Currently the prove pass uses implication queries. For each block, it
collects the set of branch conditions leading to that block, and
queries this fact table for whether any of these facts imply the
block's own branch condition (or its inverse). This works remarkably
well considering it doesn't do any deduction on these facts, but it
has various downsides:
1. It requires an implementation both of adding facts to the table and
determining implications. These are very nearly duals of each
other, but require separate implementations. Likewise, the process
of asserting facts of dominating branch conditions is very nearly
the dual of the process of querying implied branch conditions.
2. It leads to less effective use of derived facts. For example, the
prove pass currently derives facts about the relations between len
and cap, but can't make use of these unless a branch condition is
in the exact form of a derived fact. If one of these derived facts
contradicts another fact, it won't notice or make use of this.
This CL changes the approach of the prove pass to instead use
*contradiction* instead of implication. Rather than ever querying a
branch condition, it simply adds branch conditions to the fact table.
If this leads to a contradiction (specifically, it makes the fact set
unsatisfiable), that branch is impossible and can be cut. As a result,
1. We can eliminate the code for determining implications
(factsTable.get disappears entirely). Also, there is now a single
implementation of visiting and asserting branch conditions, since
we don't have to flip them around to treat them as facts in one
place and queries in another.
2. Derived facts can be used effectively. It doesn't matter *why* the
fact table is unsatisfiable; a contradiction in any of the facts is
enough.
3. As an added benefit, it's now quite easy to avoid traversing beyond
provably-unreachable blocks. In contrast, the current
implementation always visits all blocks.
The prove pass already has nearly all of the mechanism necessary to
compute unsatisfiability, which means this both simplifies the code
and makes it more powerful.
The only complication is that the current implication procedure has a
hack for dealing with the 0 <= Args[0] condition of OpIsInBounds and
OpIsSliceInBounds. We replace this with asserting the appropriate fact
when we process one of these conditions. This seems much cleaner
anyway, and works because we can now take advantage of derived facts.
This has no measurable effect on compiler performance.
Effectiveness:
There is exactly one condition in all of std and cmd that this fails
to prove that the old implementation could: (int64(^uint(0)>>1) < x)
in encoding/gob. This can never be true because x is an int, and it's
basically coincidence that the old code gets this. (For example, it
fails to prove the similar (x < ^int64(^uint(0)>>1)) condition that
immediately precedes it, and even though the conditions are logically
unrelated, it wouldn't get the second one if it hadn't first processed
the first!)
It does, however, prove a few dozen additional branches. These come
from facts that are added to the fact table about the relations
between len and cap. These were almost never queried directly before,
but could lead to contradictions, which the unsat-based approach is
able to use.
There are exactly two branches in std and cmd that this implementation
proves in the *other* direction. This sounds scary, but is okay
because both occur in already-unreachable blocks, so it doesn't matter
what we chose. Because the fact table logic is sound but incomplete,
it fails to prove that the block isn't reachable, even though it is
able to prove that both outgoing branches are impossible. We could
turn these blocks into BlockExit blocks, but it doesn't seem worth the
trouble of the extra proof effort for something that happens twice in
all of std and cmd.
Tests:
This CL updates test/prove.go to change the expected messages because
it can no longer give a "reason" why it proved or disproved a
condition. It also adds a new test of a branch it couldn't prove
before.
It mostly guts test/sliceopt.go, removing everything related to slice
bounds optimizations and moving a few relevant tests to test/prove.go.
Much of this test is actually unreachable. The new prove pass figures
this out and doesn't try to prove anything about the unreachable
parts. The output on the unreachable parts is already suspect because
anything can be proved at that point, so it's really just a regression
test for an algorithm the compiler no longer uses.
This is a step toward fixing #23354. That issue is quite easy to fix
once we can use derived facts effectively.
