cmd/internal/obj: continue to optimize ARM's constant pool

Both Keith's https://go-review.googlesource.com/c/41612/ and
and Ben's https://go-review.googlesource.com/c/41679/ optimized ARM's
constant pool. But neither was complete.

First, BIC was forgotten.
1. "BIC $0xff00ff00, Reg" can be optimized to
   "BIC $0xff000000, Reg
    BIC $0x0000ff00, Reg"
2. "BIC $0xffff00ff, Reg" can be optimized to
   "AND $0x0000ff00, Reg"
3. "AND $0xffff00ff, Reg" can be optimized to
   "BIC $0x0000ff00, Reg"

Second, break a non-ARMImmRot to the subtraction of two ARMImmRots was
left as TODO.
1. "ADD $0x00fffff0, Reg" can be optimized to
   "ADD $0x01000000, Reg
    SUB $0x00000010, Reg"
2. "SUB $0x00fffff0, Reg" can be optimized to
   "SUB $0x01000000, Reg
    ADD $0x00000010, Reg"

This patch fixes them and issue #19844.

The go1 benchmark shows improvements.

name                     old time/op    new time/op    delta
BinaryTree17-4              41.4s ± 1%     41.7s ± 1%  +0.54%  (p=0.000 n=50+49)
Fannkuch11-4                24.7s ± 1%     25.1s ± 0%  +1.70%  (p=0.000 n=50+49)
FmtFprintfEmpty-4           853ns ± 1%     852ns ± 1%    ~     (p=0.833 n=50+50)
FmtFprintfString-4         1.33µs ± 1%    1.33µs ± 1%    ~     (p=0.163 n=50+50)
FmtFprintfInt-4            1.40µs ± 1%    1.40µs ± 0%    ~     (p=0.293 n=50+35)
FmtFprintfIntInt-4         2.09µs ± 1%    2.08µs ± 1%  -0.39%  (p=0.000 n=50+49)
FmtFprintfPrefixedInt-4    2.43µs ± 1%    2.43µs ± 1%    ~     (p=0.552 n=50+50)
FmtFprintfFloat-4          4.57µs ± 1%    4.42µs ± 1%  -3.18%  (p=0.000 n=50+50)
FmtManyArgs-4              8.62µs ± 1%    8.52µs ± 0%  -1.08%  (p=0.000 n=50+50)
GobDecode-4                 101ms ± 1%     101ms ± 2%  +0.45%  (p=0.001 n=49+49)
GobEncode-4                90.7ms ± 1%    91.1ms ± 2%  +0.51%  (p=0.001 n=50+50)
Gzip-4                      4.23s ± 1%     4.21s ± 1%  -0.62%  (p=0.000 n=50+50)
Gunzip-4                    623ms ± 1%     619ms ± 0%  -0.63%  (p=0.000 n=50+42)
HTTPClientServer-4          721µs ± 5%     683µs ± 3%  -5.25%  (p=0.000 n=50+47)
JSONEncode-4                251ms ± 1%     253ms ± 1%  +0.54%  (p=0.000 n=49+50)
JSONDecode-4                941ms ± 1%     944ms ± 1%  +0.30%  (p=0.001 n=49+50)
Mandelbrot200-4            49.3ms ± 1%    49.3ms ± 0%    ~     (p=0.918 n=50+48)
GoParse-4                  47.1ms ± 1%    47.2ms ± 1%  +0.18%  (p=0.025 n=50+50)
RegexpMatchEasy0_32-4      1.23µs ± 1%    1.24µs ± 1%  +0.30%  (p=0.000 n=49+50)
RegexpMatchEasy0_1K-4      7.74µs ± 7%    7.76µs ± 5%    ~     (p=0.888 n=50+50)
RegexpMatchEasy1_32-4      1.32µs ± 1%    1.32µs ± 1%  +0.23%  (p=0.003 n=50+50)
RegexpMatchEasy1_1K-4      10.6µs ± 2%    10.5µs ± 3%  -1.29%  (p=0.000 n=49+50)
RegexpMatchMedium_32-4     2.19µs ± 1%    2.10µs ± 1%  -3.79%  (p=0.000 n=49+49)
RegexpMatchMedium_1K-4      544µs ± 0%     545µs ± 0%    ~     (p=0.123 n=41+50)
RegexpMatchHard_32-4       28.8µs ± 0%    28.8µs ± 1%    ~     (p=0.580 n=46+50)
RegexpMatchHard_1K-4        863µs ± 1%     865µs ± 1%  +0.31%  (p=0.027 n=47+50)
Revcomp-4                  82.2ms ± 2%    82.3ms ± 2%    ~     (p=0.894 n=48+49)
Template-4                  1.06s ± 1%     1.04s ± 1%  -1.18%  (p=0.000 n=50+49)
TimeParse-4                7.25µs ± 1%    7.35µs ± 0%  +1.48%  (p=0.000 n=50+50)
TimeFormat-4               13.3µs ± 1%    13.2µs ± 1%  -0.13%  (p=0.007 n=50+50)
[Geo mean]                  736µs          733µs       -0.37%

