Revert "compile: prefer an AND instead of SHR+SHL instructions"

This reverts CL 194297.

Reason for revert: introduced register allocation failures on PPC64LE builders.

Updates #33826
Updates #32781
Updates #34468

Change-Id: I7d0b55df8cdf8e7d2277f1814299b083c2692e48
Reviewed-on: https://go-review.googlesource.com/c/go/+/196957
Run-TryBot: Bryan C. Mills <bcmills@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Dmitri Shuralyov <dmitshur@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Reviewed-by: Martin Möhrmann <moehrmann@google.com>

src/cmd/compile/internal/ssa/gen/ARM64.rules

@@ -1863,8 +1863,9 @@
 (XORshiftLL <t> [c] (UBFX [bfc] x) x2) && c < 32 && t.Size() == 4 && bfc == armBFAuxInt(32-c, c)
 	-> (EXTRWconst [32-c] x2 x)
 
-// Rewrite special pairs of shifts to AND.
-// On ARM64 the bitmask can fit into an instruction.
+// Generic rules rewrite certain AND to a pair of shifts.
+// However, on ARM64 the bitmask can fit into an instruction.
+// Rewrite it back to AND.
 (SRLconst [c] (SLLconst [c] x)) && 0 < c && c < 64 -> (ANDconst [1<<uint(64-c)-1] x) // mask out high bits
 (SLLconst [c] (SRLconst [c] x)) && 0 < c && c < 64 -> (ANDconst [^(1<<uint(c)-1)] x) // mask out low bits
 
@@ -1970,8 +1971,6 @@
 	-> (BFXIL [bfc] y x)
 (ORshiftLL [sc] (UBFX [bfc] x) (SRLconst [sc] y)) && sc == getARM64BFwidth(bfc)
 	-> (BFXIL [bfc] y x)
-(ORshiftRL [rc] (ANDconst [ac] y) (SLLconst [lc] x)) && lc < rc && ac == ^((1<<uint(64-rc)-1))
-	-> (BFXIL [armBFAuxInt(rc-lc, 64-rc)] y x)
 
 // do combined loads
 // little endian loads

src/cmd/compile/internal/ssa/gen/S390X.rules

@@ -701,8 +701,6 @@
 // may need to be reworked when NIHH/OIHH are added
 (SRDconst [1] (SLDconst [1] (LGDR <t> x))) -> (LGDR <t> (LPDFR <x.Type> x))
 (LDGR <t> (SRDconst [1] (SLDconst [1] x))) -> (LPDFR (LDGR <t> x))
-(AND (MOVDconst [^(-1<<63)]) (LGDR <t> x)) -> (LGDR <t> (LPDFR <x.Type> x))
-(LDGR <t> (AND (MOVDconst [^(-1<<63)]) x)) -> (LPDFR (LDGR <t> x))
 (OR (MOVDconst [-1<<63]) (LGDR <t> x))     -> (LGDR <t> (LNDFR <x.Type> x))
 (LDGR <t> (OR (MOVDconst [-1<<63]) x))     -> (LNDFR (LDGR <t> x))
 
@@ -712,8 +710,6 @@
 // detect copysign
 (OR (SLDconst [63] (SRDconst [63] (LGDR x))) (LGDR (LPDFR <t> y))) -> (LGDR (CPSDR <t> y x))
 (OR (SLDconst [63] (SRDconst [63] (LGDR x))) (MOVDconst [c])) && c & -1<<63 == 0 -> (LGDR (CPSDR <x.Type> (FMOVDconst <x.Type> [c]) x))
-(OR (AND (MOVDconst [-1<<63]) (LGDR x)) (LGDR (LPDFR <t> y))) -> (LGDR (CPSDR <t> y x))
-(OR (AND (MOVDconst [-1<<63]) (LGDR x)) (MOVDconst [c])) && c & -1<<63 == 0 -> (LGDR (CPSDR <x.Type> (FMOVDconst <x.Type> [c]) x))
 (CPSDR y (FMOVDconst [c])) && c & -1<<63 == 0 -> (LPDFR y)
 (CPSDR y (FMOVDconst [c])) && c & -1<<63 != 0 -> (LNDFR y)
 

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

@@ -542,6 +542,14 @@
 (Slicemask (Const64 [x])) && x > 0 -> (Const64 [-1])
 (Slicemask (Const64 [0]))          -> (Const64 [0])
 
