From ee344f0dcab22151b49a1f2eff693bb904741366 Mon Sep 17 00:00:00 2001
From: Jeffrey Pfau <jeffrey@endrift.com>
Date: Tue, 15 Oct 2013 00:50:07 -0700
Subject: [PATCH] Optimize and fix IntrWait, make sure that SVC doesn't trample
 its input register r0

---
 src/gba/hle-bios.c | 30 +++++++++++------------
 src/gba/hle-bios.s | 59 +++++++++++++++++++++++-----------------------
 2 files changed, 44 insertions(+), 45 deletions(-)

diff --git a/src/gba/hle-bios.c b/src/gba/hle-bios.c
index 584279460..9b03fcfb7 100644
--- a/src/gba/hle-bios.c
+++ b/src/gba/hle-bios.c
@@ -1,24 +1,22 @@
 #include "hle-bios.h"
 
-const unsigned int hleBiosLength = 220;
+const unsigned int hleBiosLength = 196;
 const unsigned char hleBios[] = {
   0x06, 0x00, 0x00, 0xea, 0xfe, 0xff, 0xff, 0xea, 0x05, 0x00, 0x00, 0xea,
   0xfe, 0xff, 0xff, 0xea, 0xfe, 0xff, 0xff, 0xea, 0x00, 0x00, 0xa0, 0xe1,
-  0x0c, 0x00, 0x00, 0xea, 0xfe, 0xff, 0xff, 0xea, 0x02, 0xf3, 0xa0, 0xe3,
+  0x0e, 0x00, 0x00, 0xea, 0xfe, 0xff, 0xff, 0xea, 0x02, 0xf3, 0xa0, 0xe3,
   0x00, 0x00, 0x5d, 0xe3, 0x01, 0xd3, 0xa0, 0x03, 0x20, 0xd0, 0x4d, 0x02,
-  0x00, 0x40, 0x2d, 0xe9, 0x02, 0x00, 0x5e, 0xe5, 0x04, 0x00, 0x50, 0xe3,
-  0x09, 0x00, 0x00, 0x0b, 0x05, 0x00, 0x50, 0xe3, 0x07, 0x00, 0x00, 0x0b,
-  0x00, 0x40, 0xbd, 0xe8, 0x0e, 0xf0, 0xb0, 0xe1, 0x0f, 0x50, 0x2d, 0xe9,
-  0x01, 0x03, 0xa0, 0xe3, 0x00, 0xe0, 0x8f, 0xe2, 0x04, 0xf0, 0x10, 0xe5,
-  0x0f, 0x50, 0xbd, 0xe8, 0x04, 0xf0, 0x5e, 0xe2, 0x10, 0x40, 0x2d, 0xe9,
-  0x04, 0xd0, 0x4d, 0xe2, 0xb0, 0x10, 0xcd, 0xe1, 0x01, 0x43, 0xa0, 0xe3,
-  0x02, 0x4c, 0x84, 0xe2, 0xb0, 0x00, 0xd4, 0xe1, 0xb2, 0x00, 0xcd, 0xe1,
-  0xb0, 0x10, 0xdd, 0xe1, 0x01, 0x00, 0x80, 0xe1, 0xb0, 0x00, 0xc4, 0xe1,
-  0x01, 0x43, 0xa0, 0xe3, 0x1f, 0x00, 0xa0, 0xe3, 0x00, 0xf0, 0x29, 0xe1,
-  0x00, 0x00, 0xa0, 0xe3, 0x01, 0x03, 0xc4, 0xe5, 0xd3, 0x00, 0xa0, 0xe3,
-  0x00, 0xf0, 0x29, 0xe1, 0xb8, 0x00, 0x54, 0xe1, 0xb0, 0x10, 0xdd, 0xe1,
-  0x00, 0x10, 0x11, 0xe0, 0x00, 0x10, 0x21, 0x10, 0xb8, 0x10, 0x44, 0x11,
-  0xf3, 0xff, 0xff, 0x0a, 0x01, 0x43, 0xa0, 0xe3, 0x02, 0x4c, 0x84, 0xe2,
-  0xb2, 0x00, 0xdd, 0xe1, 0xb0, 0x00, 0xc4, 0xe1, 0x04, 0xd0, 0x8d, 0xe2,
+  0x04, 0x40, 0x2d, 0xe9, 0x02, 0x20, 0x5e, 0xe5, 0x04, 0x00, 0x52, 0xe3,
+  0x0b, 0x00, 0x00, 0x0b, 0x05, 0x00, 0x52, 0xe3, 0x01, 0x00, 0xa0, 0x03,
+  0x01, 0x10, 0xa0, 0x03, 0x07, 0x00, 0x00, 0x0b, 0x04, 0x40, 0xbd, 0xe8,
+  0x0e, 0xf0, 0xb0, 0xe1, 0x0f, 0x50, 0x2d, 0xe9, 0x01, 0x03, 0xa0, 0xe3,
+  0x00, 0xe0, 0x8f, 0xe2, 0x04, 0xf0, 0x10, 0xe5, 0x0f, 0x50, 0xbd, 0xe8,
+  0x04, 0xf0, 0x5e, 0xe2, 0x10, 0x40, 0x2d, 0xe9, 0x00, 0x30, 0x0f, 0xe1,
+  0x80, 0x30, 0xc3, 0xe3, 0x03, 0xf0, 0x29, 0xe1, 0x01, 0x43, 0xa0, 0xe3,
+  0x00, 0x00, 0x50, 0xe3, 0x00, 0x00, 0xa0, 0xe3, 0x01, 0x20, 0xa0, 0xe3,
+  0x00, 0x00, 0x00, 0x0a, 0x01, 0x03, 0xc4, 0xe5, 0x04, 0x22, 0xc4, 0xe5,
+  0xb8, 0x30, 0x54, 0xe1, 0x01, 0x30, 0x13, 0xe0, 0x01, 0x30, 0x23, 0x10,
+  0xb8, 0x30, 0x44, 0x11, 0x04, 0x02, 0xc4, 0xe5, 0xf7, 0xff, 0xff, 0x0a,
+  0x00, 0x00, 0x0f, 0xe1, 0x80, 0x00, 0x80, 0xe3, 0x00, 0xf0, 0x29, 0xe1,
   0x10, 0x80, 0xbd, 0xe8
 };
diff --git a/src/gba/hle-bios.s b/src/gba/hle-bios.s
index 9351ef583..d159ac578 100644
--- a/src/gba/hle-bios.s
+++ b/src/gba/hle-bios.s
@@ -18,13 +18,15 @@ swiBase:
 cmp    sp, #0
 moveq  sp, #0x04000000
 subeq  sp, #0x20
-stmfd  sp!, {lr}
-ldrb   r0, [lr, #-2]
-cmp    r0, #4
+stmfd  sp!, {r2, lr}
+ldrb   r2, [lr, #-2]
+cmp    r2, #4
 bleq   IntrWait
-cmp    r0, #5
+cmp    r2, #5
+moveq  r0, #1
+moveq  r1, #1
 bleq   IntrWait
-ldmfd  sp!, {lr}
+ldmfd  sp!, {r2, lr}
 movs   pc, lr
 
