diff options
author | Russ Cox <rsc@golang.org> | 2014-05-15 15:34:53 -0400 |
---|---|---|
committer | Russ Cox <rsc@golang.org> | 2014-05-15 15:34:53 -0400 |
commit | 17e56eaa9da22104930c0185ae5b7daa562bb17d (patch) | |
tree | 31a8c2d31468c80f3c2ba4086c1205707c56c2a1 | |
parent | 191b16d00c36c11defbe0f901bc934d2cd642e0b (diff) | |
download | go-17e56eaa9da22104930c0185ae5b7daa562bb17d.tar.gz |
cmd/gc: correct handling of globals, func args, results
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 issue 1304. (again)
Fixes issue 7944. (again)
Fixes issue 7984.
Fixes issue 7995.
LGTM=khr
R=golang-codereviews, khr
CC=golang-codereviews, iant, r
https://codereview.appspot.com/97500044
-rw-r--r-- | src/cmd/5g/reg.c | 64 | ||||
-rw-r--r-- | src/cmd/6g/reg.c | 67 | ||||
-rw-r--r-- | src/cmd/8g/reg.c | 66 | ||||
-rw-r--r-- | test/fixedbugs/issue1304.go | 23 | ||||
-rw-r--r-- | test/fixedbugs/issue7995.go | 25 | ||||
-rw-r--r-- | test/fixedbugs/issue7995b.dir/x1.go | 16 | ||||
-rw-r--r-- | test/fixedbugs/issue7995b.dir/x2.go | 10 | ||||
-rw-r--r-- | test/fixedbugs/issue7995b.go | 9 | ||||
-rw-r--r-- | test/nilptr3.go | 102 |
9 files changed, 202 insertions, 180 deletions
diff --git a/src/cmd/5g/reg.c b/src/cmd/5g/reg.c index 3eadde4cf..6129698f3 100644 --- a/src/cmd/5g/reg.c +++ b/src/cmd/5g/reg.c @@ -56,28 +56,6 @@ rcmp(const void *a1, const void *a2) return p2->varno - p1->varno; } -static void -setvar(Bits *dst, Type **args) -{ - Type *t; - Node *n; - Addr a; - Iter save; - Bits bit; - int z; - - t = structfirst(&save, args); - while(t != T) { - n = nodarg(t, 1); - a = zprog.from; - naddr(n, &a, 0); - bit = mkvar(R, &a); - for(z=0; z<BITS; z++) - dst->b[z] |= bit.b[z]; - t = structnext(&save); - } -} - void excise(Flow *r) { @@ -192,11 +170,6 @@ regopt(Prog *firstp) ovar.b[z] = 0; } - // build lists of parameters and results - setvar(&ivar, getthis(curfn->type)); - setvar(&ivar, getinarg(curfn->type)); - setvar(&ovar, getoutarg(curfn->type)); - /* * pass 1 * build aux data structure @@ -837,9 +810,6 @@ mkvar(Reg *r, Adr *a) v->nextinnode = node->opt; node->opt = v; - if(debug['R']) - print("bit=%2d et=%2E w=%d+%d %#N %D flag=%d\n", i, et, o, w, node, a, v->addr); - bit = blsh(i); if(n == D_EXTERN || n == D_STATIC) for(z=0; z<BITS; z++) @@ -848,6 +818,13 @@ mkvar(Reg *r, Adr *a) for(z=0; z<BITS; z++) params.b[z] |= bit.b[z]; + if(node->class == PPARAM) + for(z=0; z<BITS; z++) + ivar.b[z] |= bit.b[z]; + if(node->class == PPARAMOUT) + for(z=0; z<BITS; z++) + ovar.b[z] |= bit.b[z]; + // Treat values with their address taken as live at calls, // because the garbage collector's liveness analysis in ../gc/plive.c does. // These must be consistent or else we will elide stores and the garbage @@ -864,7 +841,21 @@ mkvar(Reg *r, Adr *a) // The broader := in a closure problem is mentioned in a comment in // closure.c:/^typecheckclosure and dcl.c:/^oldname. if(node->addrtaken) - setaddrs(bit); + v->addr = 1; + + // Disable registerization for globals, because: + // (1) we might panic at any time and we want the recovery code + // to see the latest values (issue 1304). + // (2) we don't know what pointers might point at them and we want + // loads via those pointers to see updated values and vice versa (issue 7995). + // + // Disable registerization for results if using defer, because the deferred func + // might