be-codegen-6800.c Optimizes addition and subtraction with signed/unsigned char.

[zu2]

Currently, addition with signed/unsigned char pushes the augend onto the stack, sign-extends the addend, and then calls a helper function.

This patch provides shorter code.

--- ../Fuzix-Compiler-Kit/be-codegen-6800.c	2024-11-18 09:16:57
+++ be-codegen-6800.c	2024-11-18 21:02:24
@@ -514,6 +514,20 @@
 			invalidate_d();
 			add_d_const(v);
 			return 1;
+		}
+		if (s == 2 && r->op == T_CAST && (r->type&~UNSIGNED) == CSHORT && is_simple(r->right)) {
+			if (r->right->type == UCHAR) {
+				op8_on_node(r->right, "add", 0);
+				puts("\tadca #0");
+				return 1;
+			}else if (r->right->type == CCHAR) {
+				unsigned l = ++label;
+				write_uni_op(r->right, "tst", 0);
+				printf("\tbpl X%u\n\tdeca\nX%u:\n",l,l);
+				op8_on_node(r->right, "add", 0);
+				puts("\tadca #0");
+				return 1;
+			}
 		}
 		return write_opd(r, "add", "adc", 0);
 	case T_MINUS:
@@ -527,6 +541,20 @@
 			if (s == 1) {
 				add_b_const(-r->value);
 				b_val -= r->value;
+				return 1;
+			}
+		}
+		if (s == 2 && r->op == T_CAST && (r->type&~UNSIGNED) == CSHORT && is_simple(r->right)) {
+			if (r->right->type == UCHAR) {
+				op8_on_node(r->right, "sub", 0);
+				puts("\tsbca #0");
+				return 1;
+			}else if (r->right->type == CCHAR) {
+				unsigned l = ++label;
+				write_uni_op(r->right, "tst", 0);
+				printf("\tbpl X%u\n\tinca\nX%u:\n",l,l);
+				op8_on_node(r->right, "sub", 0);
+				puts("\tsbca #0");
 				return 1;
 			}
 		}

the patch pass this test program.

#include	"cprintf.c"

int
main(int argc,char **argv)
{
	signed char c1, c2;
	int	i1,i2;

	for(i1=-128; i1<=127; i1++){
		for(i2=-128; i2<=127; i2++){
			c1 = i1;
			c2 = i2;
			if(c1+c2 != i1+i2){
				cprintf("%d+%d=%d (!=%d)\n",i1,i2,c1+c2,i1+i2);
				return 1;
			}
			if(c1-c2 != i1-i2){
				cprintf("%d+%d=%d (!=%d)\n",i1,i2,c1-c2,i1-i2);
				return 2;
			}
		}
	}

	return 0;
}

before

        ldb 0,x
        clra
        asrb
        rolb
        sbca #0
        pshb
        psha
        ldb 1,x
        clra
        asrb
        rolb
        sbca #0
        jsr __plus

after.

        ldb 0,x
        clra
        asrb
        rolb
        sbca #0
        tst 1,x
        bpl X1
        deca
X1:
        addb 1,x
        adca #0

This patch does not shorten the form *p - *q as memcmp uses. I am looking for a way to do this.

[zu2]

In the signed case, this results in better code. use gen_shortcut.

@@ -1393,6 +1428,38 @@
 		return do_xeqop(n, "xoreq");
 	case T_HATEQ:
 		return do_xeqop(n, "xhateq");
+	case T_PLUS:
+		if (s==2 && n->type==CSHORT
+		&&  r->op == T_CAST && r->type == CSHORT
+	    &&  l->op == T_CAST && l->type == CSHORT
+		&&  r->right->type == CCHAR
+		&&  l->right->type == CCHAR ){
+			if (is_simple(r->right)){
+				codegen_lr(l->right);
+				puts("\tclra");
+                		op8_on_node(r->right, "add", 0);
+				printf("\tbge X%d\n\tdeca\n",++label);
+				printf("X%d:\n",label);
+				return 1;
+			}
+		}
+		return 0;
+	case T_MINUS:
+		if (s==2 && n->type==CSHORT
+		&&  r->op == T_CAST && r->type == CSHORT
+	    &&  l->op == T_CAST && l->type == CSHORT
+		&&  r->right->type == CCHAR
+		&&  l->right->type == CCHAR ){
+			if (is_simple(r->right)){
+				codegen_lr(l->right);
+				puts("\tclra");
+                		op8_on_node(r->right, "sub", 0);
+				printf("\tbge X%d\n\tdeca\n",++label);
+				printf("X%d:\n",label);
+				return 1;
+			}
+		}
+		return 0;

[EtchedPixels]

Will defer until after we have byte conversions done and can see what is actually needed for best results