adding a lot better load elimination

lang/src/son.rs

@@ -22,7 +22,6 @@ use {
         fmt::{self, Debug, Display, Write},
         format_args as fa, mem,
         ops::{self, Deref},
-        usize,
     },
     hashbrown::hash_map,
     hbbytecode::DisasmError,
@@ -678,10 +677,7 @@ impl Nodes {
         loop {
             region = match self[region].kind {
                 Kind::BinOp { op: TokenKind::Add | TokenKind::Sub } => self[region].inputs[1],
-                Kind::Phi => {
+                Kind::Phi if self[region].inputs[2] == 0 => self[region].inputs[1],
-                    debug_assert_eq!(self[region].inputs[2], 0);
-                    self[region].inputs[1]
-                }
                 _ => break (self[region].aclass, region),
             };
         }
@@ -1064,12 +1060,27 @@ impl Nodes {
                 }
             }
             K::Load => {
-                if self[target].inputs.len() == 3
+                let &[_, region, store] = self[target].inputs.as_slice() else { unreachable!() };
-                    && self[self[target].inputs[2]].kind == Kind::Stre
+
-                    && self[self[target].inputs[2]].inputs[2] == self[target].inputs[1]
+                if self[store].kind == Kind::Stre
-                    && self[self[target].inputs[2]].ty == self[target].ty
+                    && self[store].inputs[2] == region
+                    && self[store].ty == self[target].ty
                 {
-                    return Some(self[self[target].inputs[2]].inputs[1]);
+                    return Some(self[store].inputs[1]);
+                }
+
+                let (index, reg) = self.aclass_index(region);
+                if index != 0 && self[reg].kind == Kind::Stck {
+                    let mut cursor = store;
+                    while cursor != MEM
+                        && self[cursor].kind == Kind::Stre
+                        && self[cursor].inputs[1] != VOID
+                    {
+                        if self[cursor].inputs[2] == region && self[cursor].ty == self[target].ty {
+                            return Some(self[cursor].inputs[1]);
+                        }
+                        cursor = self[cursor].inputs[3];
+                    }
                 }
             }
             K::Loop => {
@@ -1537,7 +1548,7 @@ pub struct Node {
 
 impl Node {
     fn is_dangling(&self) -> bool {
-        self.outputs.len() + self.lock_rc as usize == 0
+        self.outputs.len() + self.lock_rc as usize == 0 && self.kind != Kind::Arg
     }
 
     fn key(&self) -> (Kind, &[Nid], ty::Id) {

lang/tests/son_tests_aliasing_overoptimization.txt

@@ -1,16 +1,6 @@
 main:
-    ADDI64 r254, r254, -24d
+    LI64 r1, 0d
-    ADDI64 r2, r254, 16d
-    ST r2, r254, 0a, 8h
-    LI64 r5, 0d
-    LI64 r4, 2d
-    ST r5, r254, 8a, 8h
-    ST r4, r254, 16a, 8h
-    LD r10, r254, 0a, 8h
-    ST r5, r10, 0a, 8h
-    LD r1, r254, 16a, 8h
-    ADDI64 r254, r254, 24d
     JALA r0, r31, 0a
-code size: 150
+code size: 29
 ret: 0
 status: Ok(())

lang/tests/son_tests_different_types.txt

@@ -22,23 +22,9 @@ main:
     JEQ r10, r9, :2
     LI64 r1, 64d
     JMP :1
- 2: LD r4, r254, 0a, 1h
+ 2: LI64 r1, 512d
-    LD r7, r254, 1a, 1h
-    ANDI r8, r4, 255d
-    LD r6, r254, 4a, 4h
-    LD r1, r254, 2a, 1h
-    ANDI r2, r7, 255d
-    ADD32 r7, r6, r8
-    LD r5, r254, 3a, 1h
-    ANDI r6, r1, 255d
-    ADD32 r11, r7, r2
-    ANDI r9, r5, 255d
-    ADD32 r2, r11, r6
-    ADD32 r4, r2, r9
-    ADD32 r6, r4, r12
-    ANDI r1, r6, 4294967295d
  1: ADDI64 r254, r254, 12d
     JALA r0, r31, 0a
-code size: 387
+code size: 257
 ret: 512
 status: Ok(())

