ALL lines by the end of this.
llvm-svn: 214628
of normally binary shuffle instructions like PUNPCKL and MOVLHPS.
This detects cases where a single register is used for both operands
making the shuffle behave in a unary way. We detect this and adjust the
mask to use the unary form which allows the existing DAG combine for
shuffle instructions to actually work at all.
As a consequence, this uncovered a number of obvious bugs in the
existing DAG combine which are fixed. It also now canonicalizes several
shuffles even with the existing lowering. These typically are trying to
match the shuffle to the domain of the input where before we only really
modeled them with the floating point variants. All of the cases which
change to an integer shuffle here have something in the integer domain, so
there are no more or fewer domain crosses here AFAICT. Technically, it
might be better to go from a GPR directly to the floating point domain,
but detecting floating point *outputs* despite integer inputs is a lot
more code and seems unlikely to be worthwhile in practice. If folks are
seeing domain-crossing regressions here though, let me know and I can
hack something up to fix it.
Also as a consequence, a bunch of missed opportunities to form pshufb
now can be formed. Notably, splats of i8s now form pshufb.
Interestingly, this improves the existing splat lowering too. We go from
3 instructions to 1. Yes, we may tie up a register, but it seems very
likely to be worth it, especially if splatting the 0th byte (the
common case) as then we can use a zeroed register as the mask.
llvm-svn: 214625
Stop using ST registers for function returns and inline-asm instructions and use
FP registers instead. This allows removing a large amount of code in the
stackifier pass that was needed to track register liveness and handle copies
between ST and FP registers and function calls returning floating point values.
It also fixes a bug which manifests when an ST register defined by an
inline-asm instruction was live across another inline-asm instruction, as shown
in the following sequence of machine instructions:
1. INLINEASM <es:frndint> $0:[regdef], %ST0<imp-def,tied5>
2. INLINEASM <es:fldcw $0>
3. %FP0<def> = COPY %ST0
<rdar://problem/16952634>
llvm-svn: 214580
Currently when DAGCombine converts loads feeding a switch into a switch of
addresses feeding a load the new load inherits the isInvariant flag of the left
side. This is incorrect since invariant loads can be reordered in cases where it
is illegal to reoarder normal loads.
This patch adds an isInvariant parameter to getExtLoad() and updates all call
sites to pass in the data if they have it or false if they don't. It also
changes the DAGCombine to use that data to make the right decision when
creating the new load.
llvm-svn: 214449
Rename to allowsMisalignedMemoryAccess.
On R600, 8 and 16 byte accesses are mostly OK with 4-byte alignment,
and don't need to be split into multiple accesses. Vector loads with
an alignment of the element type are not uncommon in OpenCL code.
llvm-svn: 214055
instructions in the legalized DAG, and leverage it to combine long
sequences of instructions to PSHUFB.
Eventually, the other x86-instruction-specific shuffle combines will
probably all be driven out of this routine. But the real motivation is
to detect after we have fully legalized and optimized a shuffle to the
minimal number of x86 instructions whether it is profitable to replace
the chain with a fully generic PSHUFB instruction even though doing so
requires either a load from a constant pool or tying up a register with
the mask.
While the Intel manuals claim it should be used when it replaces 5 or
more instructions (!!!!) my experience is that it is actually very fast
on modern chips, and so I've gon with a much more aggressive model of
replacing any sequence of 3 or more instructions.
I've also taught it to do some basic canonicalization to special-purpose
instructions which have smaller encodings than their generic
counterparts.
There are still quite a few FIXMEs here, and I've not yet implemented
support for lowering blends with PSHUFB (where its power really shines
due to being able to zero out lanes), but this starts implementing real
PSHUFB support even when using the new, fancy shuffle lowering. =]
llvm-svn: 214042
The tale starts with r212808 which attempted to fix inversion of the low
and high bits when lowering MUL_LOHI. Sadly, that commit did not include
any positive test cases, and just removed some operations from a test
case where the actual logic being changed isn't fully visible from the
test.
What this commit did was two things. First, it reversed the low and high
results in the formation of the MERGE_VALUES node for the multiple
results. This is entirely correct.
Second it changed the shuffles for extracting the low and high
components from the i64 results of the multiplies to extract them
assuming a big-endian-style encoding of the multiply results. This
second change is wrong. There is no big-endian encoding in x86, the
results of the multiplies are normal v2i64s: when cast to v4i32, the low
i32s are at offsets 0 and 2, and the high i32s are at offsets 1 and 3.
