cmd/compile: optimize loads from readonly globals into constants

Instead of
   MOVB go.string."foo"(SB), AX
do
   MOVB $102, AX

When we know the global we're loading from is readonly, we can
do that read at compile time.

I've made this arch-dependent mostly because the cases where this
happens often are memory->memory moves, and those don't get
decomposed until lowering.

Did amd64/386/arm/arm64. Other architectures could follow.

Update #26498

Change-Id: I41b1dc831b2cd0a52dac9b97f4f4457888a46389
Reviewed-on: https://go-review.googlesource.com/c/141118
Run-TryBot: Keith Randall <[email protected]>
TryBot-Result: Gobot Gobot <[email protected]>
Reviewed-by: Josh Bleecher Snyder <[email protected]>

Author: randall77

src/cmd/compile/internal/ssa/gen/386.rules

@@ -1206,3 +1206,7 @@
 (CMPLload {sym} [off] ptr (MOVLconst [c]) mem) && validValAndOff(int64(int32(c)),off) -> (CMPLconstload {sym} [makeValAndOff(int64(int32(c)),off)] ptr mem)
 (CMPWload {sym} [off] ptr (MOVLconst [c]) mem) && validValAndOff(int64(int16(c)),off) -> (CMPWconstload {sym} [makeValAndOff(int64(int16(c)),off)] ptr mem)
 (CMPBload {sym} [off] ptr (MOVLconst [c]) mem) && validValAndOff(int64(int8(c)),off) -> (CMPBconstload {sym} [makeValAndOff(int64(int8(c)),off)] ptr mem)
+
+(MOVBload [off] {sym} (SB) _) && symIsRO(sym) -> (MOVLconst [int64(read8(sym, off))])
+(MOVWload [off] {sym} (SB) _) && symIsRO(sym) -> (MOVLconst [int64(read16(sym, off, config.BigEndian))])
+(MOVLload [off] {sym} (SB) _) && symIsRO(sym) -> (MOVLconst [int64(int32(read32(sym, off, config.BigEndian)))])

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

@@ -1073,11 +1073,11 @@
 // Fold constants into stores.
 (MOVQstore [off] {sym} ptr (MOVQconst [c]) mem) && validValAndOff(c,off) ->
 	(MOVQstoreconst [makeValAndOff(c,off)] {sym} ptr mem)
-(MOVLstore [off] {sym} ptr (MOVLconst [c]) mem) && validOff(off) ->
+(MOVLstore [off] {sym} ptr (MOV(L|Q)const [c]) mem) && validOff(off) ->
 	(MOVLstoreconst [makeValAndOff(int64(int32(c)),off)] {sym} ptr mem)
-(MOVWstore [off] {sym} ptr (MOVLconst [c]) mem) && validOff(off) ->
+(MOVWstore [off] {sym} ptr (MOV(L|Q)const [c]) mem) && validOff(off) ->
 	(MOVWstoreconst [makeValAndOff(int64(int16(c)),off)] {sym} ptr mem)
-(MOVBstore [off] {sym} ptr (MOVLconst [c]) mem) && validOff(off) ->
+(MOVBstore [off] {sym} ptr (MOV(L|Q)const [c]) mem) && validOff(off) ->
 	(MOVBstoreconst [makeValAndOff(int64(int8(c)),off)] {sym} ptr mem)
 
 // Fold address offsets into constant stores.
@@ -2487,3 +2487,8 @@
 	&& validValAndOff(0,off)
 	&& clobber(l) ->
   @l.Block (CMP(Q|L|W|B)constload {sym} [makeValAndOff(0,off)] ptr mem)
+
+(MOVBload [off] {sym} (SB) _) && symIsRO(sym) -> (MOVLconst [int64(read8(sym, off))])
+(MOVWload [off] {sym} (SB) _) && symIsRO(sym) -> (MOVLconst [int64(read16(sym, off, config.BigEndian))])
+(MOVLload [off] {sym} (SB) _) && symIsRO(sym) -> (MOVQconst [int64(read32(sym, off, config.BigEndian))])
+(MOVQload [off] {sym} (SB) _) && symIsRO(sym) -> (MOVQconst [int64(read64(sym, off, config.BigEndian))])

