cmd/compile: add arch-specific inlining for runtime.memmove

This CL add runtime.memmove inlining for AMD64 and ARM64.
According to ssa dump from testcases generic rules can't inline
memmomve properly due to one of the arguments is Phi operation. But this
Phi op will be optimized out by later optimization stages. As a result
memmove can be inlined during arch-specific rules.
The commit add new optimization rules to arch-specific rules that can
inline runtime.memmove if it possible during lowering stage.
Optimization fires 5 times in Go source-code using regabi.

Fixes #41662

Change-Id: Iaffaf4c482d068b5f0683d141863892202cc8824
Reviewed-on: https://go-review.googlesource.com/c/go/+/289151
Reviewed-by: Keith Randall <khr@golang.org>
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Go Bot <gobot@golang.org>
Trust: David Chase <drchase@google.com>

5 files changed, 155 insertions, 3 deletions

diff --git a/src/cmd/compile/internal/ssa/gen/AMD64.rules b/src/cmd/compile/internal/ssa/gen/AMD64.rules
index ec91ea1513..4cd00732fc 100644
--- a/src/cmd/compile/internal/ssa/gen/AMD64.rules
+++ b/src/cmd/compile/internal/ssa/gen/AMD64.rules
@@ -2216,3 +2216,22 @@
 (MOVOstore [dstOff] {dstSym} ptr (MOVOload [srcOff] {srcSym} (SB) _) mem) && symIsRO(srcSym) =>
   (MOVQstore [dstOff+8] {dstSym} ptr (MOVQconst [int64(read64(srcSym, int64(srcOff)+8, config.ctxt.Arch.ByteOrder))])
     (MOVQstore [dstOff] {dstSym} ptr (MOVQconst [int64(read64(srcSym, int64(srcOff), config.ctxt.Arch.ByteOrder))]) mem))
+
+// Arch-specific inlining for small or disjoint runtime.memmove
+// Match post-lowering calls, memory version.
+(SelectN [0] call:(CALLstatic {sym} s1:(MOVQstoreconst _ [sc] s2:(MOVQstore _ src s3:(MOVQstore _ dst mem)))))
+	&& sc.Val64() >= 0
+	&& isSameCall(sym, "runtime.memmove")
+	&& s1.Uses == 1 && s2.Uses == 1 && s3.Uses == 1
+	&& isInlinableMemmove(dst, src, sc.Val64(), config)
+	&& clobber(s1, s2, s3, call)
+	=> (Move [sc.Val64()] dst src mem)
+
+// Match post-lowering calls, register version.
+(SelectN [0] call:(CALLstatic {sym} dst src (MOVQconst [sz]) mem))
+	&& sz >= 0
+	&& isSameCall(sym, "runtime.memmove")
+	&& call.Uses == 1
+	&& isInlinableMemmove(dst, src, sz, config)
+	&& clobber(call)
+	=> (Move [sz] dst src mem)
diff --git a/src/cmd/compile/internal/ssa/gen/ARM64.rules b/src/cmd/compile/internal/ssa/gen/ARM64.rules
index 3d2759493e..62699f290c 100644
--- a/src/cmd/compile/internal/ssa/gen/ARM64.rules
+++ b/src/cmd/compile/internal/ssa/gen/ARM64.rules
@@ -2859,3 +2859,12 @@
 (MOVHUload [off] {sym} (SB) _) && symIsRO(sym) => (MOVDconst [int64(read16(sym, int64(off), config.ctxt.Arch.ByteOrder))])
 (MOVWUload [off] {sym} (SB) _) && symIsRO(sym) => (MOVDconst [int64(read32(sym, int64(off), config.ctxt.Arch.ByteOrder))])
 (MOVDload  [off] {sym} (SB) _) && symIsRO(sym) => (MOVDconst [int64(read64(sym, int64(off), config.ctxt.Arch.ByteOrder))])
+
+// Arch-specific inlining for small or disjoint runtime.memmove
+(SelectN [0] call:(CALLstatic {sym} s1:(MOVDstore _ (MOVDconst [sz]) s2:(MOVDstore  _ src s3:(MOVDstore {t} _ dst mem)))))
+	&& sz >= 0
+	&& isSameCall(sym, "runtime.memmove")
+	&& s1.Uses == 1 && s2.Uses == 1 && s3.Uses == 1
+	&& isInlinableMemmove(dst, src, sz, config)
