Globals, function arguments, and results are special cases in
registerization.
Globals must be flushed aggressively, because nearly any
operation can cause a panic, and the recovery code must see
the latest values. Globals also must be loaded aggressively,
because nearly any store through a pointer might be updating a
global: the compiler cannot see all the "address of"
operations on globals, especially exported globals. To
accomplish this, mark all globals as having their address
taken, which effectively disables registerization.
If a function contains a defer statement, the function results
must be flushed aggressively, because nearly any operation can
cause a panic, and the deferred code may call recover, causing
the original function to return the current values of its
function results. To accomplish this, mark all function
results as having their address taken if the function contains
any defer statements. This causes not just aggressive flushing
but also aggressive loading. The aggressive loading is
overkill but the best we can do in the current code.
Function arguments must be considered live at all safe points
in a function, because garbage collection always preserves
them: they must be up-to-date in order to be preserved
correctly. Accomplish this by marking them live at all call
sites. An earlier attempt at this marked function arguments as
having their address taken, which disabled registerization
completely, making programs slower. This CL's solution allows
registerization while preserving safety. The benchmark speedup
is caused by being able to registerize again (the earlier CL
lost the same amount).
benchmark old ns/op new ns/op delta
BenchmarkEqualPort32 61.4 56.0 -8.79%
benchmark old MB/s new MB/s speedup
BenchmarkEqualPort32 521.56 570.97 1.09x
Fixes#1304. (again)
Fixes#7944. (again)
Fixes#7984.
Fixes#7995.
LGTM=khr
R=golang-codereviews, khr
CC=golang-codereviews, iant, r
https://golang.org/cl/97500044
The inputs to a function are marked live at all times in the
liveness bitmaps, so that the garbage collector will not free
the things they point at and reuse the pointers, so that the
pointers shown in stack traces are guaranteed not to have
been recycled.
Unfortunately, no one told the register optimizer that the
inputs need to be preserved at all call sites. If a function
is done with a particular input value, the optimizer will stop
preserving it across calls. For single-word values this just
means that the value recorded might be stale. For multi-word
values like slices, the value recorded could be only partially stale:
it can happen that, say, the cap was updated but not the len,
or that the len was updated but not the base pointer.
Either of these possibilities (and others) would make the
garbage collector misinterpret memory, leading to memory
corruption.
This came up in a real program, in which the garbage collector's
'slice len ≤ slice cap' check caught the inconsistency.
Fixes#7944.
LGTM=iant
R=golang-codereviews, iant
CC=golang-codereviews, khr
https://golang.org/cl/100370045
This is joint work with Daniel Morsing.
In order for the register allocator to alias two variables, they must have the same width, stack offset, and etype. Code generation was altering a variable's etype in a few places. This prevented the variable from being moved to a register, which in turn prevented peephole optimization. This failure to alias was very common, with almost 23,000 instances just running make.bash.
This phenomenon was not visible in the register allocation debug output because the variables that failed to alias had the same name. The debugging-only change to bits.c fixes this by printing the variable number with its name.
This CL fixes the source of all etype mismatches for 6g, all but one case for 8g, and depressingly few cases for 5g. (I believe that extending CL 6819083 to 5g is a prerequisite.) Fixing the remaining cases in 8g and 5g is work for the future.
The etype mismatch fixes are:
* [gc] Slicing changed the type of the base pointer into a uintptr in order to perform arithmetic on it. Instead, support addition directly on pointers.
* [*g] OSPTR was giving type uintptr to slice base pointers; undo that. This arose, for example, while compiling copy(dst, src).
* [8g] 64 bit float conversion was assigning int64 type during codegen, overwriting the existing uint64 type.
Note that some etype mismatches are appropriate, such as a struct with a single field or an array with a single element.
With these fixes, the number of registerizations that occur while running make.bash for 6g increases ~10%. Hello world binary size shrinks ~1.5%. Running all benchmarks in the standard library show performance improvements ranging from nominal to substantive (>10%); a full comparison using 6g on my laptop is available at https://gist.github.com/josharian/8f9b5beb46667c272064. The microbenchmarks must be taken with a grain of salt; see issue 7920. The few benchmarks that show real regressions are likely due to issue 7920. I manually examined the generated code for the top few regressions and none had any assembly output changes. The few benchmarks that show extraordinary improvements are likely also due to issue 7920.
