Normalizing go/defer statements to always use functions with zero
parameters and zero results was added to escape analysis, because that
was the earliest point at which all three frontends converged. Now
that we only have the unified frontend, we can do it during typecheck,
which is where we perform all other desugaring and normalization
rewrites.
Change-Id: Iebf7679b117fd78b1dffee2974bbf85ebc923b23
Reviewed-on: https://go-review.googlesource.com/c/go/+/520260
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Dmitri Shuralyov <dmitshur@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
I previously used a clumsy hack to copy Closgen back and forth while
inlining, to handle when an inlined function contains closures, which
need to each be uniquely numbered.
The real solution was to name the closures using r.inlCaller, rather
than r.curfn. This CL adds a helper method to do exactly this.
Change-Id: I510553b5d7a8f6581ea1d21604e834fd6338cb06
Reviewed-on: https://go-review.googlesource.com/c/go/+/520339
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Dmitri Shuralyov <dmitshur@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
This CL extends escape analysis in two ways.
First, we already optimize directly called closures. For example,
given:
var x int // already stack allocated today
p := func() *int { return &x }()
we don't need to move x to the heap, because we can statically track
where &x flows. This CL extends the same idea to work for indirectly
called closures too, as long as we know everywhere that they're
called. For example:
var x int // stack allocated after this CL
f := func() *int { return &x }
p := f()
This will allow a subsequent CL to move the generation of go/defer
wrappers earlier.
Second, this CL adds tracking to detect when pointer values flow to
the pointee operand of an indirect assignment statement (i.e., flows
to p in "*p = x") or to builtins that modify memory (append, copy,
clear). This isn't utilized in the current CL, but a subsequent CL
will make use of it to better optimize string->[]byte conversions.
Updates #2205.
Change-Id: I610f9c531e135129c947684833e288ce64406f35
Reviewed-on: https://go-review.googlesource.com/c/go/+/520259
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Reviewed-by: Dmitri Shuralyov <dmitshur@google.com>
For aggregate-typed arguments passed to a call, expandCalls
decomposed them into parts in the same block where the value
was created. This is not necessarily the call block, and in
the case where stores are involved, can change the memory
leaving that block, and getting that right is problematic.
Instead, do all the expanding in the same block as the call,
which avoids the problems of (1) not being able to reorder
loads/stores across a block boundary to conform to memory
order and (2) (incorrectly, not) exposing the new memory to
consumers in other blocks. Putting it all in the same block
as the call allows reordering, and the call creates its own
new memory (which is already dealt with correctly).
Fixes#61992.
Change-Id: Icc7918f0d2dd3c480cc7f496cdcd78edeca7f297
Reviewed-on: https://go-review.googlesource.com/c/go/+/519276
Reviewed-by: Keith Randall <khr@google.com>
Run-TryBot: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
Open-coded defer slots are assigned indices upfront, so they're
logically like elements in an array. Without reassigning the indices,
we need to keep all of the elements alive so their relative offsets
are correct.
Fixes#61895.
Change-Id: Ie0191fdb33276f4e8ed0becb69086524fff022b2
Reviewed-on: https://go-review.googlesource.com/c/go/+/517856
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
CL 497276 added optimization for len(string([]byte)) by avoiding call to
slicebytetostring. However, the bytes to string expression may contain
init nodes, which need to be preserved. Otherwise, it would make the
liveness analysis confusing about the lifetime of temporary variables
created by init nodes.
Fixes#61778
Change-Id: I6d1280a7d61bcc75f11132af41bda086f084ab54
Reviewed-on: https://go-review.googlesource.com/c/go/+/516375
Run-TryBot: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
Auto-Submit: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Stop using BTSconst and friends when ORLconst can be used instead.
OR can be issued by more function units than BTS can, so it could
lead to better IPC. OR might take a few more bytes to encode, but
not a lot more.
Still use BTSconst for cases where the constant otherwise wouldn't
fit and would require a separate movabs instruction to materialize
the constant. This happens when setting bits 31-63 of 64-bit targets.
Add BTS-to-memory operations so we don't need to load/bts/store.
Fixes#61694
Change-Id: I00379608df8fb0167cb01466e97d11dec7c1596c
Reviewed-on: https://go-review.googlesource.com/c/go/+/515755
Reviewed-by: Keith Randall <khr@google.com>
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Fixes#61629
This reduce the pressure on regalloc because then the loop only keep alive
one value (the iterator) instead of the iterator and the upper bound since
the comparison now acts against an immediate, often zero which can be skipped.
