From cdc2da64d4c55a97c4507b6c9389bf4dca04695b Mon Sep 17 00:00:00 2001
From: notaz <notasas@gmail.com>
Date: Thu, 3 Feb 2022 02:04:52 +0200
Subject: [PATCH] drc: use optimized get_reg

arm32-only for now
---
 libpcsxcore/new_dynarec/linkage_arm.S | 40 +++++++++++++++++++++++++++
 libpcsxcore/new_dynarec/new_dynarec.c |  9 ++++++
 2 files changed, 49 insertions(+)

diff --git a/libpcsxcore/new_dynarec/linkage_arm.S b/libpcsxcore/new_dynarec/linkage_arm.S
index 8d9074f4..1a16aa04 100644
--- a/libpcsxcore/new_dynarec/linkage_arm.S
+++ b/libpcsxcore/new_dynarec/linkage_arm.S
@@ -840,4 +840,44 @@ FUNCTION(call_gteStall):
 	add	r10, r10, r0
 	bx	lr
 
+#ifdef HAVE_ARMV6
+
+FUNCTION(get_reg):
+	ldr	r12, [r0]
+	and	r1, r1, #0xff
+	ldr	r2, [r0, #4]
+	orr	r1, r1, r1, lsl #8
+	ldr	r3, [r0, #8]
+	orr	r1, r1, r1, lsl #16   @ searched char in every byte
+	ldrb    r0, [r0, #12]         @ last byte
+	eor	r12, r12, r1
+	eor	r2, r2, r1
+	eor	r3, r3, r1
+	cmp     r0, r1, lsr #24
+	mov	r0, #12
+	mvn	r1, #0                @ r1=~0
+	bxeq	lr
+	orr	r3, r3, #0xff000000   @ EXCLUDE_REG
+	uadd8	r0, r12, r1           @ add and set GE bits when not 0 (match)
+	mov	r12, #0
+	sel	r0, r12, r1           @ 0 if no match, else ff in some byte
+	uadd8	r2, r2, r1
+	sel	r2, r12, r1
+	uadd8	r3, r3, r1
+	sel	r3, r12, r1
+	mov	r12, #3
+	clz     r0, r0                @ 0, 8, 16, 24 or 32
+	clz     r2, r2
+	clz     r3, r3
+	sub	r0, r12, r0, lsr #3   @ 3, 2, 1, 0 or -1
+	sub	r2, r12, r2, lsr #3
+	sub	r3, r12, r3, lsr #3
+	orr	r2, r2, #4
+	orr	r3, r3, #8
+	and	r0, r0, r2
+	and	r0, r0, r3
+	bx	lr
+
+#endif /* HAVE_ARMV6 */
+
 @ vim:filetype=armasm
diff --git a/libpcsxcore/new_dynarec/new_dynarec.c b/libpcsxcore/new_dynarec/new_dynarec.c
index abb0d076..0dea9a35 100644
--- a/libpcsxcore/new_dynarec/new_dynarec.c
+++ b/libpcsxcore/new_dynarec/new_dynarec.c
@@ -35,6 +35,7 @@
 #include "../psxinterpreter.h"
 #include "../gte.h"
 #include "emu_if.h" // emulator interface
+#include "arm_features.h"
 
 #define noinline __attribute__((noinline,noclone))
 #ifndef ARRAY_SIZE
@@ -607,6 +608,12 @@ static void clear_all_regs(signed char regmap[])
   memset(regmap, -1, sizeof(regmap[0]) * HOST_REGS);
 }
 
+#if defined(__arm__) && defined(HAVE_ARMV6) && HOST_REGS == 13 && EXCLUDE_REG == 11
+
+extern signed char get_reg(const signed char regmap[], signed char r);
+
+#else
+
 static signed char get_reg(const signed char regmap[], signed char r)
 {
   int hr;
@@ -619,6 +626,8 @@ static signed char get_reg(const signed char regmap[], signed char r)
   return -1;
 }
 
+#endif
+
 static signed char get_reg_temp(const signed char regmap[])
 {
   int hr;
-- 
2.39.5