+	&& clobber(s1, s2, s3, call)
+	=> (Move [sz] dst src mem)
diff --git a/src/cmd/compile/internal/ssa/gen/generic.rules b/src/cmd/compile/internal/ssa/gen/generic.rules
index aad7600d79..5cbc70cf41 100644
--- a/src/cmd/compile/internal/ssa/gen/generic.rules
+++ b/src/cmd/compile/internal/ssa/gen/generic.rules
@@ -2065,7 +2065,7 @@
 (SelectN [0] call:(StaticCall {sym} s1:(Store _ (Const(64|32) [sz]) s2:(Store  _ src s3:(Store {t} _ dst mem)))))
 	&& sz >= 0
 	&& isSameCall(sym, "runtime.memmove")
-	&& t.IsPtr() // avoids TUINTPTR, see issue 30061
+	&& t.IsPtr() // avoids TUNSAFEPTR, see issue 30061
 	&& s1.Uses == 1 && s2.Uses == 1 && s3.Uses == 1
 	&& isInlinableMemmove(dst, src, int64(sz), config)
 	&& clobber(s1, s2, s3, call)
@@ -2076,7 +2076,7 @@
 	&& sz >= 0
 	&& call.Uses == 1 // this will exclude all calls with results
 	&& isSameCall(sym, "runtime.memmove")
-	&& dst.Type.IsPtr() // avoids TUINTPTR, see issue 30061
+	&& dst.Type.IsPtr() // avoids TUNSAFEPTR, see issue 30061
 	&& isInlinableMemmove(dst, src, int64(sz), config)
 	&& clobber(call)
 	=> (Move {dst.Type.Elem()} [int64(sz)] dst src mem)
@@ -2086,7 +2086,7 @@
 	&& sz >= 0
 	&& call.Uses == 1 // this will exclude all calls with results
 	&& isSameCall(sym, "runtime.memmove")
-	&& dst.Type.IsPtr() // avoids TUINTPTR, see issue 30061
+	&& dst.Type.IsPtr() // avoids TUNSAFEPTR, see issue 30061
 	&& isInlinableMemmove(dst, src, int64(sz), config)
 	&& clobber(call)
 	=> (Move {dst.Type.Elem()} [int64(sz)] dst src mem)
diff --git a/src/cmd/compile/internal/ssa/rewriteAMD64.go b/src/cmd/compile/internal/ssa/rewriteAMD64.go
index efb5d27145..5045ba7351 100644
--- a/src/cmd/compile/internal/ssa/rewriteAMD64.go
+++ b/src/cmd/compile/internal/ssa/rewriteAMD64.go
@@ -1038,6 +1038,8 @@ func rewriteValueAMD64(v *Value) bool {
 		return rewriteValueAMD64_OpSelect0(v)
 	case OpSelect1:
 		return rewriteValueAMD64_OpSelect1(v)
+	case OpSelectN:
+		return rewriteValueAMD64_OpSelectN(v)
 	case OpSignExt16to32:
 		v.Op = OpAMD64MOVWQSX
 		return true
@@ -32981,6 +32983,78 @@ func rewriteValueAMD64_OpSelect1(v *Value) bool {
 	}
 	return false
 }
+func rewriteValueAMD64_OpSelectN(v *Value) bool {
+	v_0 := v.Args[0]
+	b := v.Block
+	config := b.Func.Config
+	// match: (SelectN [0] call:(CALLstatic {sym} s1:(MOVQstoreconst _ [sc] s2:(MOVQstore _ src s3:(MOVQstore _ dst mem)))))
+	// cond: sc.Val64() >= 0 && isSameCall(sym, "runtime.memmove") && s1.Uses == 1 && s2.Uses == 1 && s3.Uses == 1 && isInlinableMemmove(dst, src, sc.Val64(), config) && clobber(s1, s2, s3, call)
+	// result: (Move [sc.Val64()] dst src mem)
+	for {
+		if auxIntToInt64(v.AuxInt) != 0 {
+			break
+		}
+		call := v_0
+		if call.Op != OpAMD64CALLstatic || len(call.Args) != 1 {
+			break
+		}
+		sym := auxToCall(call.Aux)
+		s1 := call.Args[0]
+		if s1.Op != OpAMD64MOVQstoreconst {
+			break
+		}
+		sc := auxIntToValAndOff(s1.AuxInt)
+		_ = s1.Args[1]
+		s2 := s1.Args[1]
+		if s2.Op != OpAMD64MOVQstore {
+			break
+		}
+		_ = s2.Args[2]
+		src := s2.Args[1]
+		s3 := s2.Args[2]
+		if s3.Op != OpAMD64MOVQstore {
+			break
+		}
+		mem := s3.Args[2]
+		dst := s3.Args[1]
+		if !(sc.Val64() >= 0 && isSameCall(sym, "runtime.memmove") && s1.Uses == 1 && s2.Uses == 1 && s3.Uses == 1 && isInlinableMemmove(dst, src, sc.Val64(), config) && clobber(s1, s2, s3, call)) {