lang/tests/son_tests_intcast_store.txt

@@ -13,17 +13,8 @@ main:
     ADDI64 r254, r254, 24d
     JALA r0, r31, 0a
 set:
-    ADDI64 r254, r254, -25d
+    CP r1, r4
-    LI8 r7, 5b
-    ST r7, r254, 0a, 1h
-    ST r4, r254, 1a, 4h
-    LI64 r11, 8d
-    ST r11, r254, 5a, 4h
-    ST r2, r254, 9a, 8h
-    ST r3, r254, 17a, 8h
-    LD r1, r254, 1a, 4h
-    ADDI64 r254, r254, 25d
     JALA r0, r31, 0a
-code size: 277
+code size: 167
 ret: 1024
 status: Ok(())

lang/tests/son_tests_struct_operators.txt

@@ -1,99 +1,37 @@
 main:
-    ADDI64 r254, r254, -152d
+    ADDI64 r254, r254, -64d
-    LI8 r1, 0b
+    LI64 r3, 1d
-    LI8 r3, 1b
+    ADDI64 r2, r254, 48d
-    ST r1, r254, 116a, 1h
-    ST r3, r254, 112a, 1h
-    ST r1, r254, 117a, 1h
-    ST r3, r254, 113a, 1h
-    ST r1, r254, 118a, 1h
-    ST r3, r254, 114a, 1h
-    ST r1, r254, 119a, 1h
-    ST r3, r254, 115a, 1h
-    LD r1, r254, 116a, 1h
-    LD r4, r254, 112a, 1h
-    ADD8 r5, r4, r1
-    LD r8, r254, 113a, 1h
-    LD r9, r254, 117a, 1h
-    ST r5, r254, 116a, 1h
-    ADD8 r12, r9, r8
-    LD r4, r254, 114a, 1h
-    LD r5, r254, 118a, 1h
-    ST r12, r254, 117a, 1h
-    ADD8 r7, r5, r4
-    ST r7, r254, 118a, 1h
-    ST r3, r254, 119a, 1h
-    LD r12, r254, 117a, 1h
-    LD r1, r254, 118a, 1h
-    ADD8 r4, r1, r12
-    LD r5, r254, 116a, 1h
-    ADD8 r7, r5, r4
-    LI8 r9, 4b
-    ADD8 r1, r7, r3
-    ANDI r1, r1, 255d
-    ANDI r9, r9, 255d
-    JEQ r1, r9, :0
-    LI64 r1, 1008d
-    JMP :1
- 0: LI64 r6, 1d
-    ADDI64 r5, r254, 96d
-    ST r6, r254, 96a, 8h
-    LI64 r9, 2d
-    ST r9, r254, 104a, 8h
-    LI64 r2, 3d
-    ADDI64 r1, r254, 80d
-    ST r2, r254, 32a, 8h
-    LI64 r6, 4d
-    LI64 r2, 0d
-    BMC r5, r1, 16h
-    ST r6, r254, 40a, 8h
-    ST r2, r254, 16a, 8h
-    LD r11, r254, 80a, 8h
-    LD r1, r254, 32a, 8h
-    ST r2, r254, 24a, 8h
-    ADD64 r4, r1, r11
-    LD r7, r254, 88a, 8h
-    LD r2, r254, 16a, 8h
-    ST r4, r254, 120a, 8h
-    ADD64 r12, r7, r6
-    SUB64 r3, r2, r1
-    ADDI64 r8, r254, 48d
-    ST r12, r254, 128a, 8h
-    SUB64 r2, r1, r11
     ST r3, r254, 48a, 8h
-    LI64 r9, -4d
+    LI64 r6, 2d
-    ST r2, r254, 136a, 8h
+    ST r6, r254, 56a, 8h
