cmd/compile: join indexed byte loads into larger loads

Fixes #14920

Change-Id: I1535dc529779e26141d92d9e2b6ba7b016590c1a
Reviewed-on: https://go-review.googlesource.com/21005
Reviewed-by: Ilya Tocar <ilya.tocar@intel.com>
Reviewed-by: Ahmed Waheed <oneofone@gmail.com>
Reviewed-by: David Chase <drchase@google.com>

diff --git a/src/cmd/compile/internal/ssa/gen/AMD64.rules b/src/cmd/compile/internal/ssa/gen/AMD64.rules
index cc210978ef0eb157f74687bbd09f4d15041caf8d..22d2e7e475bf4d9d209aaba0675a6f102fb93662 100644 (file)
--- a/src/cmd/compile/internal/ssa/gen/AMD64.rules
+++ b/src/cmd/compile/internal/ssa/gen/AMD64.rules
@@ -1260,3 +1260,22 @@
     (SHLQconst [40] (MOVBload [i+5] {s} p mem)))
     (SHLQconst [48] (MOVBload [i+6] {s} p mem)))
     (SHLQconst [56] (MOVBload [i+7] {s} p mem))) -> @x.Block (MOVQload [i] {s} p mem)
+
+(ORW              x:(MOVBloadidx1 [i]   {s} p idx mem)
+    (SHLWconst [8]  (MOVBloadidx1 [i+1] {s} p idx mem))) -> @x.Block (MOVWload [i] {s} (ADDQ <p.Type> p idx) mem)
+
+(ORL (ORL (ORL
+                  x:(MOVBloadidx1 [i]   {s} p idx mem)
+    (SHLLconst [8]  (MOVBloadidx1 [i+1] {s} p idx mem)))
+    (SHLLconst [16] (MOVBloadidx1 [i+2] {s} p idx mem)))
+    (SHLLconst [24] (MOVBloadidx1 [i+3] {s} p idx mem))) -> @x.Block (MOVLload [i] {s} (ADDQ <p.Type> p idx) mem)
+
+(ORQ (ORQ (ORQ (ORQ (ORQ (ORQ (ORQ
+                  x:(MOVBloadidx1 [i]   {s} p idx mem)
+    (SHLQconst [8]  (MOVBloadidx1 [i+1] {s} p idx mem)))
+    (SHLQconst [16] (MOVBloadidx1 [i+2] {s} p idx mem)))
+    (SHLQconst [24] (MOVBloadidx1 [i+3] {s} p idx mem)))
+    (SHLQconst [32] (MOVBloadidx1 [i+4] {s} p idx mem)))
+    (SHLQconst [40] (MOVBloadidx1 [i+5] {s} p idx mem)))
+    (SHLQconst [48] (MOVBloadidx1 [i+6] {s} p idx mem)))
+    (SHLQconst [56] (MOVBloadidx1 [i+7] {s} p idx mem))) -> @x.Block (MOVQload [i] {s} (ADDQ <p.Type> p idx) mem)

diff --git a/src/cmd/compile/internal/ssa/rewriteAMD64.go b/src/cmd/compile/internal/ssa/rewriteAMD64.go
index 8dd1b15f135a878f909e3958edbe6acc1e6481cd..2d7aa3bd80bb7402eaefb5713c037bd1d4625cec 100644 (file)
--- a/src/cmd/compile/internal/ssa/rewriteAMD64.go
+++ b/src/cmd/compile/internal/ssa/rewriteAMD64.go
@@ -10668,6 +10668,118 @@ func rewriteValueAMD64_OpAMD64ORL(v *Value, config *Config) bool {
                v0.AddArg(mem)
                return true
        }
+       // match: (ORL (ORL (ORL                   x:(MOVBloadidx1 [i]   {s} p idx mem)     (SHLLconst [8]  (MOVBloadidx1 [i+1] {s} p idx mem)))     (SHLLconst [16] (MOVBloadidx1 [i+2] {s} p idx mem)))     (SHLLconst [24] (MOVBloadidx1 [i+3] {s} p idx mem)))
+       // cond:
+       // result: @x.Block (MOVLload [i] {s} (ADDQ <p.Type> p idx) mem)
+       for {
+               v_0 := v.Args[0]
+               if v_0.Op != OpAMD64ORL {
+                       break
+               }
+               v_0_0 := v_0.Args[0]
+               if v_0_0.Op != OpAMD64ORL {
+                       break
+               }
+               x := v_0_0.Args[0]
+               if x.Op != OpAMD64MOVBloadidx1 {
+                       break
+               }
+               i := x.AuxInt
+               s := x.Aux
+               p := x.Args[0]
+               idx := x.Args[1]
+               mem := x.Args[2]
+               v_0_0_1 := v_0_0.Args[1]
+               if v_0_0_1.Op != OpAMD64SHLLconst {
+                       break
+               }
+               if v_0_0_1.AuxInt != 8 {
+                       break
+               }
+               v_0_0_1_0 := v_0_0_1.Args[0]
+               if v_0_0_1_0.Op != OpAMD64MOVBloadidx1 {
+                       break
+               }
+               if v_0_0_1_0.AuxInt != i+1 {
+                       break
+               }
+               if v_0_0_1_0.Aux != s {
+                       break
+               }
+               if p != v_0_0_1_0.Args[0] {
+                       break
+               }
+               if idx != v_0_0_1_0.Args[1] {
+                       break
+               }
+               if mem != v_0_0_1_0.Args[2] {
+                       break
+               }
+               v_0_1 := v_0.Args[1]
+               if v_0_1.Op != OpAMD64SHLLconst {
+                       break
