ifopt simplifications

3 files changed, 130 insertions, 153 deletions

diff --git a/amd64/emit.c b/amd64/emit.c
index 7290a80..8d715d0 100644
--- a/amd64/emit.c
+++ b/amd64/emit.c
@@ -12,24 +12,22 @@ struct E {
 };
 
 #define CMP(X) \
-	X(Ciule,      "be") \
-	X(Ciult,      "b")  \
-	X(Cisle,      "le") \
-	X(Cislt,      "l")  \
-	X(Cisgt,      "g")  \
-	X(Cisge,      "ge") \
-	X(Ciugt,      "a")  \
-	X(Ciuge,      "ae") \
-	X(Cieq,       "z")  \
-	X(Cine,       "nz") \
-	X(NCmpI+Cfle, "be") \
-	X(NCmpI+Cflt, "b")  \
-	X(NCmpI+Cfgt, "a")  \
-	X(NCmpI+Cfge, "ae") \
-	X(NCmpI+Cfeq, "z")  \
-	X(NCmpI+Cfne, "nz") \
-	X(NCmpI+Cfo,  "np") \
-	X(NCmpI+Cfuo, "p")
+	X(Ciule,      "be", "a") \
+	X(Ciult,      "b", "ae") \
+	X(Cisle,      "le", "g") \
+	X(Cislt,      "l", "ge") \
+	X(Cisgt,      "g", "le") \
+	X(Cisge,      "ge", "l") \
+	X(Ciugt,      "a", "be") \
+	X(Ciuge,      "ae", "b") \
+	X(Cieq,       "z", "nz") \
+	X(Cine,       "nz", "z") \
+	X(NCmpI+Cfle, "be", "a") \
+	X(NCmpI+Cflt, "b", "ae") \
+	X(NCmpI+Cfgt, "a", "be") \
+	X(NCmpI+Cfge, "ae", "b") \
+	X(NCmpI+Cfo,  "np", "p") \
+	X(NCmpI+Cfuo, "p", "np")
 
