cmd/compile: combine some extensions with loads

For cases where we already have the ops, combine
sign or zero extension with the previous load
(even if the load is larger width).

Update #15105

Change-Id: I76c5ddd69e1f900d2a17d35503083bd3b4978e48
Reviewed-on: https://go-review.googlesource.com/28190
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>

2 files changed, 312 insertions, 0 deletions

diff --git a/src/cmd/compile/internal/ssa/gen/AMD64.rules b/src/cmd/compile/internal/ssa/gen/AMD64.rules
index f4c44975a0..4c49d10924 100644
--- a/src/cmd/compile/internal/ssa/gen/AMD64.rules
+++ b/src/cmd/compile/internal/ssa/gen/AMD64.rules
@@ -738,11 +738,23 @@
 // This prevents a single load from being split into multiple loads
 // which then might return different values.  See test/atomicload.go.
 (MOVBQSX x:(MOVBload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
+(MOVBQSX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
+(MOVBQSX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
+(MOVBQSX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
 (MOVBQZX x:(MOVBload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
+(MOVBQZX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
+(MOVBQZX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
+(MOVBQZX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
 (MOVWQSX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVWQSXload <v.Type> [off] {sym} ptr mem)
+(MOVWQSX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVWQSXload <v.Type> [off] {sym} ptr mem)
+(MOVWQSX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVWQSXload <v.Type> [off] {sym} ptr mem)
 (MOVWQZX x:(MOVWload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVWload <v.Type> [off] {sym} ptr mem)
+(MOVWQZX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVWload <v.Type> [off] {sym} ptr mem)
+(MOVWQZX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVWload <v.Type> [off] {sym} ptr mem)
 (MOVLQSX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVLQSXload <v.Type> [off] {sym} ptr mem)
+(MOVLQSX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVLQSXload <v.Type> [off] {sym} ptr mem)
 (MOVLQZX x:(MOVLload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVLload <v.Type> [off] {sym} ptr mem)
+(MOVLQZX x:(MOVQload [off] {sym} ptr mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVLload <v.Type> [off] {sym} ptr mem)
 
 (MOVBQZX x:(MOVBloadidx1 [off] {sym} ptr idx mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVBloadidx1 <v.Type> [off] {sym} ptr idx mem)
 (MOVWQZX x:(MOVWloadidx1 [off] {sym} ptr idx mem)) && x.Uses == 1 && clobber(x) -> @x.Block (MOVWloadidx1 <v.Type> [off] {sym} ptr idx mem)
diff --git a/src/cmd/compile/internal/ssa/rewriteAMD64.go b/src/cmd/compile/internal/ssa/rewriteAMD64.go
