Use runtime dispatching

huff0/decompress_amd64.go

@@ -27,9 +27,6 @@ func decompress4x_main_loop_bmi2(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted,
 type decompress4x_main_loop_fn func(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted,
 	peekBits uint8, buf *byte, tbl *dEntrySingle) uint8
 
-var decompress4x_main_loop decompress4x_main_loop_fn
-var decompress4x_8b_main_loop decompress4x_main_loop_fn
-
 // decompress4x_8b_loop_x86 is an x86 assembler implementation
 // of Decompress4X when tablelog <= 8 which decodes 4 entries
 // per loop.
@@ -44,16 +41,6 @@ func decompress4x_8b_main_loop_amd64(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted,
 func decompress4x_8b_main_loop_bmi2(pbr0, pbr1, pbr2, pbr3 *bitReaderShifted,
 	peekBits uint8, buf *byte, tbl *dEntrySingle) uint8
 
-func init() {
-	if cpuinfo.HasBMI2() {
-		decompress4x_main_loop = decompress4x_main_loop_bmi2
-		decompress4x_8b_main_loop = decompress4x_8b_main_loop_bmi2
-	} else {
-		decompress4x_main_loop = decompress4x_main_loop_amd64
-		decompress4x_8b_main_loop = decompress4x_8b_main_loop_amd64
-	}
-}
-
 // fallback8BitSize is the size where using Go version is faster.
 const fallback8BitSize = 800
 
@@ -73,6 +60,9 @@ func (d *Decoder) Decompress4X(dst, src []byte) ([]byte, error) {
 	if cap(dst) < fallback8BitSize && use8BitTables {
 		return d.decompress4X8bit(dst, src)
 	}
+
+	bmi2 := cpuinfo.HasBMI2()
+
 	var br [4]bitReaderShifted
 	// Decode "jump table"
 	start := 6
@@ -120,9 +110,17 @@ func (d *Decoder) Decompress4X(dst, src []byte) ([]byte, error) {
 		}
 
 		if use8BitTables {
-			off = decompress4x_8b_main_loop(&br[0], &br[1], &br[2], &br[3], peekBits, &buf[0][0], &single[0])
+			if bmi2 {
+				off = decompress4x_8b_main_loop_bmi2(&br[0], &br[1], &br[2], &br[3], peekBits, &buf[0][0], &single[0])
+			} else {
+				off = decompress4x_8b_main_loop_amd64(&br[0], &br[1], &br[2], &br[3], peekBits, &buf[0][0], &single[0])
+			}
 		} else {
-			off = decompress4x_main_loop(&br[0], &br[1], &br[2], &br[3], peekBits, &buf[0][0], &single[0])
+			if bmi2 {
+				off = decompress4x_main_loop_bmi2(&br[0], &br[1], &br[2], &br[3], peekBits, &buf[0][0], &single[0])
+			} else {
+				off = decompress4x_main_loop_amd64(&br[0], &br[1], &br[2], &br[3], peekBits, &buf[0][0], &single[0])
+			}
 		}
 		if debug {
 			fmt.Print("DEBUG: ")