 enum {
 	SLong = 0,
@@ -72,65 +70,77 @@ static struct {
 	short cls;
 	char *fmt;
 } omap[] = {
-	{ Oadd,    Ka, "+add%k %1, %=" },
-	{ Osub,    Ka, "-sub%k %1, %=" },
-	{ Oand,    Ki, "+and%k %1, %=" },
-	{ Oor,     Ki, "+or%k %1, %=" },
-	{ Oxor,    Ki, "+xor%k %1, %=" },
-	{ Osar,    Ki, "-sar%k %B1, %=" },
-	{ Oshr,    Ki, "-shr%k %B1, %=" },
-	{ Oshl,    Ki, "-shl%k %B1, %=" },
-	{ Omul,    Ki, "+imul%k %1, %=" },
-	{ Omul,    Ks, "+mulss %1, %=" },
-	{ Omul,    Kd, "+mulsd %1, %=" },
-	{ Odiv,    Ka, "-div%k %1, %=" },
-	{ Ostorel, Ka, "movq %L0, %M1" },
-	{ Ostorew, Ka, "movl %W0, %M1" },
-	{ Ostoreh, Ka, "movw %H0, %M1" },
-	{ Ostoreb, Ka, "movb %B0, %M1" },
-	{ Ostores, Ka, "movss %S0, %M1" },
-	{ Ostored, Ka, "movsd %D0, %M1" },
-	{ Oload,   Ka, "mov%k %M0, %=" },
-	{ Oloadsw, Kl, "movslq %M0, %L=" },
-	{ Oloadsw, Kw, "movl %M0, %W=" },
-	{ Oloaduw, Ki, "movl %M0, %W=" },
-	{ Oloadsh, Ki, "movsw%k %M0, %=" },
-	{ Oloaduh, Ki, "movzw%k %M0, %=" },
-	{ Oloadsb, Ki, "movsb%k %M0, %=" },
-	{ Oloadub, Ki, "movzb%k %M0, %=" },
-	{ Oextsw,  Kl, "movslq %W0, %L=" },
-	{ Oextuw,  Kl, "movl %W0, %W=" },
-	{ Oextsh,  Ki, "movsw%k %H0, %=" },
-	{ Oextuh,  Ki, "movzw%k %H0, %=" },
-	{ Oextsb,  Ki, "movsb%k %B0, %=" },
-	{ Oextub,  Ki, "movzb%k %B0, %=" },
-
-	{ Oexts,   Kd, "cvtss2sd %0, %=" },
-	{ Otruncd, Ks, "cvtsd2ss %0, %=" },
-	{ Ostosi,  Ki, "cvttss2si%k %0, %=" },
-	{ Odtosi,  Ki, "cvttsd2si%k %0, %=" },
-	{ Oswtof,  Ka, "cvtsi2%k %W0, %=" },
-	{ Osltof,  Ka, "cvtsi2%k %L0, %=" },
-	{ Ocast,   Ki, "movq %D0, %L=" },
-	{ Ocast,   Ka, "movq %L0, %D=" },
-
-	{ Oaddr,   Ki, "lea%k %M0, %=" },
-	{ Oswap,   Ki, "xchg%k %0, %1" },
-	{ Osign,   Kl, "cqto" },
-	{ Osign,   Kw, "cltd" },
-	{ Oxdiv,   Ki, "div%k %0" },
-	{ Oxidiv,  Ki, "idiv%k %0" },
-	{ Oxcmp,   Ks, "ucomiss %S0, %S1" },
-	{ Oxcmp,   Kd, "ucomisd %D0, %D1" },
-	{ Oxcmp,   Ki, "cmp%k %0, %1" },
-	{ Oxtest,  Ki, "test%k %0, %1" },
-#define X(c, s) \
-	{ Oflag+c, Ki, "set" s " %B=\n\tmovzb%k %B=, %=" },
+	{ Oadd,     Ka, "+add%k %1, %=" },
+	{ Osub,     Ka, "-sub%k %1, %=" },
+	{ Oand,     Ki, "+and%k %1, %=" },
+	{ Oor,      Ki, "+or%k %1, %=" },
+	{ Oxor,     Ki, "+xor%k %1, %=" },
+	{ Osar,     Ki, "-sar%k %B1, %=" },
+	{ Oshr,     Ki, "-shr%k %B1, %=" },
+	{ Oshl,     Ki, "-shl%k %B1, %=" },
+	{ Omul,     Ki, "+imul%k %1, %=" },
+	{ Omul,     Ks, "+mulss %1, %=" },
+	{ Omul,     Kd, "+mulsd %1, %=" },
+	{ Odiv,     Ka, "-div%k %1, %=" },
+	{ Ostorel,  Ka, "movq %L0, %M1" },
+	{ Ostorew,  Ka, "movl %W0, %M1" },
+	{ Ostoreh,  Ka, "movw %H0, %M1" },
+	{ Ostoreb,  Ka, "movb %B0, %M1" },
+	{ Ostores,  Ka, "movss %S0, %M1" },
+	{ Ostored,  Ka, "movsd %D0, %M1" },
+	{ Oload,    Ka, "mov%k %M0, %=" },
+	{ Oloadsw,  Kl, "movslq %M0, %L=" },
+	{ Oloadsw,  Kw, "movl %M0, %W=" },
+	{ Oloaduw,  Ki, "movl %M0, %W=" },
+	{ Oloadsh,  Ki, "movsw%k %M0, %=" },
+	{ Oloaduh,  Ki, "movzw%k %M0, %=" },
+	{ Oloadsb,  Ki, "movsb%k %M0, %=" },
+	{ Oloadub,  Ki, "movzb%k %M0, %=" },
+	{ Oextsw,   Kl, "movslq %W0, %L=" },
+	{ Oextuw,   Kl, "movl %W0, %W=" },
+	{ Oextsh,   Ki, "movsw%k %H0, %=" },
+	{ Oextuh,   Ki, "movzw%k %H0, %=" },
+	{ Oextsb,   Ki, "movsb%k %B0, %=" },
+	{ Oextub,   Ki, "movzb%k %B0, %=" },
+
+	{ Oexts,    Kd, "cvtss2sd %0, %=" },
+	{ Otruncd,  Ks, "cvtsd2ss %0, %=" },
+	{ Ostosi,   Ki, "cvttss2si%k %0, %=" },
+	{ Odtosi,   Ki, "cvttsd2si%k %0, %=" },
+	{ Oswtof,   Ka, "cvtsi2%k %W0, %=" },
+	{ Osltof,   Ka, "cvtsi2%k %L0, %=" },
+	{ Ocast,    Ki, "movq %D0, %L=" },
+	{ Ocast,    Ka, "movq %L0, %D=" },
+
+	{ Oaddr,    Ki, "lea%k %M0, %=" },
+	{ Oswap,    Ki, "xchg%k %0, %1" },
+	{ Osign,    Kl, "cqto" },
+	{ Osign,    Kw, "cltd" },
+	{ Oxdiv,    Ki, "div%k %0" },
+	{ Oxidiv,   Ki, "idiv%k %0" },
+	{ Oxcmp,    Ks, "ucomiss %S0, %S1" },
+	{ Oxcmp,    Kd, "ucomisd %D0, %D1" },
+	{ Oxcmp,    Ki, "cmp%k %0, %1" },
+	{ Oxtest,   Ki, "test%k %0, %1" },
+#define X(c, s, _) \
+	{ Oflag+c,  Ki, "set" s " %B=\n\tmovzb%k %B=, %=" },
 	CMP(X)
 #undef X
+	{ Oflagfeq, Ki, "setz %B=\n\tmovzb%k %B=, %=" },
+	{ Oflagfne, Ki, "setnz %B=\n\tmovzb%k %B=, %=" },
 	{ NOp, 0, 0 }
 };
 
