cpu: more instruction pipelining

This is needed to make WNS room for fetching instructions from SRAM.

tools/as.py

@@ -92,7 +92,7 @@ def generate_ops(ops, labels, relocs):
             if isinstance(p, str):  # label ref
                 if len(params) == 1:  # branch
                     yield 14 # pc
-                    yield labels[p] - pc - 2
+                    yield labels[p] - pc - 4
                 else:  # set, allow relocs here
                     relocs.append((pc, p))
                     yield 0xff

tools/cc.py

@@ -473,10 +473,10 @@ class ShlOp(BinOp):
         return [f'set {sc1}, 1',
                 f'or {self.dest}, {self.left}, {self.left}',
                 f'sub {sc0}, {self.right}, {sc1}',
-                f'beq [pc, 6]',
+                f'beq [pc, 4]',
                 f'add {self.dest}, {self.dest}, {self.dest}',
                 f'sub {sc0}, {sc0}, {sc1}',
-                f'bneq [pc, -6]']
+                f'bneq [pc, -8]']
 
 
 class LtOp(BinOp):
@@ -485,7 +485,7 @@ class LtOp(BinOp):
         sc0 = scratches[0]
         return [f'set {self.dest}, 0',
                 f'sub {sc0}, {self.left}, {self.right}',
-                f'bneq [pc, 2]',
+                f'bneq [pc, 0]',
                 f'set {self.dest}, 1']
 
 class GtOp(LtOp):
@@ -531,7 +531,7 @@ class BoolNot(UnOp):
     def synth(self, scratches):
         return [f'set {self.dest}, 0',
                 f'cmp {self.dest}, {self.operand}',
-                f'bneq [pc, 2]',
+                f'bneq [pc, 0]',
                 f'set {self.dest}, 1']
 
 class NeqOp(BinOp):
@@ -555,7 +555,7 @@ class FnCall(AsmOp):
         sc0 = scratches[0]
         fn = self.dest_fn
 
-        return out + [f'set {sc0}, 2',
+        return out + [f'set {sc0}, 0',
                 f'add lr, pc, {sc0}',
                 f'or pc, {fn}, {fn}']
 
@@ -1375,10 +1375,11 @@ preamble = [f'_start:',
         f'set sp, 0',
         f'seth sp, {0x11}',  # 256 bytes of stack ought to be enough
         f'set r2, main',
-        f'set r3, 2',
+        f'set r3, 0',
         f'add lr, pc, r3',
         f'or pc, r2, r2',
-        f'or pc, pc, pc  // loop forever',
+        f'cmp r0, r0',
+        f'beq [pc, -4]  // loop forever',
         ]
 
 def filter_dupes(ops):