index 0f3c636469..5c685ef25f 100644
--- a/src/cmd/compile/internal/ssa/rewriteAMD64.go
+++ b/src/cmd/compile/internal/ssa/rewriteAMD64.go
@@ -3263,6 +3263,81 @@ func rewriteValueAMD64_OpAMD64MOVBQSX(v *Value, config *Config) bool {
 		v0.AddArg(mem)
 		return true
 	}
+	// match: (MOVBQSX x:(MOVWload [off] {sym} ptr mem))
+	// cond: x.Uses == 1 && clobber(x)
+	// result: @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
+	for {
+		x := v.Args[0]
+		if x.Op != OpAMD64MOVWload {
+			break
+		}
+		off := x.AuxInt
+		sym := x.Aux
+		ptr := x.Args[0]
+		mem := x.Args[1]
+		if !(x.Uses == 1 && clobber(x)) {
+			break
+		}
+		b = x.Block
+		v0 := b.NewValue0(v.Line, OpAMD64MOVBQSXload, v.Type)
+		v.reset(OpCopy)
+		v.AddArg(v0)
+		v0.AuxInt = off
+		v0.Aux = sym
+		v0.AddArg(ptr)
+		v0.AddArg(mem)
+		return true
+	}
+	// match: (MOVBQSX x:(MOVLload [off] {sym} ptr mem))
+	// cond: x.Uses == 1 && clobber(x)
+	// result: @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
+	for {
+		x := v.Args[0]
+		if x.Op != OpAMD64MOVLload {
+			break
+		}
+		off := x.AuxInt
+		sym := x.Aux
+		ptr := x.Args[0]
+		mem := x.Args[1]
+		if !(x.Uses == 1 && clobber(x)) {
+			break
+		}
+		b = x.Block
+		v0 := b.NewValue0(v.Line, OpAMD64MOVBQSXload, v.Type)
+		v.reset(OpCopy)
+		v.AddArg(v0)
+		v0.AuxInt = off
+		v0.Aux = sym
+		v0.AddArg(ptr)
+		v0.AddArg(mem)
+		return true
+	}
+	// match: (MOVBQSX x:(MOVQload [off] {sym} ptr mem))
+	// cond: x.Uses == 1 && clobber(x)
+	// result: @x.Block (MOVBQSXload <v.Type> [off] {sym} ptr mem)
+	for {
+		x := v.Args[0]
+		if x.Op != OpAMD64MOVQload {
+			break
+		}
+		off := x.AuxInt
+		sym := x.Aux
+		ptr := x.Args[0]
+		mem := x.Args[1]
+		if !(x.Uses == 1 && clobber(x)) {
+			break
+		}
+		b = x.Block
+		v0 := b.NewValue0(v.Line, OpAMD64MOVBQSXload, v.Type)
+		v.reset(OpCopy)
+		v.AddArg(v0)
+		v0.AuxInt = off
+		v0.Aux = sym
+		v0.AddArg(ptr)
+		v0.AddArg(mem)
+		return true
+	}
 	// match: (MOVBQSX (ANDLconst [c] x))
 	// cond: c & 0x80 == 0
 	// result: (ANDLconst [c & 0x7f] x)
@@ -3340,6 +3415,81 @@ func rewriteValueAMD64_OpAMD64MOVBQZX(v *Value, config *Config) bool {
 		v0.AddArg(mem)
 		return true
 	}
+	// match: (MOVBQZX x:(MOVWload [off] {sym} ptr mem))
+	// cond: x.Uses == 1 && clobber(x)
+	// result: @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
+	for {
+		x := v.Args[0]
+		if x.Op != OpAMD64MOVWload {
+			break
+		}
+		off := x.AuxInt
+		sym := x.Aux
+		ptr := x.Args[0]
+		mem := x.Args[1]
+		if !(x.Uses == 1 && clobber(x)) {
+			break
+		}
+		b = x.Block
+		v0 := b.NewValue0(v.Line, OpAMD64MOVBload, v.Type)
+		v.reset(OpCopy)
+		v.AddArg(v0)
+		v0.AuxInt = off
+		v0.Aux = sym
+		v0.AddArg(ptr)
+		v0.AddArg(mem)
+		return true
+	}
+	// match: (MOVBQZX x:(MOVLload [off] {sym} ptr mem))
+	// cond: x.Uses == 1 && clobber(x)
