cmd/internal/obj/arm64: fix the bug of incorrect handling negative offset of LDP/STP/LDPW/STPW

The current assembler will report error when the negative offset is in
the range of [-256, 0) and is not the multiples of 4/8.

The fix introduces C_NSAUTO_8, C_NSAUTO_4 and C_NAUTO4K. C_NPAUTO
includes C_NSAUTO_8 instead of C_NSAUTO, C_NAUTO4K includes C_NSAUTO_8,
C_NSAUTO_4 and C_NSAUTO. So that assembler will encode the negative offset
that is greater than -4095 and is not the multiples of 4/8 as two instructions.

Add the test cases.

Fixed #24471

Change-Id: I42f34e3b8a9fc52c9e8b41504294271aafade639
Reviewed-on: https://go-review.googlesource.com/102635
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>

src/cmd/asm/internal/asm/testdata/arm64.s

@@ -540,8 +540,12 @@ again:
 	LDP	1024(RSP), (R1, R2) // fb031091610b40a9
 	LDP.W	8(RSP), (R1, R2)    // e18bc0a9
 	LDP.P	8(RSP), (R1, R2)    // e18bc0a8
+	LDP	-31(R0), (R1, R2)   // 1b7c00d1610b40a9
+	LDP	-4(R0), (R1, R2)    // 1b1000d1610b40a9
+	LDP	-8(R0), (R1, R2)    // 01887fa9
 	LDP	x(SB), (R1, R2)
 	LDP	x+8(SB), (R1, R2)
+	LDPW	-5(R0), (R1, R2)    // 1b1400d1610b4029
 	LDPW	(R0), (R1, R2)      // 01084029
 	LDPW	4(R0), (R1, R2)     // 01884029
 	LDPW	-4(R0), (R1, R2)    // 01887f29
@@ -579,6 +583,7 @@ again:
 	STP.W	(R3, R4), 8(R5)     // a39080a9
 	STP.P	(R3, R4), 8(R5)     // a39080a8
 	STP	(R3, R4), -8(R5)    // a3903fa9
+	STP	(R3, R4), -4(R5)    // bb1000d1631300a9
 	STP	(R3, R4), 11(R0)    // 1b2c0091631300a9
 	STP	(R3, R4), 1024(R0)  // 1b001091631300a9
 	STP	(R3, R4), (RSP)     // e31300a9
@@ -595,6 +600,7 @@ again:
 	STPW.W	(R3, R4), 4(R5)     // a3908029
 	STPW.P	(R3, R4), 4(R5)     // a3908028
 	STPW	(R3, R4), -4(R5)    // a3903f29
+	STPW	(R3, R4), -5(R5)    // bb1400d163130029
 	STPW	(R3, R4), 11(R0)    // 1b2c009163130029
 	STPW	(R3, R4), 1024(R0)  // 1b00109163130029
 	STPW	(R3, R4), (RSP)     // e3130029

src/cmd/internal/obj/arm64/a.out.go

@@ -425,8 +425,11 @@ const (
 	C_LBRA
 
 	C_ZAUTO      // 0(RSP)
-	C_NPAUTO     // -512 <= x < 0, 0 mod 8
+	C_NSAUTO_8   // -256 <= x < 0, 0 mod 8
+	C_NSAUTO_4   // -256 <= x < 0, 0 mod 4
 	C_NSAUTO     // -256 <= x < 0
+	C_NPAUTO     // -512 <= x < 0, 0 mod 8
+	C_NAUTO4K    // -4095 <= x < 0
 	C_PSAUTO_8   // 0 to 255, 0 mod 8
 	C_PSAUTO_4   // 0 to 255, 0 mod 4
 	C_PSAUTO     // 0 to 255
@@ -450,9 +453,12 @@ const (
 	C_SEXT16 // 0 to 65520
 	C_LEXT
 
