cmd/compile: split 3 operand LEA in late lower pass

On newer amd64 cpus 3 operand LEA instructions are slow, CL 114655 split
them to 2 LEA instructions in genssa.

This CL make late lower pass run after addressing modes, and split 3
operand LEA in late lower pass so that we can do common-subexpression
elimination for splited LEAs.

Updates #21735

Change-Id: Ied49139c7abab655e1a14a6fd793bdf9f987d1f1
Reviewed-on: https://go-review.googlesource.com/c/go/+/440035
TryBot-Result: Gopher Robot <[email protected]>
Run-TryBot: Wayne Zuo <[email protected]>
Reviewed-by: Keith Randall <[email protected]>
Reviewed-by: Keith Randall <[email protected]>
Reviewed-by: Joedian Reid <[email protected]>

src/cmd/compile/internal/ssa/_gen/AMD64latelower.rules

@@ -0,0 +1,11 @@
+// Copyright 2022 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// split 3 operand LEA.
+// Note: Don't split pointer computations in order to avoid invalid pointers.
+(LEA(Q|L|W)1 <t> [c] {s} x y) &&  isPtr(x.Type) && c != 0 && s == nil => (ADD(Q|L|L) x (ADD(Q|L|L)const <y.Type> [c] y))
+(LEA(Q|L|W)1 <t> [c] {s} x y) && !isPtr(x.Type) && c != 0 && s == nil => (ADD(Q|L|L) y (ADD(Q|L|L)const <x.Type> [c] x))
+(LEA(Q|L|W)2 <t> [c] {s} x y) && !isPtr(t)      && c != 0 && s == nil => (ADD(Q|L|L)const [c] (LEA(Q|L|W)2 <x.Type> x y))
+(LEA(Q|L|W)4 <t> [c] {s} x y) && !isPtr(t)      && c != 0 && s == nil => (ADD(Q|L|L)const [c] (LEA(Q|L|W)4 <x.Type> x y))
+(LEA(Q|L|W)8 <t> [c] {s} x y) && !isPtr(t)      && c != 0 && s == nil => (ADD(Q|L|L)const [c] (LEA(Q|L|W)8 <x.Type> x y))

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

@@ -486,8 +486,8 @@ var passes = [...]pass{
 	{name: "insert resched checks", fn: insertLoopReschedChecks,
 		disabled: !buildcfg.Experiment.PreemptibleLoops}, // insert resched checks in loops.
 	{name: "lower", fn: lower, required: true},
-	{name: "late lower", fn: lateLower, required: true},
 	{name: "addressing modes", fn: addressingModes, required: false},
+	{name: "late lower", fn: lateLower, required: true},
 	{name: "lowered deadcode for cse", fn: deadcode}, // deadcode immediately before CSE avoids CSE making dead values live again
 	{name: "lowered cse", fn: cse},
 	{name: "elim unread autos", fn: elimUnreadAutos},

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

@@ -185,6 +185,7 @@ func NewConfig(arch string, types Types, ctxt *obj.Link, optimize, softfloat boo
 		c.RegSize = 8
 		c.lowerBlock = rewriteBlockAMD64
 		c.lowerValue = rewriteValueAMD64
+		c.lateLowerValue = rewriteValueAMD64latelower
 		c.splitLoad = rewriteValueAMD64splitload
 		c.registers = registersAMD64[:]
 		c.gpRegMask = gpRegMaskAMD64