Implement optimized lbswap and llbswap

Author: runer112

examples/standalone/math_test/autotest.json

@@ -20,6 +20,8 @@
     "key|enter",
     "hashWait|bitrev",
     "key|enter",
+    "hashWait|bswap",
+    "key|enter",
     "hashWait|popcnt",
     "key|enter",
     "hashWait|and",
@@ -87,6 +89,14 @@
         "BBADF82E"
       ]
     },
+    "bswap": {
+      "description": "bswap",
+      "start": "vram_start",
+      "size": "vram_16_size",
+      "expected_CRCs": [
+        "7709B5AE"
+      ]
+    },
     "popcnt": {
       "description": "popcnt",
       "start": "vram_start",

examples/standalone/math_test/src/main.c

@@ -206,6 +206,8 @@ DEFINE_BINOP_TYPE(u)
     static const u##UnOp unop_##name = {#name, b##name##_, s##name##_, i##name##_, l##name##_};
 #define DEFINE_UNOP_STRUCT_B_TO_LL(u, name) \
     static const u##UnOp unop_##name = {#name, b##name##_, s##name##_, i##name##_, l##name##_, ll##name##_};
+#define DEFINE_UNOP_STRUCT_S_L_LL(u, name) \
+    static const u##UnOp unop_##name = {#name, 0, s##name##_, 0, l##name##_, ll##name##_};
 
 #define DEFINE_BINOP_STRUCT_B(u, name) \
     static const u##BinOp binop_##name = {#name, b##name##_};
@@ -316,6 +318,12 @@ DEFINE_UNOP_PREFIX_FUNC_L( , bitrev, __builtin_bitreverse32)
 DEFINE_UNOP_PREFIX_FUNC_LL( , bitrev, __builtin_bitreverse64)
 DEFINE_UNOP_STRUCT_B_TO_LL( , bitrev)
 
+// Needs to be unsigned to avoid extra bits from sign extension
+DEFINE_UNOP_PREFIX_FUNC_S(u, bswap, __builtin_bswap16)
+DEFINE_UNOP_PREFIX_FUNC_L(u, bswap, __builtin_bswap32)
+DEFINE_UNOP_PREFIX_FUNC_LL(u, bswap, __builtin_bswap64)
+DEFINE_UNOP_STRUCT_S_L_LL(u, bswap)
+
 // Needs to be unsigned to avoid extra bits from sign extension
 DEFINE_UNOP_PREFIX_FUNC_B_TO_I(u, popcnt, __builtin_popcount)
 DEFINE_UNOP_PREFIX_FUNC_L(u, popcnt, __builtin_popcountl)
@@ -349,6 +357,7 @@ static const UnOp *unops[] = {
     &unop_neg,
     &unop_abs,
     &unop_bitrev,
+    (const UnOp *)&unop_bswap,
     (const UnOp *)&unop_popcnt,
 };
 

src/crt/lbswap.src

@@ -0,0 +1,15 @@
+	assume	adl=1
+
+	section	.text
+	public	__lbswap
+__lbswap:
+	push	hl			; ...[hlu]hl <- (sp)
+	inc	sp			; ...[hlu]h <- (sp)
+	ld	h, e			; uhl = [hlu]el
+	push	hl			; ...[hlu]h[hlu]el <- (sp)
+	inc	sp			; ...[hlu]h[hlu]e <- (sp)
+	ld	e, l
+	pop	hl			; euhl = lh[hlu]e
+					; ...[hlu] <- (sp)
+	inc	sp			; ... <- (sp)
+	ret

src/crt/llbswap.src

@@ -0,0 +1,16 @@
+	assume	adl=1
+
+	section	.text
+	public	__llbswap
+__llbswap:
+	push	af
+	ld	a, b			; a = b
+	ld	b, l			; bcudeuhl = lc[deu]de[hlu]hl
+	ld	l, d			; bcudeuhl = lc[deu]de[hlu]hd
+	ld	d, c			; bcudeuhl = lc[deu]ce[hlu]hd
+	ld	c, h			; bcudeuhl = lh[deu]ce[hlu]hd
+	ld	h, e			; bcudeuhl = lh[deu]ce[hlu]ed
+	ld	e, a			; bcudeuhl = lh[deu]cb[hlu]ed
+	ex	de, hl			; bcudeuhl = lh[hlu]ed[deu]cb
+	pop	af
+	ret