However, the first change wasn't enough to actually fix the bug, which
is (I assume) why the second change was also made. There was another bug
in the MERGE_VALUES formation: we weren't using a VTList, and so were
getting a single result node! When grabbing the *second* result from the
node, we got... well.. colud be anything. I think this *appeared* to
invert things, but had to be causing other problems as well.
Fortunately, I fixed the MERGE_VALUES issue in r213931, so we should
have been fine, right? NOOOPE! Because the core bug was never addressed,
the test in vector-idiv failed when I fixed the MERGE_VALUES node.
Because there are essentially no docs for this node, I had to guess at
how to fix it and tried swapping the operands, restoring the order of
the original code before r212808. While this "fixed" the test case (in
that we produced the write instructions) we were still extracting the
wrong elements of the i64s, and thus PR20355 was still broken.
This commit essentially reverts the big-endian-style extraction part of
r212808 and goes back to the original masks which were correct. Now that
the MERGE_VALUES node formation is also correct, everything works. I've
also included a more detailed test from PR20355 to make sure this stays
fixed.
llvm-svn: 214011
The clever way to implement signed multiplication with unsigned *is
already implemented* and tested and working correctly. The bug is
somewhere else. Re-investigating.
This will teach me to not scroll far enough to read the code that did
what I thought needed to be done.
llvm-svn: 214009
signed multiplication is requested. While there is not a difference in
the *low* half of the result, the *high* half (used specifically to
implement the signed division by these constants) certainly is used. The
test case I've nuked was actively asserting wrong code.
There is a delightful solution to doing signed multiplication even when
we don't have it that Richard Smith has crafted, but I'll add the
machinery back and implement that in a follow-up patch. This at least
restores correctness.
llvm-svn: 214007
address of the stack guard was being spilled to the stack.
Previously the address of the stack guard would get spilled to the stack if it
was impossible to keep it in a register. This patch introduces a new target
independent node and pseudo instruction which gets expanded post-RA to a
sequence of instructions that load the stack guard value. Register allocator
can now just remat the value when it can't keep it in a register.
<rdar://problem/12475629>
llvm-svn: 213967
SDValues, fixing the two bugs left in the regression suite.
The key for both of these was the use a single value type rather than
a VTList which caused an unintentionally single-result merge-value node.
Fix this by getting the appropriate VTList in place.
Doing this exposed that the comments in x86's code abouth how MUL_LOHI
operands are handle is wrong. The bug with the use of out-of-range
result numbers was hiding the bug about the order of operands here (as
best i can tell). There are more places where the code appears to get
this backwards still...
llvm-svn: 213931
vector operation legalization with support for custom target lowering
and fallback to expand when it fails, and use this to implement sext and
anyext load lowering for x86 in a more principled way.
Previously, the x86 backend relied on a target DAG combine to "combine
away" sextload and extload nodes prior to legalization, or would expand
them during legalization with terrible code. This is particularly
problematic because the DAG combine relies on running over non-canonical
DAG nodes at just the right time to match several common and important
patterns. It used a combine rather than lowering because we didn't have
good lowering support, and to expose some tricks being employed to more
combine phases.
With this change it becomes a proper lowering operation, the backend
marks that it can lower these nodes, and I've added support for handling
the canonical forms that don't have direct legal representations such as
sextload of a v4i8 -> v4i64 on AVX1. With this change, our test cases
for this behavior continue to pass even after the DAG combiner beigns
running more systematically over every node.
There is some noise caused by this in the test suite where we actually
use vector extends instead of subregister extraction. This doesn't
really seem like the right thing to do, but is unlikely to be a critical
regression. We do regress in one case where by lowering to the
target-specific patterns early we were able to combine away extraneous
legal math nodes. However, this regression is completely addressed by
switching to a widening based legalization which is what I'm working
toward anyways, so I've just switched the test to that mode.
Differential Revision: http://reviews.llvm.org/D4654
llvm-svn: 213897
GCC 4.8 detected a signed compare [-Wsign-compare]. Add a cast for the
destination index. Add an assert to catch a potential overflow however unlikely
it may be.
llvm-svn: 213878
When we had a vector_shuffle where we had an input from each vector, we
could miscompile it because we were assuming the input from V2 wouldn't
be moved from where it was on the vector.
Added a test case.
llvm-svn: 213826
The transform to constant fold unary operations with an AND across a
vector comparison applies when the constant is not a splat of a scalar
as well.
llvm-svn: 213800
The folding of unary operations through a vector compare and mask operation
is only safe if the unary operation result is of the same size as its input.
For example, it's not safe for [su]itofp from v4i32 to v4f64.
llvm-svn: 213799
Sanjay Patel