+// Rewrite AND of consts as shifts if possible, slightly faster for 64 bit operands
+// leading zeros can be shifted left, then right
+(And64 <t> (Const64 [y]) x) && nlz(y) + nto(y) == 64 && nto(y) >= 32
+  -> (Rsh64Ux64 (Lsh64x64 <t> x (Const64 <t> [nlz(y)])) (Const64 <t> [nlz(y)]))
+// trailing zeros can be shifted right, then left
+(And64 <t> (Const64 [y]) x) && nlo(y) + ntz(y) == 64 && ntz(y) >= 32
+  -> (Lsh64x64 (Rsh64Ux64 <t> x (Const64 <t> [ntz(y)])) (Const64 <t> [ntz(y)]))
+
 // simplifications often used for lengths.  e.g. len(s[i:i+5])==5
 (Sub(64|32|16|8) (Add(64|32|16|8) x y) x) -> y
 (Sub(64|32|16|8) (Add(64|32|16|8) x y) y) -> x

src/cmd/compile/internal/ssa/rewriteARM64.go

@@ -28467,33 +28467,6 @@ func rewriteValueARM64_OpARM64ORshiftRL_0(v *Value) bool {
 		v.AddArg(y)
 		return true
 	}
-	// match: (ORshiftRL [rc] (ANDconst [ac] y) (SLLconst [lc] x))
-	// cond: lc < rc && ac == ^((1<<uint(64-rc)-1))
-	// result: (BFXIL [armBFAuxInt(rc-lc, 64-rc)] y x)
-	for {
-		rc := v.AuxInt
-		_ = v.Args[1]
-		v_0 := v.Args[0]
-		if v_0.Op != OpARM64ANDconst {
-			break
-		}
-		ac := v_0.AuxInt
-		y := v_0.Args[0]
-		v_1 := v.Args[1]
-		if v_1.Op != OpARM64SLLconst {
-			break
-		}
-		lc := v_1.AuxInt
-		x := v_1.Args[0]
-		if !(lc < rc && ac == ^(1<<uint(64-rc)-1)) {
-			break
-		}
-		v.reset(OpARM64BFXIL)
-		v.AuxInt = armBFAuxInt(rc-lc, 64-rc)
-		v.AddArg(y)
-		v.AddArg(x)
-		return true
-	}
 	return false
 }
 func rewriteValueARM64_OpARM64RORWconst_0(v *Value) bool {