Performance results from 8g appear similar to 6g.
5g shows no performance improvements. This is not surprising, given the discussion above.
Update #7316
LGTM=rsc
R=rsc, daniel.morsing, bradfitz
CC=dave, golang-codereviews
https://golang.org/cl/91850043
Before we used line 1 of the first source file.
This should be clearer.
Fixes#4388.
LGTM=iant
R=golang-codereviews, iant
CC=golang-codereviews
https://golang.org/cl/92250044
If the ... element type contained no pointers,
then the escape analysis did not track the ... itself.
This manifested in an escaping ...byte being treated
as non-escaping.
Fixes#7934.
LGTM=iant
R=golang-codereviews, iant
CC=golang-codereviews
https://golang.org/cl/100310043
The register allocator decides which variables should be placed into registers by charging for each load/store and crediting for each use, and then selecting an allocation with minimal cost. NOPs will be eliminated, however, so using a variable in a NOP should not generate credit.
Issue 7867 arises from attempted registerization of multi-word variables because they are used in NOPs. By not crediting for that use, they will no longer be considered for registerization.
This fix could theoretically lead to better register allocation, but NOPs are rare relative to other instructions.
Fixes#7867.
LGTM=rsc
R=rsc
CC=golang-codereviews
https://golang.org/cl/94810044
Variables declared with 'var' have no sym->def.
Fixes#7794.
LGTM=rsc
R=golang-codereviews, bradfitz, rsc
CC=golang-codereviews
https://golang.org/cl/88360043
The new code is adapted from the Go 1.2 nosplit code,
but it does not have the bug reported in issue 7623:
g% go run nosplit.go
g% go1.2 run nosplit.go
BUG
rejected incorrectly:
main 0 call f; f 120
linker output:
# _/tmp/go-test-nosplit021064539
main.main: nosplit stack overflow
120 guaranteed after split check in main.main
112 on entry to main.f
-8 after main.f uses 120
g%
Fixes#6931.
Fixes#7623.
LGTM=iant
R=golang-codereviews, iant, ality
CC=golang-codereviews, r
https://golang.org/cl/88190043
Trying to make GODEBUG=gcdead=1 work with liveness
and in particular ambiguously live variables.
1. In the liveness computation, mark all ambiguously live
variables as live for the entire function, except the entry.
They are zeroed directly after entry, and we need them not
to be poisoned thereafter.
2. In the liveness computation, compute liveness (and deadness)
for all parameters, not just pointer-containing parameters.
Otherwise gcdead poisons untracked scalar parameters and results.
3. Fix liveness debugging print for -live=2 to use correct bitmaps.
(Was not updated for compaction during compaction CL.)
4. Correct varkill during map literal initialization.
Was killing the map itself instead of the inserted value temp.
5. Disable aggressive varkill cleanup for call arguments if
the call appears in a defer or go statement.
6. In the garbage collector, avoid bug scanning empty
strings. An empty string is two zeros. The multiword
code only looked at the first zero and then interpreted
the next two bits in the bitmap as an ordinary word bitmap.
For a string the bits are 11 00, so if a live string was zero
length with a 0 base pointer, the poisoning code treated
the length as an ordinary word with code 00, meaning it
needed poisoning, turning the string into a poison-length
string with base pointer 0. By the same logic I believe that
a live nil slice (bits 11 01 00) will have its cap poisoned.
Always scan full multiword struct.
7. In the runtime, treat both poison words (PoisonGC and
PoisonStack) as invalid pointers that warrant crashes.
Manual testing as follows:
- Create a script called gcdead on your PATH containing:
#!/bin/bash
GODEBUG=gcdead=1 GOGC=10 GOTRACEBACK=2 exec "$@"
- Now you can build a test and then run 'gcdead ./foo.test'.
- More importantly, you can run 'go test -short -exec gcdead std'
to run all the tests.
Fixes#7676.
While here, enable the precise scanning of slices, since that was
disabled due to bugs like these. That now works, both with and
without gcdead.
Fixes#7549.