+	// result: @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
+	for {
+		x := v.Args[0]
+		if x.Op != OpAMD64MOVLload {
+			break
+		}
+		off := x.AuxInt
+		sym := x.Aux
+		ptr := x.Args[0]
+		mem := x.Args[1]
+		if !(x.Uses == 1 && clobber(x)) {
+			break
+		}
+		b = x.Block
+		v0 := b.NewValue0(v.Line, OpAMD64MOVBload, v.Type)
+		v.reset(OpCopy)
+		v.AddArg(v0)
+		v0.AuxInt = off
+		v0.Aux = sym
+		v0.AddArg(ptr)
+		v0.AddArg(mem)
+		return true
+	}
+	// match: (MOVBQZX x:(MOVQload [off] {sym} ptr mem))
+	// cond: x.Uses == 1 && clobber(x)
+	// result: @x.Block (MOVBload <v.Type> [off] {sym} ptr mem)
+	for {
+		x := v.Args[0]
+		if x.Op != OpAMD64MOVQload {
+			break
+		}
+		off := x.AuxInt
+		sym := x.Aux
+		ptr := x.Args[0]
+		mem := x.Args[1]
+		if !(x.Uses == 1 && clobber(x)) {
+			break
+		}
+		b = x.Block
+		v0 := b.NewValue0(v.Line, OpAMD64MOVBload, v.Type)
+		v.reset(OpCopy)
+		v.AddArg(v0)
+		v0.AuxInt = off
+		v0.Aux = sym
+		v0.AddArg(ptr)
+		v0.AddArg(mem)
+		return true
+	}
 	// match: (MOVBQZX x:(MOVBloadidx1 [off] {sym} ptr idx mem))
 	// cond: x.Uses == 1 && clobber(x)
 	// result: @x.Block (MOVBloadidx1 <v.Type> [off] {sym} ptr idx mem)
@@ -4362,6 +4512,31 @@ func rewriteValueAMD64_OpAMD64MOVLQSX(v *Value, config *Config) bool {
 		v0.AddArg(mem)
 		return true
 	}
+	// match: (MOVLQSX x:(MOVQload [off] {sym} ptr mem))
+	// cond: x.Uses == 1 && clobber(x)
+	// result: @x.Block (MOVLQSXload <v.Type> [off] {sym} ptr mem)
+	for {
+		x := v.Args[0]
+		if x.Op != OpAMD64MOVQload {
+			break
+		}
+		off := x.AuxInt
+		sym := x.Aux
+		ptr := x.Args[0]
+		mem := x.Args[1]
+		if !(x.Uses == 1 && clobber(x)) {
+			break
+		}
+		b = x.Block
+		v0 := b.NewValue0(v.Line, OpAMD64MOVLQSXload, v.Type)
+		v.reset(OpCopy)
+		v.AddArg(v0)
+		v0.AuxInt = off
+		v0.Aux = sym
+		v0.AddArg(ptr)
+		v0.AddArg(mem)
+		return true
+	}
 	// match: (MOVLQSX (ANDLconst [c] x))
 	// cond: c & 0x80000000 == 0
 	// result: (ANDLconst [c & 0x7fffffff] x)
@@ -4439,6 +4614,31 @@ func rewriteValueAMD64_OpAMD64MOVLQZX(v *Value, config *Config) bool {
 		v0.AddArg(mem)
 		return true
 	}
+	// match: (MOVLQZX x:(MOVQload [off] {sym} ptr mem))
+	// cond: x.Uses == 1 && clobber(x)
+	// result: @x.Block (MOVLload <v.Type> [off] {sym} ptr mem)
+	for {
+		x := v.Args[0]
+		if x.Op != OpAMD64MOVQload {
+			break
+		}
+		off := x.AuxInt
+		sym := x.Aux
+		ptr := x.Args[0]
+		mem := x.Args[1]
+		if !(x.Uses == 1 && clobber(x)) {
+			break
+		}
+		b = x.Block
+		v0 := b.NewValue0(v.Line, OpAMD64MOVLload, v.Type)
+		v.reset(OpCopy)
+		v.AddArg(v0)
+		v0.AuxInt = off
+		v0.Aux = sym
+		v0.AddArg(ptr)
+		v0.AddArg(mem)
+		return true
+	}
 	// match: (MOVLQZX x:(MOVLloadidx1 [off] {sym} ptr idx mem))
 	// cond: x.Uses == 1 && clobber(x)