name                     old speed      new speed      delta
GobDecode-4              7.60MB/s ± 1%  7.56MB/s ± 2%  -0.46%  (p=0.001 n=49+49)
GobEncode-4              8.47MB/s ± 1%  8.42MB/s ± 2%  -0.50%  (p=0.001 n=50+50)
Gzip-4                   4.58MB/s ± 1%  4.61MB/s ± 1%  +0.59%  (p=0.000 n=50+50)
Gunzip-4                 31.2MB/s ± 1%  31.4MB/s ± 0%  +0.63%  (p=0.000 n=50+42)
JSONEncode-4             7.73MB/s ± 1%  7.69MB/s ± 1%  -0.53%  (p=0.000 n=49+50)
JSONDecode-4             2.06MB/s ± 1%  2.06MB/s ± 1%    ~     (p=0.052 n=44+50)
GoParse-4                1.23MB/s ± 0%  1.23MB/s ± 2%    ~     (p=0.526 n=26+50)
RegexpMatchEasy0_32-4    25.9MB/s ± 1%  25.9MB/s ± 1%  -0.30%  (p=0.000 n=49+50)
RegexpMatchEasy0_1K-4     132MB/s ± 7%   132MB/s ± 6%    ~     (p=0.885 n=50+50)
RegexpMatchEasy1_32-4    24.2MB/s ± 1%  24.1MB/s ± 1%  -0.22%  (p=0.003 n=50+50)
RegexpMatchEasy1_1K-4    96.4MB/s ± 2%  97.8MB/s ± 3%  +1.36%  (p=0.000 n=50+50)
RegexpMatchMedium_32-4    460kB/s ± 0%   476kB/s ± 1%  +3.43%  (p=0.000 n=49+50)
RegexpMatchMedium_1K-4   1.88MB/s ± 0%  1.88MB/s ± 0%    ~     (all equal)
RegexpMatchHard_32-4     1.11MB/s ± 0%  1.11MB/s ± 1%  +0.34%  (p=0.000 n=45+50)
RegexpMatchHard_1K-4     1.19MB/s ± 1%  1.18MB/s ± 1%  -0.34%  (p=0.033 n=50+50)
Revcomp-4                30.9MB/s ± 2%  30.9MB/s ± 2%    ~     (p=0.894 n=48+49)
Template-4               1.84MB/s ± 1%  1.86MB/s ± 2%  +1.19%  (p=0.000 n=48+50)
[Geo mean]               6.63MB/s       6.65MB/s       +0.26%


Fixes #19844.

Change-Id: I5ad16cc0b29267bb4579aca3dcc10a0b8ade1aa4
Reviewed-on: https://go-review.googlesource.com/42430
Run-TryBot: Cherry Zhang <[email protected]>
TryBot-Result: Gobot Gobot <[email protected]>
Reviewed-by: Cherry Zhang <[email protected]>

Author: benshi001

src/cmd/compile/internal/ssa/gen/ARM.rules

@@ -798,6 +798,8 @@
 // generic constant folding
 (ADDconst [c] x) && !isARMImmRot(uint32(c)) && isARMImmRot(uint32(-c)) -> (SUBconst [int64(int32(-c))] x)
 (SUBconst [c] x) && !isARMImmRot(uint32(c)) && isARMImmRot(uint32(-c)) -> (ADDconst [int64(int32(-c))] x)
+(ANDconst [c] x) && !isARMImmRot(uint32(c)) && isARMImmRot(^uint32(c)) -> (BICconst [int64(^uint32(c))] x)
+(BICconst [c] x) && !isARMImmRot(uint32(c)) && isARMImmRot(^uint32(c)) -> (ANDconst [int64(^uint32(c))] x)
 (ADDconst [c] (MOVWconst [d])) -> (MOVWconst [int64(int32(c+d))])
 (ADDconst [c] (ADDconst [d] x)) -> (ADDconst [int64(int32(c+d))] x)
 (ADDconst [c] (SUBconst [d] x)) -> (ADDconst [int64(int32(c-d))] x)

