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>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224351 91177308-0d34-0410-b5e6-96231b3b80d8
An instruction alias defined with InstAlias and an optional operand in the
middle of the AsmString field, "..${a} <operands>", would get the final
"}" printed in the instruction disassembly. This wouldn't happen if the optional
operand appeared as the last item in the AsmString which is how the current
backends avoided the problem.
There don't appear to be any tests for this part of Tablegen but it passes the
pre-commit tests. Manually tested the change by enabling the generic alias
printer in the ARM backend and checking the output.
Differential Revision: http://reviews.llvm.org/D6529
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224348 91177308-0d34-0410-b5e6-96231b3b80d8
On X86, the Intel asm parser tries to match all memory operand sizes when
none is explicitly specified. For LEA, which doesn't really have a memory
operand (just a pointer one), this results in multiple successful matches,
one for each memory size. There's no error because it's same opcode, so
really, it's just one match. However, the tablegen'd matcher function
adds opcode/operands to the passed MCInst, and this results in multiple
duplicated operands.
This commit clears the MCInst in the tablegen'd matcher function.
We sometimes clear it when the match failed, so there's no expectation of
keeping the previous content anyway.
Differential Revision: http://reviews.llvm.org/D6670
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224347 91177308-0d34-0410-b5e6-96231b3b80d8
This is a fix for PR21709 ( http://llvm.org/bugs/show_bug.cgi?id=21709 ).
When we have 2 consecutive 16-byte loads that are merged into one 32-byte vector,
we can use a single 32-byte load instead.
But we don't do this for SandyBridge / IvyBridge because they have slower 32-byte memops.
We also don't bother using 32-byte *integer* loads on a machine that only has AVX1 (btver2)
because those operands would have to be split in half anyway since there is no support for
32-byte integer math ops.
Differential Revision: http://reviews.llvm.org/D6492
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224344 91177308-0d34-0410-b5e6-96231b3b80d8
According to AVX specification:
"Most arithmetic and data processing instructions encoded using the VEX prefix and
performing memory accesses have more flexible memory alignment requirements
than instructions that are encoded without the VEX prefix. Specifically,
With the exception of explicitly aligned 16 or 32 byte SIMD load/store instructions,
most VEX-encoded, arithmetic and data processing instructions operate in
a flexible environment regarding memory address alignment, i.e. VEX-encoded
instruction with 32-byte or 16-byte load semantics will support unaligned load
operation by default. Memory arguments for most instructions with VEX prefix
operate normally without causing #GP(0) on any byte-granularity alignment
(unlike Legacy SSE instructions)."
The same for AVX-512.
This change does not affect anything right now, because only the "memop pattern fragment"
depends on FeatureVectorUAMem and it is not used in AVX patterns.
All AVX patterns are based on the "unaligned load" anyway.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224330 91177308-0d34-0410-b5e6-96231b3b80d8
It's horrible to inspect `MDNode`s in a debugger. All of their operands
that are `MDNode`s get dumped as `<badref>`, since we can't assign
metadata slots in the context of a `Metadata::dump()`. (Why not? Why
not assign numbers lazily? Because then each time you called `dump()`,
a given `MDNode` could have a different lazily assigned number.)
Fortunately, the C memory model gives us perfectly good identifiers for
`MDNode`. Add pointer addresses to the dumps, transforming this:
(lldb) e N->dump()
!{i32 662302, i32 26, <badref>, null}
(lldb) e ((MDNode*)N->getOperand(2))->dump()
!{i32 4, !"foo"}
into:
(lldb) e N->dump()
!{i32 662302, i32 26, <0x100706ee0>, null}
(lldb) e ((MDNode*)0x100706ee0)->dump()
!{i32 4, !"foo"}
and this:
(lldb) e N->dump()
0x101200248 = !{<badref>, <badref>, <badref>, <badref>, <badref>}
(lldb) e N->getOperand(0)
(const llvm::MDOperand) $0 = {
MD = 0x00000001012004e0
}
(lldb) e N->getOperand(1)
(const llvm::MDOperand) $1 = {
MD = 0x00000001012004e0
}
(lldb) e N->getOperand(2)
(const llvm::MDOperand) $2 = {
MD = 0x0000000101200058
}
(lldb) e N->getOperand(3)
(const llvm::MDOperand) $3 = {
MD = 0x00000001012004e0
}
(lldb) e N->getOperand(4)
(const llvm::MDOperand) $4 = {
MD = 0x0000000101200058
}
(lldb) e ((MDNode*)0x00000001012004e0)->dump()
!{}
(lldb) e ((MDNode*)0x0000000101200058)->dump()
!{null}
into:
(lldb) e N->dump()
!{<0x1012004e0>, <0x1012004e0>, <0x101200058>, <0x1012004e0>, <0x101200058>}
(lldb) e ((MDNode*)0x1012004e0)->dump()
!{}
(lldb) e ((MDNode*)0x101200058)->dump()
!{null}
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224325 91177308-0d34-0410-b5e6-96231b3b80d8
This test was missing a `Debug Info Version` so it's `not grep` was
passing vacuously. Update it to CHECK for something useful at the same
time so it doesn't bitrot quite so easily in the future.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224324 91177308-0d34-0410-b5e6-96231b3b80d8
The use of SP and PC in the register list for stores is deprecated on ARM
(ARM ARM A.8.8.199):
ARM deprecates the use of ARM instructions that include the SP or the PC in
the list.