This optimize things like:
for i := 0; i < n; i++ {
Or a range over a slice where the index is not used:
for _, v := range someSlice {
Or the new range over int from #61405:
for range n {
It is hit in 975 unique places while doing ./make.bash.
Change-Id: I5facff8b267a0b60ea3c1b9a58c4d74cdb38f03f
Reviewed-on: https://go-review.googlesource.com/c/go/+/512935
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Jorropo <jorropo.pgm@gmail.com>
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Auto-Submit: Keith Randall <khr@golang.org>
s==s is always true for strings. This comes up in NaN testing in
generic code, where we want x==x to compile completely away except for
float types.
Fixes#60777
Change-Id: I3ce054b5121354de2f9751b010fb409f148cb637
Reviewed-on: https://go-review.googlesource.com/c/go/+/503795
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Keith Randall <khr@golang.org>
If we load 2 values and then store those 2 loaded values, we can likely
perform that operation with a single wider load and store.
Fixes#60709
Change-Id: Ifc5f92c2f1b174c6ed82a69070f16cec6853c770
Reviewed-on: https://go-review.googlesource.com/c/go/+/502295
Reviewed-by: David Chase <drchase@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
Run-TryBot: Keith Randall <khr@golang.org>
The compiler/assembler's -S output prints relocation type
numerically, which is hard to understand. Every time I need to
count the relocation type constants to figure out which relocation
it actually is. Print the symbolic name instead.
Change-Id: I4866873bbae8b3dc0ee212609cb00280f9164243
Reviewed-on: https://go-review.googlesource.com/c/go/+/501856
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Than McIntosh <thanm@google.com>
Run-TryBot: Cherry Mui <cherryyz@google.com>
(ANDCCconst [y] (MOV.*reg x)) should only be merged when zero
extending. Otherwise, sign bits are lost on negative values.
(ANDCCconst [0xFF] (MOVBreg x)) should be simplified to a zero
extension of x. Likewise for the MOVHreg variant.
Fixes#61297
Change-Id: I04e4fd7dc6a826e870681f37506620d48393698b
Reviewed-on: https://go-review.googlesource.com/c/go/+/508775
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Paul Murphy <murp@ibm.com>
Reviewed-by: Bryan Mills <bcmills@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
User code is unlikely to be correct, but don't crash the compiler
when the offset of a pointer in an object is not a multiple of the
pointer size.
Fixes#61187
Change-Id: Ie56bfcb38556c5dd6f702ae4ec1d4534c6acd420
Reviewed-on: https://go-review.googlesource.com/c/go/+/508555
Reviewed-by: Cherry Mui <cherryyz@google.com>
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
According to RISCV manual 11.6:
FMADD x,y,z computes x*y+z and
FNMADD x,y,z => -x*y-z
FMSUB x,y,z => x*y-z
FNMSUB x,y,z => -x*y+z respectively
However our implement of SSA convert FMADD -x,y,z to FNMADD x,y,z which
is wrong and should be convert to FNMSUB according to manual.
Change-Id: Ib297bc83824e121fd7dda171ed56ea9694a4e575
Reviewed-on: https://go-review.googlesource.com/c/go/+/506575
Run-TryBot: M Zhuo <mzh@golangcn.org>
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Joedian Reid <joedian@golang.org>
Reviewed-by: Michael Munday <mike.munday@lowrisc.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
The large-function phi placement algorithm evidently doesn't like the
same pseudo-variable being used to represent expressions of varying
types.
Instead, use the same tactic as used for "valVar" (ssa.go:6585--6587),
which is to just generate a fresh marker node each time.
Maybe we could just use the OMIN/OMAX nodes themselves as the key
(like we do for OANDAND/OOROR), but that just seems needlessly risky
for negligible memory savings. Using fresh marker values each time
seems obviously safe by comparison.
Fixes#61041.
Change-Id: Ie2600c9c37b599c2e26ae01f5f8a433025d7fd08
Reviewed-on: https://go-review.googlesource.com/c/go/+/506679
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Ian Lance Taylor <iant@google.com>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
For float or string, min/max builtin performs a runtime call, so we need
to save its result to temporary variable. Otherwise, the runtime call
will clobber closure's arguments currently on the stack when passing
min/max as argument to closures.