LGTM=khr
R=khr
CC=golang-codereviews
https://golang.org/cl/83410044
1. Use n->alloc, not n->left, to hold the allocated temp being
passed from orderstmt/orderexpr to walk.
2. Treat method values the same as closures.
3. Use killed temporary for composite literal passed to
non-escaping function argument.
4. Clean temporaries promptly in if and for statements.
5. Clean temporaries promptly in select statements.
As part of this, move all the temporary-generating logic
out of select.c into order.c, so that the temporaries can
be reclaimed.
With the new temporaries, can re-enable the 1-entry
select optimization. Fixes issue 7672.
While we're here, fix a 1-line bug in select processing
turned up by the new liveness test (but unrelated; select.c:72).
Fixes#7686.
6. Clean temporaries (but not particularly promptly) in switch
and range statements.
7. Clean temporary used during convT2E/convT2I.
8. Clean temporaries promptly during && and || expressions.
---
CL 81940043 reduced the number of ambiguously live temps
in the godoc binary from 860 to 711.
CL 83090046 reduced the number from 711 to 121.
This CL reduces the number from 121 to 23.
15 the 23 that remain are in fact ambiguously live.
The final 8 could be fixed but are not trivial and
not common enough to warrant work at this point
in the release cycle.
These numbers only count ambiguously live temps,
not ambiguously live user-declared variables.
There are 18 such variables in the godoc binary after this CL,
so a total of 41 ambiguously live temps or user-declared
variables.
The net effect is that zeroing anything on entry to a function
should now be a rare event, whereas earlier it was the
common case.
This is good enough for Go 1.3, and probably good
enough for future releases too.
Fixes#7345.
LGTM=khr
R=khr
CC=golang-codereviews
https://golang.org/cl/83000048
1. In functions with heap-allocated result variables or with
defer statements, the return sequence requires more than
just a single RET instruction. There is an optimization that
arranges for all returns to jump to a single copy of the return
epilogue in this case. Unfortunately, that optimization is
fundamentally incompatible with PC-based liveness information:
it takes PCs at many different points in the function and makes
them all land at one PC, making the combined liveness information
at that target PC a mess. Disable this optimization, so that each
return site gets its own copy of the 'call deferreturn' and the
copying of result variables back from the heap.
This removes quite a few spurious 'ambiguously live' variables.
2. Let orderexpr allocate temporaries that are passed by address
to a function call and then die on return, so that we can arrange
an appropriate VARKILL.
2a. Do this for ... slices.
2b. Do this for closure structs.
2c. Do this for runtime.concatstring, which is the implementation
of large string additions. Change representation of OADDSTR to
an explicit list in typecheck to avoid reconstructing list in both
walk and order.
3. Let orderexpr allocate the temporary variable copies used for
range loops, so that they can be killed when the loop is over.
Similarly, let it allocate the temporary holding the map iterator.
CL 81940043 reduced the number of ambiguously live temps
in the godoc binary from 860 to 711.
This CL reduces the number to 121. Still more to do, but another
good checkpoint.
Update #7345
LGTM=khr
R=khr
CC=golang-codereviews
https://golang.org/cl/83090046
The new channel and map runtime routines take pointers
to values, typically temporaries. Without help, the compiler
cannot tell when those temporaries stop being needed,
because it isn't sure what happened to the pointer.
Arrange to insert explicit VARKILL instructions for these
temporaries so that the liveness analysis can avoid seeing
them as "ambiguously live".
The change is made in order.c, which was already in charge of
introducing temporaries to preserve the order-of-evaluation
guarantees. Now its job has expanded to include introducing
temporaries as needed by runtime routines, and then also
inserting the VARKILL annotations for all these temporaries,
so that their lifetimes can be shortened.
In order to do its job for the map runtime routines, order.c arranges
that all map lookups or map assignments have the form:
x = m[k]
x, y = m[k]
m[k] = x
where x, y, and k are simple variables (often temporaries).
Likewise, receiving from a channel is now always:
x = <-c
In order to provide the map guarantee, order.c is responsible for
rewriting x op= y into x = x op y, so that m[k] += z becomes
t = m[k]
t2 = t + z
m[k] = t2
While here, fix a few bugs in order.c's traversal: it was failing to
walk into select and switch case bodies, so order of evaluation
guarantees were not preserved in those situations.