-    SUB64 r7, r6, r7
+    LI64 r6, -3d
-    ST r9, r254, 56a, 8h
+    ADDI64 r5, r254, 0d
-    ADDI64 r8, r8, 16d
+    ADDI64 r11, r254, 32d
-    ST r7, r254, 144a, 8h
+    ST r6, r254, 0a, 8h
-    BMC r5, r8, 16h
+    LI64 r6, -4d
-    LD r6, r254, 120a, 8h
+    BMC r2, r11, 16h
-    LD r8, r254, 48a, 8h
-    ADD64 r9, r8, r6
-    LD r11, r254, 56a, 8h
-    LD r1, r254, 128a, 8h
-    ST r9, r254, 48a, 8h
-    ADD64 r4, r1, r11
-    LD r8, r254, 64a, 8h
-    LD r9, r254, 136a, 8h
-    ST r4, r254, 56a, 8h
-    ADD64 r12, r9, r8
-    LD r2, r254, 72a, 8h
-    ST r12, r254, 64a, 8h
-    ADD64 r12, r2, r7
-    ST r12, r254, 72a, 8h
-    LD r7, r254, 48a, 8h
-    LD r9, r254, 64a, 8h
-    ADD64 r11, r9, r7
-    LD r1, r254, 56a, 8h
-    ST r11, r254, 0a, 8h
-    ADD64 r6, r1, r12
     ST r6, r254, 8a, 8h
-    LD r7, r254, 0a, 8h
+    ADDI64 r3, r5, 16d
-    ADD64 r1, r7, r6
+    LD r9, r254, 40a, 8h
- 1: ADDI64 r254, r254, 152d
+    LI64 r8, 4d
+    LD r10, r254, 32a, 8h
+    LI64 r11, 3d
+    BMC r2, r3, 16h
+    SUB64 r4, r8, r9
+    LD r12, r254, 24a, 8h
+    ADD64 r7, r10, r11
+    LD r1, r254, 0a, 8h
+    SUB64 r8, r11, r10
+    LD r2, r254, 16a, 8h
+    ADD64 r6, r12, r4
+    ADD64 r3, r1, r7
+    ADD64 r10, r2, r8
+    ADD64 r12, r9, r6
+    ADD64 r9, r10, r3
+    ADD64 r1, r9, r12
+    ADDI64 r254, r254, 64d
     JALA r0, r31, 0a
-code size: 980
+code size: 308
 ret: 10
 status: Ok(())

lang/tests/son_tests_struct_patterns.txt

@@ -33,22 +33,19 @@ fib_iter:
     JMP :2
  1: JALA r0, r31, 0a
 main:
-    ADDI64 r254, r254, -18d
+    ADDI64 r254, r254, -24d
-    ST r31, r254, 2a, 16h
+    ST r31, r254, 0a, 24h
-    LI8 r1, 10b
+    LI64 r32, 10d
-    ST r1, r254, 0a, 1h
+    CP r2, r32
-    ST r1, r254, 1a, 1h
-    LD r5, r254, 0a, 1h
-    ANDI r2, r5, 255d
     JAL r31, r0, :fib
-    CP r32, r1
+    CP r2, r32
-    LI64 r2, 10d
+    CP r33, r1
     JAL r31, r0, :fib_iter
-    CP r4, r32
+    CP r9, r33
-    SUB64 r1, r4, r1
+    SUB64 r1, r9, r1
-    LD r31, r254, 2a, 16h
+    LD r31, r254, 0a, 24h
-    ADDI64 r254, r254, 18d
+    ADDI64 r254, r254, 24d
     JALA r0, r31, 0a
-code size: 353
+code size: 306
 ret: 0
 status: Ok(())