Change-Id: Ia48a1b9ee081310579fe474e4a61857424ff8ce8
Reviewed-on: https://go-review.googlesource.com/87478
Reviewed-by: Keith Randall <khr@golang.org>
When we do
var x []byte = ...
y := x[i:]
We can't just use y.ptr = x.ptr + i, as the new pointer may point to the
next object in memory after the backing array.
We used to fix this by doing:
y.cap = x.cap - i
delta := i
if y.cap == 0 {
delta = 0
}
y.ptr = x.ptr + delta
That generates a branch in what is otherwise straight-line code.
Better to do:
y.cap = x.cap - i
mask := (y.cap - 1) >> 63 // -1 if y.cap==0, 0 otherwise
y.ptr = x.ptr + i &^ mask
It's about the same number of instructions (~4, depending on what
parts are constant, and the target architecture), but it is all
inline. It plays nicely with CSE, and the mask can be computed
in parallel with the index (in cases where a multiply is required).
It is a minor win in both speed and space.
Change-Id: Ied60465a0b8abb683c02208402e5bb7ac0e8370f
Reviewed-on: https://go-review.googlesource.com/32022
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Adapt old test for prove's bounds check elimination.
Added missing rule to generic rules that lead to differences
between 32 and 64 bit platforms on sliceopt test.
Added debugging to prove.go that was helpful-to-necessary to
discover that missing rule.
Lowered debugging level on prove.go from 3 to 1; no idea
why it was previously 3.
Change-Id: I09de206aeb2fced9f2796efe2bfd4a59927eda0c
Reviewed-on: https://go-review.googlesource.com/23290
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
It tests the behavior of the old deleted compiler.
Fixes#17362.
Change-Id: Ia2fdec734c5cbe724a9de562ed71598f67244ab3
Reviewed-on: https://go-review.googlesource.com/30593
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
The new SSA backend modifies the ABI slightly: R0 is now a usable
general purpose register.
Fixes#16677.
Change-Id: I367435ce921e0c7e79e021c80cf8ef5d1d1466cf
Reviewed-on: https://go-review.googlesource.com/28978
Run-TryBot: Michael Munday <munday@ca.ibm.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Add the following optimizations:
- fold constants
- fold address into load/store
- simplify extensions and conditional branches
- remove nil checks
Turn on SSA on MIPS64 by default, and toggle the tests.
Fixes#16359.
Change-Id: I7f1e38c2509e22e42cd024e712990ebbe47176bd
Reviewed-on: https://go-review.googlesource.com/27870
Run-TryBot: Cherry Zhang <cherryyz@google.com>
Reviewed-by: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
This time with the cherry-pick from the proper patch of
the old CL.
Stack size increased.
Corrected NaN-comparison glitches.
Marked g register as clobbered by calls.
Fixed shared libraries.
live_ssa.go still disabled because of differences.
Presumably turning on more optimization will fix
both the stack size and the live_ssa.go glitches.
Enhanced debugging output for shared libs test.
Rebased onto master.
Updates #16010.
Change-Id: I40864faf1ef32c118fb141b7ef8e854498e6b2c4
Reviewed-on: https://go-review.googlesource.com/27159
Run-TryBot: David Chase <drchase@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Add more ARM64 optimizations:
- use hardware zero register when it is possible.
- use shifted ops.
The assembler supports shifted ops but not documented, nor knows
how to print it. This CL adds them.
- enable fast division.
This was disabled because it makes the old backend generate slower
code. But with SSA it generates faster code.
Turn on SSA by default, also adjust tests.
Change-Id: I7794479954c83bb65008dcb457bc1e21d7496da6
Reviewed-on: https://go-review.googlesource.com/26950
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
Last part of the 386 SSA port.
Modify the x86 backend to simulate SSE registers and
instructions with 387 registers and instructions.
The simulation isn't terribly performant, but it works,
and the old implementation wasn't very performant either.
Leaving to people who care about 387 to optimize if they want.
Turn on SSA backend for 386 by default.