Added tests to test/reorder2.go.
Fixes#7671.
In gc/popt's temporary-merging optimization, allow merging
of temporaries with their address taken as long as the liveness
ranges do not intersect. (There is a good chance of that now
that we have VARKILL annotations to limit the liveness range.)
Explicitly killing temporaries cuts the number of ambiguously
live temporaries that must be zeroed in the godoc binary from
860 to 711, or -17%. There is more work to be done, but this
is a good checkpoint.
Update #7345
LGTM=khr
R=khr
CC=golang-codereviews
https://golang.org/cl/81940043
1. On entry to a function, only zero the ambiguously live stack variables.
Before, we were zeroing all stack variables containing pointers.
The zeroing is pretty inefficient right now (issue 7624), but there are also
too many stack variables detected as ambiguously live (issue 7345),
and that must be addressed before deciding how to improve the zeroing code.
(Changes in 5g/ggen.c, 6g/ggen.c, 8g/ggen.c, gc/pgen.c)
Fixes#7647.
2. Make the regopt word-based liveness analysis preserve the
whole-variable liveness property expected by the garbage collection
bitmap liveness analysis. That is, if the regopt liveness decides that
one word in a struct needs to be preserved, make sure it preserves
the entire struct. This is particularly important for multiword values
such as strings, slices, and interfaces, in which all the words need
to be present in order to understand the meaning.
(Changes in 5g/reg.c, 6g/reg.c, 8g/reg.c.)
Fixes#7591.
3. Make the regopt word-based liveness analysis treat a variable
as having its address taken - which makes it preserved across
all future calls - whenever n->addrtaken is set, for consistency
with the gc bitmap liveness analysis, even if there is no machine
instruction actually taking the address. In this case n->addrtaken
is incorrect (a nicer way to put it is overconservative), and ideally
there would be no such cases, but they can happen and the two
analyses need to agree.
(Changes in 5g/reg.c, 6g/reg.c, 8g/reg.c; test in bug484.go.)
Fixes crashes found by turning off "zero everything" in step 1.
4. Remove spurious VARDEF annotations. As the comment in
gc/pgen.c explains, the VARDEF must immediately precede
the initialization. It cannot be too early, and it cannot be too late.
In particular, if a function call sits between the VARDEF and the
actual machine instructions doing the initialization, the variable
will be treated as live during that function call even though it is
uninitialized, leading to problems.
(Changes in gc/gen.c; test in live.go.)
Fixes crashes found by turning off "zero everything" in step 1.
5. Do not treat loading the address of a wide value as a signal
that the value must be initialized. Instead depend on the existence
of a VARDEF or the first actual read/write of a word in the value.
If the load is in order to pass the address to a function that does
the actual initialization, treating the load as an implicit VARDEF
causes the same problems as described in step 4.
The alternative is to arrange to zero every such value before
passing it to the real initialization function, but this is a much
easier and more efficient change.
(Changes in gc/plive.c.)
Fixes crashes found by turning off "zero everything" in step 1.
6. Treat wide input parameters with their address taken as
initialized on entry to the function. Otherwise they look
"ambiguously live" and we will try to emit code to zero them.
(Changes in gc/plive.c.)
Fixes crashes found by turning off "zero everything" in step 1.
7. An array of length 0 has no pointers, even if the element type does.
Without this change, the zeroing code complains when asked to
clear a 0-length array.
(Changes in gc/reflect.c.)
LGTM=khr
R=khr
CC=golang-codereviews
https://golang.org/cl/80160044
Revision 3ae4607a43ff introduced CONVNOP layers
to fix type checking issues arising from comparisons.
The added complexity made 8g run out of registers
when compiling an equality function in go.net/ipv6.
A similar issue occurred in test/sizeof.go on
amd64p32 with 6g.
Fixes#7405.
LGTM=khr
R=rsc, dave, iant, khr
CC=golang-codereviews
https://golang.org/cl/78100044
A too large float constant is an error.
A too small float constant is rounded to zero.
