Commit 496b8dbb authored by Cherry Zhang's avatar Cherry Zhang

cmd, runtime: remove PC_F & PC_B globals on Wasm

Following the previous CL, this removes more global variables on
Wasm.

PC_B is used mostly for intra-function jumps, and for a function
telling its callee where to start or resume. This usage can be
served by a parameter. The top level loop (wasm_pc_f_loop) uses
PC_B for resuming a function. This value is either set by gogo,
or loaded from the Go stack at function return. Instead of
loading PC_B at each function return, we could make gogo stores
PC_B at the same stack location, and let the top level loop do
the load. This way, we don't need to use global PC_B to
communicate with the top level loop, and we can replace global
PC_B with a parameter.

PC_F is similar. It is even more so in that the only reader is
the top level loop. Let the top level loop read it from the stack,
and we can get rid of PC_F entirely.

PC_F and PC_B are used less entensively as SP, so this CL has
smaller performance gain.

Running on Chrome 74.0.3729.108 on Linux/AMD64,

name                   old time/op    new time/op    delta
BinaryTree17              16.6s ± 0%     16.2s ± 1%   -2.59%  (p=0.016 n=4+5)
Fannkuch11                11.1s ± 1%     10.8s ± 0%   -2.65%  (p=0.008 n=5+5)
FmtFprintfEmpty           231ns ± 1%     217ns ± 0%   -6.06%  (p=0.008 n=5+5)
FmtFprintfString          407ns ± 3%     375ns ± 2%   -7.81%  (p=0.008 n=5+5)
FmtFprintfInt             466ns ± 2%     430ns ± 0%   -7.79%  (p=0.016 n=5+4)
FmtFprintfIntInt          719ns ± 2%     673ns ± 2%   -6.37%  (p=0.008 n=5+5)
FmtFprintfPrefixedInt     706ns ± 1%     676ns ± 3%   -4.31%  (p=0.008 n=5+5)
FmtFprintfFloat          1.01µs ± 1%    0.97µs ± 1%   -4.30%  (p=0.008 n=5+5)
FmtManyArgs              2.67µs ± 1%    2.51µs ± 1%   -5.95%  (p=0.008 n=5+5)
GobDecode                30.7ms ± 9%    31.3ms ±34%     ~     (p=0.222 n=5+5)
GobEncode                24.2ms ±23%    20.2ms ± 0%  -16.36%  (p=0.016 n=5+4)
Gzip                      852ms ± 0%     823ms ± 0%   -3.38%  (p=0.016 n=4+5)
Gunzip                    160ms ± 1%     151ms ± 1%   -5.37%  (p=0.008 n=5+5)
JSONEncode               35.7ms ± 1%    34.3ms ± 1%   -3.81%  (p=0.008 n=5+5)
JSONDecode                247ms ± 8%     254ms ± 7%     ~     (p=0.548 n=5+5)
Mandelbrot200            5.39ms ± 0%    5.41ms ± 0%   +0.42%  (p=0.008 n=5+5)
GoParse                  18.5ms ± 1%    18.3ms ± 2%     ~     (p=0.343 n=4+4)
RegexpMatchEasy0_32       424ns ± 2%     397ns ± 0%   -6.23%  (p=0.008 n=5+5)
RegexpMatchEasy0_1K      2.88µs ± 0%    2.86µs ± 1%     ~     (p=0.079 n=5+5)
RegexpMatchEasy1_32       395ns ± 2%     370ns ± 1%   -6.23%  (p=0.008 n=5+5)
RegexpMatchEasy1_1K      3.26µs ± 0%    3.19µs ± 1%   -2.06%  (p=0.008 n=5+5)
RegexpMatchMedium_32      564ns ± 1%     532ns ± 0%   -5.71%  (p=0.008 n=5+5)
RegexpMatchMedium_1K      146µs ± 2%     140µs ± 1%   -4.62%  (p=0.008 n=5+5)
RegexpMatchHard_32       8.47µs ± 1%    7.91µs ± 1%   -6.65%  (p=0.008 n=5+5)
RegexpMatchHard_1K        253µs ± 1%     236µs ± 2%   -6.66%  (p=0.008 n=5+5)
Revcomp                   1.78s ± 4%     1.76s ± 5%     ~     (p=1.000 n=5+5)
Template                  292ms ±29%     269ms ± 5%     ~     (p=0.690 n=5+5)
TimeParse                1.61µs ± 4%    1.54µs ± 1%   -4.42%  (p=0.008 n=5+5)
TimeFormat               1.66µs ± 3%    1.58µs ± 1%   -5.22%  (p=0.008 n=5+5)
[Geo mean]                232µs          221µs        -4.54%

