diff --git a/mach/proto/mcg/mcg.h b/mach/proto/mcg/mcg.h
index dd810e40d..06d444d61 100644
--- a/mach/proto/mcg/mcg.h
+++ b/mach/proto/mcg/mcg.h
@@ -112,6 +112,7 @@ extern void pass_infer_types(void);
 extern void pass_insert_moves(void);
 extern void pass_instruction_selector(void);
 extern void pass_live_vreg_analysis(void);
+extern void pass_lower_pushes(void);
 extern void pass_add_prologue_epilogue(void);
 extern void pass_register_allocator(void);
 extern void pass_remove_dead_blocks(void);
diff --git a/mach/proto/mcg/pass_lowerpushes.c b/mach/proto/mcg/pass_lowerpushes.c
new file mode 100644
index 000000000..c990f862e
--- /dev/null
+++ b/mach/proto/mcg/pass_lowerpushes.c
@@ -0,0 +1,140 @@
+#include "mcg.h"
+
+#if defined MCGG_OPTION_LOWER_PUSHES_TO_LOADS_AND_STORES
+
+/* On architectures which can't push and pop cheaply, a push typically
+ *
+ *   sub sp, sp, -4
+ *   sw r5, 0(sp)
+ *
+ * This is hugely wasteful when you want to push or multiple things
+ * at once, which happens a lot because that's how the procedure calling
+ * convention works. This code detects these runs and turns them into a
+ * single stack adjustment and then offsetted accesses via the stack
+ * pointer. In order to be efficient, the table needs to know how to
+ * handle this efficiently:
+ *
+ *   STACKADJUST(CONST(-4))
+ *   STORE.I(ADD.I(GETSP(), CONST(0)))
+ *
+ * ...otherwise the code will be *even worse*.
+ *
+ * We have to be careful, though, because after we do the adjustment,
+ * the physical stack pointer won't match em's idea of the stack pointer
+ * until the last 'push' happens. So we need to check that this is never
+ * used.
+ *
+ * With this option set, PUSH will never be seen by the instruction
+ * selector.
+ */
+
+static struct basicblock* current_bb;
+static int cursor;
+
+static bool accesses_stack_pointer_cb(struct ir* ir, void* user)
+{
+	switch (ir->opcode)
+	{
+		case IR_SETSP:
+		case IR_GETSP:
+		case IR_CALL:
+		case IR_STACKADJUST:
+			return true;
+
+		default:
+			return false;
+	}
+}
+
+static bool accesses_stack_pointer(struct ir* ir)
+{
+	return !!ir_walk(ir, accesses_stack_pointer_cb, NULL);
+}
+
+static void consider_push(struct ir* ir)
+{
+	int runstart;
+	int delta;
+	int i;
+
+	if (ir->opcode != IR_PUSH)
+	{
+		cursor++;
+		return;
+	}
+
+	/* This is the first push of a run; we'll want to insert the STACKADJUST
+	 * before this one. */
+
+	tracef('P', "found push in %s at IR index %d\n", current_bb->name, cursor);
+	runstart = cursor;
+
+	/* Now start walking forward until we find an IR which isn't a safe push.
+	 * The current IR is always safe. */
+
+	for (;;)
+	{
+		struct ir* ir;
+
+		cursor++;
+		if (cursor == current_bb->irs.count)
+			break;
+
+		ir = current_bb->irs.item[cursor];
+		if (ir->opcode != IR_PUSH)
+			break;
+		if (accesses_stack_pointer(ir))
+			break;
+
+		delta += ir->size;
+	}
+
+	tracef('P', "found end of run at IR index %d\n", cursor);
+
+	/* Now work backwards, converting each push into a stack write. */
+
+	delta = 0;
+	i = cursor - 1;
+	while (i >= runstart)
+	{
+		struct ir* ir = current_bb->irs.item[i];
+		struct ir* value_ir = ir->left;
+		assert(ir->opcode == IR_PUSH);
+
+		ir->opcode = IR_STORE;
+		ir->left = new_ir2(
+			IR_ADD, EM_pointersize,
+			new_ir0(IR_GETSP, EM_pointersize),
+			new_wordir(delta)
+		);
+		ir->left->root = ir->left->left->root = ir->left->right->root = ir->root;
+		ir->right = value_ir;
+
+		delta += ir->size;
+		i--;
+	}
+
+	/* And finally, before the place where the first push was, adjust the
+	 * stack. */
+
+	ir = new_ir1(IR_STACKADJUST, EM_pointersize, new_wordir(-delta));
+	ir->left->root = ir->root = ir;
+	array_insert(&current_bb->irs, ir, runstart);
+	cursor++;
+}
+
+void pass_lower_pushes(void)
+{
+    int i;
+
+	for (i=0; i<current_proc->blocks.count; i++)
+    {
+        current_bb = current_proc->blocks.item[i];
+
+		cursor = 0;
+		while (cursor < current_bb->irs.count)
+			consider_push(current_bb->irs.item[cursor]);
+    }
+}
+
+#endif
diff --git a/mach/proto/mcg/procedure.c b/mach/proto/mcg/procedure.c
index abd94d19b..d8b5fd18c 100644
--- a/mach/proto/mcg/procedure.c
+++ b/mach/proto/mcg/procedure.c
@@ -199,6 +199,9 @@ void procedure_compile(struct procedure* proc)
     print_blocks('3');
     pass_wire_up_return_values();
     pass_convert_stack_ops();
+    #if defined MCGG_OPTION_LOWER_PUSHES_TO_LOADS_AND_STORES
+        pass_lower_pushes();
+    #endif
     print_blocks('4');
     pass_convert_locals_to_ssa();
     print_blocks('5');
diff --git a/util/mcgg/gram.y b/util/mcgg/gram.y
index 891030a10..18b88aec5 100644
--- a/util/mcgg/gram.y
+++ b/util/mcgg/gram.y
@@ -36,6 +36,7 @@ extern int yylex(void);
 %term FRAGMENT
 %term NAMED
 %term NOTEQUALS
+%term OPTIONS
 %term PATTERNS
 %term PREFERS
 %term PRESERVED
@@ -66,11 +67,27 @@ extern int yylex(void);
 %%
 