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

@@ -5735,6 +5735,112 @@ func rewriteValuegeneric_OpAnd64_10(v *Value) bool {
 		v.AddArg(y)
 		return true
 	}
+	// match: (And64 <t> (Const64 [y]) x)
+	// cond: nlz(y) + nto(y) == 64 && nto(y) >= 32
+	// result: (Rsh64Ux64 (Lsh64x64 <t> x (Const64 <t> [nlz(y)])) (Const64 <t> [nlz(y)]))
+	for {
+		t := v.Type
+		x := v.Args[1]
+		v_0 := v.Args[0]
+		if v_0.Op != OpConst64 {
+			break
+		}
+		y := v_0.AuxInt
+		if !(nlz(y)+nto(y) == 64 && nto(y) >= 32) {
+			break
+		}
+		v.reset(OpRsh64Ux64)
+		v0 := b.NewValue0(v.Pos, OpLsh64x64, t)
+		v0.AddArg(x)
+		v1 := b.NewValue0(v.Pos, OpConst64, t)
+		v1.AuxInt = nlz(y)
+		v0.AddArg(v1)
+		v.AddArg(v0)
+		v2 := b.NewValue0(v.Pos, OpConst64, t)
+		v2.AuxInt = nlz(y)
+		v.AddArg(v2)
+		return true
+	}
+	// match: (And64 <t> x (Const64 [y]))
+	// cond: nlz(y) + nto(y) == 64 && nto(y) >= 32
+	// result: (Rsh64Ux64 (Lsh64x64 <t> x (Const64 <t> [nlz(y)])) (Const64 <t> [nlz(y)]))
+	for {
+		t := v.Type
+		_ = v.Args[1]
+		x := v.Args[0]
+		v_1 := v.Args[1]
+		if v_1.Op != OpConst64 {
+			break
+		}
+		y := v_1.AuxInt
+		if !(nlz(y)+nto(y) == 64 && nto(y) >= 32) {
+			break
+		}
+		v.reset(OpRsh64Ux64)
+		v0 := b.NewValue0(v.Pos, OpLsh64x64, t)
+		v0.AddArg(x)
+		v1 := b.NewValue0(v.Pos, OpConst64, t)
+		v1.AuxInt = nlz(y)
+		v0.AddArg(v1)
+		v.AddArg(v0)
+		v2 := b.NewValue0(v.Pos, OpConst64, t)
+		v2.AuxInt = nlz(y)
+		v.AddArg(v2)
+		return true
+	}
+	// match: (And64 <t> (Const64 [y]) x)
+	// cond: nlo(y) + ntz(y) == 64 && ntz(y) >= 32
+	// result: (Lsh64x64 (Rsh64Ux64 <t> x (Const64 <t> [ntz(y)])) (Const64 <t> [ntz(y)]))
+	for {
+		t := v.Type
+		x := v.Args[1]
+		v_0 := v.Args[0]
+		if v_0.Op != OpConst64 {
+			break
+		}
+		y := v_0.AuxInt
+		if !(nlo(y)+ntz(y) == 64 && ntz(y) >= 32) {
+			break
+		}
+		v.reset(OpLsh64x64)
+		v0 := b.NewValue0(v.Pos, OpRsh64Ux64, t)
+		v0.AddArg(x)
+		v1 := b.NewValue0(v.Pos, OpConst64, t)
+		v1.AuxInt = ntz(y)
+		v0.AddArg(v1)
+		v.AddArg(v0)
+		v2 := b.NewValue0(v.Pos, OpConst64, t)
+		v2.AuxInt = ntz(y)
+		v.AddArg(v2)
+		return true
+	}
+	// match: (And64 <t> x (Const64 [y]))
+	// cond: nlo(y) + ntz(y) == 64 && ntz(y) >= 32
+	// result: (Lsh64x64 (Rsh64Ux64 <t> x (Const64 <t> [ntz(y)])) (Const64 <t> [ntz(y)]))
+	for {
+		t := v.Type
+		_ = v.Args[1]
+		x := v.Args[0]
+		v_1 := v.Args[1]
+		if v_1.Op != OpConst64 {
+			break
+		}
+		y := v_1.AuxInt
+		if !(nlo(y)+ntz(y) == 64 && ntz(y) >= 32) {
+			break
+		}
+		v.reset(OpLsh64x64)
+		v0 := b.NewValue0(v.Pos, OpRsh64Ux64, t)
+		v0.AddArg(x)
+		v1 := b.NewValue0(v.Pos, OpConst64, t)
+		v1.AuxInt = ntz(y)
+		v0.AddArg(v1)
+		v.AddArg(v0)
+		v2 := b.NewValue0(v.Pos, OpConst64, t)
+		v2.AuxInt = ntz(y)
+		v.AddArg(v2)
+		return true
+	}
 	// match: (And64 (And64 i:(Const64 <t>) z) x)
 	// cond: (z.Op != OpConst64 && x.Op != OpConst64)
 	// result: (And64 i (And64 <t> z x))
@@ -5761,6 +5867,10 @@ func rewriteValuegeneric_OpAnd64_10(v *Value) bool {
 		v.AddArg(v0)
 		return true
 	}
+	return false
+}
+func rewriteValuegeneric_OpAnd64_20(v *Value) bool {
+	b := v.Block
 	// match: (And64 (And64 z i:(Const64 <t>)) x)
 	// cond: (z.Op != OpConst64 && x.Op != OpConst64)
 	// result: (And64 i (And64 <t> z x))
@@ -5874,10 +5984,6 @@ func rewriteValuegeneric_OpAnd64_10(v *Value) bool {
 		v.AddArg(x)
 		return true
 	}
-	return false
-}
-func rewriteValuegeneric_OpAnd64_20(v *Value) bool {
-	b := v.Block
 	// match: (And64 (Const64 <t> [c]) (And64 x (Const64 <t> [d])))
 	// cond:
 	// result: (And64 (Const64 <t> [c&d]) x)

test/codegen/math.go

@@ -81,7 +81,7 @@ func abs32(x float32) float32 {
 
 // Check that it's using integer registers
 func copysign(a, b, c float64) {
-	// amd64:"BTRQ\t[$]63","ANDQ","ORQ"
+	// amd64:"BTRQ\t[$]63","SHRQ\t[$]63","SHLQ\t[$]63","ORQ"
 	// s390x:"CPSDR",-"MOVD"         (no integer load/store)
 	// ppc64:"FCPSGN"
 	// ppc64le:"FCPSGN"
@@ -100,7 +100,7 @@ func copysign(a, b, c float64) {
 	// s390x:"LNDFR\t",-"MOVD\t"     (no integer load/store)
 	sink64[2] = math.Float64frombits(math.Float64bits(a) | 1<<63)
 
-	// amd64:"ANDQ","ORQ"
+	// amd64:-"SHLQ\t[$]1",-"SHRQ\t[$]1","SHRQ\t[$]63","SHLQ\t[$]63","ORQ"
 	// s390x:"CPSDR\t",-"MOVD\t"     (no integer load/store)
 	// ppc64:"FCPSGN"
 	// ppc64le:"FCPSGN"