src/cmd/compile/internal/ssa/rewriteARM.go

@@ -121,10 +121,11 @@ const (
 	C_PSR
 	C_FCR
 
-	C_RCON  /* 0xff rotated */
-	C_NCON  /* ~RCON */
-	C_RCON2 /* OR of two disjoint C_RCON constants */
-	C_SCON  /* 0xffff */
+	C_RCON   /* 0xff rotated */
+	C_NCON   /* ~RCON */
+	C_RCON2A /* OR of two disjoint C_RCON constants */
+	C_RCON2S /* subtraction of two disjoint C_RCON constants */
+	C_SCON   /* 0xffff */
 	C_LCON
 	C_LCONADDR
 	C_ZFCON

src/cmd/internal/obj/arm/a.out.go

@@ -16,7 +16,8 @@ var cnames5 = []string{
 	"FCR",
 	"RCON",
 	"NCON",
-	"RCON2",
+	"RCON2A",
+	"RCON2S",
 	"SCON",
 	"LCON",
 	"LCONADDR",

src/cmd/internal/obj/arm/anames5.go

@@ -88,20 +88,26 @@ var optab = []Optab{
 	{AADD, C_REG, C_NONE, C_REG, 1, 4, 0, 0, 0},
 	{AAND, C_REG, C_REG, C_REG, 1, 4, 0, 0, 0},
 	{AAND, C_REG, C_NONE, C_REG, 1, 4, 0, 0, 0},
+	{AORR, C_REG, C_REG, C_REG, 1, 4, 0, 0, 0},
+	{AORR, C_REG, C_NONE, C_REG, 1, 4, 0, 0, 0},
 	{AMOVW, C_REG, C_NONE, C_REG, 1, 4, 0, 0, 0},
 	{AMVN, C_REG, C_NONE, C_REG, 1, 4, 0, 0, 0},
 	{ACMP, C_REG, C_REG, C_NONE, 1, 4, 0, 0, 0},
 	{AADD, C_RCON, C_REG, C_REG, 2, 4, 0, 0, 0},
 	{AADD, C_RCON, C_NONE, C_REG, 2, 4, 0, 0, 0},
 	{AAND, C_RCON, C_REG, C_REG, 2, 4, 0, 0, 0},
 	{AAND, C_RCON, C_NONE, C_REG, 2, 4, 0, 0, 0},
+	{AORR, C_RCON, C_REG, C_REG, 2, 4, 0, 0, 0},
+	{AORR, C_RCON, C_NONE, C_REG, 2, 4, 0, 0, 0},
 	{AMOVW, C_RCON, C_NONE, C_REG, 2, 4, 0, 0, 0},
 	{AMVN, C_RCON, C_NONE, C_REG, 2, 4, 0, 0, 0},
 	{ACMP, C_RCON, C_REG, C_NONE, 2, 4, 0, 0, 0},
 	{AADD, C_SHIFT, C_REG, C_REG, 3, 4, 0, 0, 0},
 	{AADD, C_SHIFT, C_NONE, C_REG, 3, 4, 0, 0, 0},
 	{AAND, C_SHIFT, C_REG, C_REG, 3, 4, 0, 0, 0},
 	{AAND, C_SHIFT, C_NONE, C_REG, 3, 4, 0, 0, 0},
+	{AORR, C_SHIFT, C_REG, C_REG, 3, 4, 0, 0, 0},
+	{AORR, C_SHIFT, C_NONE, C_REG, 3, 4, 0, 0, 0},
 	{AMVN, C_SHIFT, C_NONE, C_REG, 3, 4, 0, 0, 0},
 	{ACMP, C_SHIFT, C_REG, C_NONE, 3, 4, 0, 0, 0},
 	{AMOVW, C_RACON, C_NONE, C_REG, 4, 4, REGSP, 0, 0},
@@ -136,20 +142,27 @@ var optab = []Optab{
 	{AADD, C_NCON, C_NONE, C_REG, 13, 8, 0, 0, 0},
 	{AAND, C_NCON, C_REG, C_REG, 13, 8, 0, 0, 0},
 	{AAND, C_NCON, C_NONE, C_REG, 13, 8, 0, 0, 0},
+	{AORR, C_NCON, C_REG, C_REG, 13, 8, 0, 0, 0},
+	{AORR, C_NCON, C_NONE, C_REG, 13, 8, 0, 0, 0},
 	{AMVN, C_NCON, C_NONE, C_REG, 13, 8, 0, 0, 0},
 	{ACMP, C_NCON, C_REG, C_NONE, 13, 8, 0, 0, 0},
 	{AADD, C_SCON, C_REG, C_REG, 13, 8, 0, 0, 0},
 	{AADD, C_SCON, C_NONE, C_REG, 13, 8, 0, 0, 0},
 	{AAND, C_SCON, C_REG, C_REG, 13, 8, 0, 0, 0},
 	{AAND, C_SCON, C_NONE, C_REG, 13, 8, 0, 0, 0},
+	{AORR, C_SCON, C_REG, C_REG, 13, 8, 0, 0, 0},
+	{AORR, C_SCON, C_NONE, C_REG, 13, 8, 0, 0, 0},
 	{AMVN, C_SCON, C_NONE, C_REG, 13, 8, 0, 0, 0},
 	{ACMP, C_SCON, C_REG, C_NONE, 13, 8, 0, 0, 0},
-	{AADD, C_RCON2, C_REG, C_REG, 106, 8, 0, 0, 0},
-	// TODO: RCON2: how to do AND and BIC?
+	{AADD, C_RCON2A, C_REG, C_REG, 106, 8, 0, 0, 0},
+	{AORR, C_RCON2A, C_REG, C_REG, 106, 8, 0, 0, 0},
+	{AADD, C_RCON2S, C_REG, C_REG, 107, 8, 0, 0, 0},
 	{AADD, C_LCON, C_REG, C_REG, 13, 8, 0, LFROM, 0},
 	{AADD, C_LCON, C_NONE, C_REG, 13, 8, 0, LFROM, 0},
 	{AAND, C_LCON, C_REG, C_REG, 13, 8, 0, LFROM, 0},
 	{AAND, C_LCON, C_NONE, C_REG, 13, 8, 0, LFROM, 0},
+	{AORR, C_LCON, C_REG, C_REG, 13, 8, 0, LFROM, 0},
+	{AORR, C_LCON, C_NONE, C_REG, 13, 8, 0, LFROM, 0},
 	{AMVN, C_LCON, C_NONE, C_REG, 13, 8, 0, LFROM, 0},
 	{ACMP, C_LCON, C_REG, C_NONE, 13, 8, 0, LFROM, 0},
 	{AMOVB, C_REG, C_NONE, C_REG, 1, 4, 0, 0, 0},
@@ -970,10 +983,10 @@ func immrot(v uint32) int32 {
 	return 0
 }
 
