Utilize lowMemoryUnused

src/compiler.ts

@@ -355,13 +355,20 @@ export class Compiler extends DiagnosticEmitter {
     super(program.diagnostics);
     this.program = program;
     var options = program.options;
-    this.memoryOffset = i64_new(
-      // leave space for `null`. also functions as a sentinel for erroneous stores at offset 0.
-      // note that Binaryen's asm.js output utilizes the first 8 bytes for reinterpretations (#1547)
-      max(options.memoryBase, 8)
-    );
     var module = Module.create();
     this.module = module;
+    if (options.memoryBase) {
+      this.memoryOffset = i64_new(options.memoryBase);
+      module.setLowMemoryUnused(false);
+    } else {
+      if (options.optimizeLevelHint >= 2) {
+        this.memoryOffset = i64_new(1024);
+        module.setLowMemoryUnused(true);
+      } else {
+        this.memoryOffset = i64_new(8);
+        module.setLowMemoryUnused(false);
+      }
+    }
     var featureFlags: FeatureFlags = 0;
     if (options.hasFeature(Feature.SIGN_EXTENSION)) featureFlags |= FeatureFlags.SignExt;
     if (options.hasFeature(Feature.MUTABLE_GLOBALS)) featureFlags |= FeatureFlags.MutableGloabls;

std/assembly/util/memory.ts

@@ -201,17 +201,17 @@ export function memset(dest: usize, c: u8, n: usize): void { // see: musl/src/st
   } else {
     // fill head and tail with minimal branching
     if (!n) return;
+    let dend = dest + n - 4;
     store<u8>(dest, c);
-    store<u8>(dest + n - 1, c);
+    store<u8>(dend, c, 3);
     if (n <= 2) return;
-
-    store<u8>(dest + 1, c);
-    store<u8>(dest + 2, c);
-    store<u8>(dest + n - 2, c);
-    store<u8>(dest + n - 3, c);
+    store<u8>(dest, c, 1);
+    store<u8>(dest, c, 2);
+    store<u8>(dend, c, 2);
+    store<u8>(dend, c, 1);
     if (n <= 6) return;
-    store<u8>(dest + 3, c);
-    store<u8>(dest + n - 4, c);
+    store<u8>(dest, c, 3);
+    store<u8>(dend, c);
     if (n <= 8) return;
 
     // advance pointer to align it at 4-byte boundary
@@ -223,22 +223,23 @@ export function memset(dest: usize, c: u8, n: usize): void { // see: musl/src/st
     let c32: u32 = <u32>-1 / 255 * c;
 
     // fill head/tail up to 28 bytes each in preparation
+    dend = dest + n - 28;
     store<u32>(dest, c32);
-    store<u32>(dest + n - 4, c32);
+    store<u32>(dend, c32, 24);
     if (n <= 8) return;
-    store<u32>(dest + 4, c32);
-    store<u32>(dest + 8, c32);
-    store<u32>(dest + n - 12, c32);
-    store<u32>(dest + n - 8, c32);
+    store<u32>(dest, c32, 4);
+    store<u32>(dest, c32, 8);
+    store<u32>(dend, c32, 16);
+    store<u32>(dend, c32, 20);
     if (n <= 24) return;
-    store<u32>(dest + 12, c32);
-    store<u32>(dest + 16, c32);
-    store<u32>(dest + 20, c32);
-    store<u32>(dest + 24, c32);
-    store<u32>(dest + n - 28, c32);
-    store<u32>(dest + n - 24, c32);
-    store<u32>(dest + n - 20, c32);
-    store<u32>(dest + n - 16, c32);
+    store<u32>(dest, c32, 12);
+    store<u32>(dest, c32, 16);
+    store<u32>(dest, c32, 20);
+    store<u32>(dest, c32, 24);
+    store<u32>(dend, c32);
+    store<u32>(dend, c32, 4);
+    store<u32>(dend, c32, 8);
+    store<u32>(dend, c32, 12);
 
     // align to a multiple of 8
     k = 24 + (dest & 4);
@@ -249,9 +250,9 @@ export function memset(dest: usize, c: u8, n: usize): void { // see: musl/src/st
     let c64: u64 = <u64>c32 | (<u64>c32 << 32);
     while (n >= 32) {
       store<u64>(dest, c64);
-      store<u64>(dest + 8, c64);
-      store<u64>(dest + 16, c64);
-      store<u64>(dest + 24, c64);
+      store<u64>(dest, c64, 8);
+      store<u64>(dest, c64, 16);
+      store<u64>(dest, c64, 24);
       n -= 32;
       dest += 32;
     }

tests/compiler/abi.optimized.wat

@@ -2,7 +2,7 @@
  (type $none_=>_none (func))
  (type $none_=>_i32 (func (result i32)))
  (memory $0 1)
- (data (i32.const 16) "\0c\00\00\00\01\00\00\00\01\00\00\00\0c\00\00\00a\00b\00i\00.\00t\00s")
+ (data (i32.const 1024) "\0c\00\00\00\01\00\00\00\01\00\00\00\0c\00\00\00a\00b\00i\00.\00t\00s")
  (global $abi/condition (mut i32) (i32.const 0))
  (global $abi/y (mut i32) (i32.const 0))
  (export "memory" (memory $0))

tests/compiler/assert-nonnull.optimized.wat

@@ -5,10 +5,10 @@
  (type $i32_i32_i32_i32_=>_none (func (param i32 i32 i32 i32)))
  (import "env" "abort" (func $~lib/builtins/abort (param i32 i32 i32 i32)))
  (memory $0 1)
- (data (i32.const 16) "\"\00\00\00\01\00\00\00\01\00\00\00\"\00\00\00a\00s\00s\00e\00r\00t\00-\00n\00o\00n\00n\00u\00l\00l\00.\00t\00s")
- (data (i32.const 80) "$\00\00\00\01\00\00\00\01\00\00\00$\00\00\00I\00n\00d\00e\00x\00 \00o\00u\00t\00 \00o\00f\00 \00r\00a\00n\00g\00e")
- (data (i32.const 144) "\1a\00\00\00\01\00\00\00\01\00\00\00\1a\00\00\00~\00l\00i\00b\00/\00a\00r\00r\00a\00y\00.\00t\00s")
- (data (i32.const 192) "^\00\00\00\01\00\00\00\01\00\00\00^\00\00\00E\00l\00e\00m\00e\00n\00t\00 \00t\00y\00p\00e\00 \00m\00u\00s\00t\00 \00b\00e\00 \00n\00u\00l\00l\00a\00b\00l\00e\00 \00i\00f\00 \00a\00r\00r\00a\00y\00 \00i\00s\00 \00h\00o\00l\00e\00y")
+ (data (i32.const 1024) "\"\00\00\00\01\00\00\00\01\00\00\00\"\00\00\00a\00s\00s\00e\00r\00t\00-\00n\00o\00n\00n\00u\00l\00l\00.\00t\00s")
+ (data (i32.const 1088) "$\00\00\00\01\00\00\00\01\00\00\00$\00\00\00I\00n\00d\00e\00x\00 \00o\00u\00t\00 \00o\00f\00 \00r\00a\00n\00g\00e")
+ (data (i32.const 1152) "\1a\00\00\00\01\00\00\00\01\00\00\00\1a\00\00\00~\00l\00i\00b\00/\00a\00r\00r\00a\00y\00.\00t\00s")