recover and return, causing the current values to be used. + if(node->class == PEXTERN || (hasdefer && node->class == PPARAMOUT)) + v->addr = 1; + + if(debug['R']) + print("bit=%2d et=%2E w=%d+%d %#N %D flag=%d\n", i, et, o, w, node, a, v->addr); return bit; @@ -961,17 +952,6 @@ prop(Reg *r, Bits ref, Bits cal) ref.b[z] = 0; } break; - - default: - // Work around for issue 1304: - // flush modified globals before each instruction. - for(z=0; z<BITS; z++) { - cal.b[z] |= externs.b[z]; - // issue 4066: flush modified return variables in case of panic - if(hasdefer) - cal.b[z] |= ovar.b[z]; - } - break; } for(z=0; z<BITS; z++) { ref.b[z] = (ref.b[z] & ~r1->set.b[z]) | diff --git a/src/cmd/6g/reg.c b/src/cmd/6g/reg.c index a1f0c756a..919a07d7b 100644 --- a/src/cmd/6g/reg.c +++ b/src/cmd/6g/reg.c @@ -55,28 +55,6 @@ rcmp(const void *a1, const void *a2) } static void -setvar(Bits *dst, Type **args) -{ - Type *t; - Node *n; - Addr a; - Iter save; - Bits bit; - int z; - - t = structfirst(&save, args); - while(t != T) { - n = nodarg(t, 1); - a = zprog.from; - naddr(n, &a, 0); - bit = mkvar(R, &a); - for(z=0; z<BITS; z++) - dst->b[z] |= bit.b[z]; - t = structnext(&save); - } -} - -static void setaddrs(Bits bit) { int i, n; @@ -178,11 +156,6 @@ regopt(Prog *firstp) ovar.b[z] = 0; } - // build lists of parameters and results - setvar(&ivar, getthis(curfn->type)); - setvar(&ivar, getinarg(curfn->type)); - setvar(&ovar, getoutarg(curfn->type)); - /* * pass 1 * build aux data structure @@ -690,11 +663,6 @@ mkvar(Reg *r, Adr *a) v->nextinnode = node->opt; node->opt = v; - if(debug['R']) - print("bit=%2d et=%2E w=%lld+%lld %#N %D flag=%d\n", i, et, o, w, node, a, v->addr); - - ostats.nvar++; - bit = blsh(i); if(n == D_EXTERN || n == D_STATIC) for(z=0; z<BITS; z++) @@ -703,6 +671,13 @@ mkvar(Reg *r, Adr *a) for(z=0; z<BITS; z++) params.b[z] |= bit.b[z]; + if(node->class == PPARAM) + for(z=0; z<BITS; z++) + ivar.b[z] |= bit.b[z]; + if(node->class == PPARAMOUT) + for(z=0; z<BITS; z++) + ovar.b[z] |= bit.b[z]; + // Treat values with their address taken as live at calls, // because the garbage collector's liveness analysis in ../gc/plive.c does. // These must be consistent or else we will elide stores and the garbage @@ -719,7 +694,22 @@ mkvar(Reg *r, Adr *a) // The broader := in a closure problem is mentioned in a comment in // closure.c:/^typecheckclosure and dcl.c:/^oldname. if(node->addrtaken) - setaddrs(bit); + v->addr = 1; + + // Disable registerization for globals, because: + // (1) we might panic at any time and we want the recovery code + // to see the latest values (issue 1304). + // (2) we don't know what pointers might point at them and we want + // loads via those pointers to see updated values and vice versa (issue 7995). + // + // Disable registerization for results if using defer, because the deferred func + // might recover and return, causing the current values to be used. + if(node->class == PEXTERN || (hasdefer && node->class == PPARAMOUT)) + v->addr = 1; + + if(debug['R']) + print("bit=%2d et=%2E w=%lld+%lld %#N %D flag=%d\n", i, et, o, w, node, a, v->addr); + ostats.nvar++; return bit; @@ -816,17 +806,6 @@ prop(Reg *r, Bits ref, Bits cal) ref.b[z] = 0; } break; - - default: - // Work around for issue 1304: - // flush modified globals before each instruction. - for(z=0; z<BITS; z++) { - cal.b[z] |= externs.b[z]; - // issue 