Fixes#16358
Change-Id: I678fb59132620b2c47e993c1c10c4c21135f70c0
Reviewed-on: https://go-review.googlesource.com/25271
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
It's not a new backend, just a PtrSize==4 modification
of the existing AMD64 backend.
Change-Id: Icc63521a5cf4ebb379f7430ef3f070894c09afda
Reviewed-on: https://go-review.googlesource.com/25586
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>
NaCl code runs in sandbox and there are restrictions for its
instruction uses
(https://developer.chrome.com/native-client/reference/sandbox_internals/arm-32-bit-sandbox).
Like the legacy backend, on NaCl,
- don't use R9, which is used as NaCl's "thread pointer".
- don't use Duff's device.
- don't use indexed load/stores.
- the assembler rewrites DIV/MOD to runtime calls, which on NaCl
clobbers R12, so R12 is marked as clobbered for DIV/MOD.
- other restrictions are satisfied by the assembler.
Enable SSA specific tests on nacl/arm, and disable non-SSA ones.
Updates #15365.
Change-Id: I9262693ec6756b89ca29d3ae4e52a96fe5403b02
Reviewed-on: https://go-review.googlesource.com/24859
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
As Josh mentioned in CL 24716, there has been requests for using SSA
for ARM. SSA can still be disabled by setting -ssa=0 for cmd/compile,
or partially enabled with GOSSAFUNC, GOSSAPKG, and GOSSAHASH.
Not enable SSA by default on NaCl, which is not supported yet.
Enable SSA-specific tests on ARM: live_ssa.go and nilptr3_ssa.go;
disable non-SSA tests: live.go, nilptr3.go, and slicepot.go.
Updates #15365.
Change-Id: Ic2ca8d166aeca8517b9d262a55e92f2130683a16
Reviewed-on: https://go-review.googlesource.com/23953
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Josh Bleecher Snyder <josharian@gmail.com>
Reviewed-by: David Chase <drchase@google.com>
Some tests disabled, some bifurcated into _ssa and not,
with appropriate logging added to compiler.
"tests/live.go" in particular needs attention.
SSA-specific testing removed, since it's all SSA now.
Added "-run_skips" option to tests/run.go to simplify
checking whether a test still fails (or how it fails)
on a skipped platform.
The compiler now compiles with SSA by default.
If you don't want SSA, specify GOSSAHASH=n (or N) as
an environment variable. Function names ending in "_ssa"
are always SSA-compiled.
GOSSAFUNC=fname retains its "SSA for fname, log to ssa.html"
GOSSAPKG=pkg only has an effect when GOSSAHASH=n
GOSSAHASH=10101 etc retains its name-hash-matching behavior
for purposes of debugging.
See #13068
Change-Id: I8217bfeb34173533eaeb391b5f6935483c7d6b43
Reviewed-on: https://go-review.googlesource.com/16299
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: David Chase <drchase@google.com>
The code generated for x = append(x, v) is roughly:
t := x
if len(t)+1 > cap(t) {
t = grow(t)
}
t[len(t)] = v
len(t)++
x = t
We used to generate this code as Go pseudocode during walk.
Generate it instead as actual instructions during gen.
Doing so lets us apply a few optimizations. The most important
is that when, as in the above example, the source slice and the
destination slice are the same, the code can instead do:
t := x
if len(t)+1 > cap(t) {
t = grow(t)
x = {base(t), len(t)+1, cap(t)}
} else {
len(x)++
}
t[len(t)] = v
That is, in the fast path that does not reallocate the array,
only the updated length needs to be written back to x,
not the array pointer and not the capacity. This is more like
what you'd write by hand in C. It's faster in general, since
the fast path elides two of the three stores, but it's especially
faster when the form of x is such that the base pointer write
would turn into a write barrier. No write, no barrier.
