Implement some asm optimized memcpy/memmove checks.

    TODO: memcmp

3 files changed, 46 insertions, 10 deletions

diff --git a/bench/bld.sub b/bench/bld.sub
index a6117ba..484226d 100644
--- a/bench/bld.sub
+++ b/bench/bld.sub
@@ -31,6 +31,14 @@ bin regex-match =
 	lib @/lib/sys:sys
 	lib @/lib/regex:regex
 ;;
+
+bin many-memcpy =
+	memcpy.myr
+	lib @/lib/std:std
+	lib @/lib/sys:sys
+;;
+
+# benchmark runner
 bin runbench =
 	runbench.myr
 	lib @/lib/std:std
@@ -45,4 +53,5 @@ cmd benchit =
 		bigfactorial
 		mandelbrot
 		regex-match
+		many-memcpy
 ;;
diff --git a/bench/copious-allocs.myr b/bench/copious-allocs.myr
index 1237b7f..7b02594 100644
--- a/bench/copious-allocs.myr
+++ b/bench/copious-allocs.myr
@@ -5,32 +5,31 @@ type blob = struct
 ;;
 
 const main = {
-	var i, j
 	var a : blob#[10000]
 
-	for j = 0; j < 100; j++
+	for var j = 0; j < 100; j++
 		/* alloc forwards, dealloc forwards */
-		for i = 0; i < a.len; i++
+		for var i = 0; i < a.len; i++
 			a[i] = std.alloc()
 		;;
-		for i = 0; i < a.len; i++
+		for var i = 0; i < a.len; i++
 			std.free(a[i])
 		;;
 
 		/* alloc forwards, dealloc backwards */
-		for i = 0; i < a.len; i++
+		for var i = 0; i < a.len; i++
 			a[i] = std.alloc()
 		;;
-		for i = a.len; i > 0; i--
+		for var i = a.len; i > 0; i--
 			std.free(a[i - 1])
 		;;
 
 		/* alloc forwards, dealloc randomly */
-		for i = 0; i < a.len; i++
+		for var i = 0; i < a.len; i++
 			a[i] = std.alloc()
 		;;
 		shuffle(a[:])
-		for i = a.len; i > 0; i--
+		for var i = a.len; i > 0; i--
 			std.free(a[i - 1])
 		;;
 	;;
@@ -39,10 +38,11 @@ const main = {
 const shuffle = {a
 	var t
 	var rng
-	var i, j
+	var j
 
+	/* we want determinism for benchmarking */
 	rng = std.mksrng(123)
-	for i = 0; i < a.len - 1; i++
+	for var i = 0; i < a.len - 1; i++
 		j = std.rngrand(rng, i, a.len)
 		t = a[j]
 		a[j] = a[i]
diff --git a/bench/many-memcpy.myr b/bench/many-memcpy.myr
new file mode 100644
index 0000000..94a5486
--- /dev/null
+++ b/bench/many-memcpy.myr
@@ -0,0 +1,27 @@
+use std
+
+const main = {
+	var a : uint64[100000]
+
+	for var j = 0; j < 100; j++
+		/* independent copies forward */
+		for var i = 0; i < 10; i++
+			std.slcp(a[:a.len/2-1], a[a.len/2+1:])
+		;;
+		/* independent copies backward */
+		for var i = 0; i < 10; i++
+			std.slcp(a[:a.len/2-1], a[a.len/2+1:])
+		;;
+
+		/* dependent copies forward */
+		for var i = 0; i < 10; i++
+			std.slcp(a[:a.len/2+1000], a[a.len/2-1000:])
+		;;
+		/* dependent copies backward */
+		for var i = 0; i < 10; i++
+			std.slcp(a[a.len/2-1000:], a[:a.len/2+1000])
+		;;
+	;;
+}
+
+