4066: flush modified return variables in case of panic - if(hasdefer) - cal.b[z] |= ovar.b[z]; - } - break; } for(z=0; z<BITS; z++) { ref.b[z] = (ref.b[z] & ~r1->set.b[z]) | diff --git a/src/cmd/8g/reg.c b/src/cmd/8g/reg.c index 046011c90..ed019f937 100644 --- a/src/cmd/8g/reg.c +++ b/src/cmd/8g/reg.c @@ -55,28 +55,6 @@ rcmp(const void *a1, const void *a2) } static void -setvar(Bits *dst, Type **args) -{ - Type *t; - Node *n; - Addr a; - Iter save; - Bits bit; - int z; - - t = structfirst(&save, args); - while(t != T) { - n = nodarg(t, 1); - a = zprog.from; - naddr(n, &a, 0); - bit = mkvar(R, &a); - for(z=0; z<BITS; z++) - dst->b[z] |= bit.b[z]; - t = structnext(&save); - } -} - -static void setaddrs(Bits bit) { int i, n; @@ -148,11 +126,6 @@ regopt(Prog *firstp) ovar.b[z] = 0; } - // build lists of parameters and results - setvar(&ivar, getthis(curfn->type)); - setvar(&ivar, getinarg(curfn->type)); - setvar(&ovar, getoutarg(curfn->type)); - /* * pass 1 * build aux data structure @@ -656,10 +629,6 @@ mkvar(Reg *r, Adr *a) v->nextinnode = node->opt; node->opt = v; - if(debug['R']) - print("bit=%2d et=%2E w=%d+%d %#N %D flag=%d\n", i, et, o, w, node, a, v->addr); - ostats.nvar++; - bit = blsh(i); if(n == D_EXTERN || n == D_STATIC) for(z=0; z<BITS; z++) @@ -667,6 +636,13 @@ mkvar(Reg *r, Adr *a) if(n == D_PARAM) for(z=0; z<BITS; z++) params.b[z] |= bit.b[z]; + + if(node->class == PPARAM) + for(z=0; z<BITS; z++) + ivar.b[z] |= bit.b[z]; + if(node->class == PPARAMOUT) + for(z=0; z<BITS; z++) + ovar.b[z] |= bit.b[z]; // Treat values with their address taken as live at calls, // because the garbage collector's liveness analysis in ../gc/plive.c does. @@ -684,7 +660,22 @@ mkvar(Reg *r, Adr *a) // The broader := in a closure problem is mentioned in a comment in // closure.c:/^typecheckclosure and dcl.c:/^oldname. if(node->addrtaken) - setaddrs(bit); + v->addr = 1; + + // Disable registerization for globals, because: + // (1) we might panic at any time and we want the recovery code + // to see the latest values (issue 1304). + // (2) we don't know what pointers might point at them and we want + // loads via those pointers to see updated values and vice versa (issue 7995). + // + // Disable registerization for results if using defer, because the deferred func + // might recover and return, causing the current values to be used. + if(node->class == PEXTERN || (hasdefer && node->class == PPARAMOUT)) + v->addr = 1; + + if(debug['R']) + print("bit=%2d et=%2E w=%d+%d %#N %D flag=%d\n", i, et, o, w, node, a, v->addr); + ostats.nvar++; return bit; @@ -781,17 +772,6 @@ prop(Reg *r, Bits ref, Bits cal) ref.b[z] = 0; } break; - - default: - // Work around for issue 1304: - // flush modified globals before each instruction. - for(z=0; z<BITS; z++) { - cal.b[z] |= externs.b[z]; - // issue 4066: flush modified return variables in case of panic - if(hasdefer) - cal.b[z] |= ovar.b[z]; - } - break; } for(z=0; z<BITS; z++) { ref.b[z] = (ref.b[z] & ~r1->set.b[z]) | diff --git a/test/fixedbugs/issue1304.go b/test/fixedbugs/issue1304.go new file mode 100644 index 000000000..1206e1840 --- /dev/null +++ b/test/fixedbugs/issue1304.go @@ -0,0 +1,23 @@ +// run + +// Copyright 2014 The Go Authors. All rights reserved. +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +package main + +var a = 1 + +func