Prepare to use a specialized calling convention: return globasp

3 files changed, 178 insertions, 135 deletions

diff --git a/fib_eqI_b.ll b/fib_eqI_b.ll
index 72e08e1..a6a684b 100644
--- a/fib_eqI_b.ll
+++ b/fib_eqI_b.ll
@@ -1,4 +1,4 @@
-define {i64*,i64*} @e_fib_eqI_b_nfib(i64* %hp.0, i64** %globasp, i64* %n) align 8
+define {i64*,i64**,i64*} @e_fib_eqI_b_nfib(i64* %hp.0, i64** %globasp.0, i64* %n) align 8
     prefix {i8*, i64} {i8* inttoptr (i64 0 to i8*), i64 1} {
   %astack = alloca i64*, i64 10
   %asp.000 = getelementptr i64*, i64** %astack
@@ -8,15 +8,17 @@ define {i64*,i64*} @e_fib_eqI_b_nfib(i64* %hp.0, i64** %globasp, i64* %n) align
   %asp.001 = getelementptr i64*, i64** %asp.000, i64 1
   store i64* %n, i64** %asp.001
 
-  %t.0 = bitcast {i64*,i64*}(i64**,i64*)* @_cycle_in_spine to i64*
+  %t.0 = bitcast {i64*,i64**,i64*}(i64*,i64**)* @_cycle_in_spine to i64*
   %asp.002 = call i64** @push_node(i64* %t.0, i64 1, i64** %asp.001)
 
   %t.1 = call i64* @peek_a(i64** %asp.002)
   %asp.003 = call i64** @pop_a1(i64** %asp.002)
-  %globasp.0 = call i64** @_push_local_astack(i64** %asp.003, i64** %asp.000, i64** %globasp)
-  %t.2 = call {i64*,i64} @ea1(i64* %hp.0, i64** %globasp.0, i64* %t.1)
-  %hp.1 = extractvalue {i64*,i64} %t.2, 0
-  %t.2.0 = extractvalue {i64*,i64} %t.2, 1
+  %globasp.1 = call i64** @_push_local_astack(i64** %asp.003, i64** %asp.000, i64** %globasp.0)
+  %t.2 = call {i64*,i64**,i64} @ea1(i64* %hp.0, i64** %globasp.1, i64* %t.1)
+  %hp.1 = extractvalue {i64*,i64**,i64} %t.2, 0
+  %globasp.2 = extractvalue {i64*,i64**,i64} %t.2, 1
+  %t.2.0 = extractvalue {i64*,i64**,i64} %t.2, 2
+  %globasp.3 = call i64** @_pop_global_astack(i64** %asp.003, i64** %asp.000, i64** %globasp.2)
   %bsp.001 = call i64* @pushI(i64 %t.2.0, i64* %bsp.000)
 
   %asp.004 = call i64** @fillI_b(i64 0, i64 -0, i64* %bsp.001, i64** %asp.003)
@@ -25,12 +27,13 @@ define {i64*,i64*} @e_fib_eqI_b_nfib(i64* %hp.0, i64** %globasp, i64* %n) align
   %t.3 = call i64* @peek_a(i64** %asp.004)
   %asp.005 = call i64** @pop_a1(i64** %asp.004)
 
-  %ret.0 = insertvalue {i64*,i64*} undef, i64* %hp.1, 0
-  %ret.1 = insertvalue {i64*,i64*} %ret.0, i64* %t.3, 1
-  ret {i64*,i64*} %ret.1
+  %ret.0 = insertvalue {i64*,i64**,i64*} undef, i64* %hp.1, 0
+  %ret.1 = insertvalue {i64*,i64**,i64*} %ret.0, i64** %globasp.3, 1
+  %ret.2 = insertvalue {i64*,i64**,i64*} %ret.1, i64* %t.3, 2
+  ret {i64*,i64**,i64*} %ret.2
 }
 
-define private {i64*,i64} @ea1(i64* %hp.0, i64** %globasp, i64* %n) {
+define private {i64*,i64**,i64} @ea1(i64* %hp.0, i64** %globasp.0, i64* %n) {
   %astack = alloca i64*, i64 10
   %asp.000 = getelementptr i64*, i64** %astack, i64 0
   %bstack = alloca i64, i64 10
@@ -39,22 +42,26 @@ define private {i64*,i64} @ea1(i64* %hp.0, i64** %globasp, i64* %n) {
   %asp.001 = getelementptr i64*, i64** %asp.000, i64 1
   store i64* %n, i64** %asp.001
 
-  %hp.1 = call i64* @jsr_eval(i64 -0, i64** %asp.001, i64** %asp.000, i64** %globasp, i64* %hp.0)
+  %j.0 = call {i64*,i64**} @jsr_eval(i64 -0, i64** %asp.001, i64** %asp.000, i64** %globasp.0, i64* %hp.0)
+  %hp.1 = extractvalue {i64*,i64**} %j.0, 0
+  %globasp.1 = extractvalue {i64*,i64**} %j.0, 1
   %bsp.001 = call i64* @pushI_a(i64 -0, i64* %bsp.000, i64** %asp.001)
   %asp.002 = call i64** @pop_a1(i64** %asp.001)
 
   %t.0 = call i64 @peek_b(i64* %bsp.001)
   %bsp.002 = call i64* @pop_b1(i64* %bsp.001)
-  %t.1 = tail call {i64*,i64} @s1(i64* %hp.1, i64** %globasp, i64 %t.0)
-  %hp.2 = extractvalue {i64*,i64} %t.1, 0
-  %t.2 = extractvalue {i64*,i64} %t.1, 1
-
-  %ret.0 = insertvalue {i64*,i64} undef, i64* %hp.2, 0
-  %ret.1 = insertvalue {i64*,i64} %ret.0, i64 %t.2, 1
-  ret {i64*,i64} %ret.1
+  %t.1 = tail call {i64*,i64**,i64} @s1(i64* %hp.1, i64** %globasp.1, i64 %t.0)
+  %hp.2 = extractvalue {i64*,i64**,i64} %t.1, 0
+  %globasp.2 = extractvalue {i64*,i64**,i64} %t.1, 1
+  %t.2 = extractvalue {i64*,i64**,i64} %t.1, 2
+
+  %ret.0 = insertvalue {i64*,i64**,i64} undef, i64* %hp.2, 0
+  %ret.1 = insertvalue {i64*,i64**,i64} %ret.0, i64** %globasp.2, 1
+  %ret.2 = insertvalue {i64*,i64**,i64} %ret.1, i64 %t.2, 2
+  ret {i64*,i64**,i64} %ret.2
 }
 
-define private {i64*,i64} @s1(i64* %hp.0, i64** %globasp, i64 %arg) #0 {
+define private {i64*,i64**,i64} @s1(i64* %hp.0, i64** %globasp.0, i64 %arg) #0 {
   %astack = alloca i64*, i64 10000
   %asp.000 = getelementptr i64*, i64** %astack
   %bstack = alloca i64, i64 10000
@@ -82,9 +89,10 @@ case.1:
   %r.1 = call i64 @peek_b(i64* %bsp.101)
   %bsp.102 = call i64* @pop_b1(i64* %bsp.101)
 
-  %ret.0 = insertvalue {i64*,i64} undef, i64* %hp.0, 0
-  %ret.1 = insertvalue {i64*,i64} %ret.0, i64 %r.1, 1
-  ret {i64*,i64} %ret.1
+  %ret.0 = insertvalue {i64*,i64**,i64} undef, i64* %hp.0, 0
+  %ret.1 = insertvalue {i64*,i64**,i64} %ret.0, i64** %globasp.0, 1
+  %ret.2 = insertvalue {i64*,i64**,i64} %ret.1, i64 %r.1, 2
+  ret {i64*,i64**,i64} %ret.2
 
 case.2:
   %bsp.200 = call i64* @pop_b1(i64* %bsp.005)
@@ -93,9 +101,10 @@ case.2:
   %r.2 = call i64 @peek_b(i64* %bsp.201)
   %bsp.202 = call i64* @pop_b1(i64* %bsp.201)
 
-  %ret.2 = insertvalue {i64*,i64} undef, i64* %hp.0, 0
-  %ret.3 = insertvalue {i64*,i64} %ret.2, i64 %r.2, 1
-  ret {i64*,i64} %ret.3
+  %ret.3 = insertvalue {i64*,i64**,i64} undef, i64* %hp.0, 0
+  %ret.4 = insertvalue {i64*,i64**,i64} %ret.3, i64** %globasp.0, 1
+  %ret.5 = insertvalue {i64*,i64**,i64} %ret.4, i64 %r.2, 2
+  ret {i64*,i64**,i64} %ret.5
 
 case.3:
   %bsp.300 = call i64* @pushI(i64 1, i64* %bsp.005)
@@ -104,9 +113,10 @@ case.3:
 
   %arg.0 = call i64 @peek_b(i64* %bsp.302)
   %bsp.302.0 = call i64* @pop_b1(i64* %bsp.302)
-  %t.4 = call {i64*,i64} @s1(i64* %hp.0, i64** %globasp, i64 %arg.0)
-  %hp.1 = extractvalue {i64*,i64} %t.4, 0
-  %t.5 = extractvalue {i64*,i64} %t.4, 1
+  %t.4 = call {i64*,i64**,i64} @s1(i64* %hp.0, i64** %globasp.0, i64 %arg.0)
+  %hp.1 = extractvalue {i64*,i64**,i64} %t.4, 0
+  %globasp.1 = extractvalue {i64*,i64**,i64} %t.4, 1
+  %t.5 = extractvalue {i64*,i64**,i64} %t.4, 2
   %bsp.303 = call i64* @pushI(i64 %t.5, i64* %bsp.302.0)
 
   %bsp.304 = call i64* @pushI(i64 2, i64* %bsp.303)
@@ -115,9 +125,10 @@ case.3:
 
   %arg.1 = call i64 @peek_b(i64* %bsp.306)
   %bsp.306.0 = call i64* @pop_b1(i64* %bsp.306)
-  %t.6 = call {i64*,i64} @s1(i64* %hp.1, i64** %globasp, i64 %arg.1)
-  %hp.2 = extractvalue {i64*,i64} %t.6, 0
-  %t.7 = extractvalue {i64*,i64} %t.6, 1
+  %t.6 = call {i64*,i64**,i64} @s1(i64* %hp.1, i64** %globasp.1, i64 %arg.1)
+  %hp.2 = extractvalue {i64*,i64**,i64} %t.6, 0
+  %globasp.2 = extractvalue {i64*,i64**,i64} %t.6, 1
+  %t.7 = extractvalue {i64*,i64**,i64} %t.6, 2
   %bsp.307 = call i64* @pushI(i64 %t.7, i64* %bsp.306.0)
 
   %bsp.308 = call i64* @update_b(i64 1, i64 2, i64* %bsp.307)
@@ -127,14 +138,15 @@ case.3:
   %r.3 = call i64 @peek_b(i64* %bsp.310)
   %bsp.311 = call i64* @pop_b1(i64* %bsp.310)
 
-  %ret.4 = insertvalue {i64*,i64} undef, i64* %hp.2, 0
-  %ret.5 = insertvalue {i64*,i64} %ret.4, i64 %r.3, 1
-  ret {i64*,i64} %ret.5
+  %ret.6 = insertvalue {i64*,i64**,i64} undef, i64* %hp.2, 0
+  %ret.7 = insertvalue {i64*,i64**,i64} %ret.6, i64** %globasp.2, 1
+  %ret.8 = insertvalue {i64*,i64**,i64} %ret.7, i64 %r.3, 2
+  ret {i64*,i64**,i64} %ret.8
 }
 
 define i64 @main() {
   %hp.0 = call i64* @_alloc_heap(i64 10000)
-  %astack = call i64** @_alloc_a_stack(i64 10000)
+  %asp.0 = call i64** @_alloc_a_stack(i64 10000)
 
   %n.0 = getelementptr i64, i64* %hp.0
   %INT.0 = ptrtoint {i64, i64, i64*, i64, i8, i8, i8}* @INT to i64
@@ -145,15 +157,16 @@ define i64 @main() {
   %hp.1 = getelementptr i64, i64* %hp.0, i64 2
 
   %n.1 = getelementptr i64, i64* %hp.1
-  store i64 ptrtoint ({i64*,i64*}(i64*,i64**,i64*)* @e_fib_eqI_b_nfib to i64), i64* %n.1
+  store i64 ptrtoint ({i64*,i64**,i64*}(i64*,i64**,i64*)* @e_fib_eqI_b_nfib to i64), i64* %n.1
   %t.0 = ptrtoint i64* %n.0 to i64
   %n.1.0 = getelementptr i64, i64* %n.1, i64 1
   store i64 %t.0, i64* %n.1.0
   %hp.2 = getelementptr i64, i64* %hp.0, i64 3
 
-  %t.1 = call {i64*,i64*} @e_fib_eqI_b_nfib(i64* %hp.2, i64** %astack, i64* %n.1)
-  %hp.3 = extractvalue {i64*,i64*} %t.1, 0
-  %n.2 = extractvalue {i64*,i64*} %t.1, 1
+  %t.1 = call {i64*,i64**,i64*} @e_fib_eqI_b_nfib(i64* %hp.2, i64** %asp.0, i64* %n.1)
+  %hp.3 = extractvalue {i64*,i64**,i64*} %t.1, 0
+  %asp.1 = extractvalue {i64*,i64**,i64*} %t.1, 1
+  %n.2 = extractvalue {i64*,i64**,i64*} %t.1, 2
   %n.3 = getelementptr i64, i64* %n.2, i64 1
 
   %r = load i64, i64* %n.3
diff --git a/nfib_nsa.ll b/nfib_nsa.ll
index b49ceae..62da012 100644
--- a/nfib_nsa.ll
+++ b/nfib_nsa.ll
@@ -1,7 +1,7 @@
-define {i64*,i64*} @n5(
-    i64* inreg align 8 nocapture nofree nonnull %hp.0,
-    i64** inreg align 8 noalias nocapture nofree nonnull %globasp.0,
-    i64* inreg align 8 nocapture nofree dereferenceable(3) %n)
+define {i64*,i64**,i64*} @n5(
+    i64* align 8 nocapture nofree nonnull %hp.0,
+    i64** nocapture nofree nonnull %globasp.0,
+    i64* align 8 nocapture nofree dereferenceable(3) %n)
     nofree willreturn nosync nounwind argmemonly align 8
     prefix {i8*, i64} {i8* inttoptr (i64 0 to i8*), i64 0} {
   %astack = alloca i64*, i64 5
@@ -12,12 +12,13 @@ define {i64*,i64*} @n5(
   %asp.001 = getelementptr i64*, i64** %asp.000, i64 1
   store i64* %n, i64** %asp.001
 
-  %t.0 = bitcast {i64*,i64*}(i64**,i64*)* @_cycle_in_spine to i64*
+  %t.0 = bitcast {i64*,i64**,i64*}(i64*,i64**)* @_cycle_in_spine to i64*
   %asp.002 = call i64** @push_node(i64* %t.0, i64 0, i64** %asp.001)
 
-  %t.1 = call {i64*,i64} @ea5(i64* %hp.0, i64** %globasp.0)
-  %hp.1 = extractvalue {i64*,i64} %t.1, 0
-  %t.2 = extractvalue {i64*,i64} %t.1, 1
+  %t.1 = call {i64*,i64**,i64} @ea5(i64* %hp.0, i64** %globasp.0)
+  %hp.1 = extractvalue {i64*,i64**,i64} %t.1, 0
+  %globasp.1 = extractvalue {i64*,i64**,i64} %t.1, 1
+  %t.2 = extractvalue {i64*,i64**,i64} %t.1, 2
   %bsp.001 = call i64* @pushI(i64 %t.2, i64* %bsp.000)
 
   %asp.003 = call i64** @fillI_b(i64 0, i64 -0, i64* %bsp.001, i64** %asp.002)
@@ -26,22 +27,23 @@ define {i64*,i64*} @n5(
   %t.3 = call i64* @peek_a(i64** %asp.003)
   %asp.004 = call i64** @pop_a1(i64** %asp.003)
 
-  %ret.0 = insertvalue {i64*, i64*} undef, i64* %hp.1, 0
-  %ret.1 = insertvalue {i64*, i64*} %ret.0, i64* %t.3, 1
-  ret {i64*, i64*} %ret.1
+  %ret.0 = insertvalue {i64*,i64**,i64*} undef, i64* %hp.1, 0
+  %ret.1 = insertvalue {i64*,i64**,i64*} %ret.0, i64** %globasp.1, 1
+  %ret.2 = insertvalue {i64*,i64**,i64*} %ret.1, i64* %t.3, 2
+  ret {i64*,i64**,i64*} %ret.2
 }
 
-define private {i64*,i64} @ea5(
-    i64* inreg align 8 nocapture nofree nonnull %hp.0,
-    i64** inreg align 8 noalias nocapture nofree nonnull %globasp.0)
+define private {i64*,i64**,i64} @ea5(
+    i64* align 8 nocapture nofree nonnull %hp.0,
+    i64** align 8 noalias nocapture nofree nonnull %globasp.0)
     nofree willreturn nosync nounwind argmemonly {
-  %ret.0 = tail call {i64*,i64} @s5(i64* %hp.0, i64** %globasp.0)
-  ret {i64*,i64} %ret.0
+  %ret.0 = tail call {i64*,i64**,i64} @s5(i64* %hp.0, i64** %globasp.0)
+  ret {i64*,i64**,i64} %ret.0
 }
 
-define private {i64*,i64} @s5(
-    i64* inreg align 8 nocapture nofree nonnull %hp.0,
-    i64** inreg align 8 noalias nocapture nofree nonnull %globasp.0)
+define private {i64*,i64**,i64} @s5(
+    i64* align 8 nocapture nofree nonnull %hp.0,
+    i64** align 8 noalias nocapture nofree nonnull %globasp.0)
     nofree willreturn nosync nounwind argmemonly {
   %astack = alloca i64*, i64 5
   %asp.000 = getelementptr i64*, i64** %astack
@@ -54,14 +56,14 @@ define private {i64*,i64} @s5(
 
   %t.1 = call i64* @peek_a(i64** %asp.001)
   %asp.002 = call i64** @pop_a1(i64** %asp.001)
-  %t.2 = tail call {i64*,i64} @s4(i64* %hp.1, i64** %globasp.0, i64* %t.1)
-  ret {i64*,i64} %t.2
+  %t.2 = tail call {i64*,i64**,i64} @s4(i64* %hp.1, i64** %globasp.0, i64* %t.1)
+  ret {i64*,i64**,i64} %t.2
 }
 
-define inreg {i64*,i64} @s4(
-    i64* inreg align 8 nocapture nofree nonnull %hp.0,
-    i64** inreg align 8 noalias nocapture nofree nonnull %globasp.0,
-    i64* inreg align 8 nocapture nofree dereferenceable(3) %n)
+define {i64*,i64**,i64} @s4(
+    i64* align 8 nocapture nofree nonnull %hp.0,
+    i64** align 8 noalias nocapture nofree nonnull %globasp.0,
+    i64* align 8 nocapture nofree dereferenceable(3) %n)
     nofree willreturn nosync nounwind argmemonly {
   %astack = alloca i64*, i64 5
   %asp.000 = getelementptr i64*, i64** %astack
@@ -72,7 +74,9 @@ define inreg {i64*,i64} @s4(
   store i64* %n, i64** %asp.001
 
   %asp.002 = call i64** @push_a(i64 -0, i64** %asp.001)
-  %hp.1 = call i64* @jsr_eval(i64 -0, i64** %asp.002, i64** %asp.000, i64** %globasp.0, i64* %hp.0)
+  %j.0 = call {i64*,i64**} @jsr_eval(i64 -0, i64** %asp.002, i64** %asp.000, i64** %globasp.0, i64* %hp.0)
+  %hp.1 = extractvalue {i64*,i64**} %j.0, 0
+  %globasp.1 = extractvalue {i64*,i64**} %j.0, 1
   %bsp.001 = call i64* @pushI_a(i64 -1, i64* %bsp.000, i64** %asp.002)
   %asp.003 = call i64** @pop_a1(i64** %asp.002)
   %bsp.002 = call i64* @pushI(i64 2, i64* %bsp.001)
@@ -94,9 +98,10 @@ l.0:
   %t.2 = call i64 @peek_b(i64* %bsp.100)
   %bsp.101 = call i64* @pop_b1(i64* %bsp.100)
 
-  %ret.0 = insertvalue {i64*,i64} undef, i64* %hp.1, 0
-  %ret.1 = insertvalue {i64*,i64} %ret.0, i64 %t.2, 1
-  ret {i64*,i64} %ret.1
+  %ret.0 = insertvalue {i64*,i64**,i64} undef, i64* %hp.1, 0
+  %ret.1 = insertvalue {i64*,i64**,i64} %ret.0, i64** %globasp.1, 1
+  %ret.2 = insertvalue {i64*,i64**,i64} %ret.1, i64 %t.2, 2
+  ret {i64*,i64**,i64} %ret.2
 
 else.1:
   %t.3 = call {i64**,i64*} @buildI(i64 1, i64** %asp.003, i64* %hp.1)
@@ -105,16 +110,17 @@ else.1:
 
   %asp.201 = call i64** @push_a(i64 -1, i64** %asp.200)
 
-  %t.4 = bitcast {i64*,i64*}(i64*,i64**,i64*)* @n2 to i64*
-  %t.5 = call {i64**,i64*} @build(i64* %t.4, i64 2, i64** %asp.201, i64** %asp.000, i64** %globasp.0, i64* %hp.2)
+  %t.4 = bitcast {i64*,i64**,i64*}(i64*,i64**,i64*)* @n2 to i64*
+  %t.5 = call {i64**,i64*} @build(i64* %t.4, i64 2, i64** %asp.201, i64** %asp.000, i64** %globasp.1, i64* %hp.2)
   %asp.202 = extractvalue {i64**,i64*} %t.5, 0
   %hp.3 = extractvalue {i64**,i64*} %t.5, 1
 
   %t.6 = call i64* @peek_a(i64** %asp.202)
   %asp.203 = call i64** @pop_a1(i64** %asp.202)
-  %t.7 = call {i64*,i64} @s4(i64* %hp.3, i64** %globasp.0, i64* %t.6)
-  %hp.4 = extractvalue {i64*,i64} %t.7, 0
-  %t.8 = extractvalue {i64*,i64} %t.7, 1
+  %t.7 = call {i64*,i64**,i64} @s4(i64* %hp.3, i64** %globasp.1, i64* %t.6)
+  %hp.4 = extractvalue {i64*,i64**,i64} %t.7, 0
+  %globasp.2 = extractvalue {i64*,i64**,i64} %t.7, 1
+  %t.8 = extractvalue {i64*,i64**,i64} %t.7, 2
   %bsp.200 = call i64* @pushI(i64 %t.8, i64* %bsp.008)
 
   %t.9 = call {i64**,i64*} @buildI(i64 2, i64** %asp.203, i64* %hp.4)
@@ -123,16 +129,17 @@ else.1:
 
   %asp.205 = call i64** @push_a(i64 -1, i64** %asp.204)
 
-  %t.10 = bitcast {i64*,i64*}(i64*,i64**,i64*)* @n2 to i64*
-  %t.11 = call {i64**,i64*} @build(i64* %t.10, i64 2, i64** %asp.205, i64** %asp.000, i64** %globasp.0, i64* %hp.5)
+  %t.10 = bitcast {i64*,i64**,i64*}(i64*,i64**,i64*)* @n2 to i64*
+  %t.11 = call {i64**,i64*} @build(i64* %t.10, i64 2, i64** %asp.205, i64** %asp.000, i64** %globasp.2, i64* %hp.5)
   %asp.206 = extractvalue {i64**,i64*} %t.11, 0
   %hp.6 = extractvalue {i64**,i64*} %t.11, 1
 
   %t.12 = call i64* @peek_a(i64** %asp.206)
   %asp.207 = call i64** @pop_a1(i64** %asp.206)
-  %t.13 = call {i64*,i64} @s4(i64* %hp.6, i64** %globasp.0, i64* %t.12)
-  %hp.7 = extractvalue {i64*,i64} %t.13, 0
-  %t.14 = extractvalue {i64*,i64} %t.13, 1
+  %t.13 = call {i64*,i64**,i64} @s4(i64* %hp.6, i64** %globasp.2, i64* %t.12)
+  %hp.7 = extractvalue {i64*,i64**,i64} %t.13, 0
+  %globasp.3 = extractvalue {i64*,i64**,i64} %t.13, 1
+  %t.14 = extractvalue {i64*,i64**,i64} %t.13, 2
   %bsp.201 = call i64* @pushI(i64 %t.14, i64* %bsp.200)
 
   %bsp.202 = call i64* @addI(i64* %bsp.201)
@@ -146,15 +153,16 @@ else.1:
   %t.15 = call i64 @peek_b(i64* %bsp.208)
   %bsp.209 = call i64* @pop_b1(i64* %bsp.208)
 
-  %ret.2 = insertvalue {i64*,i64} undef, i64* %hp.7, 0
-  %ret.3 = insertvalue {i64*,i64} %ret.2, i64 %t.15, 1
-  ret {i64*,i64} %ret.3
+  %ret.3 = insertvalue {i64*,i64**,i64} undef, i64* %hp.7, 0
+  %ret.4 = insertvalue {i64*,i64**,i64} %ret.3, i64** %globasp.3, 1
+  %ret.5 = insertvalue {i64*,i64**,i64} %ret.4, i64 %t.15, 2
+  ret {i64*,i64**,i64} %ret.5
 }
 
-define {i64*,i64*} @n2(
-    i64* inreg align 8 nocapture nofree nonnull %hp.0,
-    i64** inreg align 8 noalias nocapture nofree nonnull %globasp.0,
-    i64* inreg align 8 nocapture nofree dereferenceable(3) %n)
+define {i64*,i64**,i64*} @n2(
+    i64* align 8 nocapture nofree nonnull %hp.0,
+    i64** align 8 noalias nocapture nofree nonnull %globasp.0,
+    i64* align 8 nocapture nofree dereferenceable(3) %n)
     nofree willreturn nosync nounwind argmemonly align 8
     prefix {i8*, i64} {i8* inttoptr (i64 0 to i8*), i64 2} {
   %astack = alloca i64*, i64 5
@@ -165,7 +173,7 @@ define {i64*,i64*} @n2(
   %asp.001 = getelementptr i64*, i64** %asp.000, i64 1
   store i64* %n, i64** %asp.001
 
-  %t.0 = bitcast {i64*,i64*}(i64**,i64*)* @_cycle_in_spine to i64*
+  %t.0 = bitcast {i64*,i64**,i64*}(i64*,i64**)* @_cycle_in_spine to i64*
   %asp.002 = call i64** @push_node(i64* %t.0, i64 2, i64** %asp.001)
 
   %t.1 = call i64* @peek_a(i64** %asp.002)
@@ -173,9 +181,11 @@ define {i64*,i64*} @n2(
   %t.2 = call i64* @peek_a(i64** %asp.003)
   %asp.004 = call i64** @pop_a1(i64** %asp.003)
   %globasp.1 = call i64** @_push_local_astack(i64** %asp.004, i64** %asp.000, i64** %globasp.0)
-  %t.3 = call {i64*,i64} @ea2(i64* %hp.0, i64** %globasp.1, i64* %t.1, i64* %t.2)
-  %hp.1 = extractvalue {i64*,i64} %t.3, 0
-  %t.4 = extractvalue {i64*,i64} %t.3, 1
+  %t.3 = call {i64*,i64**,i64} @ea2(i64* %hp.0, i64** %globasp.1, i64* %t.1, i64* %t.2)
+  %hp.1 = extractvalue {i64*,i64**,i64} %t.3, 0
+  %globasp.2 = extractvalue {i64*,i64**,i64} %t.3, 1
+  %t.4 = extractvalue {i64*,i64**,i64} %t.3, 2
+  %globasp.3 = call i64** @_pop_global_astack(i64** %asp.004, i64** %asp.000, i64** %globasp.2)
   %bsp.001 = call i64* @pushI(i64 %t.4, i64* %bsp.000)
 
   %asp.005 = call i64** @fillI_b(i64 0, i64 -0, i64* %bsp.001, i64** %asp.004)
@@ -184,16 +194,17 @@ define {i64*,i64*} @n2(
   %t.5 = call i64* @peek_a(i64** %asp.005)
   %asp.006 = call i64** @pop_a1(i64** %asp.005)
 
-  %ret.0 = insertvalue {i64*, i64*} undef, i64* %hp.0, 0
-  %ret.1 = insertvalue {i64*, i64*} %ret.0, i64* %t.5, 1
-  ret {i64*, i64*} %ret.1
+  %ret.0 = insertvalue {i64*,i64**,i64*} undef, i64* %hp.0, 0
+  %ret.1 = insertvalue {i64*,i64**,i64*} %ret.0, i64** %globasp.3, 1
+  %ret.2 = insertvalue {i64*,i64**,i64*} %ret.1, i64* %t.5, 2
+  ret {i64*,i64**,i64*} %ret.2
 }
 
-define private {i64*,i64} @ea2(
-    i64* inreg align 8 nocapture nofree nonnull %hp.0,
-    i64** inreg align 8 noalias nocapture nofree nonnull %globasp.0,
-    i64* inreg align 8 nofree nonnull %arg1,
-    i64* inreg align 8 nofree nonnull %arg2)
+define private {i64*,i64**,i64} @ea2(
+    i64* align 8 nocapture nofree nonnull %hp.0,
+    i64** align 8 noalias nocapture nofree nonnull %globasp.0,
+    i64* align 8 nofree nonnull %arg1,
+    i64* align 8 nofree nonnull %arg2)
     nofree willreturn nosync nounwind argmemonly {
   %astack = alloca i64*, i64 5
   %asp.000 = getelementptr i64*, i64** %astack
@@ -205,8 +216,12 @@ define private {i64*,i64} @ea2(
   %asp.002 = getelementptr i64*, i64** %asp.001, i64 1
   store i64* %arg1, i64** %asp.002
 
-  %hp.1 = call i64* @jsr_eval(i64 -1, i64** %asp.002, i64** %asp.000, i64** %globasp.0, i64* %hp.0)
-  %hp.2 = call i64* @jsr_eval(i64 -0, i64** %asp.002, i64** %asp.000, i64** %globasp.0, i64* %hp.1)
+  %j.0 = call {i64*,i64**} @jsr_eval(i64 -1, i64** %asp.002, i64** %asp.000, i64** %globasp.0, i64* %hp.0)
+  %hp.1 = extractvalue {i64*,i64**} %j.0, 0
+  %globasp.1 = extractvalue {i64*,i64**} %j.0, 1
+  %j.1 = call {i64*,i64**} @jsr_eval(i64 -0, i64** %asp.002, i64** %asp.000, i64** %globasp.1, i64* %hp.1)
+  %hp.2 = extractvalue {i64*,i64**} %j.1, 0
+  %globasp.2 = extractvalue {i64*,i64**} %j.1, 1
   %bsp.001 = call i64* @pushI_a(i64 -1, i64* %bsp.000, i64** %asp.002)
   %bsp.002 = call i64* @pushI_a(i64 -0, i64* %bsp.001, i64** %asp.002)
   %asp.003 = call i64** @pop_a1(i64** %asp.002)
@@ -217,13 +232,13 @@ define private {i64*,i64} @ea2(
   %t.1 = call i64 @peek_b(i64* %bsp.003)
   %bsp.004 = call i64* @pop_b1(i64* %bsp.003)
 
-  %ret = tail call {i64*,i64} @s2(i64* %hp.2, i64** %globasp.0, i64 %t.0, i64 %t.1)
-  ret {i64*,i64} %ret
+  %ret = tail call {i64*,i64**,i64} @s2(i64* %hp.2, i64** %globasp.2, i64 %t.0, i64 %t.1)
+  ret {i64*,i64**,i64} %ret
 }
 
-define private {i64*,i64} @s2(
-    i64* inreg align 8 nocapture nofree nonnull %hp.0,
-    i64** inreg align 8 noalias nocapture nofree nonnull %globasp.0,
+define private {i64*,i64**,i64} @s2(
+    i64* align 8 nocapture nofree nonnull %hp.0,
+    i64** align 8 noalias nocapture nofree nonnull %globasp.0,
     i64 inreg %arg1,
     i64 inreg %arg2)
     nofree willreturn nosync nounwind argmemonly {
@@ -240,22 +255,24 @@ define private {i64*,i64} @s2(
   %t.0 = call i64 @peek_b(i64* %bsp.003)
   %bsp.004 = call i64* @pop_b1(i64* %bsp.003)
 
-  %ret.0 = insertvalue {i64*,i64} undef, i64* %hp.0, 0
-  %ret.1 = insertvalue {i64*,i64} %ret.0, i64 %t.0, 1
-  ret {i64*,i64} %ret.1
+  %ret.0 = insertvalue {i64*,i64**,i64} undef, i64* %hp.0, 0
+  %ret.1 = insertvalue {i64*,i64**,i64} %ret.0, i64** %globasp.0, 1
+  %ret.2 = insertvalue {i64*,i64**,i64} %ret.1, i64 %t.0, 2
+  ret {i64*,i64**,i64} %ret.2
 }
 
 define i64 @main() {
   %hp.0 = call i64* @_alloc_heap(i64 100000000)
-  %astack = call i64** @_alloc_a_stack(i64 10000)
+  %asp.0 = call i64** @_alloc_a_stack(i64 10000)
 
   %n.0 = getelementptr i64, i64* %hp.0
-  store i64 ptrtoint ({i64*,i64*}(i64*,i64**,i64*)* @n5 to i64), i64* %n.0
+  store i64 ptrtoint ({i64*,i64**,i64*}(i64*,i64**,i64*)* @n5 to i64), i64* %n.0
   %hp.1 = getelementptr i64, i64* %hp.0, i64 3
 
-  %t.0 = call {i64*,i64*} @n5(i64* %hp.1, i64** %astack, i64* %n.0)
-  %hp.2 = extractvalue {i64*,i64*} %t.0, 0
-  %n.1 = extractvalue {i64*,i64*} %t.0, 1
+  %t.0 = call {i64*,i64**,i64*} @n5(i64* %hp.1, i64** %asp.0, i64* %n.0)
+  %hp.2 = extractvalue {i64*,i64**,i64*} %t.0, 0
+  %asp.1 = extractvalue {i64*,i64**,i64*} %t.0, 1
+  %n.1 = extractvalue {i64*,i64**,i64*} %t.0, 2
 
   %n.1.0 = getelementptr i64, i64* %n.1, i64 1
   %r = load i64, i64* %n.1.0
diff --git a/rts.ll b/rts.ll
index c625a9f..c0f4a86 100644
--- a/rts.ll
+++ b/rts.ll
@@ -53,19 +53,24 @@ loop.0:
   br label %loop
 }
 
-; TODO remove only the used positions
-define private void @_cleanup_global_astack(i64** %globasp.0) noinline optnone {
-  %globasp.0.1 = getelementptr i64*, i64** %globasp.0, i64 1
-  %globasp.0.2 = getelementptr i64*, i64** %globasp.0, i64 2
-  %globasp.0.3 = getelementptr i64*, i64** %globasp.0, i64 3
-  %globasp.0.4 = getelementptr i64*, i64** %globasp.0, i64 4
-  %globasp.0.5 = getelementptr i64*, i64** %globasp.0, i64 5
-  store i64* undef, i64** %globasp.0.1
-  store i64* undef, i64** %globasp.0.2
-  store i64* undef, i64** %globasp.0.3
-  store i64* undef, i64** %globasp.0.4
-  store i64* undef, i64** %globasp.0.5
-  ret void
+define private i64** @_pop_global_astack(i64** %asp, i64** %aspstart.0, i64** %globasp.0) noinline optnone {
+entry:
+  br label %loop
+
+loop:
+  %aspstart = phi i64** [%aspstart.0, %entry], [%aspstart.1, %loop.0]
+  %globasp = phi i64** [%globasp.0, %entry], [%globasp.1, %loop.0]
+  %t.0 = icmp eq i64** %aspstart, %asp
+  br i1 %t.0, label %done, label %loop.0
+done:
+  ret i64** %globasp
+
+loop.0:
+  %n = load i64*, i64** %globasp
+  %aspstart.1 = getelementptr i64*, i64** %aspstart, i64 1
+  store i64* %n, i64** %aspstart.1
+  %globasp.1 = getelementptr i64*, i64** %globasp, i64 -1
+  br label %loop
 }
 
 define private i64* @addI(i64* %bsp.0) #0 {
@@ -155,7 +160,7 @@ define private i64** @fillI_b(i64 %b_offset, i64 %a_offset, i64* %bsp.0, i64** %
   ret i64** %asp.0
 }
 
-define private i64* @jsr_eval(i64 %n, i64** %asp.0, i64** %aspstart, i64** %globasp.0, i64* %hp.0) #0 {
+define private {i64*,i64**} @jsr_eval(i64 %n, i64** %asp.0, i64** %aspstart, i64** %globasp.0, i64* %hp.0) #0 {
   %asp.1 = getelementptr i64*, i64** %asp.0, i64 %n
 
   %n.0 = load i64*, i64** %asp.1
@@ -166,17 +171,23 @@ define private i64* @jsr_eval(i64 %n, i64** %asp.0, i64** %aspstart, i64** %glob
   br i1 %t.1, label %eval, label %done
 
 done:
-  ret i64* %hp.0
+  %r.0 = insertvalue {i64*,i64**} undef, i64* %hp.0, 0
+  %r.1 = insertvalue {i64*,i64**} %r.0, i64** %globasp.0, 1
+  ret {i64*,i64**} %r.1
 
 eval:
   %globasp.1 = call i64** @_push_local_astack(i64** %asp.0, i64** %aspstart, i64** %globasp.0)
-  %nentry.0 = inttoptr i64 %d.0 to {i64*, i64*} (i64*, i64**, i64*)*
-  %t.2 = call {i64*, i64*} %nentry.0(i64* %hp.0, i64** %globasp.1, i64* %n.0)
-  call void @_cleanup_global_astack(i64** %globasp.0)
-  %hp.1 = extractvalue {i64*, i64*} %t.2, 0
-  %n.1 = extractvalue {i64*, i64*} %t.2, 1
-  store i64* %n.1, i64** %asp.1
-  ret i64* %hp.1
+  %nentry.0 = inttoptr i64 %d.0 to {i64*,i64**,i64*} (i64*, i64**, i64*)*
+  %t.2 = call {i64*,i64**,i64*} %nentry.0(i64* %hp.0, i64** %globasp.1, i64* %n.0)
+  %hp.1 = extractvalue {i64*,i64**,i64*} %t.2, 0
+  %globasp.2 = extractvalue {i64*,i64**,i64*} %t.2, 1
+  %n.1 = extractvalue {i64*,i64**,i64*} %t.2, 2
+  %globasp.3 = call i64** @_pop_global_astack(i64** %asp.0, i64** %aspstart, i64** %globasp.2)
+  store i64* %n.1, i64** %asp.1 ; TODO this may not be needed?
+
+  %r.2 = insertvalue {i64*,i64**} undef, i64* %hp.1, 0
+  %r.3 = insertvalue {i64*,i64**} %r.2, i64** %globasp.3, 1
+  ret {i64*,i64**} %r.3
 }
 
 define private i64* @ltI(i64* %bsp.0) #0 {
@@ -307,7 +318,9 @@ define private i64 @peek_b(i64* %bsp) #0 {
 @INT = constant {i64, i64, i64*, i64, i8, i8, i8}
   {i64 0, i64 0, i64* inttoptr (i64 0 to i64*), i64 3, i8 73, i8 78, i8 84}, align 8
 
-define {i64*,i64*} @_cycle_in_spine(i64** %globasp, i64* %hp.0) align 8 prefix {i8*, i64} {i8* inttoptr (i64 0 to i8*), i64 0} {
+define {i64*,i64**,i64*} @_cycle_in_spine(i64* %hp, i64** %globasp)
+    align 8
+    prefix {i8*, i64} {i8* inttoptr (i64 0 to i8*), i64 0} {
   call i64 (i8*, ...) @printf(i8* getelementptr inbounds ([16 x i8], [16 x i8]* @printf_cycle, i64 0, i64 0))
   unreachable
 }