Mishandled the complex addressing mode in masks<>(SB)(CX*8)
as a casualty of the ARM work. Fix by backing all the flows up to
the state where registerIndirect is always called with the input
sitting on the opening paren.
With this, build passes for me with linux-arm, linux-386, and linux-amd64.
Change-Id: I7cae69a6fa9b635c79efd93850bd1e744b22bc79
Reviewed-on: https://go-review.googlesource.com/4964
Reviewed-by: Russ Cox <rsc@golang.org>
A consequence of the ARM work overlooked that SP is a real register
on x86, so we need to detect it specially.
This will be done better soon, but this is a fast fix for the build.
Change-Id: Ia30d111c3f42a5f0b5f4eddd4cc4d8b10470c14f
Reviewed-on: https://go-review.googlesource.com/4963
Reviewed-by: Russ Cox <rsc@golang.org>
The tools have been fixed to not do this, but verifyAsm depends on this
being fixed.
TBR=rsc
Change-Id: Ia8968cc803b3498dfa2f98188c6ed1cf2e11c66d
Reviewed-on: https://go-review.googlesource.com/4962
Reviewed-by: Rob Pike <r@golang.org>
There are many peculiarites of the ARM architecture that require work:
condition codes, new instructions, new instruction arg counts, and more.
Rewrite the parser to do a cleaner job, flowing left to right through the
sequence of elements of an operand.
Add ARM to arch.
Add ARM-specific details to the arch in a new file, internal/arch/arm.
These are probably better kept away from the "portable" asm. However
there are some pieces, like MRC, that are hard to disentangle. They
can be cleaned up later.
Change-Id: I8c06aedcf61f8a3960a406c094e168182d21b972
Reviewed-on: https://go-review.googlesource.com/4923
Reviewed-by: Russ Cox <rsc@golang.org>
Because text/scanner hides the spaces, the lexer treated
#define A(x)
and
#define A (x)
the same, but they are not: the first is an argument with macros, the
second is a simple one-word macro whose definition contains parentheses.
Fix this by noticing the relative column number as we move from A to (.
Hacky but simple.
Also add a helper to recognize the peculiar ARM shifted register operators.
Change-Id: I2cad22f5f1e11d8dad40ad13955793d178afb3ae
Reviewed-on: https://go-review.googlesource.com/4872
Reviewed-by: Russ Cox <rsc@golang.org>
Make cmd/ld a real library invoked by the individual linkers.
There are no reverse symbol references anymore
(symbols referred to in cmd/ld but defined in cmd/5l etc).
This means that in principle we could do an automatic
conversion of these to Go, as a stopgap until cmd/link is done
or as a replacement for cmd/link.
Change-Id: I4a94570257a3a7acc31601bfe0fad9dea0aea054
Reviewed-on: https://go-review.googlesource.com/4649
Reviewed-by: Rob Pike <r@golang.org>
- remove a few uses of ? :
- rename variables named len
- rewrite a few gotos as nested switches
- move goto targets to scope allowed by Go
- use consistent return type of anyregalloc
(was int or int32 in different places)
- remove unused nr variable in agen
- include proper headers in generated builtin1.c
- avoid strange sized %E formats (%-6E, %2E)
- change gengcmask argument from uint8[16] to uint8*
(diagnosed by c2go; not an array in any real sense).
- replace #ifdef XXX with comment block in 5g/peep.c
- expand and remove FAIL macro from 5g
- expand and remove noimpl macro from 9g
- print regalloc errors to stdout in 8g
(only use of fprint(2, ...) in all compilers)
Still producing bit-for-bit identical output.
Change-Id: Id46efcd2a89241082b234f63f375b66f2754d695
Reviewed-on: https://go-review.googlesource.com/4646
Reviewed-by: Austin Clements <austin@google.com>
In mparith, all the a1-- are problematic. Rewrite it all without pointers.
It's clearer anyway.
In popt, v is problematic because it is used both as a fixed pointer
(v = byvar[i]) and as a moving pointer (v = var; v++) aka slice.
Eliminate pointer movement.
Tested that this still produces bit-for-bit output for 'go build -a std'
compared to d260756 (current master).