+               }
+               if v_0_1.AuxInt != 16 {
+                       break
+               }
+               v_0_1_0 := v_0_1.Args[0]
+               if v_0_1_0.Op != OpAMD64MOVBloadidx1 {
+                       break
+               }
+               if v_0_1_0.AuxInt != i+2 {
+                       break
+               }
+               if v_0_1_0.Aux != s {
+                       break
+               }
+               if p != v_0_1_0.Args[0] {
+                       break
+               }
+               if idx != v_0_1_0.Args[1] {
+                       break
+               }
+               if mem != v_0_1_0.Args[2] {
+                       break
+               }
+               v_1 := v.Args[1]
+               if v_1.Op != OpAMD64SHLLconst {
+                       break
+               }
+               if v_1.AuxInt != 24 {
+                       break
+               }
+               v_1_0 := v_1.Args[0]
+               if v_1_0.Op != OpAMD64MOVBloadidx1 {
+                       break
+               }
+               if v_1_0.AuxInt != i+3 {
+                       break
+               }
+               if v_1_0.Aux != s {
+                       break
+               }
+               if p != v_1_0.Args[0] {
+                       break
+               }
+               if idx != v_1_0.Args[1] {
+                       break
+               }
+               if mem != v_1_0.Args[2] {
+                       break
+               }
+               b = x.Block
+               v0 := b.NewValue0(v.Line, OpAMD64MOVLload, config.fe.TypeUInt32())
+               v.reset(OpCopy)
+               v.AddArg(v0)
+               v0.AuxInt = i
+               v0.Aux = s
+               v1 := b.NewValue0(v.Line, OpAMD64ADDQ, p.Type)
+               v1.AddArg(p)
+               v1.AddArg(idx)
+               v0.AddArg(v1)
+               v0.AddArg(mem)
+               return true
+       }
        return false
 }
 func rewriteValueAMD64_OpAMD64ORLconst(v *Value, config *Config) bool {
@@ -10974,6 +11086,238 @@ func rewriteValueAMD64_OpAMD64ORQ(v *Value, config *Config) bool {
                v0.AddArg(mem)
                return true
        }
+       // match: (ORQ (ORQ (ORQ (ORQ (ORQ (ORQ (ORQ                   x:(MOVBloadidx1 [i]   {s} p idx mem)     (SHLQconst [8]  (MOVBloadidx1 [i+1] {s} p idx mem)))     (SHLQconst [16] (MOVBloadidx1 [i+2] {s} p idx mem)))     (SHLQconst [24] (MOVBloadidx1 [i+3] {s} p idx mem)))     (SHLQconst [32] (MOVBloadidx1 [i+4] {s} p idx mem)))     (SHLQconst [40] (MOVBloadidx1 [i+5] {s} p idx mem)))     (SHLQconst [48] (MOVBloadidx1 [i+6] {s} p idx mem)))     (SHLQconst [56] (MOVBloadidx1 [i+7] {s} p idx mem)))
+       // cond:
+       // result: @x.Block (MOVQload [i] {s} (ADDQ <p.Type> p idx) mem)
+       for {
+               v_0 := v.Args[0]
+               if v_0.Op != OpAMD64ORQ {
+                       break
+               }
+               v_0_0 := v_0.Args[0]
+               if v_0_0.Op != OpAMD64ORQ {
+                       break
+               }
+               v_0_0_0 := v_0_0.Args[0]
+               if v_0_0_0.Op != OpAMD64ORQ {
+                       break
+               }
+               v_0_0_0_0 := v_0_0_0.Args[0]
+               if v_0_0_0_0.Op != OpAMD64ORQ {
+                       break
+               }
+               v_0_0_0_0_0 := v_0_0_0_0.Args[0]
+               if v_0_0_0_0_0.Op != OpAMD64ORQ {
+                       break
+               }
+               v_0_0_0_0_0_0 := v_0_0_0_0_0.Args[0]
+               if v_0_0_0_0_0_0.Op != OpAMD64ORQ {
+                       break
+               }
+               x := v_0_0_0_0_0_0.Args[0]
+               if x.Op != OpAMD64MOVBloadidx1 {
+                       break
+               }
+               i := x.AuxInt
+               s := x.Aux
+               p := x.Args[0]
+               idx := x.Args[1]
+               mem := x.Args[2]
+               v_0_0_0_0_0_0_1 := v_0_0_0_0_0_0.Args[1]
+               if v_0_0_0_0_0_0_1.Op != OpAMD64SHLQconst {
+                       break
+               }
+               if v_0_0_0_0_0_0_1.AuxInt != 8 {
+                       break