name old mean new mean delta
BinaryTree17 5.68s × (0.97,1.04) 5.81s × (0.98,1.03) +2.35% (p=0.023)
Fannkuch11 4.41s × (0.98,1.03) 4.35s × (1.00,1.00) ~ (p=0.090)
FmtFprintfEmpty 92.7ns × (0.91,1.16) 86.0ns × (0.94,1.11) -7.31% (p=0.038)
FmtFprintfString 281ns × (0.96,1.08) 276ns × (0.98,1.04) ~ (p=0.219)
FmtFprintfInt 288ns × (0.97,1.06) 274ns × (0.98,1.06) -4.94% (p=0.002)
FmtFprintfIntInt 493ns × (0.97,1.04) 506ns × (0.99,1.01) +2.65% (p=0.009)
FmtFprintfPrefixedInt 423ns × (0.97,1.04) 391ns × (0.99,1.01) -7.52% (p=0.000)
FmtFprintfFloat 598ns × (0.99,1.01) 566ns × (0.99,1.01) -5.27% (p=0.000)
FmtManyArgs 1.89µs × (0.98,1.05) 1.91µs × (0.99,1.01) ~ (p=0.231)
GobDecode 14.8ms × (0.98,1.03) 15.3ms × (0.99,1.02) +3.01% (p=0.000)
GobEncode 12.3ms × (0.98,1.01) 11.5ms × (0.97,1.03) -5.93% (p=0.000)
Gzip 656ms × (0.99,1.05) 645ms × (0.99,1.01) ~ (p=0.055)
Gunzip 142ms × (1.00,1.00) 142ms × (1.00,1.00) -0.32% (p=0.034)
HTTPClientServer 91.2µs × (0.97,1.04) 90.5µs × (0.97,1.04) ~ (p=0.468)
JSONEncode 32.6ms × (0.97,1.08) 32.0ms × (0.98,1.03) ~ (p=0.190)
JSONDecode 114ms × (0.97,1.05) 114ms × (0.99,1.01) ~ (p=0.887)
Mandelbrot200 6.11ms × (0.98,1.04) 6.04ms × (1.00,1.01) ~ (p=0.167)
GoParse 6.66ms × (0.97,1.04) 6.47ms × (0.97,1.05) -2.81% (p=0.014)
RegexpMatchEasy0_32 159ns × (0.99,1.00) 171ns × (0.93,1.07) +7.19% (p=0.002)
RegexpMatchEasy0_1K 538ns × (1.00,1.01) 550ns × (0.98,1.01) +2.30% (p=0.000)
RegexpMatchEasy1_32 138ns × (1.00,1.00) 135ns × (0.99,1.02) -1.60% (p=0.000)
RegexpMatchEasy1_1K 869ns × (0.99,1.01) 879ns × (1.00,1.01) +1.08% (p=0.000)
RegexpMatchMedium_32 252ns × (0.99,1.01) 243ns × (1.00,1.00) -3.71% (p=0.000)
RegexpMatchMedium_1K 72.7µs × (1.00,1.00) 70.3µs × (1.00,1.00) -3.34% (p=0.000)
RegexpMatchHard_32 3.85µs × (1.00,1.00) 3.82µs × (1.00,1.01) -0.81% (p=0.000)
RegexpMatchHard_1K 118µs × (1.00,1.00) 117µs × (1.00,1.00) -0.56% (p=0.000)
Revcomp 920ms × (0.97,1.07) 917ms × (0.97,1.04) ~ (p=0.808)
Template 129ms × (0.98,1.03) 114ms × (0.99,1.01) -12.06% (p=0.000)
TimeParse 619ns × (0.99,1.01) 622ns × (0.99,1.01) ~ (p=0.062)
TimeFormat 661ns × (0.98,1.04) 665ns × (0.99,1.01) ~ (p=0.524)
See next CL for combination with a similar optimization for slice.
The benchmarks that are slower in this CL are still faster overall
with the combination of the two.
Change-Id: I2a7421658091b2488c64741b4db15ab6c3b4cb7e
Reviewed-on: https://go-review.googlesource.com/9812
Reviewed-by: David Chase <drchase@google.com>