-// immrot2 returns bits encoding the immediate constant fields of two instructions,
+// immrot2a returns bits encoding the immediate constant fields of two instructions,
 // such that the encoded constants x, y satisfy x|y==v, x&y==0.
 // Returns 0,0 if no such decomposition of v exists.
-func immrot2(v uint32) (uint32, uint32) {
+func immrot2a(v uint32) (uint32, uint32) {
 	for i := uint(1); i < 32; i++ {
 		m := uint32(1<<i - 1)
 		if x, y := immrot(v&m), immrot(v&^m); x != 0 && y != 0 {
@@ -985,6 +998,32 @@ func immrot2(v uint32) (uint32, uint32) {
 	return 0, 0
 }
 
+// immrot2s returns bits encoding the immediate constant fields of two instructions,
+// such that the encoded constants y, x satisfy y-x==v, y&x==0.
+// Returns 0,0 if no such decomposition of v exists.
+func immrot2s(v uint32) (uint32, uint32) {
+	if immrot(v) == 0 {
+		return v, 0
+	}
+	// suppose v in the form of {leading 00, upper effective bits, lower 8 effective bits, trailing 00}
+	// omit trailing 00
+	var i uint32
+	for i = 2; i < 32; i += 2 {
+		if v&(1<<i-1) != 0 {
+			break
+		}
+	}
+	// i must be <= 24, then adjust i just above lower 8 effective bits of v
+	i += 6
+	// let x = {the complement of lower 8 effective bits, trailing 00}, y = x + v
+	x := 1<<i - v&(1<<i-1)
+	y := v + x
+	if y, x = uint32(immrot(y)), uint32(immrot(x)); y != 0 && x != 0 {
+		return y, x
+	}
+	return 0, 0
+}
+
 func immaddr(v int32) int32 {
 	if v >= 0 && v <= 0xfff {
 		return v&0xfff | 1<<24 | 1<<23 /* pre indexing */ /* pre indexing, up */
@@ -1159,8 +1198,11 @@ func (c *ctxt5) aclass(a *obj.Addr) int {
 			if uint32(c.instoffset) <= 0xffff && objabi.GOARM == 7 {
 				return C_SCON
 			}
-			if x, y := immrot2(uint32(c.instoffset)); x != 0 && y != 0 {
-				return C_RCON2
+			if x, y := immrot2a(uint32(c.instoffset)); x != 0 && y != 0 {
+				return C_RCON2A
+			}
+			if y, x := immrot2s(uint32(c.instoffset)); x != 0 && y != 0 {
+				return C_RCON2S
 			}
 			return C_LCON
 
