diff --git a/mandel-core.s b/mandel-core.s
index 2e1c8b2..17ecfc4 100644
--- a/mandel-core.s
+++ b/mandel-core.s
@@ -558,8 +558,6 @@ input_max:
 
 ; input: arg1, arg2 as fixed4.12
 ; output: dest as fixed8.24
-; patch point jsr at 16 bytes in
-imul16_patch_offset = 16
 .macro imul16 dest, arg1, arg2
     copy16 FR0, arg1  ; 12 cyc
     copy16 FR1, arg2  ; 12 cyc
@@ -569,13 +567,11 @@ imul16_patch_offset = 16
 
 ; input: arg as fixed4.12
 ; output: dest as fixed8.24
-; patch point jsr at 8 bytes in
-sqr16_patch_offset = 8
-.macro sqr16 dest, arg
-    copy16 FR0, arg   ; 12 cyc
-    jsr sqr16_func    ; ? cyc
-    copy32 dest, FR2  ; 24 cyc
-.endmacro
+;.macro sqr16 dest, arg
+;    copy16 FR0, arg   ; 12 cyc
+;    jsr sqr16_func    ; ? cyc
+;    copy32 dest, FR2  ; 24 cyc
+;.endmacro
 
 ; input: arg as u8
 ; output: dest as u16
@@ -912,15 +908,15 @@ inner_loop:
     rts
 .endproc
 
-.proc sqr16_func
-    sqr16_impl FR2, FR0, 0
-    rts
-.endproc
+;.proc sqr16_func
+;    sqr16_impl FR2, FR0, 0
+;    rts
+;.endproc
 
-.proc sqr16xe_func
-    sqr16_impl FR2, FR0, 1
-    rts
-.endproc
+;.proc sqr16xe_func
+;    sqr16_impl FR2, FR0, 1
+;    rts
+;.endproc
 
 ; 11-27 cycles
 .macro round16 arg
@@ -1160,16 +1156,11 @@ keep_going:
     shift_round_16 zy, 3
 
     ; zx_2 = zx * zx
-fixup_sqr16_1:
-    sqr16 zx_2, zx + 2
-
     ; zy_2 = zy * zy
-fixup_sqr16_2:
-    sqr16 zy_2, zy + 2
-
     ; zx_zy = zx * zy
-fixup_imul16_1:
-    imul16 zx_zy, zx + 2, zy + 2
+fixup_mandelbrot_hotspot:
+    jmp mandelbrot_hotspot
+after_mandelbrot_hotspot:
 
     ; dist = zx_2 + zy_2
     add32 dist, zx_2, zy_2
@@ -1265,6 +1256,27 @@ next:
 
 .endproc
 
+.macro mandelbrot_hotspot_impl xe 
+    ; zx_2 = zx * zx
+    sqr16_impl zx_2, zx + 2, xe
+
+    ; zy_2 = zy * zy
+    sqr16_impl zy_2, zy + 2, xe
+
+    ; zx_zy = zx * zy
+    imul16_impl zx_zy, zx + 2, zy + 2, xe
+
+    jmp mandelbrot::after_mandelbrot_hotspot
+.endmacro
+
+.proc mandelbrot_hotspot
+    mandelbrot_hotspot_impl 0
+.endproc
+
+.proc mandelbrot_hotspot_xe
+    mandelbrot_hotspot_impl 1
+.endproc
+
 .macro scale_zoom dest
     ; clobbers X, flags
     .local cont
@@ -2192,22 +2204,21 @@ init:
     sta imul16_func
     lda #.lobyte(imul16xe_func)
     sta imul16_func + 1
-    sta mandelbrot::fixup_imul16_1 + imul16_patch_offset + 1
     lda #.hibyte(imul16xe_func)
     sta imul16_func + 2
-    sta mandelbrot::fixup_imul16_1 + imul16_patch_offset + 2
 
     ; ditto for sqr16_func -> sqr16xe_func
-    lda #$4c ; 'jmp' opcode
-    sta sqr16_func
-    lda #.lobyte(sqr16xe_func)
-    sta sqr16_func + 1
-    sta mandelbrot::fixup_sqr16_1 + sqr16_patch_offset + 1
-    sta mandelbrot::fixup_sqr16_2 + sqr16_patch_offset + 1
-    lda #.hibyte(sqr16xe_func)
-    sta sqr16_func + 2
-    sta mandelbrot::fixup_sqr16_1 + sqr16_patch_offset + 2
-    sta mandelbrot::fixup_sqr16_2 + sqr16_patch_offset + 2
+    ;lda #$4c ; 'jmp' opcode
+    ;sta sqr16_func
+    ;lda #.lobyte(sqr16xe_func)
+    ;sta sqr16_func + 1
+    ;lda #.hibyte(sqr16xe_func)
+    ;sta sqr16_func + 2
+
+    lda #.lobyte(mandelbrot_hotspot_xe)
+    sta mandelbrot::fixup_mandelbrot_hotspot + 1
+    lda #.hibyte(mandelbrot_hotspot_xe)
+    sta mandelbrot::fixup_mandelbrot_hotspot + 2
 
 
     ; create the lookup table