name                   old speed      new speed      delta
GobDecode              25.0MB/s ± 8%  25.1MB/s ±27%     ~     (p=0.222 n=5+5)
GobEncode              32.8MB/s ±21%  38.0MB/s ± 0%  +15.84%  (p=0.016 n=5+4)
Gzip                   22.8MB/s ± 0%  23.6MB/s ± 0%   +3.49%  (p=0.016 n=4+5)
Gunzip                  121MB/s ± 1%   128MB/s ± 1%   +5.68%  (p=0.008 n=5+5)
JSONEncode             54.4MB/s ± 1%  56.5MB/s ± 1%   +3.97%  (p=0.008 n=5+5)
JSONDecode             7.88MB/s ± 8%  7.65MB/s ± 8%     ~     (p=0.548 n=5+5)
GoParse                3.07MB/s ± 8%  3.00MB/s ±22%     ~     (p=0.579 n=5+5)
RegexpMatchEasy0_32    75.6MB/s ± 2%  80.5MB/s ± 0%   +6.58%  (p=0.008 n=5+5)
RegexpMatchEasy0_1K     356MB/s ± 0%   358MB/s ± 1%     ~     (p=0.095 n=5+5)
RegexpMatchEasy1_32    81.1MB/s ± 2%  86.5MB/s ± 1%   +6.69%  (p=0.008 n=5+5)
RegexpMatchEasy1_1K     314MB/s ± 0%   320MB/s ± 0%   +2.10%  (p=0.008 n=5+5)
RegexpMatchMedium_32   1.77MB/s ± 1%  1.88MB/s ± 0%   +6.09%  (p=0.016 n=5+4)
RegexpMatchMedium_1K   6.99MB/s ± 2%  7.33MB/s ± 1%   +4.83%  (p=0.008 n=5+5)
RegexpMatchHard_32     3.78MB/s ± 1%  4.04MB/s ± 1%   +7.04%  (p=0.008 n=5+5)
RegexpMatchHard_1K     4.04MB/s ± 1%  4.33MB/s ± 2%   +7.17%  (p=0.008 n=5+5)
Revcomp                 143MB/s ± 4%   145MB/s ± 5%     ~     (p=1.000 n=5+5)
Template               6.77MB/s ±24%  7.22MB/s ± 5%     ~     (p=0.690 n=5+5)
[Geo mean]             24.4MB/s       25.4MB/s        +4.18%

