cmd/compile: don't use duffcopy and duffzero on Plan 9

The ssa compiler uses the duffcopy and duffzero functions,
which rely on the MOVUPS instructions.

However, this doesn't work on Plan 9, since floating point
operations are not allowed in the note handler.

This change disables the use of duffcopy and duffzero
on Plan 9 in the ssa compiler.

Updates #14605.

Change-Id: I017f8ff83de00eabaf7e146b4344a863db1dfddc
Reviewed-on: https://go-review.googlesource.com/20171Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: David du Colombier <0intro@gmail.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>

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

@@ -22,6 +22,7 @@ type Config struct {
 	HTML         *HTMLWriter                // html writer, for debugging
 	ctxt         *obj.Link                  // Generic arch information
 	optimize     bool                       // Do optimization
+	noDuffDevice bool                       // Don't use Duff's device
 	curFunc      *Func
 
 	// TODO: more stuff. Compiler flags of interest, ...
@@ -122,6 +123,12 @@ func NewConfig(arch string, fe Frontend, ctxt *obj.Link, optimize bool) *Config
 	c.ctxt = ctxt
 	c.optimize = optimize
 
+	// Don't use Duff's device on Plan 9, because floating
+	// point operations are not allowed in note handler.
+	if obj.Getgoos() == "plan9" {
+		c.noDuffDevice = true
+	}
+
 	// Assign IDs to preallocated values/blocks.
 	for i := range c.values {
 		c.values[i].ID = ID(i)

src/cmd/compile/internal/ssa/gen/AMD64.rules

@@ -328,7 +328,7 @@
 		(MOVOstore dst (MOVOload src mem) mem))
 
 // Medium copying uses a duff device.
-(Move [size] dst src mem) && size >= 32 && size <= 16*64 && size%16 == 0 ->
+(Move [size] dst src mem) && size >= 32 && size <= 16*64 && size%16 == 0 && !config.noDuffDevice ->
 	(DUFFCOPY [14*(64-size/16)] dst src mem)
 // 14 and 64 are magic constants.  14 is the number of bytes to encode:
 //	MOVUPS	(SI), X0
@@ -338,7 +338,7 @@
 // and 64 is the number of such blocks. See src/runtime/duff_amd64.s:duffcopy.
 
 // Large copying uses REP MOVSQ.
-(Move [size] dst src mem) && size > 16*64 && size%8 == 0 ->
+(Move [size] dst src mem) && (size > 16*64 || config.noDuffDevice) && size%8 == 0 ->
 	(REPMOVSQ dst src (MOVQconst [size/8]) mem)
 
 (Not x) -> (XORBconst [1] x)
@@ -869,13 +869,13 @@
 				(MOVQstoreconst [0] destptr mem))))
 
 // Medium zeroing uses a duff device.
-(Zero [size] destptr mem) && size <= 1024 && size%8 == 0 && size%16 != 0 ->
+(Zero [size] destptr mem) && size <= 1024 && size%8 == 0 && size%16 != 0 && !config.noDuffDevice ->
 	(Zero [size-8] (ADDQconst [8] destptr) (MOVQstore destptr (MOVQconst [0]) mem))
-(Zero [size] destptr mem) && size <= 1024 && size%16 == 0 ->
+(Zero [size] destptr mem) && size <= 1024 && size%16 == 0 && !config.noDuffDevice ->
 	(DUFFZERO [duffStart(size)] (ADDQconst [duffAdj(size)] destptr) (MOVOconst [0]) mem)
 
 // Large zeroing uses REP STOSQ.
-(Zero [size] destptr mem) && size > 1024 && size%8 == 0 ->
+(Zero [size] destptr mem) && (size > 1024 || (config.noDuffDevice && size > 32)) && size%8 == 0 ->
 	(REPSTOSQ destptr (MOVQconst [size/8]) (MOVQconst [0]) mem)
 
 // Absorb InvertFlags into branches.