+			break
+		}
+		v.reset(OpMove)
+		v.AuxInt = int64ToAuxInt(sc.Val64())
+		v.AddArg3(dst, src, mem)
+		return true
+	}
+	// match: (SelectN [0] call:(CALLstatic {sym} dst src (MOVQconst [sz]) mem))
+	// cond: sz >= 0 && isSameCall(sym, "runtime.memmove") && call.Uses == 1 && isInlinableMemmove(dst, src, sz, config) && clobber(call)
+	// result: (Move [sz] dst src mem)
+	for {
+		if auxIntToInt64(v.AuxInt) != 0 {
+			break
+		}
+		call := v_0
+		if call.Op != OpAMD64CALLstatic || len(call.Args) != 4 {
+			break
+		}
+		sym := auxToCall(call.Aux)
+		mem := call.Args[3]
+		dst := call.Args[0]
+		src := call.Args[1]
+		call_2 := call.Args[2]
+		if call_2.Op != OpAMD64MOVQconst {
+			break
+		}
+		sz := auxIntToInt64(call_2.AuxInt)
+		if !(sz >= 0 && isSameCall(sym, "runtime.memmove") && call.Uses == 1 && isInlinableMemmove(dst, src, sz, config) && clobber(call)) {
+			break
+		}
+		v.reset(OpMove)
+		v.AuxInt = int64ToAuxInt(sz)
+		v.AddArg3(dst, src, mem)
+		return true
+	}
+	return false
+}
 func rewriteValueAMD64_OpSlicemask(v *Value) bool {
 	v_0 := v.Args[0]
 	b := v.Block
diff --git a/src/cmd/compile/internal/ssa/rewriteARM64.go b/src/cmd/compile/internal/ssa/rewriteARM64.go
index 0ba3951df5..3cdc4d36cb 100644
--- a/src/cmd/compile/internal/ssa/rewriteARM64.go
+++ b/src/cmd/compile/internal/ssa/rewriteARM64.go
@@ -984,6 +984,8 @@ func rewriteValueARM64(v *Value) bool {
 		return rewriteValueARM64_OpSelect0(v)
 	case OpSelect1:
 		return rewriteValueARM64_OpSelect1(v)
+	case OpSelectN:
+		return rewriteValueARM64_OpSelectN(v)
 	case OpSignExt16to32:
 		v.Op = OpARM64MOVHreg
 		return true
@@ -25983,6 +25985,54 @@ func rewriteValueARM64_OpSelect1(v *Value) bool {
 	}
 	return false
 }
+func rewriteValueARM64_OpSelectN(v *Value) bool {
+	v_0 := v.Args[0]
+	b := v.Block
+	config := b.Func.Config
+	// match: (SelectN [0] call:(CALLstatic {sym} s1:(MOVDstore _ (MOVDconst [sz]) s2:(MOVDstore _ src s3:(MOVDstore {t} _ dst mem)))))
+	// cond: sz >= 0 && isSameCall(sym, "runtime.memmove") && s1.Uses == 1 && s2.Uses == 1 && s3.Uses == 1 && isInlinableMemmove(dst, src, sz, config) && clobber(s1, s2, s3, call)
+	// result: (Move [sz] dst src mem)
+	for {
+		if auxIntToInt64(v.AuxInt) != 0 {
+			break
+		}
+		call := v_0
+		if call.Op != OpARM64CALLstatic {
+			break
+		}
+		sym := auxToCall(call.Aux)
+		s1 := call.Args[0]
+		if s1.Op != OpARM64MOVDstore {
+			break
+		}
+		_ = s1.Args[2]
+		s1_1 := s1.Args[1]
+		if s1_1.Op != OpARM64MOVDconst {
+			break
+		}
+		sz := auxIntToInt64(s1_1.AuxInt)
+		s2 := s1.Args[2]
+		if s2.Op != OpARM64MOVDstore {
+			break
+		}
+		_ = s2.Args[2]
+		src := s2.Args[1]
+		s3 := s2.Args[2]
+		if s3.Op != OpARM64MOVDstore {
+			break
+		}
+		mem := s3.Args[2]
+		dst := s3.Args[1]
+		if !(sz >= 0 && isSameCall(sym, "runtime.memmove") && s1.Uses == 1 && s2.Uses == 1 && s3.Uses == 1 && isInlinableMemmove(dst, src, sz, config) && clobber(s1, s2, s3, call)) {
+			break
+		}
+		v.reset(OpMove)
+		v.AuxInt = int64ToAuxInt(sz)
+		v.AddArg3(dst, src, mem)
+		return true
+	}
+	return false
+}
 func rewriteValueARM64_OpSlicemask(v *Value) bool {
 	v_0 := v.Args[0]
 	b := v.Block