optimize hline function, update wasmtime

platform/Cargo.lock

@@ -146,7 +146,7 @@ checksum = "7a81dae078cea95a014a339291cec439d2f232ebe854a9d672b796c6afafa9b7"
 [[package]]
 name = "curlywas"
 version = "0.1.0"
-source = "git+https://github.com/exoticorn/curlywas.git?rev=0a0d90c#0a0d90c8013f1d7856a9b71a19d3d006300f31f9"
+source = "git+https://github.com/exoticorn/curlywas.git?rev=0e7ea50#0e7ea508cd0e76836283ae68a44c9097df83c8ac"
 dependencies = [
  "anyhow",
  "ariadne",

platform/Cargo.toml

@@ -6,7 +6,7 @@ edition = "2021"
 # See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
 
 [dependencies]
-curlywas = { git="https://github.com/exoticorn/curlywas.git", rev="0a0d90c" }
+curlywas = { git="https://github.com/exoticorn/curlywas.git", rev="0e7ea50" }
 uw8-tool = { path="../uw8-tool" }
 anyhow = "1"
 lodepng = "3.4"

platform/src/platform.cwa

@@ -62,11 +62,7 @@ export fn cls(col: i32) {
     textCursorX = 0;
     textCursorY = 0;
     outputChannel = 0;
-    col = (col & 255) * 0x1010101;
-    loop pixels {
-        i!120 = col;
-        branch_if (i := i + 4) < 320*240: pixels;
-    }
+    memory.fill(120, col, 320*240);
 }
 
 export fn setPixel(x: i32, y: i32, col: i32) {
@@ -90,12 +86,70 @@ fn clamp(v: i32, min: i32, max: i32) -> i32 {
 export fn hline(x1: i32, x2: i32, y: i32, col: i32) {
     x1 = clamp(x1, 0, 320);
     x2 = clamp(x2, 0, 320);
-    if x1 < x2 & y #< 240 {
+    if  y #>= 240 {
+        return;
+    }
+    let word_start = (x1 + 3) & -4;
+    let word_end = x2 & -4;
+    if word_end > word_start {
+        col = (col & 255) * 0x1010101;
+        let ptr = y * 320 + x1;
+        let end = ptr + word_start - x1;
+        if ptr + 2 <= end {
+            ptr?120 = col;
+            ptr?121 = col;
+            ptr += 2;
+        }
+        if ptr < end {
+            ptr?120 = col;
+            ptr += 1;
+        }
+        end += word_end - word_start;
+        loop words {
+            if ptr + 16 <= end {
+                ptr!120 = col;
+                ptr!124 = col;
+                ptr!128 = col;
+                ptr!132 = col;
+                ptr += 16;
+                branch words;
+            }
+            if ptr + 8 <= end {
+                ptr!120 = col;
+                ptr!124 = col;
+                ptr += 8;
+            }
+            if ptr < end {
+                ptr!120 = col;
+                ptr += 4;
+            }
+        }
+        end += x2 - word_end;
+        if ptr + 2 <= end {
+            ptr?120 = col;
+            ptr?121 = col;
+            ptr += 2;
+        }
+        if ptr < end {
+            ptr?120 = col;
+        }
+    } else {
         let ptr = y * 320 + x1;
         let end = ptr + x2 - x1;
-        loop pixels {
+        if ptr + 4 <= end {
+            ptr?120 = col;
+            ptr?121 = col;
+            ptr?122 = col;
+            ptr?123 = col;
+            ptr += 4;
+        }
+        if ptr + 2 <= end {
+            ptr?120 = col;
+            ptr?121 = col;
+            ptr += 2;
+        }
+        if ptr < end {
             ptr?120 = col;
-            branch_if (ptr := ptr + 1) < end: pixels;
         }
     }
 }
@@ -548,13 +602,6 @@ export fn endFrame() {
     68!4 = 68!0;
 }
 
-fn memclr(base: i32, size: i32) {
-    loop bytes {
-        (base + (size := size - 1))?0 = 0;
-        branch_if size: bytes;
-    }
-}
-
 start fn setup() {
     let i: i32 = 12*16*3-1;
     let avg: f32;
@@ -587,9 +634,9 @@ start fn setup() {
         branch_if (i := i - 1) >= 0: expand_sweetie;
     }
 
-    memclr(0, 64);
-    memclr(112, 8);
-    memclr(0x14000, 0x2c000);
+    memory.fill(0, 0, 64);
+    memory.fill(112, 0, 8);
+    memory.fill(0x14000, 0, 0x2c000);
 
 
     cls(0);