Change-Id: Ib80716e62992aec28b2c4a96af280c278f83aa49
Reviewed-on: https://go-review.googlesource.com/c/go/+/173980
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: default avatarRichard Musiol <neelance@gmail.com>
parent 7ed7669c
......@@ -250,9 +250,7 @@ const (
const (
// globals
REG_PC_F = obj.RBaseWasm + iota
REG_PC_B
REG_SP // SP is currently 32-bit, until 64-bit memory operations are available
REG_SP = obj.RBaseWasm + iota // SP is currently 32-bit, until 64-bit memory operations are available
REG_CTXT
REG_g
// RET* are used by runtime.return0 and runtime.reflectcall. These functions pass return values in registers.
......@@ -296,9 +294,11 @@ const (
REG_F14
REG_F15
REG_PC_B // also first parameter, i32
MAXREG
MINREG = REG_PC_F
MINREG = REG_SP
REGSP = REG_SP
REGCTXT = REG_CTXT
REGG = REG_g
......
......@@ -16,8 +16,6 @@ import (
)
var Register = map[string]int16{
"PC_F": REG_PC_F,
"PC_B": REG_PC_B,
"SP": REG_SP,
"CTXT": REG_CTXT,
"g": REG_g,
......@@ -60,6 +58,8 @@ var Register = map[string]int16{
"F13": REG_F13,
"F14": REG_F14,
"F15": REG_F15,
"PC_B": REG_PC_B,
}
var registerNames []string
......@@ -368,20 +368,31 @@ func preprocess(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
break
}
// reset PC_B to function entry
p = appendp(p, AI32Const, constAddr(0))
p = appendp(p, ASet, regAddr(REG_PC_B))
// low-level WebAssembly call to function
switch jmp.To.Type {
case obj.TYPE_MEM:
if !notUsePC_B[jmp.To.Sym.Name] {
// Set PC_B parameter to function entry.
p = appendp(p, AI32Const, constAddr(0))
}
p = appendp(p, ACall, jmp.To)
case obj.TYPE_NONE:
// (target PC is on stack)
p = appendp(p, AI32WrapI64)
p = appendp(p, AI32Const, constAddr(16)) // only needs PC_F bits (16-31), PC_B bits (0-15) are zero
p = appendp(p, AI32ShrU)
// Set PC_B parameter to function entry.
// We need to push this before pushing the target PC_F,
// so temporarily pop PC_F, using our REG_PC_B as a
// scratch register, and push it back after pushing 0.
p = appendp(p, ASet, regAddr(REG_PC_B))
p = appendp(p, AI32Const, constAddr(0))
p = appendp(p, AGet, regAddr(REG_PC_B))
p = appendp(p, ACallIndirect)
default:
panic("bad target for JMP")
}
......@@ -419,20 +430,31 @@ func preprocess(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
})
p = appendp(p, AI64Store, constAddr(0))
// reset PC_B to function entry
p = appendp(p, AI32Const, constAddr(0))
p = appendp(p, ASet, regAddr(REG_PC_B))
// low-level WebAssembly call to function
switch call.To.Type {
case obj.TYPE_MEM:
if !notUsePC_B[call.To.Sym.Name] {
// Set PC_B parameter to function entry.
p = appendp(p, AI32Const, constAddr(0))
}
p = appendp(p, ACall, call.To)
case obj.TYPE_NONE:
// (target PC is on stack)
p = appendp(p, AI32WrapI64)
p = appendp(p, AI32Const, constAddr(16)) // only needs PC_F bits (16-31), PC_B bits (0-15) are zero
p = appendp(p, AI32ShrU)
// Set PC_B parameter to function entry.
// We need to push this before pushing the target PC_F,
// so temporarily pop PC_F, using our PC_B as a
// scratch register, and push it back after pushing 0.
p = appendp(p, ASet, regAddr(REG_PC_B))
p = appendp(p, AI32Const, constAddr(0))
p = appendp(p, AGet, regAddr(REG_PC_B))
p = appendp(p, ACallIndirect)
default:
panic("bad target for CALL")
}
......@@ -465,7 +487,13 @@ func preprocess(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
// jump to before the call if jmpdefer has reset the return address to the call's PC
if call.To.Sym == deferreturn {
p = appendp(p, AGet, regAddr(REG_PC_B))
// get PC_B from -8(SP)
p = appendp(p, AGet, regAddr(REG_SP))
p = appendp(p, AI32Const, constAddr(8))
p = appendp(p, AI32Sub)
p = appendp(p, AI32Load16U, constAddr(0))
p = appendp(p, ATee, regAddr(REG_PC_B))
p = appendp(p, AI32Const, constAddr(call.Pc))
p = appendp(p, AI32Eq)
p = appendp(p, ABrIf, constAddr(0))
......@@ -487,9 +515,8 @@ func preprocess(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
}
if ret.To.Type == obj.TYPE_MEM {
// reset PC_B to function entry