main() { + defer func() { + recover() + if a != 2 { + println("BUG a =", a) + } + }() + a = 2 + b := a - a + c := 4 + a = c / b + a = 3 +} diff --git a/test/fixedbugs/issue7995.go b/test/fixedbugs/issue7995.go new file mode 100644 index 000000000..05f116823 --- /dev/null +++ b/test/fixedbugs/issue7995.go @@ -0,0 +1,25 @@ +// run + +// Copyright 2014 The Go Authors. All rights reserved. +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +// Issue 7995: globals not flushed quickly enough. + +package main + +import "fmt" + +var ( + p = 1 + q = &p +) + +func main() { + p = 50 + *q = 100 + s := fmt.Sprintln(p, *q) + if s != "100 100\n" { + println("BUG:", s) + } +} diff --git a/test/fixedbugs/issue7995b.dir/x1.go b/test/fixedbugs/issue7995b.dir/x1.go new file mode 100644 index 000000000..075911b92 --- /dev/null +++ b/test/fixedbugs/issue7995b.dir/x1.go @@ -0,0 +1,16 @@ +package x1 + +import "fmt" + +var P int + +var b bool + +func F(x *int) string { + if b { // avoid inlining + F(x) + } + P = 50 + *x = 100 + return fmt.Sprintln(P, *x) +} diff --git a/test/fixedbugs/issue7995b.dir/x2.go b/test/fixedbugs/issue7995b.dir/x2.go new file mode 100644 index 000000000..eea23eabb --- /dev/null +++ b/test/fixedbugs/issue7995b.dir/x2.go @@ -0,0 +1,10 @@ +package main + +import "./x1" + +func main() { + s := x1.F(&x1.P) + if s != "100 100\n" { + println("BUG:", s) + } +} diff --git a/test/fixedbugs/issue7995b.go b/test/fixedbugs/issue7995b.go new file mode 100644 index 000000000..2f57371e3 --- /dev/null +++ b/test/fixedbugs/issue7995b.go @@ -0,0 +1,9 @@ +// rundir + +// Copyright 2014 The Go Authors. All rights reserved. +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +// Issue 7995: globals not flushed quickly enough. + +package ignored diff --git a/test/nilptr3.go b/test/nilptr3.go index 08597a02d..2757daef0 100644 --- a/test/nilptr3.go +++ b/test/nilptr3.go @@ -17,7 +17,7 @@ type Struct struct { type BigStruct struct { X int Y float64 - A [1<<20]int + A [1 << 20]int Z string } @@ -29,93 +29,94 @@ type Empty1 struct { } var ( - intp *int - arrayp *[10]int - array0p *[0]int - bigarrayp *[1<<26]int - structp *Struct + intp *int + arrayp *[10]int + array0p *[0]int + bigarrayp *[1 << 26]int + structp *Struct bigstructp *BigStruct - emptyp *Empty - empty1p *Empty1 + emptyp *Empty + empty1p *Empty1 ) func f1() { _ = *intp // ERROR "generated nil check" - + // This one should be removed but the block copy needs // to be turned into its own pseudo-op in order to see // the indirect. _ = *arrayp // ERROR "generated nil check" - - // 0-byte indirect doesn't suffice + + // 0-byte indirect doesn't suffice. + // we don't registerize globals, so there are no removed repeated nil checks. + _ = *array0p // ERROR "generated nil check" _ = *array0p // ERROR "generated nil check" - _ = *array0p // ERROR "removed repeated nil check" 386 - _ = *intp // ERROR "removed repeated nil check" - _ = *arrayp // ERROR "removed repeated nil check" + _ = *intp // ERROR "generated nil check" + _ = *arrayp // ERROR "generated nil check" _ = *structp // ERROR "generated nil check" - _ = *emptyp // ERROR "generated nil check" - _ = *arrayp // ERROR "removed repeated nil check" + _ = *emptyp // ERROR "generated nil check" + _ = *arrayp // ERROR "generated nil check" } func f2() { var ( - intp *int - arrayp *[10]int - array0p *[0]int - bigarrayp *[1<<20]int - structp *Struct + intp *int + arrayp *[10]int + array0p *[0]int + bigarrayp *[1 << 20]int + structp *Struct bigstructp *BigStruct - emptyp *Empty - empty1p *Empty1 + emptyp *Empty + empty1p *Empty1 ) - _ = *intp // ERROR "generated nil check" - _ = *arrayp // ERROR "generated nil check" - _ = *array0p // ERROR "generated nil check" - _ = *array0p // ERROR "removed repeated nil check" - _ = *intp // ERROR "removed repeated nil check" - _ = *arrayp // ERROR "removed repeated nil check" - _ = *structp // ERROR "generated nil check" - _ = *emptyp // ERROR "generated nil check" - _ = *arrayp // ERROR "removed repeated nil check" - _ = *bigarrayp // ERROR "generated nil check" ARM removed nil check before indirect!! + _ = *intp // ERROR "generated nil check" + _ = *arrayp // ERROR "generated nil check" + _ = *array0p // ERROR "generated nil check" + _ = *array0p // ERROR "removed repeated nil check" + _ = *intp // ERROR "removed repeated nil check" + _ = *arrayp // ERROR "removed repeated nil check" + _ = *structp // ERROR "generated nil check" + _ = *emptyp // ERROR "generated nil check" + _ = *arrayp // ERROR "removed repeated nil check" + _ = *bigarrayp // ERROR "generated nil check" ARM removed nil check before indirect!! _ = *bigstructp // ERROR "generated nil check" - _ = *empty1p // ERROR "generated nil check" + _ = *empty1p // ERROR "generated nil check" } func fx10k() *[10000]int -var b bool +var b bool func f3(x *[10000]int) { // Using a huge type and huge offsets so the compiler // does not expect the memory hardware to fault. _ = x[9999] // ERROR "generated nil check" - + for { if x[9999] != 0 { // ERROR "generated nil check" break } } - - x = fx10k() + + x = fx10k() _ = x[9999] // ERROR "generated nil check" if b { _ = x[9999] // ERROR "removed repeated nil check" } else { _ = x[9999] // ERROR "removed repeated nil check" - } + } _ = x[9999] // ERROR "generated nil check" - x = fx10k() + x = fx10k() if b { _ = x[9999] // ERROR "generated nil check" } else { _ = x[9999] // ERROR "generated nil check" - } + } _ = x[9999] // ERROR "generated nil check" - + fx10k() // This one is a bit redundant, if we figured out that // x wasn't going to change across the function call. @@ -145,7 +146,7 @@ func f3b() { _ = &x[9] // ERROR "removed repeated nil check" } -func fx10() *[10]int +func fx10() *[10]int func f4(x *[10]int) { // Most of these have no checks because a real memory reference follows, @@ -153,14 +154,14 @@ func f4(x *[10]int) { // in the first unmapped page of memory. _ = x[9] // ERROR "removed nil check before indirect" - + for { if x[9] != 0 { // ERROR "removed nil check before indirect" break } } - - x = fx10() + + x = fx10() _ = x[9] // ERROR "removed nil check before indirect" if b { _ = x[9] // ERROR "removed nil check before indirect" @@ -169,17 +170,17 @@ func f4(x *[10]int) { } _ = x[9] // ERROR "removed nil check before indirect" - x = fx10() + x = fx10() if b { _ = x[9] // ERROR "removed nil check before indirect" } else { _ = &x[9] // ERROR "generated nil check" - } + } _ = x[9] // ERROR "removed nil check before indirect" - + fx10() _ = x[9] // ERROR "removed nil check before indirect" - + x = fx10() y := fx10() _ = &x[9] // ERROR "generated nil check" @@ -188,4 +189,3 @@ func f4(x *[10]int) { x = y _ = &x[9] // ERROR "removed repeated nil check" } - |