+               }
+               v_0_0_0_0_0_0_1_0 := v_0_0_0_0_0_0_1.Args[0]
+               if v_0_0_0_0_0_0_1_0.Op != OpAMD64MOVBloadidx1 {
+                       break
+               }
+               if v_0_0_0_0_0_0_1_0.AuxInt != i+1 {
+                       break
+               }
+               if v_0_0_0_0_0_0_1_0.Aux != s {
+                       break
+               }
+               if p != v_0_0_0_0_0_0_1_0.Args[0] {
+                       break
+               }
+               if idx != v_0_0_0_0_0_0_1_0.Args[1] {
+                       break
+               }
+               if mem != v_0_0_0_0_0_0_1_0.Args[2] {
+                       break
+               }
+               v_0_0_0_0_0_1 := v_0_0_0_0_0.Args[1]
+               if v_0_0_0_0_0_1.Op != OpAMD64SHLQconst {
+                       break
+               }
+               if v_0_0_0_0_0_1.AuxInt != 16 {
+                       break
+               }
+               v_0_0_0_0_0_1_0 := v_0_0_0_0_0_1.Args[0]
+               if v_0_0_0_0_0_1_0.Op != OpAMD64MOVBloadidx1 {
+                       break
+               }
+               if v_0_0_0_0_0_1_0.AuxInt != i+2 {
+                       break
+               }
+               if v_0_0_0_0_0_1_0.Aux != s {
+                       break
+               }
+               if p != v_0_0_0_0_0_1_0.Args[0] {
+                       break
+               }
+               if idx != v_0_0_0_0_0_1_0.Args[1] {
+                       break
+               }
+               if mem != v_0_0_0_0_0_1_0.Args[2] {
+                       break
+               }
+               v_0_0_0_0_1 := v_0_0_0_0.Args[1]
+               if v_0_0_0_0_1.Op != OpAMD64SHLQconst {
+                       break
+               }
+               if v_0_0_0_0_1.AuxInt != 24 {
+                       break
+               }
+               v_0_0_0_0_1_0 := v_0_0_0_0_1.Args[0]
+               if v_0_0_0_0_1_0.Op != OpAMD64MOVBloadidx1 {
+                       break
+               }
+               if v_0_0_0_0_1_0.AuxInt != i+3 {
+                       break
+               }
+               if v_0_0_0_0_1_0.Aux != s {
+                       break
+               }
+               if p != v_0_0_0_0_1_0.Args[0] {
+                       break
+               }
+               if idx != v_0_0_0_0_1_0.Args[1] {
+                       break
+               }
+               if mem != v_0_0_0_0_1_0.Args[2] {
+                       break
+               }
+               v_0_0_0_1 := v_0_0_0.Args[1]
+               if v_0_0_0_1.Op != OpAMD64SHLQconst {
+                       break
+               }
+               if v_0_0_0_1.AuxInt != 32 {
+                       break
+               }
+               v_0_0_0_1_0 := v_0_0_0_1.Args[0]
+               if v_0_0_0_1_0.Op != OpAMD64MOVBloadidx1 {
+                       break
+               }
+               if v_0_0_0_1_0.AuxInt != i+4 {
+                       break
+               }
+               if v_0_0_0_1_0.Aux != s {
+                       break
+               }
+               if p != v_0_0_0_1_0.Args[0] {
+                       break
+               }
+               if idx != v_0_0_0_1_0.Args[1] {
+                       break
+               }
+               if mem != v_0_0_0_1_0.Args[2] {
+                       break
+               }
+               v_0_0_1 := v_0_0.Args[1]
+               if v_0_0_1.Op != OpAMD64SHLQconst {
+                       break
+               }
+               if v_0_0_1.AuxInt != 40 {
+                       break
+               }
+               v_0_0_1_0 := v_0_0_1.Args[0]
+               if v_0_0_1_0.Op != OpAMD64MOVBloadidx1 {
+                       break
+               }
+               if v_0_0_1_0.AuxInt != i+5 {
+                       break
+               }
+               if v_0_0_1_0.Aux != s {
+                       break
+               }
+               if p != v_0_0_1_0.Args[0] {
+                       break
+               }
+               if idx != v_0_0_1_0.Args[1] {
+                       break
+               }
+               if mem != v_0_0_1_0.Args[2] {
+                       break