Change-Id: I1a1bed0f98b594c3864fe95075dd95f9b52113e0
Reviewed-on: https://go-review.googlesource.com/4645
Reviewed-by: Austin Clements <austin@google.com>
Otherwise the exported variable collides with the type Arch.
While we're here, remove arch.dumpit (now in portable code)
and add arch.defframe (forgotten originally, somehow).
Change-Id: I1b3a7dd7e96c5f632dba7cd6c1217b42a2004d72
Reviewed-on: https://go-review.googlesource.com/4644
Reviewed-by: Austin Clements <austin@google.com>
If the Go source says x.y, and x is undefined, today we get
undefined: x
Change to:
undefined: x in x.y
Change-Id: I8ea95503bd469ea933c6bcbd675b7122a5d454f3
Reviewed-on: https://go-review.googlesource.com/4643
Reviewed-by: Austin Clements <austin@google.com>
Even with debugmerge = 1, the debugging output only happens
with the -v command-line flag. This is useful because it gets added
in automatically when debugging things like registerization with -R -v.
Change-Id: I9a5c7f562507b72e8e2fe2686fd07d069721345a
Reviewed-on: https://go-review.googlesource.com/4641
Reviewed-by: Austin Clements <austin@google.com>
Noticed last week.
Just saw a strange build failure in the revised rcmp (called by qsort on region)
and this fixed it.
Submitting first to avoid finding out which of my pending CLs tickled the
problem.
Change-Id: I4cafd611e2bf8e813e57ad0025e48bde5ae54359
Reviewed-on: https://go-review.googlesource.com/4830
Reviewed-by: Russ Cox <rsc@golang.org>
When the compiler echoes back an expression, it shows the
generated yacc expression. Change the generated code to
use a slice so that $3 shows up as yyDollar[3] in such messages.
Consider changing testdata/expr/expr.y to say:
$$.Sub(float64($1), $3)
(The float64 conversion is incorrect.)
Before:
expr.y:70[expr.go:486]: cannot convert exprS[exprpt - 2].num (type *big.Rat) to type float64
After:
expr.y:70[expr.go:492]: cannot convert exprDollar[1].num (type *big.Rat) to type float64
Change-Id: I74e494069df588e62299d1fccb282f3658d8f8f4
Reviewed-on: https://go-review.googlesource.com/4630
Reviewed-by: Rob Pike <r@golang.org>
Currently we always create context objects for closures that capture variables.
However, it is completely unnecessary for direct calls of closures
(whether it is func()(), defer func()() or go func()()).
This change transforms any OCALLFUNC(OCLOSURE) to normal function call.
Closed variables become function arguments.
This transformation is especially beneficial for go func(),
because we do not need to allocate context object on heap.
But it makes direct closure calls a bit faster as well (see BenchmarkClosureCall).
On implementation level it required to introduce yet another compiler pass.
However, the pass iterates only over xtop, so it should not be an issue.
Transformation consists of two parts: closure transformation and call site
transformation. We can't run these parts on different sides of escape analysis,
because tree state is inconsistent. We can do both parts during typecheck,
we don't know how to capture variables and don't have call site.
We can't do both parts during walk of OCALLFUNC, because we can walk
OCLOSURE body earlier.
So now capturevars pass only decides how to capture variables
(this info is required for escape analysis). New transformclosure
pass, that runs just before order/walk, does all transformations
of a closure. And later walk of OCALLFUNC(OCLOSURE) transforms call site.
benchmark old ns/op new ns/op delta
BenchmarkClosureCall 4.89 3.09 -36.81%
BenchmarkCreateGoroutinesCapture 1634 1294 -20.81%
benchmark old allocs new allocs delta
BenchmarkCreateGoroutinesCapture 6 2 -66.67%
benchmark old bytes new bytes delta
BenchmarkCreateGoroutinesCapture 176 48 -72.73%
Change-Id: Ic85e1706e18c3235cc45b3c0c031a9c1cdb7a40e
Reviewed-on: https://go-review.googlesource.com/4050
Reviewed-by: Russ Cox <rsc@golang.org>
The only remaining uses of four spaces instead of a tab is
when the line is too long (e.g. type Package).
