cmd/compile: fix load combining from a non-zero pointer offset

When the pointer offset is non-zero in the small loads, we need to add the offset
when converting to the larger load.

Fixes #14694

Change-Id: I5ba8bcb3b9ce26c7fae0c4951500b9ef0fed54cd
Reviewed-on: https://go-review.googlesource.com/20333Reviewed-by: Keith Randall <khr@golang.org>

src/cmd/compile/internal/gc/testdata/arith_ssa.go

@@ -14,6 +14,40 @@ const (
 	y = 0x0fffFFFF
 )
 
+//go:noinline
+func parseLE64(b []byte) uint64 {
+	// skip the first two bytes, and parse the remaining 8 as a uint64
+	return uint64(b[2]) | uint64(b[3])<<8 | uint64(b[4])<<16 | uint64(b[5])<<24 |
+		uint64(b[6])<<32 | uint64(b[7])<<40 | uint64(b[8])<<48 | uint64(b[9])<<56
+}
+
+//go:noinline
+func parseLE32(b []byte) uint32 {
+	return uint32(b[2]) | uint32(b[3])<<8 | uint32(b[4])<<16 | uint32(b[5])<<24
+}
+
+//go:noinline
+func parseLE16(b []byte) uint16 {
+	return uint16(b[2]) | uint16(b[3])<<8
+}
+
+// testLoadCombine tests for issue #14694 where load combining didn't respect the pointer offset.
+func testLoadCombine() {
+	testData := []byte{0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, 0x09}
+	if want, got := uint64(0x0908070605040302), parseLE64(testData); want != got {
+		println("testLargeConst add failed, wanted", want, "got", got)
+		failed = true
+	}
+	if want, got := uint32(0x05040302), parseLE32(testData); want != got {
+		println("testLargeConst add failed, wanted", want, "got", got)
+		failed = true
+	}
+	if want, got := uint16(0x0302), parseLE16(testData); want != got {
+		println("testLargeConst add failed, wanted", want, "got", got)
+		failed = true
+	}
+}
+
 //go:noinline
 func invalidAdd_ssa(x uint32) uint32 {
 	return x + y + y + y + y + y + y + y + y + y + y + y + y + y + y + y + y + y
@@ -431,6 +465,7 @@ func main() {
 	testArithConstShift()
 	testArithRshConst()
 	testLargeConst()
+	testLoadCombine()
 
 	if failed {
 		panic("failed")

src/cmd/compile/internal/ssa/gen/AMD64.rules

@@ -1167,13 +1167,13 @@
 // There are many ways these combinations could occur.  This is
 // designed to match the way encoding/binary.LittleEndian does it.
 (ORW                (MOVBQZXload [i]   {s} p mem)
-     (SHLWconst [8] (MOVBQZXload [i+1] {s} p mem))) -> (MOVWload p mem)
+     (SHLWconst [8] (MOVBQZXload [i+1] {s} p mem))) -> (MOVWload (ADDQconst [i] p) mem)
 
 (ORL (ORL (ORL
                     (MOVBQZXload [i]   {s} p mem)
     (SHLLconst [8]  (MOVBQZXload [i+1] {s} p mem)))
     (SHLLconst [16] (MOVBQZXload [i+2] {s} p mem)))
-    (SHLLconst [24] (MOVBQZXload [i+3] {s} p mem))) -> (MOVLload p mem)
+    (SHLLconst [24] (MOVBQZXload [i+3] {s} p mem))) -> (MOVLload (ADDQconst [i] p) mem)
 