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

@@ -1544,3 +1544,7 @@
 (GE (CMPconst [0] l:(XORshiftLLreg x y z)) yes no) && l.Uses==1 -> (GE (TEQshiftLLreg x y z) yes no)
 (GE (CMPconst [0] l:(XORshiftRLreg x y z)) yes no) && l.Uses==1 -> (GE (TEQshiftRLreg x y z) yes no)
 (GE (CMPconst [0] l:(XORshiftRAreg x y z)) yes no) && l.Uses==1 -> (GE (TEQshiftRAreg x y z) yes no)
+
+(MOVBUload [off] {sym} (SB) _) && symIsRO(sym) -> (MOVWconst [int64(read8(sym, off))])
+(MOVHUload [off] {sym} (SB) _) && symIsRO(sym) -> (MOVWconst [int64(read16(sym, off, config.BigEndian))])
+(MOVWload [off] {sym} (SB) _) && symIsRO(sym) -> (MOVWconst [int64(int32(read32(sym, off, config.BigEndian)))])

src/cmd/compile/internal/ssa/gen/ARM64.rules

@@ -2957,3 +2957,8 @@
 (FSUBD a (FNMULD x y)) -> (FMADDD a x y)
 (FSUBS (FNMULS x y) a) -> (FNMADDS a x y)
 (FSUBD (FNMULD x y) a) -> (FNMADDD a x y)
+
+(MOVBUload [off] {sym} (SB) _) && symIsRO(sym) -> (MOVDconst [int64(read8(sym, off))])
+(MOVHUload [off] {sym} (SB) _) && symIsRO(sym) -> (MOVDconst [int64(read16(sym, off, config.BigEndian))])
+(MOVWUload [off] {sym} (SB) _) && symIsRO(sym) -> (MOVDconst [int64(read32(sym, off, config.BigEndian))])
+(MOVDload  [off] {sym} (SB) _) && symIsRO(sym) -> (MOVDconst [int64(read64(sym, off, config.BigEndian))])

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

@@ -7,7 +7,9 @@ package ssa
 import (
 	"cmd/compile/internal/types"
 	"cmd/internal/obj"
+	"cmd/internal/objabi"
 	"cmd/internal/src"
+	"encoding/binary"
 	"fmt"
 	"io"
 	"math"
@@ -1090,3 +1092,45 @@ func needRaceCleanup(sym interface{}, v *Value) bool {
 	}
 	return true
 }
+
+// symIsRO reports whether sym is a read-only global.
+func symIsRO(sym interface{}) bool {
+	lsym := sym.(*obj.LSym)
+	return lsym.Type == objabi.SRODATA && len(lsym.R) == 0
+}
+
+// read8 reads one byte from the read-only global sym at offset off.
+func read8(sym interface{}, off int64) uint8 {
+	lsym := sym.(*obj.LSym)
+	return lsym.P[off]
+}
+
+// read16 reads two bytes from the read-only global sym at offset off.
+func read16(sym interface{}, off int64, bigEndian bool) uint16 {
+	lsym := sym.(*obj.LSym)
+	if bigEndian {
+		return binary.BigEndian.Uint16(lsym.P[off:])
+	} else {
+		return binary.LittleEndian.Uint16(lsym.P[off:])
+	}
+}
+
+// read32 reads four bytes from the read-only global sym at offset off.
+func read32(sym interface{}, off int64, bigEndian bool) uint32 {
+	lsym := sym.(*obj.LSym)
+	if bigEndian {
+		return binary.BigEndian.Uint32(lsym.P[off:])
+	} else {
+		return binary.LittleEndian.Uint32(lsym.P[off:])
+	}
+}
+
+// read64 reads eight bytes from the read-only global sym at offset off.
+func read64(sym interface{}, off int64, bigEndian bool) uint64 {
+	lsym := sym.(*obj.LSym)
+	if bigEndian {
+		return binary.BigEndian.Uint64(lsym.P[off:])
+	} else {
+		return binary.LittleEndian.Uint64(lsym.P[off:])
+	}
+}