Fixes#9809
Change-Id: Ifffd3639aa9264e795686ef1879a7686f182d2e5
Reviewed-on: https://go-review.googlesource.com/4182
Reviewed-by: Andrew Gerrand <adg@golang.org>
Using a zero register results in shorter, faster code.
5g already did this. Bring 6g, 8g, and 9g up to speed.
Reduces godoc binary size by 0.29% using 6g.
This CL includes cosmetic changes to 5g and 8g.
With those cosmetic changes included, componentgen is now
character-for-character equivalent across the four architectures.
Change-Id: I0e13dd48374bad830c725b117a1c86d4197d390c
Reviewed-on: https://go-review.googlesource.com/2606
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
Fix a flipped nil check.
The flipped check prevented componentgen
from zeroing a non-cadable nl.
This fix reduces the number of non-SB LEAQs
in godoc from 35323 to 34920 (-1.1%).
Update #1914
Change-Id: I15ea303068835f606f883ddf4a2bb4cb2287e9ae
Reviewed-on: https://go-review.googlesource.com/2605
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
Consider an interface value i of type I and concrete value c of type C.
Prior to this CL, i==c was evaluated as
I(c) == i
Evaluating I(c) can allocate.
This CL changes the evaluation of i==c to
x, ok := i.(C); ok && x == c
The new generated code is shorter and does not allocate directly.
If C is small, as it is in every instance in the stdlib,
the new code also uses less stack space
and makes one runtime call instead of two.
If C is very large, the original implementation is used.
The cutoff for "very large" is 1<<16,
following the stack vs heap cutoff used elsewhere.
This kind of comparison occurs in 38 places in the stdlib,
mostly in the net and os packages.
benchmark old ns/op new ns/op delta
BenchmarkEqEfaceConcrete 29.5 7.92 -73.15%
BenchmarkEqIfaceConcrete 32.1 7.90 -75.39%
BenchmarkNeEfaceConcrete 29.9 7.90 -73.58%
BenchmarkNeIfaceConcrete 35.9 7.90 -77.99%
Fixes#9370.
Change-Id: I7c4555950bcd6406ee5c613be1f2128da2c9a2b7
Reviewed-on: https://go-review.googlesource.com/2096
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
When compiling the stdlib most of the calls
to sgen are for exactly 2 or 3 words:
85% for 6g and 70% for 8g.
Special case them for performance.
This optimization is not relevant to 5g and 9g.
6g
benchmark old ns/op new ns/op delta
BenchmarkCopyFat16 3.25 0.82 -74.77%
BenchmarkCopyFat24 5.47 0.95 -82.63%
8g
benchmark old ns/op new ns/op delta
BenchmarkCopyFat8 3.84 2.42 -36.98%
BenchmarkCopyFat12 4.94 2.15 -56.48%
Change-Id: I8bc60b453f12597dfd916df2d072a7d5fc33ab85
Reviewed-on: https://go-review.googlesource.com/2607
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
When possible, generate nodl/nodr directly into DI/SI
rather than going through a temporary register.
CX has already been saved; use it during trailing bytes cleanup.
Change-Id: I4ec6209bcc5d3bfdc927c5c132009bd8d791ada3
Reviewed-on: https://go-review.googlesource.com/2608
Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Josh Bleecher Snyder <josharian@gmail.com>
Dump frames of functions.
Add function name and var width to output.
Change-Id: Ida06b8def96178fa550ca90836eb4a2509b9e13f
Reviewed-on: https://go-review.googlesource.com/3870
Reviewed-by: Russ Cox <rsc@golang.org>
typedslicecopy is another write barrier that is not
understood by racewalk. It seems quite complex to handle it
in the compiler, so instead just instrument it in runtime.
Update #9796
Change-Id: I0eb6abf3a2cd2491a338fab5f7da22f01bf7e89b
Reviewed-on: https://go-review.googlesource.com/4370
Reviewed-by: Russ Cox <rsc@golang.org>
Walk calls it outervalue, racewalk calls it basenod,
isstack does it manually and slightly differently.