Fixes#60990
Change-Id: I1397800f815ec7853182868678d0f760b22afff2
Reviewed-on: https://go-review.googlesource.com/c/go/+/506115
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Auto-Submit: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Our large-function phi placement algorithm is incompatible with phi
opcodes already existing in the SSA representation. Instead, use simple
variable assignments and have the phi placement algorithm place the phis
we need for min/max.
Turns out the small-function phi placement algorithm doesn't have this
sensitivity, so this bug only occurs in large functions (>500 basic blocks).
Maybe we should document/check that no phis are present when we start
phi placement (regardless of size). Leaving for a potential separate CL.
We should probably also fix the placement algorithm to handle existing
phis correctly. But this CL is probably a lot smaller/safer than
messing with phi placement.
Fixes#60982
Change-Id: I59ba7f506c72b22bc1485099a335d96315ebef67
Reviewed-on: https://go-review.googlesource.com/c/go/+/505756
Reviewed-by: Keith Randall <khr@google.com>
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
CL 410344 fixed missing method value wrapper, by visiting body of
wrapper function after applying inlining pass.
CL 492017 allow more inlining of functions that construct closures,
which ends up making the wrapper function now inlineable, but can
contain closure nodes that couldn't be inlined. These closures body may
contain OMETHVALUE nodes that we never seen, thus we need to scan
closures body for finding them.
Fixes#60945
Change-Id: Ia1e31420bb172ff87d7321d2da2989ef23e6ebb6
Reviewed-on: https://go-review.googlesource.com/c/go/+/505255
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Than McIntosh <thanm@google.com>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Run-TryBot: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Auto-Submit: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Change-Id: Idd872c5b90dbca564ed8a37bb3683e642142ae63
Reviewed-on: https://go-review.googlesource.com/c/go/+/505015
Reviewed-by: Than McIntosh <thanm@google.com>
Reviewed-by: Ian Lance Taylor <iant@google.com>
Auto-Submit: Ian Lance Taylor <iant@google.com>
Run-TryBot: Ian Lance Taylor <iant@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Ian Lance Taylor <iant@google.com>
Currently ArenaNew expects the type passed in to be a *T and it returns
a *T. This does not match the function's documentation.
Since this is an experiment, change ArenaNew to match the documentation.
This more closely aligns ArenaNew with arena.New. (Takes a type T,
returns a *T value.)
Note that this is a breaking change. However, as far as pkg.go.dev can
tell, there's exactly one package using it in the open source world.
Also, add smoke test for the exported API, which is just a wrapper
around the internal API. Clearly there's enough room for error here that
it should be tested, but we don't need thorough tests at this layer
because that already exists in the runtime. We just need to make sure it
basically works.
Fixes#60528.
Change-Id: I673cc4609378380ef80648b0c2eb2928e73f49c9
Reviewed-on: https://go-review.googlesource.com/c/go/+/501860
Run-TryBot: Michael Knyszek <mknyszek@google.com>
Reviewed-by: Ian Lance Taylor <iant@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
types2 have already errored about any spec-required overflows, and
division by zero. CL 469595 unintentionally fixed typecheck not to error
about overflows, but zero division is still be checked during tcArith.
This causes unsafe operations with variable size failed to compile,
instead of raising runtime error.
Fixes#60601
Change-Id: I7bea2821099556835c920713397f7c5d8a4025ac
Reviewed-on: https://go-review.googlesource.com/c/go/+/501735
Auto-Submit: Cuong Manh Le <cuong.manhle.vn@gmail.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Don't use the line number of the argument itself, as that may be from
arbitrarily earlier in the function.
Fixes#60673
Change-Id: Ifc0a2aaae221a256be3a4b0b2e04849bae4b79d7
Reviewed-on: https://go-review.googlesource.com/c/go/+/502656
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: David Chase <drchase@google.com>
Reviewed-by: Keith Randall <khr@google.com>
CL 496257 adds min/max builtins, which may appear as argument to a
function call, so it will be tested by mayCall. But those ops are not
handled by mayCall, causes the compiler crashes.
Fixes#60582
Change-Id: I729f10bf62b4aad39ffcb1433f576e74d09fdd9a
Reviewed-on: https://go-review.googlesource.com/c/go/+/500575
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: Dmitri Shuralyov <dmitshur@google.com>
Auto-Submit: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Reviewed-by: Keith Randall <khr@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Cuong Manh Le <cuong.manhle.vn@gmail.com>
HAUTO should be handled the same way as other stack offsets for
adding to constant pool. Add the missing cases.