Fixes#7419
Update #6902
LGTM=iant
R=golang-codereviews, iant
CC=golang-codereviews
https://golang.org/cl/76730046
The lowering to runtime calls introduces hidden pointers to the
arguments of select clauses. When implicit conversions were
involved it could end up with incompatible pointers. Since the
pointed-to types have the same representation, we can introduce a
forced conversion.
Fixes#6847.
LGTM=rsc
R=rsc, iant, khr
CC=golang-codereviews
https://golang.org/cl/72380043
The garbage collector uses type information to guide the
traversal of the heap. If it sees a field that should be a string,
it marks the object pointed at by the string data pointer as
visited but does not bother to look at the data, because
strings contain bytes, not pointers.
If you save s[len(s):] somewhere, though, the string data pointer
actually points just beyond the string data; if the string data
were exactly the size of an allocated block, the string data
pointer would actually point at the next block. It is incorrect
to mark that next block as visited and not bother to look at
the data, because the next block may be some other type
entirely.
The fix is to ignore strings with zero length during collection:
they are empty and can never become non-empty: the base
pointer will never be used again. The handling of slices already
does this (but using cap instead of len).
This was not a bug in Go 1.2, because until January all string
allocations included a trailing NUL byte not included in the
length, so s[len(s):] still pointed inside the string allocation
(at the NUL).
This bug was causing the crashes in test/run.go. Specifically,
the parsing of a regexp in package regexp/syntax allocated a
[]syntax.Inst with rounded size 1152 bytes. In fact it
allocated many such slices, because during the processing of
test/index2.go it creates thousands of regexps that are all
approximately the same complexity. That takes a long time, and
test/run works on other tests in other goroutines. One such
other test is chan/perm.go, which uses an 1152-byte source
file. test/run reads that file into a []byte and then calls
strings.Split(string(src), "\n"). The string(src) creates an
1152-byte string - and there's a very good chance of it
landing next to one of the many many regexp slices already
allocated - and then because the file ends in a \n,
strings.Split records the tail empty string as the final
element in the slice. A garbage collection happens at this
point, the collection finds that string before encountering
the []syntax.Inst data it now inadvertently points to, and the
[]syntax.Inst data is not scanned for the pointers that it
contains. Each syntax.Inst contains a []rune, those are
missed, and the backing rune arrays are freed for reuse. When
the regexp is later executed, the runes being searched for are
no longer runes at all, and there is no match, even on text
that should match.
On 64-bit machines the pointer in the []rune inside the
syntax.Inst is larger (along with a few other pointers),
pushing the []syntax.Inst backing array into a larger size
class, avoiding the collision with chan/perm.go's
inadvertently sized file.
I expect this was more prevalent on OS X than on Linux or
Windows because those managed to run faster or slower and
didn't overlap index2.go with chan/perm.go as often. On the
ARM systems, we only run one errorcheck test at a time, so
index2 and chan/perm would never overlap.
It is possible that this bug is the root cause of other crashes
as well. For now we only know it is the cause of the test/run crash.
Many thanks to Dmitriy for help debugging.
Fixes#7344.
Fixes#7455.
LGTM=r, dvyukov, dave, iant
R=golang-codereviews, dave, r, dvyukov, delpontej, iant
CC=golang-codereviews, khr
https://golang.org/cl/74250043
Some of the errorcheck tests have many many identical regexps.
Use a map to avoid storing the compiled form many many times
in memory. Change the filterRe to a simple string to avoid
the expense of those regexps as well.
Cuts the time for run.go on index2.go by almost 50x.
Noticed during debugging of issue 7344.
LGTM=bradfitz
R=bradfitz, josharian
CC=golang-codereviews
https://golang.org/cl/74380043
The byte that r is or'd into is already 0x7, so the failure to zero r only
impacts the generated machine code if the register is > 7.
Fixes#7044.
LGTM=dave, minux.ma, rsc
R=dave, minux.ma, bradfitz, rsc
CC=golang-codereviews
https://golang.org/cl/73730043
The cached computed interface tables are indexed by the interface
types, not by the unnamed underlying interfaces
To preserve the invariants expected by interface comparison, an
itab generated for an interface type must not be used for a value
of a different interface type even if the representation is identical.
Fixes#7207.
LGTM=rsc
R=rsc, iant, khr
CC=golang-codereviews
https://golang.org/cl/69210044