 (ORQ (ORQ (ORQ (ORQ (ORQ (ORQ (ORQ
                     (MOVBQZXload [i]   {s} p mem)
@@ -1183,4 +1183,4 @@
     (SHLQconst [32] (MOVBQZXload [i+4] {s} p mem)))
     (SHLQconst [40] (MOVBQZXload [i+5] {s} p mem)))
     (SHLQconst [48] (MOVBQZXload [i+6] {s} p mem)))
-    (SHLQconst [56] (MOVBQZXload [i+7] {s} p mem))) -> (MOVQload p mem)
+    (SHLQconst [56] (MOVBQZXload [i+7] {s} p mem))) -> (MOVQload (ADDQconst [i] p) mem)

src/cmd/compile/internal/ssa/rewriteAMD64.go

@@ -9465,7 +9465,7 @@ func rewriteValueAMD64_OpAMD64ORL(v *Value, config *Config) bool {
 	}
 	// match: (ORL (ORL (ORL                     (MOVBQZXload [i]   {s} p mem)     (SHLLconst [8]  (MOVBQZXload [i+1] {s} p mem)))     (SHLLconst [16] (MOVBQZXload [i+2] {s} p mem)))     (SHLLconst [24] (MOVBQZXload [i+3] {s} p mem)))
 	// cond:
-	// result: (MOVLload p mem)
+	// result: (MOVLload (ADDQconst [i] p) mem)
 	for {
 		if v.Args[0].Op != OpAMD64ORL {
 			break
@@ -9544,7 +9544,10 @@ func rewriteValueAMD64_OpAMD64ORL(v *Value, config *Config) bool {
 			break
 		}
 		v.reset(OpAMD64MOVLload)
-		v.AddArg(p)
+		v0 := b.NewValue0(v.Line, OpAMD64ADDQconst, config.fe.TypeUInt64())
+		v0.AuxInt = i
+		v0.AddArg(p)
+		v.AddArg(v0)
 		v.AddArg(mem)
 		return true
 	}
@@ -9646,7 +9649,7 @@ func rewriteValueAMD64_OpAMD64ORQ(v *Value, config *Config) bool {
 	}
 	// match: (ORQ (ORQ (ORQ (ORQ (ORQ (ORQ (ORQ                     (MOVBQZXload [i]   {s} p mem)     (SHLQconst [8]  (MOVBQZXload [i+1] {s} p mem)))     (SHLQconst [16] (MOVBQZXload [i+2] {s} p mem)))     (SHLQconst [24] (MOVBQZXload [i+3] {s} p mem)))     (SHLQconst [32] (MOVBQZXload [i+4] {s} p mem)))     (SHLQconst [40] (MOVBQZXload [i+5] {s} p mem)))     (SHLQconst [48] (MOVBQZXload [i+6] {s} p mem)))     (SHLQconst [56] (MOVBQZXload [i+7] {s} p mem)))
 	// cond:
-	// result: (MOVQload p mem)
+	// result: (MOVQload (ADDQconst [i] p) mem)
 	for {
 		if v.Args[0].Op != OpAMD64ORQ {
 			break
@@ -9821,7 +9824,10 @@ func rewriteValueAMD64_OpAMD64ORQ(v *Value, config *Config) bool {
 			break
 		}
 		v.reset(OpAMD64MOVQload)
-		v.AddArg(p)
+		v0 := b.NewValue0(v.Line, OpAMD64ADDQconst, config.fe.TypeUInt64())
+		v0.AuxInt = i
+		v0.AddArg(p)
+		v.AddArg(v0)
 		v.AddArg(mem)
 		return true
 	}
@@ -9915,7 +9921,7 @@ func rewriteValueAMD64_OpAMD64ORW(v *Value, config *Config) bool {
 	}
 	// match: (ORW                (MOVBQZXload [i]   {s} p mem)      (SHLWconst [8] (MOVBQZXload [i+1] {s} p mem)))
 	// cond:
-	// result: (MOVWload p mem)
+	// result: (MOVWload (ADDQconst [i] p) mem)
 	for {
 		if v.Args[0].Op != OpAMD64MOVBQZXload {
 			break
@@ -9946,7 +9952,10 @@ func rewriteValueAMD64_OpAMD64ORW(v *Value, config *Config) bool {
 			break
 		}
 		v.reset(OpAMD64MOVWload)
-		v.AddArg(p)
+		v0 := b.NewValue0(v.Line, OpAMD64ADDQconst, config.fe.TypeUInt64())
+		v0.AuxInt = i
+		v0.AddArg(p)
+		v.AddArg(v0)
 		v.AddArg(mem)
 		return true
 	}