 	// result: @x.Block (MOVLloadidx1 <v.Type> [off] {sym} ptr idx mem)
@@ -8156,6 +8356,56 @@ func rewriteValueAMD64_OpAMD64MOVWQSX(v *Value, config *Config) bool {
 		v0.AddArg(mem)
 		return true
 	}
+	// match: (MOVWQSX x:(MOVLload [off] {sym} ptr mem))
+	// cond: x.Uses == 1 && clobber(x)
+	// result: @x.Block (MOVWQSXload <v.Type> [off] {sym} ptr mem)
+	for {
+		x := v.Args[0]
+		if x.Op != OpAMD64MOVLload {
+			break
+		}
+		off := x.AuxInt
+		sym := x.Aux
+		ptr := x.Args[0]
+		mem := x.Args[1]
+		if !(x.Uses == 1 && clobber(x)) {
+			break
+		}
+		b = x.Block
+		v0 := b.NewValue0(v.Line, OpAMD64MOVWQSXload, v.Type)
+		v.reset(OpCopy)
+		v.AddArg(v0)
+		v0.AuxInt = off
+		v0.Aux = sym
+		v0.AddArg(ptr)
+		v0.AddArg(mem)
+		return true
+	}
+	// match: (MOVWQSX x:(MOVQload [off] {sym} ptr mem))
+	// cond: x.Uses == 1 && clobber(x)
+	// result: @x.Block (MOVWQSXload <v.Type> [off] {sym} ptr mem)
+	for {
+		x := v.Args[0]
+		if x.Op != OpAMD64MOVQload {
+			break
+		}
+		off := x.AuxInt
+		sym := x.Aux
+		ptr := x.Args[0]
+		mem := x.Args[1]
+		if !(x.Uses == 1 && clobber(x)) {
+			break
+		}
+		b = x.Block
+		v0 := b.NewValue0(v.Line, OpAMD64MOVWQSXload, v.Type)
+		v.reset(OpCopy)
+		v.AddArg(v0)
+		v0.AuxInt = off
+		v0.Aux = sym
+		v0.AddArg(ptr)
+		v0.AddArg(mem)
+		return true
+	}
 	// match: (MOVWQSX (ANDLconst [c] x))
 	// cond: c & 0x8000 == 0
 	// result: (ANDLconst [c & 0x7fff] x)
@@ -8233,6 +8483,56 @@ func rewriteValueAMD64_OpAMD64MOVWQZX(v *Value, config *Config) bool {
 		v0.AddArg(mem)
 		return true
 	}
+	// match: (MOVWQZX x:(MOVLload [off] {sym} ptr mem))
+	// cond: x.Uses == 1 && clobber(x)
+	// result: @x.Block (MOVWload <v.Type> [off] {sym} ptr mem)
+	for {
+		x := v.Args[0]
+		if x.Op != OpAMD64MOVLload {
+			break
+		}
+		off := x.AuxInt
+		sym := x.Aux
+		ptr := x.Args[0]
+		mem := x.Args[1]
+		if !(x.Uses == 1 && clobber(x)) {
+			break
+		}
+		b = x.Block
+		v0 := b.NewValue0(v.Line, OpAMD64MOVWload, v.Type)
+		v.reset(OpCopy)
+		v.AddArg(v0)
+		v0.AuxInt = off
+		v0.Aux = sym
+		v0.AddArg(ptr)
+		v0.AddArg(mem)
+		return true
+	}
+	// match: (MOVWQZX x:(MOVQload [off] {sym} ptr mem))
+	// cond: x.Uses == 1 && clobber(x)
+	// result: @x.Block (MOVWload <v.Type> [off] {sym} ptr mem)
+	for {
+		x := v.Args[0]
+		if x.Op != OpAMD64MOVQload {
+			break
+		}
+		off := x.AuxInt
+		sym := x.Aux
+		ptr := x.Args[0]
+		mem := x.Args[1]
+		if !(x.Uses == 1 && clobber(x)) {
+			break
+		}
+		b = x.Block
+		v0 := b.NewValue0(v.Line, OpAMD64MOVWload, v.Type)
+		v.reset(OpCopy)
+		v.AddArg(v0)
+		v0.AuxInt = off
+		v0.Aux = sym
+		v0.AddArg(ptr)
+		v0.AddArg(mem)
+		return true
+	}
 	// match: (MOVWQZX x:(MOVWloadidx1 [off] {sym} ptr idx mem))
 	// cond: x.Uses == 1 && clobber(x)
 	// result: @x.Block (MOVWloadidx1 <v.Type> [off] {sym} ptr idx mem)