diff --git a/src/cmd/compile/internal/ssa/_gen/AMD64.rules b/src/cmd/compile/internal/ssa/_gen/AMD64.rules
index f9adaf8fc4..4988af6626 100644
--- a/src/cmd/compile/internal/ssa/_gen/AMD64.rules
+++ b/src/cmd/compile/internal/ssa/_gen/AMD64.rules
@@ -1602,6 +1602,7 @@
 
 // CPUID feature: BMI1.
 (AND(Q|L) x (NOT(Q|L) y))               && buildcfg.GOAMD64 >= 3 => (ANDN(Q|L) x y)
+(SUB(Q|L) x (AND(Q|L) x y))             && buildcfg.GOAMD64 >= 3 => (ANDN(Q|L) x y)
 (AND(Q|L) x (NEG(Q|L) x))               && buildcfg.GOAMD64 >= 3 => (BLSI(Q|L) x)
 (XOR(Q|L) x (ADD(Q|L)const [-1] x))     && buildcfg.GOAMD64 >= 3 => (BLSMSK(Q|L) x)
 (AND(Q|L) <t> x (ADD(Q|L)const [-1] x)) && buildcfg.GOAMD64 >= 3 => (Select0 <t> (BLSR(Q|L) x))
diff --git a/src/cmd/compile/internal/ssa/_gen/generic.rules b/src/cmd/compile/internal/ssa/_gen/generic.rules
index d70496c278..ea5a7f6e44 100644
--- a/src/cmd/compile/internal/ssa/_gen/generic.rules
+++ b/src/cmd/compile/internal/ssa/_gen/generic.rules
@@ -708,6 +708,9 @@
 (Or(64|32|16|8) x (Or(64|32|16|8) x y)) => (Or(64|32|16|8) x y)
 (Xor(64|32|16|8) x (Xor(64|32|16|8) x y)) => y
 
+// x-(x&y) == x&^y
+(Sub(64|32|16|8) x (And(64|32|16|8) <t> x y)) => (And(64|32|16|8) x (Com(64|32|16|8) <t> y))
+
 // Fold comparisons with numeric bounds
 (Less(64|32|16|8)U _ (Const(64|32|16|8) [0]))  => (ConstBool [false])
 (Leq(64|32|16|8)U (Const(64|32|16|8) [0]) _)   => (ConstBool [true])
diff --git a/src/cmd/compile/internal/ssa/rewriteAMD64.go b/src/cmd/compile/internal/ssa/rewriteAMD64.go
index d59a29af93..8f501c19d2 100644
--- a/src/cmd/compile/internal/ssa/rewriteAMD64.go
+++ b/src/cmd/compile/internal/ssa/rewriteAMD64.go
@@ -40193,6 +40193,31 @@ func rewriteValueAMD64_OpAMD64SUBL(v *Value) bool {
 		v.AddArg3(x, ptr, mem)
 		return true
 	}
+	// match: (SUBL x (ANDL x y))
+	// cond: buildcfg.GOAMD64 >= 3
+	// result: (ANDNL x y)
+	for {
+		x := v_0
+		if v_1.Op != OpAMD64ANDL {
+			break
+		}
+		_ = v_1.Args[1]
+		v_1_0 := v_1.Args[0]
+		v_1_1 := v_1.Args[1]
+		for _i0 := 0; _i0 <= 1; _i0, v_1_0, v_1_1 = _i0+1, v_1_1, v_1_0 {
+			if x != v_1_0 {
+				continue
+			}
+			y := v_1_1
+			if !(buildcfg.GOAMD64 >= 3) {
+				continue
+			}
+			v.reset(OpAMD64ANDNL)
+			v.AddArg2(x, y)
+			return true
+		}
+		break
+	}
 	return false
 }
 func rewriteValueAMD64_OpAMD64SUBLconst(v *Value) bool {
@@ -40435,6 +40460,31 @@ func rewriteValueAMD64_OpAMD64SUBQ(v *Value) bool {
 		v.AddArg3(x, ptr, mem)
 		return true
 	}
+	// match: (SUBQ x (ANDQ x y))
+	// cond: buildcfg.GOAMD64 >= 3
+	// result: (ANDNQ x y)
+	for {
+		x := v_0
+		if v_1.Op != OpAMD64ANDQ {
+			break
+		}
+		_ = v_1.Args[1]
+		v_1_0 := v_1.Args[0]
+		v_1_1 := v_1.Args[1]
+		for _i0 := 0; _i0 <= 1; _i0, v_1_0, v_1_1 = _i0+1, v_1_1, v_1_0 {
+			if x != v_1_0 {
+				continue
+			}
+			y := v_1_1
+			if !(buildcfg.GOAMD64 >= 3) {
+				continue
+			}
+			v.reset(OpAMD64ANDNQ)
+			v.AddArg2(x, y)
+			return true
+		}
+		break
+	}
 	return false
 }
 func rewriteValueAMD64_OpAMD64SUBQborrow(v *Value) bool {
diff --git a/src/cmd/compile/internal/ssa/rewritegeneric.go b/src/cmd/compile/internal/ssa/rewritegeneric.go
index 89d58f7db0..24d80d9aa7 100644