// Set PC_B parameter to function entry.
p = appendp(p, AI32Const, constAddr(0))
p = appendp(p, ASet, regAddr(REG_PC_B))
// low-level WebAssembly call to function
p = appendp(p, ACall, ret.To)
......@@ -497,16 +524,6 @@ func preprocess(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
break
}
// read return PC_F from Go stack
p = appendp(p, AGet, regAddr(REG_SP))
p = appendp(p, AI32Load16U, constAddr(2))
p = appendp(p, ASet, regAddr(REG_PC_F))
// read return PC_B from Go stack
p = appendp(p, AGet, regAddr(REG_SP))
p = appendp(p, AI32Load16U, constAddr(0))
p = appendp(p, ASet, regAddr(REG_PC_B))
// SP += 8
p = appendp(p, AGet, regAddr(REG_SP))
p = appendp(p, AI32Const, constAddr(8))
......@@ -771,16 +788,38 @@ func countRegisters(s *obj.LSym) (numI, numF int16) {
return
}
// Most of the Go functions has a single parameter (PC_B) in
// Wasm ABI. This is a list of exceptions.
var notUsePC_B = map[string]bool{
"_rt0_wasm_js": true,
"wasm_export_run": true,
"wasm_export_resume": true,
"wasm_export_getsp": true,
"wasm_pc_f_loop": true,
"runtime.wasmMove": true,
"runtime.wasmZero": true,
"runtime.wasmDiv": true,
"runtime.wasmTruncS": true,
"runtime.wasmTruncU": true,
"runtime.gcWriteBarrier": true,
"cmpbody": true,
"memeqbody": true,
"memcmp": true,
"memchr": true,
}
func assemble(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
w := new(bytes.Buffer)
hasLocalSP := false
hasPC_B := false
var r0, f0 int16
// Function starts with declaration of locals: numbers and types.
// Some functions use a special calling convention.
switch s.Name {
case "wasm_export_run", "runtime.wasmMove", "runtime.wasmZero", "runtime.wasmDiv", "runtime.wasmTruncS", "runtime.wasmTruncU", "memeqbody":
case "_rt0_wasm_js", "wasm_export_run", "wasm_export_resume", "wasm_export_getsp", "wasm_pc_f_loop",
"runtime.wasmMove", "runtime.wasmZero", "runtime.wasmDiv", "runtime.wasmTruncS", "runtime.wasmTruncU", "memeqbody":
writeUleb128(w, 0) // number of sets of locals
case "memchr", "memcmp":
writeUleb128(w, 1) // number of sets of locals
......@@ -797,9 +836,10 @@ func assemble(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
default:
// Normal calling convention: No WebAssembly parameters. First local variable is local SP cache.
hasLocalSP = true
hasPC_B = true
numI, numF := countRegisters(s)
r0 = 1
f0 = 1 + numI
r0 = 2
f0 = 2 + numI
numTypes := 1
if numI > 0 {
......@@ -826,6 +866,7 @@ func assemble(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
// Copy SP from its global variable into a local variable. Accessing a local variable is more efficient.
updateLocalSP(w)
}
for p := s.Func.Text; p != nil; p = p.Link {
switch p.As {
case AGet:
......@@ -836,10 +877,16 @@ func assemble(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
switch {
case reg == REG_SP && hasLocalSP:
w.WriteByte(0x20) // local.get
writeUleb128(w, 0) // local SP
case reg >= REG_PC_F && reg <= REG_PAUSE:
writeUleb128(w, 1) // local SP
case reg >= REG_SP && reg <= REG_PAUSE:
w.WriteByte(0x23) // global.get
writeUleb128(w, uint64(reg-REG_PC_F))
writeUleb128(w, uint64(reg-REG_SP))
case reg == REG_PC_B:
if !hasPC_B {
panic(fmt.Sprintf("PC_B is not used in %s", s.Name))
}
w.WriteByte(0x20) // local.get (i32)
writeUleb128(w, 0) // local PC_B
case reg >= REG_R0 && reg <= REG_R15:
w.WriteByte(0x20) // local.get (i64)
writeUleb128(w, uint64(r0+(reg-REG_R0)))
......@@ -857,21 +904,26 @@ func assemble(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
}
reg := p.To.Reg
switch {
case reg >= REG_PC_F && reg <= REG_PAUSE:
case reg >= REG_SP && reg <= REG_PAUSE:
if reg == REG_SP && hasLocalSP {
w.WriteByte(0x22) // local.tee
writeUleb128(w, 0) // local SP
writeUleb128(w, 1) // local SP
}
w.WriteByte(0x24) // global.set
writeUleb128(w, uint64(reg-REG_PC_F))
case reg >= REG_R0 && reg <= REG_F15:
writeUleb128(w, uint64(reg-REG_SP))
case reg >= REG_R0 && reg <= REG_PC_B:
if p.Link.As == AGet && p.Link.From.Reg == reg {
w.WriteByte(0x22) // local.tee
p = p.Link
} else {
w.WriteByte(0x21) // local.set