-	C_ZOREG  // 0(R)
-	C_NPOREG // must mirror NPAUTO, etc
+	C_ZOREG    // 0(R)
+	C_NSOREG_8 // must mirror C_NSAUTO_8, etc
+	C_NSOREG_4
 	C_NSOREG
+	C_NPOREG
+	C_NOREG4K
 	C_PSOREG_8
 	C_PSOREG_4
 	C_PSOREG

src/cmd/internal/obj/arm64/anames7.go

@@ -37,8 +37,11 @@ var cnames7 = []string{
 	"SBRA",
 	"LBRA",
 	"ZAUTO",
-	"NPAUTO",
+	"NSAUTO_8",
+	"NSAUTO_4",
 	"NSAUTO",
+	"NPAUTO",
+	"NAUTO4K",
 	"PSAUTO_8",
 	"PSAUTO_4",
 	"PSAUTO",
@@ -61,8 +64,11 @@ var cnames7 = []string{
 	"SEXT16",
 	"LEXT",
 	"ZOREG",
-	"NPOREG",
+	"NSOREG_8",
+	"NSOREG_4",
 	"NSOREG",
+	"NPOREG",
+	"NOREG4K",
 	"PSOREG_8",
 	"PSOREG_4",
 	"PSOREG",

src/cmd/internal/obj/arm64/asm7.go

@@ -473,6 +473,9 @@ var optab = []Optab{
 	{ALDP, C_UAUTO4K, C_NONE, C_PAIR, 74, 8, REGSP, 0, 0},
 	{ALDP, C_UAUTO4K, C_NONE, C_PAIR, 74, 8, REGSP, 0, C_XPRE},
 	{ALDP, C_UAUTO4K, C_NONE, C_PAIR, 74, 8, REGSP, 0, C_XPOST},
+	{ALDP, C_NAUTO4K, C_NONE, C_PAIR, 74, 8, REGSP, 0, 0},
+	{ALDP, C_NAUTO4K, C_NONE, C_PAIR, 74, 8, REGSP, 0, C_XPRE},
+	{ALDP, C_NAUTO4K, C_NONE, C_PAIR, 74, 8, REGSP, 0, C_XPOST},
 	{ALDP, C_LAUTO, C_NONE, C_PAIR, 75, 12, REGSP, LFROM, 0},
 	{ALDP, C_LAUTO, C_NONE, C_PAIR, 75, 12, REGSP, LFROM, C_XPRE},
 	{ALDP, C_LAUTO, C_NONE, C_PAIR, 75, 12, REGSP, LFROM, C_XPOST},
@@ -485,6 +488,9 @@ var optab = []Optab{
 	{ALDP, C_UOREG4K, C_NONE, C_PAIR, 74, 8, 0, 0, 0},
 	{ALDP, C_UOREG4K, C_NONE, C_PAIR, 74, 8, 0, 0, C_XPRE},
 	{ALDP, C_UOREG4K, C_NONE, C_PAIR, 74, 8, 0, 0, C_XPOST},
+	{ALDP, C_NOREG4K, C_NONE, C_PAIR, 74, 8, 0, 0, 0},
+	{ALDP, C_NOREG4K, C_NONE, C_PAIR, 74, 8, 0, 0, C_XPRE},
+	{ALDP, C_NOREG4K, C_NONE, C_PAIR, 74, 8, 0, 0, C_XPOST},
 	{ALDP, C_LOREG, C_NONE, C_PAIR, 75, 12, 0, LFROM, 0},
 	{ALDP, C_LOREG, C_NONE, C_PAIR, 75, 12, 0, LFROM, C_XPRE},
 	{ALDP, C_LOREG, C_NONE, C_PAIR, 75, 12, 0, LFROM, C_XPOST},
@@ -499,6 +505,9 @@ var optab = []Optab{
 	{ASTP, C_PAIR, C_NONE, C_UAUTO4K, 76, 8, REGSP, 0, 0},
 	{ASTP, C_PAIR, C_NONE, C_UAUTO4K, 76, 8, REGSP, 0, C_XPRE},
 	{ASTP, C_PAIR, C_NONE, C_UAUTO4K, 76, 8, REGSP, 0, C_XPOST},
+	{ASTP, C_PAIR, C_NONE, C_NAUTO4K, 76, 12, REGSP, 0, 0},
+	{ASTP, C_PAIR, C_NONE, C_NAUTO4K, 76, 12, REGSP, 0, C_XPRE},
+	{ASTP, C_PAIR, C_NONE, C_NAUTO4K, 76, 12, REGSP, 0, C_XPOST},
 	{ASTP, C_PAIR, C_NONE, C_LAUTO, 77, 12, REGSP, LTO, 0},
 	{ASTP, C_PAIR, C_NONE, C_LAUTO, 77, 12, REGSP, LTO, C_XPRE},
 	{ASTP, C_PAIR, C_NONE, C_LAUTO, 77, 12, REGSP, LTO, C_XPOST},
@@ -511,59 +520,74 @@ var optab = []Optab{
 	{ASTP, C_PAIR, C_NONE, C_UOREG4K, 76, 8, 0, 0, 0},
 	{ASTP, C_PAIR, C_NONE, C_UOREG4K, 76, 8, 0, 0, C_XPRE},
 	{ASTP, C_PAIR, C_NONE, C_UOREG4K, 76, 8, 0, 0, C_XPOST},
+	{ASTP, C_PAIR, C_NONE, C_NOREG4K, 76, 8, 0, 0, 0},
+	{ASTP, C_PAIR, C_NONE, C_NOREG4K, 76, 8, 0, 0, C_XPRE},
+	{ASTP, C_PAIR, C_NONE, C_NOREG4K, 76, 8, 0, 0, C_XPOST},
 	{ASTP, C_PAIR, C_NONE, C_LOREG, 77, 12, 0, LTO, 0},
 	{ASTP, C_PAIR, C_NONE, C_LOREG, 77, 12, 0, LTO, C_XPRE},
 	{ASTP, C_PAIR, C_NONE, C_LOREG, 77, 12, 0, LTO, C_XPOST},
 	{ASTP, C_PAIR, C_NONE, C_ADDR, 87, 12, 0, 0, 0},
 
 	// differ from LDP/STP for C_NSAUTO_4/C_PSAUTO_4/C_NSOREG_4/C_PSOREG_4
-	{ALDPW, C_NSAUTO, C_NONE, C_PAIR, 66, 4, REGSP, 0, 0},
-	{ALDPW, C_NSAUTO, C_NONE, C_PAIR, 66, 4, REGSP, 0, C_XPRE},
-	{ALDPW, C_NSAUTO, C_NONE, C_PAIR, 66, 4, REGSP, 0, C_XPOST},
+	{ALDPW, C_NSAUTO_4, C_NONE, C_PAIR, 66, 4, REGSP, 0, 0},
+	{ALDPW, C_NSAUTO_4, C_NONE, C_PAIR, 66, 4, REGSP, 0, C_XPRE},
+	{ALDPW, C_NSAUTO_4, C_NONE, C_PAIR, 66, 4, REGSP, 0, C_XPOST},
 	{ALDPW, C_PSAUTO_4, C_NONE, C_PAIR, 66, 4, REGSP, 0, 0},
 	{ALDPW, C_PSAUTO_4, C_NONE, C_PAIR, 66, 4, REGSP, 0, C_XPRE},
 	{ALDPW, C_PSAUTO_4, C_NONE, C_PAIR, 66, 4, REGSP, 0, C_XPOST},
 	{ALDPW, C_UAUTO4K, C_NONE, C_PAIR, 74, 8, REGSP, 0, 0},
 	{ALDPW, C_UAUTO4K, C_NONE, C_PAIR, 74, 8, REGSP, 0, C_XPRE},
 	{ALDPW, C_UAUTO4K, C_NONE, C_PAIR, 74, 8, REGSP, 0, C_XPOST},
+	{ALDPW, C_NAUTO4K, C_NONE, C_PAIR, 74, 8, REGSP, 0, 0},
+	{ALDPW, C_NAUTO4K, C_NONE, C_PAIR, 74, 8, REGSP, 0, C_XPRE},
+	{ALDPW, C_NAUTO4K, C_NONE, C_PAIR, 74, 8, REGSP, 0, C_XPOST},
 	{ALDPW, C_LAUTO, C_NONE, C_PAIR, 75, 12, REGSP, LFROM, 0},
 	{ALDPW, C_LAUTO, C_NONE, C_PAIR, 75, 12, REGSP, LFROM, C_XPRE},
 	{ALDPW, C_LAUTO, C_NONE, C_PAIR, 75, 12, REGSP, LFROM, C_XPOST},
-	{ALDPW, C_NSOREG, C_NONE, C_PAIR, 66, 4, 0, 0, 0},
-	{ALDPW, C_NSOREG, C_NONE, C_PAIR, 66, 4, 0, 0, C_XPRE},
-	{ALDPW, C_NSOREG, C_NONE, C_PAIR, 66, 4, 0, 0, C_XPOST},
+	{ALDPW, C_NSOREG_4, C_NONE, C_PAIR, 66, 4, 0, 0, 0},
+	{ALDPW, C_NSOREG_4, C_NONE, C_PAIR, 66, 4, 0, 0, C_XPRE},
+	{ALDPW, C_NSOREG_4, C_NONE, C_PAIR, 66, 4, 0, 0, C_XPOST},
 	{ALDPW, C_PSOREG_4, C_NONE, C_PAIR, 66, 4, 0, 0, 0},
 	{ALDPW, C_PSOREG_4, C_NONE, C_PAIR, 66, 4, 0, 0, C_XPRE},
 	{ALDPW, C_PSOREG_4, C_NONE, C_PAIR, 66, 4, 0, 0, C_XPOST},
 	{ALDPW, C_UOREG4K, C_NONE, C_PAIR, 74, 8, 0, 0, 0},
 	{ALDPW, C_UOREG4K, C_NONE, C_PAIR, 74, 8, 0, 0, C_XPRE},
 	{ALDPW, C_UOREG4K, C_NONE, C_PAIR, 74, 8, 0, 0, C_XPOST},
+	{ALDPW, C_NOREG4K, C_NONE, C_PAIR, 74, 8, 0, 0, 0},
+	{ALDPW, C_NOREG4K, C_NONE, C_PAIR, 74, 8, 0, 0, C_XPRE},
+	{ALDPW, C_NOREG4K, C_NONE, C_PAIR, 74, 8, 0, 0, C_XPOST},
 	{ALDPW, C_LOREG, C_NONE, C_PAIR, 75, 12, 0, LFROM, 0},
 	{ALDPW, C_LOREG, C_NONE, C_PAIR, 75, 12, 0, LFROM, C_XPRE},
 	{ALDPW, C_LOREG, C_NONE, C_PAIR, 75, 12, 0, LFROM, C_XPOST},
 	{ALDPW, C_ADDR, C_NONE, C_PAIR, 88, 12, 0, 0, 0},
 
-	{ASTPW, C_PAIR, C_NONE, C_NSAUTO, 67, 4, REGSP, 0, 0},
-	{ASTPW, C_PAIR, C_NONE, C_NSAUTO, 67, 4, REGSP, 0, C_XPRE},
-	{ASTPW, C_PAIR, C_NONE, C_NSAUTO, 67, 4, REGSP, 0, C_XPOST},
+	{ASTPW, C_PAIR, C_NONE, C_NSAUTO_4, 67, 4, REGSP, 0, 0},
+	{ASTPW, C_PAIR, C_NONE, C_NSAUTO_4, 67, 4, REGSP, 0, C_XPRE},
+	{ASTPW, C_PAIR, C_NONE, C_NSAUTO_4, 67, 4, REGSP, 0, C_XPOST},
 	{ASTPW, C_PAIR, C_NONE, C_PSAUTO_4, 67, 4, REGSP, 0, 0},
 	{ASTPW, C_PAIR, C_NONE, C_PSAUTO_4, 67, 4, REGSP, 0, C_XPRE},
 	{ASTPW, C_PAIR, C_NONE, C_PSAUTO_4, 67, 4, REGSP, 0, C_XPOST},
 	{ASTPW, C_PAIR, C_NONE, C_UAUTO4K, 76, 8, REGSP, 0, 0},
 	{ASTPW, C_PAIR, C_NONE, C_UAUTO4K, 76, 8, REGSP, 0, C_XPRE},
 	{ASTPW, C_PAIR, C_NONE, C_UAUTO4K, 76, 8, REGSP, 0, C_XPOST},
+	{ASTPW, C_PAIR, C_NONE, C_NAUTO4K, 76, 12, REGSP, 0, 0},
+	{ASTPW, C_PAIR, C_NONE, C_NAUTO4K, 76, 12, REGSP, 0, C_XPRE},
+	{ASTPW, C_PAIR, C_NONE, C_NAUTO4K, 76, 12, REGSP, 0, C_XPOST},
 	{ASTPW, C_PAIR, C_NONE, C_LAUTO, 77, 12, REGSP, LTO, 0},
 	{ASTPW, C_PAIR, C_NONE, C_LAUTO, 77, 12, REGSP, LTO, C_XPRE},
 	{ASTPW, C_PAIR, C_NONE, C_LAUTO, 77, 12, REGSP, LTO, C_XPOST},
-	{ASTPW, C_PAIR, C_NONE, C_NSOREG, 67, 4, 0, 0, 0},
-	{ASTPW, C_PAIR, C_NONE, C_NSOREG, 67, 4, 0, 0, C_XPRE},
-	{ASTPW, C_PAIR, C_NONE, C_NSOREG, 67, 4, 0, 0, C_XPOST},
+	{ASTPW, C_PAIR, C_NONE, C_NSOREG_4, 67, 4, 0, 0, 0},
+	{ASTPW, C_PAIR, C_NONE, C_NSOREG_4, 67, 4, 0, 0, C_XPRE},
+	{ASTPW, C_PAIR, C_NONE, C_NSOREG_4, 67, 4, 0, 0, C_XPOST},
 	{ASTPW, C_PAIR, C_NONE, C_PSOREG_4, 67, 4, 0, 0, 0},
 	{ASTPW, C_PAIR, C_NONE, C_PSOREG_4, 67, 4, 0, 0, C_XPRE},
 	{ASTPW, C_PAIR, C_NONE, C_PSOREG_4, 67, 4, 0, 0, C_XPOST},
 	{ASTPW, C_PAIR, C_NONE, C_UOREG4K, 76, 8, 0, 0, 0},
 	{ASTPW, C_PAIR, C_NONE, C_UOREG4K, 76, 8, 0, 0, C_XPRE},
 	{ASTPW, C_PAIR, C_NONE, C_UOREG4K, 76, 8, 0, 0, C_XPOST},
+	{ASTPW, C_PAIR, C_NONE, C_NOREG4K, 76, 8, 0, 0, 0},
+	{ASTPW, C_PAIR, C_NONE, C_NOREG4K, 76, 8, 0, 0, C_XPRE},
+	{ASTPW, C_PAIR, C_NONE, C_NOREG4K, 76, 8, 0, 0, C_XPOST},
 	{ASTPW, C_PAIR, C_NONE, C_LOREG, 77, 12, 0, LTO, 0},
 	{ASTPW, C_PAIR, C_NONE, C_LOREG, 77, 12, 0, LTO, C_XPRE},
 	{ASTPW, C_PAIR, C_NONE, C_LOREG, 77, 12, 0, LTO, C_XPOST},
@@ -996,8 +1020,11 @@ func (c *ctxt7) addpool(p *obj.Prog, a *obj.Addr) {
 		C_UAUTO16K_8,
 		C_UAUTO16K,
 		C_UAUTO32K,
+		C_NSAUTO_8,
+		C_NSAUTO_4,
 		C_NSAUTO,
 		C_NPAUTO,
+		C_NAUTO4K,
 		C_LAUTO,
 		C_PPOREG,
 		C_PSOREG,
@@ -1013,8 +1040,11 @@ func (c *ctxt7) addpool(p *obj.Prog, a *obj.Addr) {
 		C_UOREG16K_8,
 		C_UOREG16K,
 		C_UOREG32K,
+		C_NSOREG_8,
+		C_NSOREG_4,
 		C_NSOREG,
 		C_NPOREG,
+		C_NOREG4K,
 		C_LOREG,
 		C_LACON,
 		C_LCON,
@@ -1221,12 +1251,21 @@ func autoclass(l int64) int {
 	}
 
 	if l < 0 {
+		if l >= -256 && (l&7) == 0 {
+			return C_NSAUTO_8
+		}
+		if l >= -256 && (l&3) == 0 {
+			return C_NSAUTO_4
+		}
 		if l >= -256 {
 			return C_NSAUTO
 		}
 		if l >= -512 && (l&7) == 0 {
 			return C_NPAUTO
 		}
+		if l >= -4095 {
+			return C_NAUTO4K
+		}
 		return C_LAUTO
 	}
 
@@ -1649,63 +1688,113 @@ func cmp(a int, b int) bool {
 			return true
 		}
 
+	case C_NSAUTO_4:
+		if b == C_NSAUTO_8 {
+			return true
+		}
+
+	case C_NSAUTO:
+		switch b {
+		case C_NSAUTO_4, C_NSAUTO_8:
+			return true
+		}
+
+	case C_NPAUTO:
+		switch b {
+		case C_NSAUTO_8:
+			return true
+		}
+
+	case C_NAUTO4K:
+		switch b {
+		case C_NSAUTO_8, C_NSAUTO_4, C_NSAUTO, C_NPAUTO:
+			return true
+		}
+
 	case C_PSAUTO_8:
 		if b == C_ZAUTO {
 			return true
 		}
 
 	case C_PSAUTO_4:
-		if b == C_ZAUTO || b == C_PSAUTO_8 {
+		switch b {
+		case C_ZAUTO, C_PSAUTO_8:
 			return true
 		}
 
 	case C_PSAUTO:
-		if b == C_ZAUTO || b == C_PSAUTO_8 || b == C_PSAUTO_4 {
+		switch b {
+		case C_ZAUTO, C_PSAUTO_8, C_PSAUTO_4:
 			return true
 		}
 
 	case C_PPAUTO:
-		if b == C_ZAUTO || b == C_PSAUTO_8 {
+		switch b {
+		case C_ZAUTO, C_PSAUTO_8:
 			return true
 		}
 
 	case C_UAUTO4K:
 		switch b {
-		case C_ZAUTO, C_PSAUTO, C_PSAUTO_4, C_PSAUTO_8, C_PPAUTO, C_UAUTO4K_2, C_UAUTO4K_4, C_UAUTO4K_8:
+		case C_ZAUTO, C_PSAUTO, C_PSAUTO_4, C_PSAUTO_8,
+			C_PPAUTO, C_UAUTO4K_2, C_UAUTO4K_4, C_UAUTO4K_8:
 			return true
 		}
 
 	case C_UAUTO8K:
 		switch b {
-		case C_ZAUTO, C_PSAUTO, C_PSAUTO_4, C_PSAUTO_8, C_PPAUTO, C_UAUTO4K_2, C_UAUTO4K_4, C_UAUTO4K_8, C_UAUTO8K_4, C_UAUTO8K_8:
+		case C_ZAUTO, C_PSAUTO, C_PSAUTO_4, C_PSAUTO_8, C_PPAUTO,
+			C_UAUTO4K_2, C_UAUTO4K_4, C_UAUTO4K_8, C_UAUTO8K_4, C_UAUTO8K_8:
 			return true
 		}
 
 	case C_UAUTO16K:
 		switch b {
-		case C_ZAUTO, C_PSAUTO, C_PSAUTO_4, C_PSAUTO_8, C_PPAUTO, C_UAUTO4K_4, C_UAUTO4K_8, C_UAUTO8K_4, C_UAUTO8K_8, C_UAUTO16K_8:
+		case C_ZAUTO, C_PSAUTO, C_PSAUTO_4, C_PSAUTO_8, C_PPAUTO,
+			C_UAUTO4K_4, C_UAUTO4K_8, C_UAUTO8K_4, C_UAUTO8K_8, C_UAUTO16K_8:
 			return true
 		}
 
 	case C_UAUTO32K:
 		switch b {
-		case C_ZAUTO, C_PSAUTO, C_PSAUTO_4, C_PSAUTO_8, C_PPAUTO, C_UAUTO4K_8, C_UAUTO8K_8, C_UAUTO16K_8:
+		case C_ZAUTO, C_PSAUTO, C_PSAUTO_4, C_PSAUTO_8,
+			C_PPAUTO, C_UAUTO4K_8, C_UAUTO8K_8, C_UAUTO16K_8:
 			return true
 		}
 
-	case C_NPAUTO:
-		return cmp(C_NSAUTO, b)
-
 	case C_LAUTO:
 		switch b {
-		case C_ZAUTO, C_PSAUTO, C_PSAUTO_4, C_PSAUTO_8, C_PPAUTO,
+		case C_ZAUTO, C_NSAUTO, C_NSAUTO_4, C_NSAUTO_8, C_NPAUTO,
+			C_NAUTO4K, C_PSAUTO, C_PSAUTO_4, C_PSAUTO_8, C_PPAUTO,
 			C_UAUTO4K, C_UAUTO4K_2, C_UAUTO4K_4, C_UAUTO4K_8,
 			C_UAUTO8K, C_UAUTO8K_4, C_UAUTO8K_8,
 			C_UAUTO16K, C_UAUTO16K_8,
 			C_UAUTO32K:
 			return true
 		}
-		return cmp(C_NPAUTO, b)
+
+	case C_NSOREG_4:
+		if b == C_NSOREG_8 {
+			return true
+		}
+
+	case C_NSOREG:
+		switch b {
+		case C_NSOREG_4, C_NSOREG_8:
+			return true
+		}
+
+	case C_NPOREG:
+		switch b {
+		case C_NSOREG_8:
+			return true
+		}
+
+	case C_NOREG4K:
+		switch b {
+		case C_NSOREG_8, C_NSOREG_4, C_NSOREG, C_NPOREG:
+			return true
+		}
 
 	case C_PSOREG_4:
 		switch b {
@@ -1727,41 +1816,44 @@ func cmp(a int, b int) bool {
 
 	case C_UOREG4K:
 		switch b {
-		case C_ZOREG, C_PSOREG_4, C_PSOREG_8, C_PSOREG, C_PPOREG, C_UOREG4K_2, C_UOREG4K_4, C_UOREG4K_8:
+		case C_ZOREG, C_PSOREG_4, C_PSOREG_8, C_PSOREG,
+			C_PPOREG, C_UOREG4K_2, C_UOREG4K_4, C_UOREG4K_8:
 			return true
 		}
 
 	case C_UOREG8K:
 		switch b {
-		case C_ZOREG, C_PSOREG_4, C_PSOREG_8, C_PSOREG, C_PPOREG, C_UOREG4K_2, C_UOREG4K_4, C_UOREG4K_8, C_UOREG8K_4, C_UOREG8K_8:
+		case C_ZOREG, C_PSOREG_4, C_PSOREG_8, C_PSOREG,
+			C_PPOREG, C_UOREG4K_2, C_UOREG4K_4, C_UOREG4K_8,
+			C_UOREG8K_4, C_UOREG8K_8:
 			return true
 		}
 
 	case C_UOREG16K:
 		switch b {
-		case C_ZOREG, C_PSOREG_4, C_PSOREG_8, C_PSOREG, C_PPOREG, C_UOREG4K_4, C_UOREG4K_8, C_UOREG8K_4, C_UOREG8K_8, C_UOREG16K_8:
+		case C_ZOREG, C_PSOREG_4, C_PSOREG_8, C_PSOREG,
+			C_PPOREG, C_UOREG4K_4, C_UOREG4K_8, C_UOREG8K_4,
+			C_UOREG8K_8, C_UOREG16K_8:
 			return true
 		}
 
 	case C_UOREG32K:
 		switch b {
-		case C_ZOREG, C_PSOREG_4, C_PSOREG_8, C_PSOREG, C_PPOREG, C_UOREG4K_8, C_UOREG8K_8, C_UOREG16K_8:
+		case C_ZOREG, C_PSOREG_4, C_PSOREG_8, C_PSOREG,
+			C_PPOREG, C_UOREG4K_8, C_UOREG8K_8, C_UOREG16K_8:
 			return true
 		}
 
-	case C_NPOREG:
-		return cmp(C_NSOREG, b)
-
 	case C_LOREG:
 		switch b {
-		case C_ZOREG, C_PSOREG_4, C_PSOREG_8, C_PSOREG, C_PPOREG,
+		case C_ZOREG, C_NSOREG, C_NSOREG_4, C_NSOREG_8, C_NPOREG,
+			C_NOREG4K, C_PSOREG_4, C_PSOREG_8, C_PSOREG, C_PPOREG,
 			C_UOREG4K, C_UOREG4K_2, C_UOREG4K_4, C_UOREG4K_8,
 			C_UOREG8K, C_UOREG8K_4, C_UOREG8K_8,
 			C_UOREG16K, C_UOREG16K_8,
 			C_UOREG32K:
 			return true
 		}
-		return cmp(C_NPOREG, b)
 
 	case C_LBRA:
 		if b == C_SBRA {
@@ -3614,20 +3706,29 @@ func (c *ctxt7) asmout(p *obj.Prog, o *Optab, out []uint32) {
 		o1 |= (uint32(imm5&0x1f) << 16) | (uint32(rf&31) << 5) | uint32(rt&31)
 