src/cmd/compile/internal/ssa/rewriteAMD64.go

@@ -8811,14 +8811,14 @@ func rewriteValueAMD64_OpMove(v *Value, config *Config) bool {
 		return true
 	}
 	// match: (Move [size] dst src mem)
-	// cond: size >= 32 && size <= 16*64 && size%16 == 0
+	// cond: size >= 32 && size <= 16*64 && size%16 == 0 && !config.noDuffDevice
 	// result: (DUFFCOPY [14*(64-size/16)] dst src mem)
 	for {
 		size := v.AuxInt
 		dst := v.Args[0]
 		src := v.Args[1]
 		mem := v.Args[2]
-		if !(size >= 32 && size <= 16*64 && size%16 == 0) {
+		if !(size >= 32 && size <= 16*64 && size%16 == 0 && !config.noDuffDevice) {
 			break
 		}
 		v.reset(OpAMD64DUFFCOPY)
@@ -8829,14 +8829,14 @@ func rewriteValueAMD64_OpMove(v *Value, config *Config) bool {
 		return true
 	}
 	// match: (Move [size] dst src mem)
-	// cond: size > 16*64 && size%8 == 0
+	// cond: (size > 16*64 || config.noDuffDevice) && size%8 == 0
 	// result: (REPMOVSQ dst src (MOVQconst [size/8]) mem)
 	for {
 		size := v.AuxInt
 		dst := v.Args[0]
 		src := v.Args[1]
 		mem := v.Args[2]
-		if !(size > 16*64 && size%8 == 0) {
+		if !((size > 16*64 || config.noDuffDevice) && size%8 == 0) {
 			break
 		}
 		v.reset(OpAMD64REPMOVSQ)
@@ -13693,13 +13693,13 @@ func rewriteValueAMD64_OpZero(v *Value, config *Config) bool {
 		return true
 	}
 	// match: (Zero [size] destptr mem)
-	// cond: size <= 1024 && size%8 == 0 && size%16 != 0
+	// cond: size <= 1024 && size%8 == 0 && size%16 != 0 && !config.noDuffDevice
 	// result: (Zero [size-8] (ADDQconst [8] destptr) (MOVQstore destptr (MOVQconst [0]) mem))
 	for {
 		size := v.AuxInt
 		destptr := v.Args[0]
 		mem := v.Args[1]
-		if !(size <= 1024 && size%8 == 0 && size%16 != 0) {
+		if !(size <= 1024 && size%8 == 0 && size%16 != 0 && !config.noDuffDevice) {
 			break
 		}
 		v.reset(OpZero)
@@ -13718,13 +13718,13 @@ func rewriteValueAMD64_OpZero(v *Value, config *Config) bool {
 		return true
 	}
 	// match: (Zero [size] destptr mem)
-	// cond: size <= 1024 && size%16 == 0
+	// cond: size <= 1024 && size%16 == 0 && !config.noDuffDevice
 	// result: (DUFFZERO [duffStart(size)] (ADDQconst [duffAdj(size)] destptr) (MOVOconst [0]) mem)
 	for {
 		size := v.AuxInt
 		destptr := v.Args[0]
 		mem := v.Args[1]
-		if !(size <= 1024 && size%16 == 0) {
+		if !(size <= 1024 && size%16 == 0 && !config.noDuffDevice) {
 			break
 		}
 		v.reset(OpAMD64DUFFZERO)
@@ -13740,13 +13740,13 @@ func rewriteValueAMD64_OpZero(v *Value, config *Config) bool {
 		return true
 	}
 	// match: (Zero [size] destptr mem)
-	// cond: size > 1024 && size%8 == 0
+	// cond: (size > 1024 || (config.noDuffDevice && size > 32)) && size%8 == 0
 	// result: (REPSTOSQ destptr (MOVQconst [size/8]) (MOVQconst [0]) mem)
 	for {
 		size := v.AuxInt
 		destptr := v.Args[0]
 		mem := v.Args[1]
-		if !(size > 1024 && size%8 == 0) {
+		if !((size > 1024 || (config.noDuffDevice && size > 32)) && size%8 == 0) {
 			break
 		}
 		v.reset(OpAMD64REPSTOSQ)