--- a/src/cmd/compile/internal/ssa/rewritegeneric.go
+++ b/src/cmd/compile/internal/ssa/rewritegeneric.go
@@ -35967,6 +35967,30 @@ func rewriteValuegeneric_OpSub16(v *Value) bool {
 		}
 		break
 	}
+	// match: (Sub16 x (And16 <t> x y))
+	// result: (And16 x (Com16 <t> y))
+	for {
+		x := v_0
+		if v_1.Op != OpAnd16 {
+			break
+		}
+		t := v_1.Type
+		_ = v_1.Args[1]
+		v_1_0 := v_1.Args[0]
+		v_1_1 := v_1.Args[1]
+		for _i0 := 0; _i0 <= 1; _i0, v_1_0, v_1_1 = _i0+1, v_1_1, v_1_0 {
+			if x != v_1_0 {
+				continue
+			}
+			y := v_1_1
+			v.reset(OpAnd16)
+			v0 := b.NewValue0(v.Pos, OpCom16, t)
+			v0.AddArg(y)
+			v.AddArg2(x, v0)
+			return true
+		}
+		break
+	}
 	// match: (Sub16 (Add16 x y) x)
 	// result: y
 	for {
@@ -36340,6 +36364,30 @@ func rewriteValuegeneric_OpSub32(v *Value) bool {
 		}
 		break
 	}
+	// match: (Sub32 x (And32 <t> x y))
+	// result: (And32 x (Com32 <t> y))
+	for {
+		x := v_0
+		if v_1.Op != OpAnd32 {
+			break
+		}
+		t := v_1.Type
+		_ = v_1.Args[1]
+		v_1_0 := v_1.Args[0]
+		v_1_1 := v_1.Args[1]
+		for _i0 := 0; _i0 <= 1; _i0, v_1_0, v_1_1 = _i0+1, v_1_1, v_1_0 {
+			if x != v_1_0 {
+				continue
+			}
+			y := v_1_1
+			v.reset(OpAnd32)
+			v0 := b.NewValue0(v.Pos, OpCom32, t)
+			v0.AddArg(y)
+			v.AddArg2(x, v0)
+			return true
+		}
+		break
+	}
 	// match: (Sub32 (Add32 x y) x)
 	// result: y
 	for {
@@ -36737,6 +36785,30 @@ func rewriteValuegeneric_OpSub64(v *Value) bool {
 		}
 		break
 	}
+	// match: (Sub64 x (And64 <t> x y))
+	// result: (And64 x (Com64 <t> y))
+	for {
+		x := v_0
+		if v_1.Op != OpAnd64 {
+			break
+		}
+		t := v_1.Type
+		_ = v_1.Args[1]
+		v_1_0 := v_1.Args[0]
+		v_1_1 := v_1.Args[1]
+		for _i0 := 0; _i0 <= 1; _i0, v_1_0, v_1_1 = _i0+1, v_1_1, v_1_0 {
+			if x != v_1_0 {
+				continue
+			}
+			y := v_1_1
+			v.reset(OpAnd64)
+			v0 := b.NewValue0(v.Pos, OpCom64, t)
+			v0.AddArg(y)
+			v.AddArg2(x, v0)
+			return true
+		}
+		break
+	}
 	// match: (Sub64 (Add64 x y) x)
 	// result: y
 	for {
@@ -37134,6 +37206,30 @@ func rewriteValuegeneric_OpSub8(v *Value) bool {
 		}
 		break
 	}
+	// match: (Sub8 x (And8 <t> x y))
+	// result: (And8 x (Com8 <t> y))
+	for {
+		x := v_0
+		if v_1.Op != OpAnd8 {
+			break
+		}
+		t := v_1.Type
+		_ = v_1.Args[1]
+		v_1_0 := v_1.Args[0]
+		v_1_1 := v_1.Args[1]
+		for _i0 := 0; _i0 <= 1; _i0, v_1_0, v_1_1 = _i0+1, v_1_1, v_1_0 {
+			if x != v_1_0 {
+				continue
+			}
+			y := v_1_1
+			v.reset(OpAnd8)
+			v0 := b.NewValue0(v.Pos, OpCom8, t)
+			v0.AddArg(y)
+			v.AddArg2(x, v0)
+			return true
+		}
+		break
+	}
 	// match: (Sub8 (Add8 x y) x)
 	// result: y
 	for {
diff --git a/test/codegen/bmi.go b/test/codegen/bmi.go
index c50f15a7e0..3c00dd328b 100644
--- a/test/codegen/bmi.go
+++ b/test/codegen/bmi.go
@@ -207,3 +207,15 @@ func shlrx32_load(x []uint32, i int, s uint32) uint32 {
 	s = x[i+1] << s
 	return s
 }
+
+func XSubXandYAndn64(x, y uint64) uint64 {
+	// amd64/v1:"ANDQ" "NOTQ" -"SUBQ"
+	// amd64/v3:"ANDNQ" -"SUBQ" -"ANDQ"
+	return x - (x & y)
+}
+
+func XSubXandYAndn32(x, y uint32) uint32 {
+	// amd64/v1:"ANDL" "NOTL" -"SUBL"
+	// amd64/v3:"ANDNL" -"SUBL" -"ANDL"
+	return x - (x & y)
+}