 	case 74:
-		//	add $O, R, Rtmp
+		//	add $O, R, Rtmp or sub $O, R, Rtmp
 		//	ldp (Rtmp), (R1, R2)
 		r := int(p.From.Reg)
 		if r == obj.REG_NONE {
 			r = int(o.param)
 		}
 		if r == obj.REG_NONE {
-			c.ctxt.Diag("invalid ldp source: %v\n", p)
+			c.ctxt.Diag("invalid ldp source: %v", p)
 		}
 		v := int32(c.regoff(&p.From))
-		if v < 0 || v > 4095 {
-			c.ctxt.Diag("offset out of range%v\n", p)
+
+		if v > 0 {
+			if v > 4095 {
+				c.ctxt.Diag("offset out of range: %v", p)
+			}
+			o1 = c.oaddi(p, int32(c.opirr(p, AADD)), v, r, REGTMP)
+		}
+		if v < 0 {
+			if v < -4095 {
+				c.ctxt.Diag("offset out of range: %v", p)
+			}
+			o1 = c.oaddi(p, int32(c.opirr(p, ASUB)), -v, r, REGTMP)
 		}
-		o1 = c.oaddi(p, int32(c.opirr(p, AADD)), v, r, REGTMP)
 		o2 |= c.opldpstp(p, o, 0, uint32(REGTMP), uint32(p.To.Reg), uint32(p.To.Offset), 1)
 
 	case 75:
@@ -3639,7 +3740,7 @@ func (c *ctxt7) asmout(p *obj.Prog, o *Optab, out []uint32) {
 			r = int(o.param)
 		}
 		if r == obj.REG_NONE {
-			c.ctxt.Diag("invalid ldp source: %v\n", p)
+			c.ctxt.Diag("invalid ldp source: %v", p)
 		}
 		o1 = c.omovlit(AMOVD, p, &p.From, REGTMP)
 		o2 = c.opxrrr(p, AADD, false)
@@ -3649,20 +3750,28 @@ func (c *ctxt7) asmout(p *obj.Prog, o *Optab, out []uint32) {
 		o3 |= c.opldpstp(p, o, 0, uint32(REGTMP), uint32(p.To.Reg), uint32(p.To.Offset), 1)
 
 	case 76:
-		//	add $O, R, Rtmp
+		//	add $O, R, Rtmp or sub $O, R, Rtmp
 		//	stp (R1, R2), (Rtmp)
 		r := int(p.To.Reg)
 		if r == obj.REG_NONE {
 			r = int(o.param)
 		}
 		if r == obj.REG_NONE {
-			c.ctxt.Diag("invalid stp destination: %v\n", p)
+			c.ctxt.Diag("invalid stp destination: %v", p)
 		}
 		v := int32(c.regoff(&p.To))
-		if v < 0 || v > 4095 {
-			c.ctxt.Diag("offset out of range%v\n", p)
+		if v > 0 {
+			if v > 4095 {
+				c.ctxt.Diag("offset out of range: %v", p)
+			}
+			o1 = c.oaddi(p, int32(c.opirr(p, AADD)), v, r, REGTMP)
+		}
+		if v < 0 {
+			if v < -4095 {
+				c.ctxt.Diag("offset out of range: %v", p)
+			}
+			o1 = c.oaddi(p, int32(c.opirr(p, ASUB)), -v, r, REGTMP)
 		}
-		o1 = c.oaddi(p, int32(c.opirr(p, AADD)), v, r, REGTMP)
 		o2 |= c.opldpstp(p, o, 0, uint32(REGTMP), uint32(p.From.Reg), uint32(p.From.Offset), 0)
 
 	case 77:
@@ -3674,7 +3783,7 @@ func (c *ctxt7) asmout(p *obj.Prog, o *Optab, out []uint32) {
 			r = int(o.param)
 		}
 		if r == obj.REG_NONE {
-			c.ctxt.Diag("invalid stp destination: %v\n", p)
+			c.ctxt.Diag("invalid stp destination: %v", p)
 		}
 		o1 = c.omovlit(AMOVD, p, &p.To, REGTMP)
 		o2 = c.opxrrr(p, AADD, false)