}
if reg <= REG_R15 {
if reg == REG_PC_B {
if !hasPC_B {
panic(fmt.Sprintf("PC_B is not used in %s", s.Name))
}
writeUleb128(w, 0) // local PC_B
} else if reg <= REG_R15 {
writeUleb128(w, uint64(r0+(reg-REG_R0)))
} else {
writeUleb128(w, uint64(f0+(reg-REG_F0)))
......@@ -887,6 +939,12 @@ func assemble(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
}
reg := p.To.Reg
switch {
case reg == REG_PC_B:
if !hasPC_B {
panic(fmt.Sprintf("PC_B is not used in %s", s.Name))
}
w.WriteByte(0x22) // local.tee (i32)
writeUleb128(w, 0) // local PC_B
case reg >= REG_R0 && reg <= REG_R15:
w.WriteByte(0x22) // local.tee (i64)
writeUleb128(w, uint64(r0+(reg-REG_R0)))
......@@ -1036,10 +1094,10 @@ func assemble(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
}
func updateLocalSP(w *bytes.Buffer) {
w.WriteByte(0x23) // global.get
writeUleb128(w, uint64(REG_SP-REG_PC_F)) // SP
w.WriteByte(0x21) // local.set
writeUleb128(w, 0) // local SP
w.WriteByte(0x23) // global.get
writeUleb128(w, 0) // global SP
w.WriteByte(0x21) // local.set
writeUleb128(w, 1) // local SP
}
func align(as obj.As) uint64 {
......
......@@ -102,10 +102,11 @@ func asmb2(ctxt *ld.Link) {
}
types := []*wasmFuncType{
// For normal Go functions the return value is
// For normal Go functions, the single parameter is PC_B,
// the return value is
// 0 if the function returned normally or
// 1 if the stack needs to be unwound.
{Results: []byte{I32}},
{Params: []byte{I32}, Results: []byte{I32}},
}
// collect host imports (functions that get imported from the WebAssembly host, usually JavaScript)
......@@ -320,16 +321,14 @@ func writeGlobalSec(ctxt *ld.Link) {
sizeOffset := writeSecHeader(ctxt, sectionGlobal)
globalRegs := []byte{
I32, // 0: PC_F
I32, // 1: PC_B
I32, // 2: SP
I64, // 3: CTXT
I64, // 4: g
I64, // 5: RET0
I64, // 6: RET1
I64, // 7: RET2
I64, // 8: RET3
I32, // 9: PAUSE
I32, // 0: SP
I64, // 1: CTXT
I64, // 2: g
I64, // 3: RET0
I64, // 4: RET1
I64, // 5: RET2
I64, // 6: RET3
I32, // 7: PAUSE
}
writeUleb128(ctxt.Out, uint64(len(globalRegs))) // number of globals
......
......@@ -37,17 +37,12 @@ TEXT runtime·gogo(SB), NOSPLIT, $0-8
MOVD gobuf_g(R0), g
MOVD gobuf_sp(R0), SP
// Put target PC at -8(SP), wasm_pc_f_loop will pick it up
Get SP
I32Const $8
I32Sub
I64Load gobuf_pc(R0)
I32WrapI64
I32Const $16
I32ShrU
Set PC_F
I64Load gobuf_pc(R0)
I64Const $0xFFFF
I64And
I32WrapI64
Set PC_B
I64Store $0
MOVD gobuf_ret(R0), RET0
MOVD gobuf_ctxt(R0), CTXT
......
......@@ -31,14 +31,9 @@ TEXT wasm_export_run(SB),NOSPLIT,$0
I64ExtendI32U
I64Store $8
I32Const $runtime·rt0_go(SB)
I32Const $16
I32ShrU
Set PC_F
I32Const $0
Set PC_B
I32Const $0 // entry PC_B
Call runtime·rt0_go(SB)
Drop
Call wasm_pc_f_loop(SB)
Return
......@@ -46,14 +41,9 @@ TEXT wasm_export_run(SB),NOSPLIT,$0
// wasm_export_resume gets called from JavaScript. It resumes the execution of Go code until it needs to wait for
// an event.
TEXT wasm_export_resume(SB),NOSPLIT,$0
I32Const $runtime·handleEvent(SB)
I32Const $16
I32ShrU
Set PC_F
I32Const $0
Set PC_B
Call runtime·handleEvent(SB)
Drop
Call wasm_pc_f_loop(SB)
Return
......@@ -63,15 +53,30 @@ TEXT wasm_pc_f_loop(SB),NOSPLIT,$0
// The WebAssembly stack may unwind, e.g. when switching goroutines.
// The Go stack on the linear memory is then used to jump to the correct functions
// with this loop, without having to restore the full WebAssembly stack.
loop:
Loop
Get PC_F
CallIndirect $0
Drop
Get PAUSE
I32Eqz
BrIf loop
// It is expected to have a pending call before entering the loop, so check PAUSE first.
Get PAUSE
I32Eqz
If
loop:
Loop
// Get PC_B & PC_F from -8(SP)
Get SP
I32Const $8
I32Sub
I32Load16U $0 // PC_B
Get SP
I32Const $8
I32Sub
I32Load16U $2 // PC_F
CallIndirect $0
Drop
Get PAUSE
I32Eqz
BrIf loop
End
End
I32Const $0
......@@ -91,6 +96,7 @@ TEXT runtime·pause(SB), NOSPLIT, $0-8
RETUNWIND
TEXT runtime·exit(SB), NOSPLIT, $0-4
I32Const $0
Call runtime·wasmExit(SB)
Drop
I32Const $1
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment