saving

lang/src/son.rs

@@ -943,11 +943,51 @@ impl Nodes {
                     return Some(self.new_node(self[lhs].ty, Kind::Stre, vc));
                 }
             }
+            K::Stck => {
+                if let &[mut a, mut b] = self[target].outputs.as_slice() {
+                    if self[a].kind == Kind::Load {
+                        mem::swap(&mut a, &mut b);
+                    }
+
+                    if matches!(self[a].kind, Kind::Call { .. })
+                        && self[a].inputs.last() == Some(&target)
+                        && self[b].kind == Kind::Load
+                        && let &[store] = self[b].outputs.as_slice()
+                        && self[store].kind == Kind::Stre
+                    {
+                        let len = self[a].inputs.len();
+                        let stre = self[store].inputs[3];
+                        if stre != MEM {
+                            self[a].inputs.push(stre);
+                            self[a].inputs.swap(len - 2, len - 1);
+                            self[stre].outputs.push(a);
+                        }
+                        return Some(self[store].inputs[2]);
+                    }
+                }
+            }
             K::Stre => {
                 let &[_, value, region, store, ..] = self[target].inputs.as_slice() else {
                     unreachable!()
                 };
 
+                if self[value].kind == Kind::Load && self[value].inputs[1] == region {
+                    return Some(store);
+                }
+
+                let mut cursor = target;
+                while self[cursor].kind == Kind::Stre
+                    && self[cursor].inputs[1] != VOID
+                    && let &[next_store] = self[cursor].outputs.as_slice()
+                {
+                    if self[next_store].inputs[2] == region
+                        && self[next_store].ty == self[target].ty
+                    {
+                        return Some(store);
+                    }
+                    cursor = next_store;
+                }
+
                 'eliminate: {
                     if self[target].outputs.is_empty() {
                         break 'eliminate;

lang/tests/son_tests_generic_types.txt

@@ -1,20 +1,17 @@
 deinit:
-    ADDI64 r254, r254, -48d
+    ADDI64 r254, r254, -16d
-    ST r31, r254, 24a, 24h
+    ST r31, r254, 0a, 16h
-    LD r5, r2, 16a, 8h
     CP r32, r2
+    LD r5, r2, 16a, 8h
     LI64 r4, 8d
     MUL64 r3, r5, r4
     CP r5, r32
     LD r2, r5, 0a, 8h
     JAL r31, r0, :free
-    ADDI64 r33, r254, 0d
+    CP r1, r32
-    CP r1, r33
     JAL r31, r0, :new
-    CP r2, r32
+    LD r31, r254, 0a, 16h
-    BMC r33, r2, 24h
+    ADDI64 r254, r254, 16d
-    LD r31, r254, 24a, 24h
-    ADDI64 r254, r254, 48d
     JALA r0, r31, 0a
 free:
     CP r10, r2
@@ -26,23 +23,21 @@ free:
     ECA 
     JALA r0, r31, 0a
 main:
-    ADDI64 r254, r254, -80d
+    ADDI64 r254, r254, -48d
-    ST r31, r254, 48a, 32h
+    ST r31, r254, 24a, 24h
-    ADDI64 r32, r254, 24d
+    ADDI64 r32, r254, 0d
     CP r1, r32
     JAL r31, r0, :new
-    ADDI64 r33, r254, 0d
-    BMC r32, r33, 24h
     LI64 r3, 69d
-    CP r2, r33
+    CP r2, r32
     JAL r31, r0, :push
-    LD r12, r254, 0a, 8h
+    LD r9, r254, 0a, 8h
-    LD r34, r12, 0a, 8h
+    LD r33, r9, 0a, 8h
-    CP r2, r33
+    CP r2, r32
     JAL r31, r0, :deinit
-    CP r1, r34
+    CP r1, r33
-    LD r31, r254, 48a, 32h
+    LD r31, r254, 24a, 24h
-    ADDI64 r254, r254, 80d
+    ADDI64 r254, r254, 48d
     JALA r0, r31, 0a
 malloc:
     CP r9, r2
@@ -126,6 +121,6 @@ push:
  4: LD r31, r254, 0a, 72h
     ADDI64 r254, r254, 72d
     JALA r0, r31, 0a
-code size: 980
+code size: 945
 ret: 69
 status: Ok(())

lang/tests/son_tests_overwrite_aliasing_overoptimization.txt

@@ -1,25 +1,22 @@
 main:
-    ADDI64 r254, r254, -72d
+    ADDI64 r254, r254, -56d
-    ST r31, r254, 56a, 16h
+    ST r31, r254, 40a, 16h
-    ADDI64 r32, r254, 0d
+    LI64 r4, 4d
+    ADDI64 r3, r254, 24d
+    ADDI64 r6, r254, 0d
+    ST r4, r254, 24a, 8h
+    LI64 r32, 1d
+    ST r32, r254, 32a, 8h
+    ST r32, r254, 16a, 8h
+    BMC r3, r6, 16h
     JAL r31, r0, :opaque
     ST r1, r254, 0a, 16h
-    LI64 r6, 4d
+    LD r4, r254, 8a, 8h
-    ADDI64 r5, r254, 40d
+    ADD64 r7, r4, r32
-    ADDI64 r8, r254, 16d
+    LD r5, r254, 0a, 8h
-    ST r6, r254, 40a, 8h
+    SUB64 r1, r5, r7
-    LI64 r7, 1d
+    LD r31, r254, 40a, 16h
-    ST r7, r254, 48a, 8h
+    ADDI64 r254, r254, 56d
-    ST r7, r254, 32a, 8h
-    BMC r5, r8, 16h
-    BMC r32, r8, 16h
-    LD r7, r254, 24a, 8h
-    LD r9, r254, 32a, 8h
-    ADD64 r11, r9, r7
-    LD r9, r254, 16a, 8h
-    SUB64 r1, r9, r11
-    LD r31, r254, 56a, 16h
-    ADDI64 r254, r254, 72d
     JALA r0, r31, 0a
 opaque:
     ADDI64 r254, r254, -16d
@@ -31,6 +28,6 @@ opaque:
     LD r1, r2, 0a, 16h
     ADDI64 r254, r254, 16d
     JALA r0, r31, 0a
-code size: 339
+code size: 310
 ret: 0
 status: Ok(())

lang/tests/son_tests_returning_global_struct.txt

@@ -1,29 +1,27 @@
 main:
-    ADDI64 r254, r254, -24d
+    ADDI64 r254, r254, -12d
-    ST r31, r254, 8a, 16h
+    ST r31, r254, 4a, 8h
-    ADDI64 r32, r254, 4d
+    ADDI64 r2, r254, 0d
     JAL r31, r0, :random_color
-    ST r1, r254, 4a, 4h
+    ST r1, r254, 0a, 4h
-    ADDI64 r5, r254, 0d
+    LD r5, r254, 0a, 1h
-    BMC r32, r5, 4h
+    LD r8, r254, 1a, 1h
-    LD r8, r254, 0a, 1h
+    LD r12, r254, 2a, 1h
-    LD r11, r254, 1a, 1h
+    ANDI r9, r5, 255d
-    LD r3, r254, 2a, 1h
+    ANDI r1, r8, 255d
-    ANDI r12, r8, 255d
+    LD r6, r254, 3a, 1h
-    ANDI r4, r11, 255d
+    ANDI r5, r12, 255d
-    LD r9, r254, 3a, 1h
+    ADD64 r4, r1, r9
-    ANDI r8, r3, 255d
+    ANDI r10, r6, 255d
-    ADD64 r7, r4, r12
+    ADD64 r9, r4, r5
-    ANDI r1, r9, 255d
+    ADD64 r1, r9, r10
-    ADD64 r12, r7, r8
+    LD r31, r254, 4a, 8h
-    ADD64 r1, r12, r1
+    ADDI64 r254, r254, 12d
-    LD r31, r254, 8a, 16h
-    ADDI64 r254, r254, 24d
     JALA r0, r31, 0a
 random_color:
     LRA r1, r0, :white
     LD r1, r1, 0a, 4h
     JALA r0, r31, 0a
-code size: 257
+code size: 241
 ret: 1020
 status: Ok(())

lang/tests/son_tests_small_struct_assignment.txt

@@ -1,14 +1,12 @@
 main:
     ADDI64 r254, r254, -4d
-    LRA r1, r0, :black
+    LRA r2, r0, :white
     ADDI64 r3, r254, 0d
-    LRA r5, r0, :white
+    BMC r2, r3, 4h
-    BMC r1, r3, 4h
+    LD r6, r254, 3a, 1h
-    BMC r5, r3, 4h
+    ANDI r1, r6, 255d
-    LD r9, r254, 3a, 1h
-    ANDI r1, r9, 255d
     ADDI64 r254, r254, 4d
     JALA r0, r31, 0a
-code size: 108
+code size: 92
 ret: 255
 status: Ok(())

lang/tests/son_tests_structs.txt

@@ -1,24 +1,23 @@
 main:
-    ADDI64 r254, r254, -72d
+    ADDI64 r254, r254, -56d
-    ST r31, r254, 48a, 24h
+    ST r31, r254, 32a, 24h
     LI64 r3, 4d
-    ADDI64 r2, r254, 32d
+    ADDI64 r2, r254, 16d
-    ST r3, r254, 32a, 8h
+    ST r3, r254, 16a, 8h
     LI64 r32, 3d
-    ST r32, r254, 40a, 8h
+    ST r32, r254, 24a, 8h
-    ADDI64 r33, r254, 16d
+    ADDI64 r33, r254, 0d
     LD r3, r2, 0a, 16h
     JAL r31, r0, :odher_pass
-    ST r1, r254, 16a, 16h
+    ST r1, r254, 0a, 16h
-    ADDI64 r2, r254, 0d
+    LD r2, r254, 8a, 8h
-    BMC r33, r2, 16h
+    JNE r2, r32, :0
-    LD r4, r254, 8a, 8h
+    CP r2, r33
-    JNE r4, r32, :0
     JAL r31, r0, :pass
     JMP :1
  0: LI64 r1, 0d
- 1: LD r31, r254, 48a, 24h
+ 1: LD r31, r254, 32a, 24h
-    ADDI64 r254, r254, 72d
+    ADDI64 r254, r254, 56d
     JALA r0, r31, 0a
 odher_pass:
     ADDI64 r254, r254, -16d
@@ -30,6 +29,6 @@ odher_pass:
 pass:
     LD r1, r2, 0a, 8h
     JALA r0, r31, 0a
-code size: 318
+code size: 305
 ret: 4
 status: Ok(())

lang/tests/son_tests_wide_ret.txt

@@ -1,19 +1,17 @@
 main:
-    ADDI64 r254, r254, -48d
+    ADDI64 r254, r254, -24d
-    ST r31, r254, 32a, 16h
+    ST r31, r254, 16a, 8h
-    ADDI64 r32, r254, 16d
+    ADDI64 r3, r254, 0d
     LI64 r4, 0d
     CP r3, r4
     JAL r31, r0, :maina
-    ST r1, r254, 16a, 16h
+    ST r1, r254, 0a, 16h
-    ADDI64 r7, r254, 0d
+    LD r8, r254, 12a, 1h
-    BMC r32, r7, 16h
+    LD r9, r254, 3a, 1h
-    LD r11, r254, 12a, 1h
+    SUB8 r11, r9, r8
-    LD r12, r254, 3a, 1h
+    ANDI r1, r11, 255d
-    SUB8 r2, r12, r11
+    LD r31, r254, 16a, 8h
-    ANDI r1, r2, 255d
+    ADDI64 r254, r254, 24d
-    LD r31, r254, 32a, 16h
-    ADDI64 r254, r254, 48d
     JALA r0, r31, 0a
 maina:
     ADDI64 r254, r254, -36d
@@ -50,6 +48,6 @@ small_struct:
     LD r1, r3, 0a, 4h
     ADDI64 r254, r254, 4d
     JALA r0, r31, 0a
-code size: 514
+code size: 498
 ret: 2
 status: Ok(())