Provide a deprecation warning from the assembler in the case that the syntax is
ever seen.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224319 91177308-0d34-0410-b5e6-96231b3b80d8
The PowerPC backend, somewhat embarrassingly, did not generate an
optimal-length sequence of instructions for a 32-bit bswap. While adding a
pattern for the bswap intrinsic to fix this would not have been terribly
difficult, doing so would not have addressed the real problem: we had been
generating poor code for many bit-permuting operations (by which I mean things
like byte-swap that permute the bits of one or more inputs around in various
ways). Here are some initial steps toward solving this deficiency.
Bit-permuting operations are represented, at the SDAG level, using ISD::ROTL,
SHL, SRL, AND and OR (mostly with constant second operands). Looking back
through these operations, we can build up a description of the bits in the
resulting value in terms of bits of one or more input values (and constant
zeros). For each bit, we compute the rotation amount from the original value,
and then group consecutive (value, rotation factor) bits into groups. Groups
sharing these attributes are then collected and sorted, and we can then
instruction select the entire permutation using a combination of masked
rotations (rlwinm), imm ands (andi/andis), and masked rotation inserts
(rlwimi).
The result is that instead of lowering an i32 bswap as:
rlwinm 5, 3, 24, 16, 23
rlwinm 4, 3, 24, 0, 7
rlwimi 4, 3, 8, 8, 15
rlwimi 5, 3, 8, 24, 31
rlwimi 4, 5, 0, 16, 31
we now produce:
rlwinm 4, 3, 8, 0, 31
rlwimi 4, 3, 24, 16, 23
rlwimi 4, 3, 24, 0, 7
and for the 'test6' example in the PowerPC/README.txt file:
unsigned test6(unsigned x) {
return ((x & 0x00FF0000) >> 16) | ((x & 0x000000FF) << 16);
}
we used to produce:
lis 4, 255
rlwinm 3, 3, 16, 0, 31
ori 4, 4, 255
and 3, 3, 4
and now we produce:
rlwinm 4, 3, 16, 24, 31
rlwimi 4, 3, 16, 8, 15
and, as a nice bonus, this fixes the FIXME in
test/CodeGen/PowerPC/rlwimi-and.ll.
This commit does not include instruction-selection for i64 operations, those
will come later.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224318 91177308-0d34-0410-b5e6-96231b3b80d8
This changes subrange calculation to calculate subranges sequentially
instead of in parallel. The code is easier to understand that way and
addresses the code review issues raised about LiveOutData being
hard to understand/needing more comments by removing them :)
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224313 91177308-0d34-0410-b5e6-96231b3b80d8
Debug info marks the first instruction without the FrameSetup flag
as being the end of the function prologue. Any CFI instructions in the
middle of the function prologue would cause debug info to end the prologue
too early and worse, attach the line number of the CFI instruction, which
incidentally is often 0.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224294 91177308-0d34-0410-b5e6-96231b3b80d8
isKnownPredicate.
The motivation for this change is to optimize away checks in loops
like this:
limit = min(t, len)
for (i = 0 to limit)
if (i >= len || i < 0) throw_array_of_of_bounds();
a[i] = ...
Differential Revision: http://reviews.llvm.org/D6635
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224285 91177308-0d34-0410-b5e6-96231b3b80d8
Summary: x86 allows either ordering for the LOCK and DATA16 prefixes, but using GCC+GAS leads to different code generation than using LLVM. This change matches the order that GAS emits the x86 prefixes when a semicolon isn't used in inline assembly (see tc-i386.c comment before define LOCK_PREFIX), and helps simplify tooling that operates on the instruction's byte sequence (such as NaCl's validator). This change shouldn't have any performance impact.
Test Plan: ninja check
Reviewers: craig.topper, jvoung
Subscribers: jfb, llvm-commits
Differential Revision: http://reviews.llvm.org/D6630
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224283 91177308-0d34-0410-b5e6-96231b3b80d8
r223763 was made to work around a temporary issue where a user of the
JIT was passing down a declaration (incorrectly). This shouldn't
occur, so assert rather than silently continue.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224277 91177308-0d34-0410-b5e6-96231b3b80d8
This changes subrange calculation to calculate subranges sequentially
instead of in parallel. The code is easier to understand that way and
addresses the code review issues raised about LiveOutData being
hard to understand/needing more comments by removing them :)
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@224272 91177308-0d34-0410-b5e6-96231b3b80d8