@@ -1226,13 +1268,12 @@ func (c *ctxt5) oplook(p *obj.Prog) *Optab {
 		a2 = C_REG
 	}
 
-	// If Scond != 0, we must use the constant pool instead of
-	// splitting the instruction in two. The most common reason is
-	// .S (flag updating) instructions. There may be others.
-	if a1 == C_RCON2 && p.Scond != 0 {
+	// If current instruction has a .S suffix (flags update),
+	// we must use the constant pool instead of splitting it.
+	if (a1 == C_RCON2A || a1 == C_RCON2S) && p.Scond&C_SBIT != 0 {
 		a1 = C_LCON
 	}
-	if a3 == C_RCON2 && p.Scond != 0 {
+	if (a3 == C_RCON2A || a3 == C_RCON2S) && p.Scond&C_SBIT != 0 {
 		a3 = C_LCON
 	}
 
@@ -1266,7 +1307,7 @@ func cmp(a int, b int) bool {
 	}
 	switch a {
 	case C_LCON:
-		if b == C_RCON || b == C_NCON || b == C_SCON || b == C_RCON2 {
+		if b == C_RCON || b == C_NCON || b == C_SCON || b == C_RCON2A || b == C_RCON2S {
 			return true
 		}
 
@@ -1406,16 +1447,14 @@ func buildop(ctxt *obj.Link) {
 			log.Fatalf("bad code")
 
 		case AADD:
-			opset(AEOR, r0)
 			opset(ASUB, r0)
 			opset(ARSB, r0)
 			opset(AADC, r0)
 			opset(ASBC, r0)
 			opset(ARSC, r0)
-			opset(AORR, r0)
 
-		case AAND:
-			opset(AAND, r0)
+		case AORR:
+			opset(AEOR, r0)
 			opset(ABIC, r0)
 
 		case ACMP:
@@ -1541,6 +1580,7 @@ func buildop(ctxt *obj.Link) {
 			ALDREXD,
 			ASTREXD,
 			APLD,
+			AAND,
 			obj.AUNDEF,
 			obj.AFUNCDATA,
 			obj.APCDATA,
@@ -1609,11 +1649,11 @@ func (c *ctxt5) asmout(p *obj.Prog, o *Optab, out []uint32) {
 		c.aclass(&p.From)
 		r := int(p.Reg)
 		rt := int(p.To.Reg)
-		x, y := immrot2(uint32(c.instoffset))
+		x, y := immrot2a(uint32(c.instoffset))
 		var as2 obj.As
 		switch p.As {
-		case AADD, ASUB, AORR, AEOR:
-			as2 = p.As // ADD, SUB, ORR, EOR
+		case AADD, ASUB, AORR, AEOR, ABIC:
+			as2 = p.As // ADD, SUB, ORR, EOR, BIC
 		case ARSB:
 			as2 = AADD // RSB -> RSB/ADD pair
 		case AADC:
@@ -1632,6 +1672,35 @@ func (c *ctxt5) asmout(p *obj.Prog, o *Optab, out []uint32) {
 		o1 |= x
 		o2 |= y
 
+	case 107: /* op $I,R,R where I can be decomposed into 2 immediates */
+		c.aclass(&p.From)
+		r := int(p.Reg)
+		rt := int(p.To.Reg)
+		y, x := immrot2s(uint32(c.instoffset))
+		var as2 obj.As
+		switch p.As {
+		case AADD:
+			as2 = ASUB // ADD -> ADD/SUB pair
+		case ASUB:
+			as2 = AADD // SUB -> SUB/ADD pair
+		case ARSB:
+			as2 = ASUB // RSB -> RSB/SUB pair
+		case AADC:
+			as2 = ASUB // ADC -> ADC/SUB pair
+		case ASBC:
+			as2 = AADD // SBC -> SBC/ADD pair
+		case ARSC:
+			as2 = ASUB // RSC -> RSC/SUB pair
+		default:
+			c.ctxt.Diag("unknown second op for %v", p)
+		}
+		o1 = c.oprrr(p, p.As, int(p.Scond))
+		o2 = c.oprrr(p, as2, int(p.Scond))
+		o1 |= (uint32(r)&15)<<16 | (uint32(rt)&15)<<12
+		o2 |= (uint32(rt)&15)<<16 | (uint32(rt)&15)<<12
+		o1 |= y
+		o2 |= x
+
 	case 3: /* add R<<[IR],[R],R */
 		o1 = c.mov(p)