From 1d41674dc1fdb0798729b4bd3beac2c44d189442 Mon Sep 17 00:00:00 2001
From: Myles Busig <mdbusig@gmail.com>
Date: Thu, 4 Apr 2024 04:18:36 -0600
Subject: [PATCH] Fix udiv100000 and udiv1000000000 ASM macros changing the
 source register

---
 include/{mtl/armv4t => armv4t/mtl}/asm/math.s | 15 ++++++++-------
 src/armv4t/math.s                             |  2 +-
 src/armv4t/to.s                               |  6 +++---
 3 files changed, 12 insertions(+), 11 deletions(-)
 rename include/{mtl/armv4t => armv4t/mtl}/asm/math.s (86%)

diff --git a/include/mtl/armv4t/asm/math.s b/include/armv4t/mtl/asm/math.s
similarity index 86%
rename from include/mtl/armv4t/asm/math.s
rename to include/armv4t/mtl/asm/math.s
index a3f38b2..e3a5c30 100644
--- a/include/mtl/armv4t/asm/math.s
+++ b/include/armv4t/mtl/asm/math.s
@@ -41,12 +41,13 @@
  *   k = 17
  *
  * This division uses the output produces by clang for a division by 100000.
- * I don't understand why it works, but it does.
+ * I don't understand why it works, but it does. It also needs one extra
+ * temporary register to preserve the value of rx.
  */
-.macro udiv100000 rd, rx, rt
-	lsr	\rx, $5
+.macro udiv100000 rd, rx, rt, rtt
+	lsr	\rtt, \rx, $5
 	ldr	\rt, =0xA7C5AC5
-	umull	\rt, \rd, \rx, \rt
+	umull	\rt, \rd, \rtt, \rt
 	lsrs	\rd, $7
 .endm
 
@@ -71,10 +72,10 @@
 /*
  * Same situation as udiv100000
  */
-.macro udiv1000000000 rd, rx, rt
-	lsr	\rx, $9
+.macro udiv1000000000 rd, rx, rt, rtt
+	lsr	\rtt, \rx, $9
 	ldr	\rt, =0x44B83
-	umull	\rt, \rd, \rx, \rt
+	umull	\rt, \rd, \rtt, \rt
 	lsrs	\rd, $7
 .endm
 
diff --git a/src/armv4t/math.s b/src/armv4t/math.s
index b7d33d9..6b21bae 100644
--- a/src/armv4t/math.s
+++ b/src/armv4t/math.s
@@ -1,6 +1,6 @@
 .syntax unified
 
-.include "mtl/armv4t/asm/math.s"
+.include "mtl/asm/math.s"
 
 .section .iwram, "ax", %progbits
 .arm
diff --git a/src/armv4t/to.s b/src/armv4t/to.s
index 1546a23..98fb6e1 100644
--- a/src/armv4t/to.s
+++ b/src/armv4t/to.s
@@ -1,6 +1,6 @@
 .syntax unified
 
-.include "mtl/armv4t/asm/math.s"
+.include "mtl/asm/math.s"
 
 .section .iwram, "ax", %progbits
 .arm
@@ -133,7 +133,7 @@ push	{r4-r6}
 	add	r3, $1
 .Ldigits_10:
 	// Divide by 10^9 to get the most significant digit
-	udiv1000000000	r5, r0, r12
+	udiv1000000000	r5, r0, r12, r6
 	// Write the digit to the string
 	add	r6, r5, '0'
 	strb	r6, [r1], $1
@@ -164,7 +164,7 @@ push	{r4-r6}
 	mul	r6, r5
 	sub	r0, r6
 .Ldigits_6:
-	udiv100000	r5, r0, r12
+	udiv100000	r5, r0, r12, r6
 	add	r6, r5, '0'
 	strb	r6, [r1], $1
 	ldr	r6, =100000