From 04945edd40fff4d66321a4f98c1bb070b6356008 Mon Sep 17 00:00:00 2001
From: =?utf8?q?Alexandru=20Mo=C8=99oi?= <mosoi@google.com>
Date: Mon, 4 Apr 2016 19:23:41 +0200
Subject: [PATCH] cmd/compile: replaces ANDQ with MOV?ZX
MIME-Version: 1.0
Content-Type: text/plain; charset=utf8
Content-Transfer-Encoding: 8bit

Where possible replace ANDQ with MOV?ZX.
Takes care that we don't regress wrt bounds checking,
for example [1000]int{}[i&255].

According to "Intel 64 and IA-32 Architectures Optimization Reference
Manual" Section: "3.5.1.13 Zero-Latency MOV Instructions"
MOV?ZX instructions have zero latency on newer processors.

Updates #15105

Change-Id: I63539fdbc5812d5563aa1ebc49eca035bd307997
Reviewed-on: https://go-review.googlesource.com/21508
Reviewed-by: Айнар Гарипов <gugl.zadolbal@gmail.com>
Reviewed-by: David Chase <drchase@google.com>
---
 src/cmd/compile/internal/ssa/gen/AMD64.rules |  8 ++
 src/cmd/compile/internal/ssa/rewriteAMD64.go | 81 ++++++++++++++++++++
 2 files changed, 89 insertions(+)

diff --git a/src/cmd/compile/internal/ssa/gen/AMD64.rules b/src/cmd/compile/internal/ssa/gen/AMD64.rules
index 4ad0f883b0..b37720eb39 100644
--- a/src/cmd/compile/internal/ssa/gen/AMD64.rules
+++ b/src/cmd/compile/internal/ssa/gen/AMD64.rules
@@ -587,6 +587,11 @@
 (CMPB x (MOVBconst [c])) -> (CMPBconst x [c])
 (CMPB (MOVBconst [c]) x) -> (InvertFlags (CMPBconst x [c]))
 
+// Using MOVBQZX instead of ANDQ is cheaper.
+(ANDQconst [0xFF] x) -> (MOVBQZX x)
+(ANDQconst [0xFFFF] x) -> (MOVWQZX x)
+(ANDQconst [0xFFFFFFFF] x) -> (MOVLQZX x)
+
 // strength reduction
 // Assumes that the following costs from https://gmplib.org/~tege/x86-timing.pdf:
 //    1 - addq, shlq, leaq, negq
@@ -1093,6 +1098,9 @@
 (CMPBconst (MOVBconst [x]) [y]) && int8(x)>int8(y) && uint8(x)>uint8(y) -> (FlagGT_UGT)
 
 // Other known comparisons.
+(CMPQconst (MOVBQZX _) [c]) && 0xFF < c -> (FlagLT_ULT)
+(CMPQconst (MOVWQZX _) [c]) && 0xFFFF < c -> (FlagLT_ULT)
+(CMPQconst (MOVLQZX _) [c]) && 0xFFFFFFFF < c -> (FlagLT_ULT)
 (CMPQconst (ANDQconst _ [m]) [n]) && 0 <= m && m < n -> (FlagLT_ULT)
 (CMPLconst (ANDLconst _ [m]) [n]) && 0 <= int32(m) && int32(m) < int32(n) -> (FlagLT_ULT)
 (CMPWconst (ANDWconst _ [m]) [n]) && 0 <= int16(m) && int16(m) < int16(n) -> (FlagLT_ULT)
diff --git a/src/cmd/compile/internal/ssa/rewriteAMD64.go b/src/cmd/compile/internal/ssa/rewriteAMD64.go
index 11c2de391c..a1d1e4edd9 100644
--- a/src/cmd/compile/internal/ssa/rewriteAMD64.go
+++ b/src/cmd/compile/internal/ssa/rewriteAMD64.go
@@ -1838,6 +1838,42 @@ func rewriteValueAMD64_OpAMD64ANDQconst(v *Value, config *Config) bool {
 		v.AddArg(x)
 		return true
 	}
+	// match: (ANDQconst [0xFF] x)
+	// cond:
+	// result: (MOVBQZX x)
+	for {
+		if v.AuxInt != 0xFF {
+			break
+		}
+		x := v.Args[0]
+		v.reset(OpAMD64MOVBQZX)
+		v.AddArg(x)
+		return true
+	}
+	// match: (ANDQconst [0xFFFF] x)
+	// cond:
+	// result: (MOVWQZX x)
+	for {
+		if v.AuxInt != 0xFFFF {
+			break
+		}
+		x := v.Args[0]
+		v.reset(OpAMD64MOVWQZX)
+		v.AddArg(x)
+		return true
+	}
+	// match: (ANDQconst [0xFFFFFFFF] x)
+	// cond:
+	// result: (MOVLQZX x)
+	for {
+		if v.AuxInt != 0xFFFFFFFF {
+			break
+		}
+		x := v.Args[0]
+		v.reset(OpAMD64MOVLQZX)
+		v.AddArg(x)
+		return true
+	}
 	// match: (ANDQconst [0] _)
 	// cond:
 	// result: (MOVQconst [0])
@@ -3026,6 +3062,51 @@ func rewriteValueAMD64_OpAMD64CMPQconst(v *Value, config *Config) bool {
 		v.reset(OpAMD64FlagGT_UGT)
 		return true
 	}
+	// match: (CMPQconst (MOVBQZX _) [c])
+	// cond: 0xFF < c
+	// result: (FlagLT_ULT)
+	for {
+		v_0 := v.Args[0]
+		if v_0.Op != OpAMD64MOVBQZX {
+			break
+		}
+		c := v.AuxInt
+		if !(0xFF < c) {
+			break
+		}
+		v.reset(OpAMD64FlagLT_ULT)
+		return true
+	}
+	// match: (CMPQconst (MOVWQZX _) [c])
+	// cond: 0xFFFF < c
+	// result: (FlagLT_ULT)
+	for {
+		v_0 := v.Args[0]
+		if v_0.Op != OpAMD64MOVWQZX {
+			break
+		}
+		c := v.AuxInt
+		if !(0xFFFF < c) {
+			break
+		}
+		v.reset(OpAMD64FlagLT_ULT)
+		return true
+	}
+	// match: (CMPQconst (MOVLQZX _) [c])
+	// cond: 0xFFFFFFFF < c
+	// result: (FlagLT_ULT)
+	for {
+		v_0 := v.Args[0]
+		if v_0.Op != OpAMD64MOVLQZX {
+			break
+		}
+		c := v.AuxInt
+		if !(0xFFFFFFFF < c) {
+			break
+		}
+		v.reset(OpAMD64FlagLT_ULT)
+		return true
+	}
 	// match: (CMPQconst (ANDQconst _ [m]) [n])
 	// cond: 0 <= m && m < n
 	// result: (FlagLT_ULT)
-- 
2.48.1