+ (data (i32.const 1200) "^\00\00\00\01\00\00\00\01\00\00\00^\00\00\00E\00l\00e\00m\00e\00n\00t\00 \00t\00y\00p\00e\00 \00m\00u\00s\00t\00 \00b\00e\00 \00n\00u\00l\00l\00a\00b\00l\00e\00 \00i\00f\00 \00a\00r\00r\00a\00y\00 \00i\00s\00 \00h\00o\00l\00e\00y")
  (table $0 1 funcref)
  (global $~argumentsLength (mut i32) (i32.const 0))
  (export "__setArgumentsLength" (func $~setArgumentsLength))
@@ -30,7 +30,7 @@
   i32.eqz
   if
    i32.const 0
-   i32.const 32
+   i32.const 1040
    i32.const 2
    i32.const 9
    call $~lib/builtins/abort
@@ -43,7 +43,7 @@
   i32.eqz
   if
    i32.const 0
-   i32.const 32
+   i32.const 1040
    i32.const 11
    i32.const 9
    call $~lib/builtins/abort
@@ -59,7 +59,7 @@
   i32.eqz
   if
    i32.const 0
-   i32.const 32
+   i32.const 1040
    i32.const 15
    i32.const 9
    call $~lib/builtins/abort
@@ -72,7 +72,7 @@
   i32.eqz
   if
    i32.const 0
-   i32.const 32
+   i32.const 1040
    i32.const 19
    i32.const 9
    call $~lib/builtins/abort
@@ -83,8 +83,8 @@
   i32.load offset=12
   i32.ge_u
   if
-   i32.const 96
-   i32.const 160
+   i32.const 1104
+   i32.const 1168
    i32.const 93
    i32.const 41
    call $~lib/builtins/abort
@@ -96,8 +96,8 @@
   local.tee $0
   i32.eqz
   if
-   i32.const 208
-   i32.const 160
+   i32.const 1216
+   i32.const 1168
    i32.const 97
    i32.const 39
    call $~lib/builtins/abort
@@ -111,8 +111,8 @@
   i32.load offset=12
   i32.ge_u
   if
-   i32.const 96
-   i32.const 160
+   i32.const 1104
+   i32.const 1168
    i32.const 93
    i32.const 41
    call $~lib/builtins/abort
@@ -129,7 +129,7 @@
   i32.eqz
   if
    i32.const 0
-   i32.const 32
+   i32.const 1040
    i32.const 23
    i32.const 9
    call $~lib/builtins/abort
@@ -157,7 +157,7 @@
    return
   end
   i32.const 0
-  i32.const 32
+  i32.const 1040
   i32.const 27
   i32.const 9
   call $~lib/builtins/abort
@@ -183,7 +183,7 @@
    return
   end
   i32.const 0
-  i32.const 32
+  i32.const 1040
   i32.const 31
   i32.const 9
   call $~lib/builtins/abort
@@ -204,7 +204,7 @@
   i32.eqz
   if
    i32.const 0
-   i32.const 32
+   i32.const 1040
    i32.const 39
    i32.const 12
    call $~lib/builtins/abort
@@ -226,7 +226,7 @@
   i32.eqz
   if
    i32.const 0
-   i32.const 32
+   i32.const 1040
    i32.const 44
    i32.const 9
    call $~lib/builtins/abort
@@ -252,7 +252,7 @@
   i32.eqz
   if
    i32.const 0
-   i32.const 32
+   i32.const 1040
    i32.const 52
    i32.const 9
    call $~lib/builtins/abort