+               }
+               v_0_1 := v_0.Args[1]
+               if v_0_1.Op != OpAMD64SHLQconst {
+                       break
+               }
+               if v_0_1.AuxInt != 48 {
+                       break
+               }
+               v_0_1_0 := v_0_1.Args[0]
+               if v_0_1_0.Op != OpAMD64MOVBloadidx1 {
+                       break
+               }
+               if v_0_1_0.AuxInt != i+6 {
+                       break
+               }
+               if v_0_1_0.Aux != s {
+                       break
+               }
+               if p != v_0_1_0.Args[0] {
+                       break
+               }
+               if idx != v_0_1_0.Args[1] {
+                       break
+               }
+               if mem != v_0_1_0.Args[2] {
+                       break
+               }
+               v_1 := v.Args[1]
+               if v_1.Op != OpAMD64SHLQconst {
+                       break
+               }
+               if v_1.AuxInt != 56 {
+                       break
+               }
+               v_1_0 := v_1.Args[0]
+               if v_1_0.Op != OpAMD64MOVBloadidx1 {
+                       break
+               }
+               if v_1_0.AuxInt != i+7 {
+                       break
+               }
+               if v_1_0.Aux != s {
+                       break
+               }
+               if p != v_1_0.Args[0] {
+                       break
+               }
+               if idx != v_1_0.Args[1] {
+                       break
+               }
+               if mem != v_1_0.Args[2] {
+                       break
+               }
+               b = x.Block
+               v0 := b.NewValue0(v.Line, OpAMD64MOVQload, config.fe.TypeUInt64())
+               v.reset(OpCopy)
+               v.AddArg(v0)
+               v0.AuxInt = i
+               v0.Aux = s
+               v1 := b.NewValue0(v.Line, OpAMD64ADDQ, p.Type)
+               v1.AddArg(p)
+               v1.AddArg(idx)
+               v0.AddArg(v1)
+               v0.AddArg(mem)
+               return true
+       }
        return false
 }
 func rewriteValueAMD64_OpAMD64ORQconst(v *Value, config *Config) bool {
@@ -11110,6 +11454,58 @@ func rewriteValueAMD64_OpAMD64ORW(v *Value, config *Config) bool {
                v0.AddArg(mem)
                return true
        }
+       // match: (ORW              x:(MOVBloadidx1 [i]   {s} p idx mem)     (SHLWconst [8]  (MOVBloadidx1 [i+1] {s} p idx mem)))
+       // cond:
+       // result: @x.Block (MOVWload [i] {s} (ADDQ <p.Type> p idx) mem)
+       for {
+               x := v.Args[0]
+               if x.Op != OpAMD64MOVBloadidx1 {
+                       break
+               }
+               i := x.AuxInt
+               s := x.Aux
+               p := x.Args[0]
+               idx := x.Args[1]
+               mem := x.Args[2]
+               v_1 := v.Args[1]
+               if v_1.Op != OpAMD64SHLWconst {
+                       break
+               }
+               if v_1.AuxInt != 8 {
+                       break
+               }
+               v_1_0 := v_1.Args[0]
+               if v_1_0.Op != OpAMD64MOVBloadidx1 {
+                       break
+               }
+               if v_1_0.AuxInt != i+1 {
+                       break
+               }
+               if v_1_0.Aux != s {
+                       break
+               }
+               if p != v_1_0.Args[0] {
+                       break
+               }
+               if idx != v_1_0.Args[1] {
+                       break
+               }
+               if mem != v_1_0.Args[2] {
+                       break
+               }
+               b = x.Block
+               v0 := b.NewValue0(v.Line, OpAMD64MOVWload, config.fe.TypeUInt16())
+               v.reset(OpCopy)
+               v.AddArg(v0)
+               v0.AuxInt = i
+               v0.Aux = s
+               v1 := b.NewValue0(v.Line, OpAMD64ADDQ, p.Type)
+               v1.AddArg(p)
+               v1.AddArg(idx)
+               v0.AddArg(v1)
+               v0.AddArg(mem)
+               return true
+       }
        return false
 }
 func rewriteValueAMD64_OpAMD64ORWconst(v *Value, config *Config) bool {