 irqBase:
@@ -37,32 +39,31 @@ subs   pc, lr, #4
 
 IntrWait:
 stmfd  sp!, {r4, lr}
-add    sp, #-4
-strh   r1, [sp, #0]
+# Save inputs
+mrs    r3, cpsr
+bic    r3, #0x80
+msr    cpsr, r3
+# Pull current interrupts enabled and add the ones we need
 mov    r4, #0x04000000
-add    r4, #0x200
-ldrh   r0, [r4, #0]
-strh   r0, [sp, #2]
-ldrh   r1, [sp, #0]
-orr    r0, r1
-strh   r0, [r4, #0x0]
-mov    r4, #0x04000000
-IntrWaitLoop:
-mov    r0, #0x1F
-msr    cpsr, r0
+# See if we want to return immediately
+cmp    r0, #0
 mov    r0, #0
+mov    r2, #1
+beq    .L1
+# Halt
+.L0:
 strb   r0, [r4, #0x301]
-mov    r0, #0xD3
+.L1:
+# Check which interrupts were acknowledged
+strb   r2, [r4, #0x204]
+ldrh   r3, [r4, #-8]
+ands   r3, r1
+eorne  r3, r1
+strneh r3, [r4, #-8]
+strb   r0, [r4, #0x204]
+beq .L0
+#Restore state
+mrs    r0, cpsr
+orr    r0, #0x80
 msr    cpsr, r0
-ldrh   r0, [r4, #-8]
-ldrh   r1, [sp, #0]
-ands   r1, r0
-eorne  r1, r0
-strneh r1, [r4, #-8]
-beq    IntrWaitLoop
-mov    r4, #0x04000000
-add    r4, #0x200
-ldrh   r0, [sp, #2]
-strh   r0, [r4, #0]
-add    sp, #4
 ldmfd  sp!, {r4, pc}