tests/compiler/binary.optimized.wat

@@ -3,8 +3,8 @@
  (type $f32_=>_f32 (func (param f32) (result f32)))
  (type $f64_=>_f64 (func (param f64) (result f64)))
  (memory $0 1)
- (data (i32.const 17) "\01\00\00\01\00\00\00\03\00\00\00\00\01\00\00\be\f3\f8y\eca\f6?\190\96[\c6\fe\de\bf=\88\afJ\edq\f5?\a4\fc\d42h\0b\db\bf\b0\10\f0\f09\95\f4?{\b7\1f\n\8bA\d7\bf\85\03\b8\b0\95\c9\f3?{\cfm\1a\e9\9d\d3\bf\a5d\88\0c\19\0d\f3?1\b6\f2\f3\9b\1d\d0\bf\a0\8e\0b{\"^\f2?\f0z;\1b\1d|\c9\bf?4\1aJJ\bb\f1?\9f<\af\93\e3\f9\c2\bf\ba\e5\8a\f0X#\f1?\\\8dx\bf\cb`\b9\bf\a7\00\99A?\95\f0?\ce_G\b6\9do\aa\bf\00\00\00\00\00\00\f0?\00\00\00\00\00\00\00\00\acG\9a\fd\8c`\ee?=\f5$\9f\ca8\b3?\a0j\02\1f\b3\a4\ec?\ba\918T\a9v\c4?\e6\fcjW6 \eb?\d2\e4\c4J\0b\84\ce?-\aa\a1c\d1\c2\e9?\1ce\c6\f0E\06\d4?\edAx\03\e6\86\e8?\f8\9f\1b,\9c\8e\d8?bHS\f5\dcg\e7?\cc{\b1N\a4\e0\dc?")
- (data (i32.const 289) "\01\00\00\01\00\00\00\04\00\00\00\00\01\00\00\00\00\00\00\00\00\f0?t\85\15\d3\b0\d9\ef?\0f\89\f9lX\b5\ef?Q[\12\d0\01\93\ef?{Q}<\b8r\ef?\aa\b9h1\87T\ef?8bunz8\ef?\e1\de\1f\f5\9d\1e\ef?\15\b71\n\fe\06\ef?\cb\a9:7\a7\f1\ee?\"4\12L\a6\de\ee?-\89a`\08\ce\ee?\'*6\d5\da\bf\ee?\82O\9dV+\b4\ee?)TH\dd\07\ab\ee?\85U:\b0~\a4\ee?\cd;\7ff\9e\a0\ee?t_\ec\e8u\9f\ee?\87\01\ebs\14\a1\ee?\13\ceL\99\89\a5\ee?\db\a0*B\e5\ac\ee?\e5\c5\cd\b07\b7\ee?\90\f0\a3\82\91\c4\ee?]%>\b2\03\d5\ee?\ad\d3Z\99\9f\e8\ee?G^\fb\f2v\ff\ee?\9cR\85\dd\9b\19\ef?i\90\ef\dc 7\ef?\87\a4\fb\dc\18X\ef?_\9b{3\97|\ef?\da\90\a4\a2\af\a4\ef?@En[v\d0\ef?")
+ (data (i32.const 1025) "\01\00\00\01\00\00\00\03\00\00\00\00\01\00\00\be\f3\f8y\eca\f6?\190\96[\c6\fe\de\bf=\88\afJ\edq\f5?\a4\fc\d42h\0b\db\bf\b0\10\f0\f09\95\f4?{\b7\1f\n\8bA\d7\bf\85\03\b8\b0\95\c9\f3?{\cfm\1a\e9\9d\d3\bf\a5d\88\0c\19\0d\f3?1\b6\f2\f3\9b\1d\d0\bf\a0\8e\0b{\"^\f2?\f0z;\1b\1d|\c9\bf?4\1aJJ\bb\f1?\9f<\af\93\e3\f9\c2\bf\ba\e5\8a\f0X#\f1?\\\8dx\bf\cb`\b9\bf\a7\00\99A?\95\f0?\ce_G\b6\9do\aa\bf\00\00\00\00\00\00\f0?\00\00\00\00\00\00\00\00\acG\9a\fd\8c`\ee?=\f5$\9f\ca8\b3?\a0j\02\1f\b3\a4\ec?\ba\918T\a9v\c4?\e6\fcjW6 \eb?\d2\e4\c4J\0b\84\ce?-\aa\a1c\d1\c2\e9?\1ce\c6\f0E\06\d4?\edAx\03\e6\86\e8?\f8\9f\1b,\9c\8e\d8?bHS\f5\dcg\e7?\cc{\b1N\a4\e0\dc?")
+ (data (i32.const 1297) "\01\00\00\01\00\00\00\04\00\00\00\00\01\00\00\00\00\00\00\00\00\f0?t\85\15\d3\b0\d9\ef?\0f\89\f9lX\b5\ef?Q[\12\d0\01\93\ef?{Q}<\b8r\ef?\aa\b9h1\87T\ef?8bunz8\ef?\e1\de\1f\f5\9d\1e\ef?\15\b71\n\fe\06\ef?\cb\a9:7\a7\f1\ee?\"4\12L\a6\de\ee?-\89a`\08\ce\ee?\'*6\d5\da\bf\ee?\82O\9dV+\b4\ee?)TH\dd\07\ab\ee?\85U:\b0~\a4\ee?\cd;\7ff\9e\a0\ee?t_\ec\e8u\9f\ee?\87\01\ebs\14\a1\ee?\13\ceL\99\89\a5\ee?\db\a0*B\e5\ac\ee?\e5\c5\cd\b07\b7\ee?\90\f0\a3\82\91\c4\ee?]%>\b2\03\d5\ee?\ad\d3Z\99\9f\e8\ee?G^\fb\f2v\ff\ee?\9cR\85\dd\9b\19\ef?i\90\ef\dc 7\ef?\87\a4\fb\dc\18X\ef?_\9b{3\97|\ef?\da\90\a4\a2\af\a4\ef?@En[v\d0\ef?")
  (global $binary/b (mut i32) (i32.const 0))
  (global $binary/i (mut i32) (i32.const 0))
  (global $binary/I (mut i64) (i64.const 0))

tests/compiler/bool.optimized.wat

@@ -1,5 +1,5 @@
 (module
  (memory $0 1)
- (data (i32.const 16) "\0e\00\00\00\01\00\00\00\01\00\00\00\0e\00\00\00b\00o\00o\00l\00.\00t\00s")
+ (data (i32.const 1024) "\0e\00\00\00\01\00\00\00\01\00\00\00\0e\00\00\00b\00o\00o\00l\00.\00t\00s")
  (export "memory" (memory $0))
 )