From 3a85d802ee63cce871963b28de50fd9ffd8900ac Mon Sep 17 00:00:00 2001
From: Alexander von Gluck IV <kallisti5@unixzen.com>
Date: Mon, 27 Jan 2020 22:23:13 -0600
Subject: [PATCH] libroot/riscv64: Fix build.  Add MIT optimized bswap from rv8

Change-Id: I4d0e1ec34d1568594cac0f7378f87852c997ab81
Reviewed-on: https://review.haiku-os.org/c/haiku/+/2166
Reviewed-by: Alex von Gluck IV <kallisti5@unixzen.com>
Reviewed-by: Adrien Destugues <pulkomandy@gmail.com>
---
 src/system/libroot/os/arch/riscv64/Jamfile    |  2 +
 .../libroot/os/arch/riscv64/byteorder.S       | 68 ++++++++++++++-----
 2 files changed, 53 insertions(+), 17 deletions(-)

diff --git a/src/system/libroot/os/arch/riscv64/Jamfile b/src/system/libroot/os/arch/riscv64/Jamfile
index 64aa1a8422..e16a245868 100644
--- a/src/system/libroot/os/arch/riscv64/Jamfile
+++ b/src/system/libroot/os/arch/riscv64/Jamfile
@@ -13,6 +13,8 @@ for architectureObject in [ MultiArchSubDirSetup riscv64 ] {
 		SEARCH_SOURCE += [ FDirName $(SUBDIR) $(DOTDOT) generic ] ;
 
 		MergeObject <$(architecture)>os_arch_$(TARGET_ARCH).o :
+			byteorder.S
+
 			stack_frame.c
 			system_time.c
 
diff --git a/src/system/libroot/os/arch/riscv64/byteorder.S b/src/system/libroot/os/arch/riscv64/byteorder.S
index 9bded18ee4..06fa75ed35 100644
--- a/src/system/libroot/os/arch/riscv64/byteorder.S
+++ b/src/system/libroot/os/arch/riscv64/byteorder.S
@@ -1,30 +1,41 @@
 /*
- * Copyright 2009, Johannes Wischert. All rights reserved.
+ * Copyright 2018-2020, Haiku, Inc. All rights reserved.
  * Distributed under the terms of the MIT License.
+ *
+ * Optimized byteswap functions from the rv8 test
+ * suite licensed under the MIT License
+ * https://github.com/rv8-io/rv8/tree/master/src/test
+ *
  */
 
 #include <asm_defs.h>
 
 .text
 
-/*
- * These aren't needed on gcc4+
- *  uint16 __swap_int16(uint16 value)
- *  uint32 __swap_int32(uint32 value)
- *  uint64 __swap_int64(uint64 value)
- */
-
-/* TODO: The following functions can surely be optimized. A simple optimization
- * would be to define macros with the contents of the __swap_int{32,64}
- * functions and use those instead of calling the functions.
- */
-
 /* float __swap_float(float value)
  */
 FUNCTION(__swap_float):
 		// Assumes single precision
-		b		__swap_int32
-		//rts
+	1:	auipc   a4, %pcrel_hi(__bswap64_c1)
+		ld      a4, %pcrel_lo(1b)(a4)
+		slli    a5, a0, 8
+		and     a5, a5, a4
+		srli    a0, a0, 8
+		srli    a4, a4, 8
+		and     a0, a0, a4
+		or      a5, a5, a0
+	1:	auipc   a4, %pcrel_hi(__bswap64_c2)
+		ld      a4, %pcrel_lo(1b)(a4)
+		slli    a0, a5, 16
+		and     a0, a0, a4
+		srli    a5, a5, 16
+		srli    a4, a4, 16
+		and     a5, a5, a4
+		or      a5, a5, a0
+		slli    a0, a5, 32
+		srli    a5, a5, 32
+		or      a0, a0, a5
+		ret
 FUNCTION_END(__swap_float)
 
 
@@ -32,7 +43,30 @@ FUNCTION_END(__swap_float)
  */
 FUNCTION(__swap_double):
 		// Assumes double is int64 on RV64
-		b		__swap_int64
-		//rts
+	1:	auipc   a4, %pcrel_hi(__bswap64_c1)
+		ld      a4, %pcrel_lo(1b)(a4)
+		slli    a5, a0, 8
+		and     a5, a5, a4
+		srli    a0, a0, 8
+		srli    a4, a4, 8
+		and     a0, a0, a4
+		or      a5, a5, a0
+	1:	auipc   a4, %pcrel_hi(__bswap64_c2)
+		ld      a4, %pcrel_lo(1b)(a4)
+		slli    a0, a5, 16
+		and     a0, a0, a4
+		srli    a5, a5, 16
+		srli    a4, a4, 16
+		and     a5, a5, a4
+		or      a5, a5, a0
+		slli    a0, a5, 32
+		srli    a5, a5, 32
+		or      a0, a0, a5
+		ret
 FUNCTION_END(__swap_double)
 
+
+.section .rodata
+__bswap32_c1: .4byte 0xFF00FF00ULL
+__bswap64_c1: .8byte 0xFF00FF00FF00FF00ULL
+__bswap64_c2: .8byte 0xFFFF0000FFFF0000ULL