Fixes#57955.
Change-Id: If7fc82cafb2bbf0a6121e73e353b8825cb36b5bc
Reviewed-on: https://go-review.googlesource.com/c/go/+/463138
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
Run-TryBot: Cherry Mui <cherryyz@google.com>
Some testdir tests fail if GOEXPERIMENT=cgocheck2 is set. Fix this by
skipping these tests.
Change-Id: I58d4ef0cceb86bcf93220b4a44de9b9dc4879b16
Reviewed-on: https://go-review.googlesource.com/c/go/+/499675
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Austin Clements <austin@google.com>
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Two interface types that are assignable don't have to be identical;
specifically, if they are defined types, they can be different
defined types. If those defined types specify type parameters which
are never used, do not infer a type argument based on the instantiation
of a matching defined type.
Adjusted three existing tests where we inferred type arguments incorrectly.
Fixes#60377.
Change-Id: I91fb207235424b3cbc42b5fd93eee619e7541cb7
Reviewed-on: https://go-review.googlesource.com/c/go/+/498315
Auto-Submit: Robert Griesemer <gri@google.com>
Reviewed-by: Robert Findley <rfindley@google.com>
Reviewed-by: Robert Griesemer <gri@google.com>
Run-TryBot: Robert Griesemer <gri@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
sparse conditional constant propagation can discover optimization opportunities that cannot be found by just combining constant folding and constant propagation and dead code elimination separately.
Updates #59399
Change-Id: Ia954e906480654a6f0cc065d75b5912f96f36b2e
GitHub-Last-Rev: 90fc02db99
GitHub-Pull-Request: golang/go#59575
Reviewed-on: https://go-review.googlesource.com/c/go/+/483875
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Michael Pratt <mpratt@google.com>
Run-TryBot: Keith Randall <khr@golang.org>
CL 415241 and CL 411935 break tests into unified/nounified variants, for
compatibility with old frontend while developing unified IR. Now the old
frontend was gone, so moving those tests back to the original files.
Change-Id: Iecdcd4e6ee33c723f6ac02189b0be26248e15f0f
Reviewed-on: https://go-review.googlesource.com/c/go/+/497275
Run-TryBot: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Auto-Submit: Cuong Manh Le <cuong.manhle.vn@gmail.com>
In Go 1.17, cmd/compile gained the ability to inline calls to
functions that contain function literals (aka "closures"). This was
implemented by duplicating the function literal body and emitting a
second LSym, because in general it might be optimized better than the
original function literal.
However, the second LSym was named simply as any other function
literal appearing literally in the enclosing function would be named.
E.g., if f has a closure "f.funcX", and f is inlined into g, we would
create "g.funcY" (N.B., X and Y need not be the same.). Users then
have no idea this function originally came from f.
With this CL, the inlined call stack is incorporated into the clone
LSym's name: instead of "g.funcY", it's named "g.f.funcY".
In the future, it seems desirable to arrange for the clone's name to
appear exactly as the original name, so stack traces remain the same
as when -l or -d=inlfuncswithclosures are used. But it's unclear
whether the linker supports that today, or whether any downstream
tooling would be confused by this.
Updates #60324.
Change-Id: Ifad0ccef7e959e72005beeecdfffd872f63982f8
Reviewed-on: https://go-review.googlesource.com/c/go/+/497137
Reviewed-by: Michael Pratt <mpratt@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Retrying the original CL with a small modification. The original CL
did not handle the case of reading an itab out of a dictionary
correctly. When we read an itab out of a dictionary, we must treat
the type inside that itab as maybe being put in an interface.
Original CL: 486895
Revert CL: 490156
Change-Id: Id2dc1699d184cd8c63dac83986a70b60b4e6cbd7
Reviewed-on: https://go-review.googlesource.com/c/go/+/491495
Reviewed-by: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
This enables the implementation for proposal #58671, which is
a likely accept. By enabling it early we get a bit extra soak
time for this feature. The change can be reverted trivially, if
need be.
For #58671.
Change-Id: Id6c27515e45ff79f4f1d2fc1706f3f672ccdd1ab
Reviewed-on: https://go-review.googlesource.com/c/go/+/495955
Run-TryBot: Robert Griesemer <gri@google.com>
Reviewed-by: Robert Griesemer <gri@google.com>
Auto-Submit: Robert Griesemer <gri@google.com>
Reviewed-by: Robert Findley <rfindley@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
In the types1 universe, we only need to represent value types. For
interfaces, this means we only need to worry about pure interfaces. A
pure interface can embed a union type, but the overall union must be
equivalent to "any".