Change-Id: Id5b5d32b8faf143fe9d34bd08457bfab6fb33daa
Reviewed-on: https://go-review.googlesource.com/3745
Reviewed-by: Russ Cox <rsc@golang.org>
Support the following conversions in escape analysis:
[]rune("foo")
[]byte("foo")
string([]rune{})
If the result does not escape, allocate temp buffer on stack
and pass it to runtime functions.
Change-Id: I1d075907eab8b0109ad7ad1878104b02b3d5c690
Reviewed-on: https://go-review.googlesource.com/3590
Reviewed-by: Russ Cox <rsc@golang.org>
Now:
0x0000 00000 (/tmp/x.s:2) MULLU R6,R3,(R7, R6)
The space is a little odd but I'd rather fix the usual printing to add spaces
than delete that one. But in a different CL, once C is gone.
Change-Id: I344e0b06eedaaf53cd79d370fa13c444a1e69c81
Reviewed-on: https://go-review.googlesource.com/4647
Reviewed-by: Rob Pike <r@golang.org>
(In non-Go print formats, the 016 includes the leading 0x prefix.
No one noticed, but we were printing hex numbers with a minimum
of 30 digits, not 32.)
Change-Id: I10ff7a51a567ad7c8440418ac034be9e4b2d6bc1
Reviewed-on: https://go-review.googlesource.com/4592
Reviewed-by: Austin Clements <austin@google.com>
This matches all the other pseudo-packages.
The line was simply forgotten.
Change-Id: I278f6cbcfc883ea7efad07f99fc8c853b9b5d274
Reviewed-on: https://go-review.googlesource.com/4591
Reviewed-by: Austin Clements <austin@google.com>
Otherwise different qsort implementations might result
in different sort orders and therefore different compiled
object files.
Change-Id: Ie783ba55a55af06941307e150b0c406e0a8128b0
Reviewed-on: https://go-review.googlesource.com/4590
Reviewed-by: Austin Clements <austin@google.com>
It does not convert to Go well.
Being able to do this just once, instead of 4 times, was the primary
motivation for all the recent refactoring (not that it wasn't overdue).
Still bit-for-bit identical.
Change-Id: Ia01f17948441bf64fa78ec4226f0bb40af0bbaab
Reviewed-on: https://go-review.googlesource.com/3962
Reviewed-by: Austin Clements <austin@google.com>
Now there is only one registerizer shared among all the systems.
There are some unfortunate special cases based on arch.thechar
in reg.c, to preserve bit-for-bit compatibility during the refactoring.
Most are probably bugs one way or another and should be revisited.
Change-Id: I153b435c0eaa05bbbeaf8876822eeb6dedaae3cf
Reviewed-on: https://go-review.googlesource.com/3883
Reviewed-by: Austin Clements <austin@google.com>
gc/order.c rewrites OASOP nodes into ordinary assignments.
The back ends never see them anymore.
Change-Id: I268ac8bdc92dccd7123110a21f99ada3ceeb2baa
Reviewed-on: https://go-review.googlesource.com/3882
Reviewed-by: Austin Clements <austin@google.com>
This isn't everything, but it's a start.
Still producing bit-identical compiler output.
The semantics of the old back ends is preserved,
even when they are probably buggy.
There are some TODOs in gc/gsubr.c to
remove special cases to preserve bugs in 5g and 8g.
Change-Id: I28ae295fbfc94ef9df43e13ab96bd6fc2f194bc4
Reviewed-on: https://go-review.googlesource.com/3802
Reviewed-by: Austin Clements <austin@google.com>
Currently, if there is a VERSION.cache, running make.bash will set
runtime.theVersion to the revision as of the *last* make.bash run
instead of the current make.bash run.
For example,
$ git rev-parse --short HEAD
5c4a86d
$ ./make.bash
...
$ cat ../VERSION.cache
devel +5c4a86d Tue Feb 10 01:46:30 2015 +0000
$ git checkout a1dbb92
$ ./make.bash
...
$ go version
go version devel +5c4a86d Tue Feb 10 01:46:30 2015 +0000 linux/amd64
$ ./make.bash
...