 spec
-    : REGISTERS registers 
+    : optionaloptions
+      REGISTERS registers
       DECLARATIONS declarations
       PATTERNS patterns
 	;
 
+optionaloptions
+	: /* nothing */
+	| OPTIONS options
+	;
+
+options
+	: /* nothing */
+	| options option ';'
+	| option ';'
+	;
+
+option
+	: ID                              { option($1); }
+	;
+
 registers
     : /* nothing */
     | registers register ';'
diff --git a/util/mcgg/iburg.c b/util/mcgg/iburg.c
index 9979627c0..cd5be61af 100644
--- a/util/mcgg/iburg.c
+++ b/util/mcgg/iburg.c
@@ -836,6 +836,11 @@ static void emitheader(void)
 	printh("#define MCG_DEFS_H\n\n");
 }
 
+void option(const char* o)
+{
+	printh("#define MCGG_OPTION_%s\n", o);
+}
+
 /* computekids - compute paths to kids in tree t */
 static char* computekids(Tree node, const char* v, char* bp, int* ip)
 {
diff --git a/util/mcgg/iburg.h b/util/mcgg/iburg.h
index 739641d4d..199d1aa83 100644
--- a/util/mcgg/iburg.h
+++ b/util/mcgg/iburg.h
@@ -72,6 +72,7 @@ struct regattr
 	int number;            /* identifying number */
 };
 
+extern void option(const char* name);
 extern struct reg* makereg(const char* name);
 extern void setregnames(struct reg* reg, struct stringlist* names);
 extern void addregattr(struct reg* reg, const char* regattr);
diff --git a/util/mcgg/ir.dat b/util/mcgg/ir.dat
index 4b7c4e920..d32e642ac 100644
--- a/util/mcgg/ir.dat
+++ b/util/mcgg/ir.dat
@@ -1,7 +1,7 @@
-# Types:
+# The second column indicates the type inference rules:
 #
 #   I, F, L, D: int, float, long, double
-#   i, F: int, float; promoted to long, double based on size
+#   i, f: int, float; promoted to long, double based on size
 #   .: ignore this parameter
 #   ?: pull/push types from other ? parameters
 
diff --git a/util/mcgg/scan.l b/util/mcgg/scan.l
index 9d06d5b38..5e61f31f9 100644
--- a/util/mcgg/scan.l
+++ b/util/mcgg/scan.l
@@ -40,6 +40,7 @@ static int braces = 0;
 "DECLARATIONS"              return DECLARATIONS;
 "PATTERNS"                  return PATTERNS;
 "REGISTERS"                 return REGISTERS;
+"OPTIONS"                   return OPTIONS;
 "aliases"                   return ALIASES;
 "corrupted"                 return CORRUPTED;
 "cost"                      return COST;