In go.dev/cl/458619, we changed the types1 reader to return "any", but
to incorporate a consistency check to make sure this is valid.
Unfortunately, a pure interface can actually still reference impure
interfaces, and in general this is hard to check precisely without
reimplementing a lot of types2 data structures and logic into types1.
We haven't had any other reports of this check failing since 1.20, so
it seems simplest to just suppress for now.
Fixes#60117.
Change-Id: I5053faafe2d1068c6d438b2193347546bf5330cd
Reviewed-on: https://go-review.googlesource.com/c/go/+/495455
Run-TryBot: Matthew Dempsky <mdempsky@google.com>
Auto-Submit: Matthew Dempsky <mdempsky@google.com>
Reviewed-by: Cuong Manh Le <cuong.manhle.vn@gmail.com>
Reviewed-by: Keith Randall <khr@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
Be more liberal about expanding the OR tree. Handle any tree shape
instead of a fully left or right associative tree.
Also remove tail feature, it isn't ever needed.
Change-Id: If16bebef94b952a604d6069e9be3d9129994cb6f
Reviewed-on: https://go-review.googlesource.com/c/go/+/494056
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: Ryan Berger <ryanbberger@gmail.com>
Reviewed-by: Keith Randall <khr@google.com>
Reviewed-by: David Chase <drchase@google.com>
[This is a roll-forward of CL 458755, which was reverted due to make.bash
being broken on GOAMD64=v3. But it turned out that the problem was caused
by wrong bswap/load rewrite rules, and it was fixed in CL 492616.]
This CL enhances the tighten pass. Previously if a value has memory arg,
then the tighten pass won't move it, actually if the memory state is
consistent among definition and use block, we can move the value. This
CL optimizes this case. This is useful for the following situation:
b1:
x = load(...mem)
if(...) goto b2 else b3
b2:
use(x)
b3:
some_op_not_use_x
For the micro-benchmark mentioned in #56620, the performance improvement
is about 15%.
There's no noticeable performance change in the go1 benchmark.
Fixes#56620
Change-Id: I36ea68bed384986cd3ae81cb9e6efe84bb213adc
Reviewed-on: https://go-review.googlesource.com/c/go/+/492895
Reviewed-by: Keith Randall <khr@golang.org>
TryBot-Result: Gopher Robot <gobot@golang.org>
Reviewed-by: Heschi Kreinick <heschi@google.com>
Reviewed-by: Keith Randall <khr@google.com>
Run-TryBot: Eric Fang <eric.fang@arm.com>
This modifies some existing rules to allow more prefixed instructions
to be generated when using GOPPC64=power10. Some rules also check
if PCRel is available, which is currently supported for linux/ppc64le
and linux/ppc64 (internal linking only).
Prior to p10, DS-offset loads and stores had a 16 bit size limit for
the offset field. If the offset of the data for load or store was
beyond this range then an indexed load or store would be selected by
the rules.
In p10 the assembler can generate prefixed instructions in this case,
but does not if an indexed instruction was selected during the lowering
pass.
This allows many more cases to use prefixed loads or stores, reducing
function sizes and improving performance in some cases where the code
change happens in key loops.
For example in strconv BenchmarkAppendQuoteRune before:
12c5e4: 15 00 10 06 pla r10,1425660
12c5e8: fc c0 40 39
12c5ec: 00 00 6a e8 ld r3,0(r10)
12c5f0: 10 00 aa e8 ld r5,16(r10)
After this change:
12a828: 15 00 10 04 pld r3,1433272
12a82c: b8 de 60 e4
12a830: 15 00 10 04 pld r5,1433280
12a834: c0 de a0 e4
Performs better in the second case.
A testcase was added to verify that the rules correctly select a load or
store based on the offset and whether power10 or earlier.
Change-Id: I4335fed0bd9b8aba8a4f84d69b89f819cc464846
Reviewed-on: https://go-review.googlesource.com/c/go/+/477398
Reviewed-by: Heschi Kreinick <heschi@google.com>
Reviewed-by: Archana Ravindar <aravind5@in.ibm.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Lynn Boger <laboger@linux.vnet.ibm.com>
Reviewed-by: Paul Murphy <murp@ibm.com>