$ go version
go version devel +a1dbb92 Tue Feb 10 02:31:27 2015 +0000 linux/amd64
This happens because go tool dist reads the potentially stale
VERSION.cache into goversion during early initialization; then cleans,
which deletes VERSION.cache; then builds the runtime using the stale
revision read in to goversion. It isn't until make later in the build
process, when make.bash invokes go tool dist again, that VERSION.cache
gets updated with the current revision.
To address this, simply don't bother fetching the version until go
tool dist needs it and don't bother caching the value in memory. This
is more robust since it interacts with cleaning in the expected ways.
Futhermore, there's no downside to eliminating the in-memory cache;
the file system cache is perfectly reasonable for the whole three
times make.bash consults it.
Change-Id: I8c480100e56bb2db0816e8a088177004d9e87973
Reviewed-on: https://go-review.googlesource.com/4540
Reviewed-by: Russ Cox <rsc@golang.org>
Also clean up the branch code a bit
TBR=rsc
Change-Id: I209dea750db3a6769e7ccd79bb65c4d809aba152
Reviewed-on: https://go-review.googlesource.com/4530
Reviewed-by: Rob Pike <r@golang.org>
- obj: add a missing setting of the context for a generated JMP instruction
- asm: correct the encoding of mode (R)(R*scale)
- asm: fix a silly bug in the test for macro recursion.
- asm: accept address mode sym(R)(R*8); was an oversight
Change-Id: I27112eaaa1faa0d2ba97e414f0571b70733ea087
Reviewed-on: https://go-review.googlesource.com/4502
Reviewed-by: Russ Cox <rsc@golang.org>
Container symbols shouldn't be considered as functions in the functab.
Having them present probably messes up function lookup, as you might get
the descriptor of the container instead of the descriptor of the actual
function on the stack. It also messed up the findfunctab because these
entries caused off-by-one errors in how functab entries were counted.
Normal code is not affected - it only changes (& hopefully fixes) the
behavior for libraries linked as a unit, like:
net
runtime/cgo
runtime/race
Fixes#9804
Change-Id: I81e036e897571ac96567d59e1f1d7f058ca75e85
Reviewed-on: https://go-review.googlesource.com/4290
Reviewed-by: Russ Cox <rsc@golang.org>
I am an idiot but the failure to implement this means we can decide
exactly what its design should be for 1.5
Change-Id: Ie2b025fcd899d306ddeddd09d1d0e8f9a99ab7a8
Reviewed-on: https://go-review.googlesource.com/4291
Reviewed-by: Minux Ma <minux@golang.org>
Fixes#9732Fixes#9819
Rather than detecting vfp support via catching SIGILL signals,
parse the contents of /proc/cpuinfo.
As the GOARM values for NaCl and freebsd are hard coded, this parsing
logic only needs to support linux/arm.
This change also fixes the nacl/arm build which is broken because the
first stage of nacltest.bash is executed with GOARM=5, embedding that
into 5g.
The second stage of nacltest.bash correctly detects GOARM=7, but this is
ignored as we pass --no-clean at that point, and thus do not replace
the compiler.
Lastyly, include a fix to error message in nacltest.bash
Change-Id: I13f306ff07a99b44b493fade72ac00d0d5097e1c
Reviewed-on: https://go-review.googlesource.com/3981
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>
These illegal addressing modes were caught downstream in the assembler
or link library, but we can give a better error message upstream.
Change-Id: Ib30ef4d94d5d8d44900276592edd7997e6f91e55
Reviewed-on: https://go-review.googlesource.com/4260
Reviewed-by: Russ Cox <rsc@golang.org>
Considerable rewriting of the parser and assembler (code generator)
but it's simpler and shorter now. The internal Addr type is gone; so
is the package that held it. Parsing of operands goes directly into
obj.Addrs now.
There is a horrible hack regarding register pairs. It uses the Class
field to store the second register since it needs _some_ place to
put it but none is provided in the API. An alternative would be nice
but this works for now.
Once again creates identical .6 and .8 files as the old assembler.
Change-Id: I8207d6dfdfdb5bbed0bd870cb34ee0fe61c2fbfd
Reviewed-on: https://go-review.googlesource.com/4062
Reviewed-by: Russ Cox <rsc@golang.org>