diff --git a/src/arm/cpu/arm/psr_transfer.zig b/src/arm/cpu/arm/psr_transfer.zig
index c97af6b..3dbf962 100644
--- a/src/arm/cpu/arm/psr_transfer.zig
+++ b/src/arm/cpu/arm/psr_transfer.zig
@@ -61,7 +61,7 @@ pub fn control(comptime InstrFn: type, comptime I: bool, comptime op: u6) InstrF
 
                     cpu.r[rd] = @bitCast(if (U) left -| right else left +| right);
 
-                    if (cpu.r[rd] == if (U) 0x8000_0000 else 0x7FFF_FFFF) cpu.cpsr.q.set();
+                    if (cpu.r[rd] == 0x8000_0000 or cpu.r[rd] == 0x7FFF_FFFF) cpu.cpsr.q.set();
                 },
                 0b10_0101, 0b11_0101 => { // QDADD / QDSUB
                     const U = op >> 4 & 1 == 1;
@@ -76,7 +76,7 @@ pub fn control(comptime InstrFn: type, comptime I: bool, comptime op: u6) InstrF
                     const left: i32 = @bitCast(cpu.r[rm]);
 
                     cpu.r[rd] = @bitCast(if (U) left -| product else left +| product);
-                    if (cpu.r[rd] == if (U) 0x800_0000 else 0x7FFF_FFFF) cpu.cpsr.q.set();
+                    if (cpu.r[rd] == 0x8000_0000 or cpu.r[rd] == 0x7FFF_FFFF) cpu.cpsr.q.set();
                 },
                 0b01_0111 => cpu.panic("TODO: handle BKPT", .{}),
                 0b00_1000, 0b00_1010, 0b00_1100, 0b00_1110 => { // SMLA<x><y>
@@ -111,16 +111,18 @@ pub fn control(comptime InstrFn: type, comptime I: bool, comptime op: u6) InstrF
                     const right: i32 = @as(i16, @bitCast(@as(u16, @truncate(cpu.r[rs] >> 16 * Y))));
 
                     // TODO: de-clutter this lmao
-                    cpu.r[rd_lo] = @bitCast(@as(i32, @bitCast(cpu.r[rd_lo])) + (left * right));
+                    const rdlo_val: i32 = @bitCast(cpu.r[rd_lo]);
+                    const product = left * right;
 
+                    // WHY DOESN'T THIS WORK????????
+
+                    cpu.r[rd_lo] = @bitCast(rdlo_val + product);
                     cpu.r[rd_hi] = blk: {
-                        const _left: i32 = @bitCast(cpu.r[rd_hi]);
-                        const _mid: i32 = if (left * right < 0) @bitCast(@as(u32, 0xFFFF_FFFF)) else 0;
+                        // FIXME: I think this has to do with correcting sign values?
+                        const offset_thing: i32 = @bitCast(if (product < 0) 0xFFFF_FFFF else @as(u32, 0));
 
-                        // FIXME: chances are the read from rd_lo here is incorrect
-                        const _right: i32 = @addWithOverflow(@as(i32, @bitCast(cpu.r[rd_lo])), left * right)[1];
-
-                        break :blk @bitCast(_left + _mid + _right);
+                        const rdhi_val: i32 = @bitCast(cpu.r[rd_hi]);
+                        break :blk @bitCast(rdhi_val + offset_thing + @addWithOverflow(rdlo_val, product)[1]);
                     };
                 },
                 0b01_1000, 0b01_1100 => { // SMLAW<y>
@@ -172,6 +174,16 @@ pub fn control(comptime InstrFn: type, comptime I: bool, comptime op: u6) InstrF
             }
         }
 
+        // TODO: make generic on any integer?
+        inline fn carryFrom(left: i32, right: i32) u1 {
+            const _left: u32 = @bitCast(left);
+            const _right: u32 = @bitCast(right);
+
+            const sum = @as(u64, _left) + @as(u64, _right);
+
+            return @intCast(sum >> 32 & 1);
+        }
+
         inline fn msr(comptime R: bool, comptime imm: bool, cpu: *Arm32, opcode: u32) void {
             const field_mask: u4 = @truncate(opcode >> 16 & 0xF);
             const rm_idx = opcode & 0xF;