+static char cmov[][2][16] = {
+#define X(c, s0, s1) \
+	[c] = { \
+		"cmov" s0 " %0, %=", \
+		"cmov" s1 " %1, %=", \
+	},
+	CMP(X)
+#undef X
+};
+
 static char *rname[][4] = {
 	[RAX] = {"rax", "eax", "ax", "al"},
 	[RBX] = {"rbx", "ebx", "bx", "bl"},
@@ -401,6 +411,8 @@ emitins(Ins i, E *e)
 
 	switch (i.op) {
 	default:
+		if (isxsel(i.op))
+			goto case_Oxsel;
 	Table:
 		/* most instructions are just pulled out of
 		 * the table omap[], some special cases are
@@ -576,44 +588,16 @@ emitins(Ins i, E *e)
 	case Odbgloc:
 		emitdbgloc(i.arg[0].val, i.arg[1].val, e->f);
 		break;
-	case Oxselieq:
-	case Oxseline:
-	case Oxselisge:
-	case Oxselisgt:
-	case Oxselisle:
-	case Oxselislt:
-	case Oxseliuge:
-	case Oxseliugt:
-	case Oxseliule:
-	case Oxseliult:
-	case Oxselfeq:
-	case Oxselfge:
-	case Oxselfgt:
-	case Oxselfle:
-	case Oxselflt:
-	case Oxselfne:
-	case Oxselfo:
-	case Oxselfuo:
-	{
-		// TODO - how to do this "properly"?
-		static char *F0[] = {
-			 "z", "nz", "ge",  "g", "le",  "l", "ae",  "a", "be",  "b",
-			"nz", "ae",  "a", "be",  "b",  "nz",  "p", "np"
-		};
-		static char *F1[] = {
-			"nz", "z",  "l", "le",  "g", "ge",  "b", "be",  "a", "ae",
-			 "z", "b", "be",  "a", "ae",  "z", "p",  "np"
-		};
-		char ins[16];
-		sprintf(ins, "cmov%s %%1, %%=", F1[i.op-Oxselieq]);
+	case_Oxsel:
 		if (req(i.to, i.arg[1]))
-			sprintf(ins, "cmov%s %%0, %%=", F0[i.op-Oxselieq]);
-		else if (!req(i.to, i.arg[0]))
-			emitf("mov %0, %=", &i, e);
-		emitf(ins, &i, e);
+			emitf(cmov[i.op-Oxsel][0], &i, e);
+		else {
+			if (!req(i.to, i.arg[0]))
+				emitf("mov %0, %=", &i, e);
+			emitf(cmov[i.op-Oxsel][1], &i, e);
+		}
 		break;
 	}
-	}
 }
 
 static void
@@ -641,7 +625,7 @@ void
 amd64_emitfn(Fn *fn, FILE *f)
 {
 	static char *ctoa[] = {
-	#define X(c, s) [c] = s,
+	#define X(c, s, _) [c] = s,
 		CMP(X)
 	#undef X
 	};
diff --git a/amd64/isel.c b/amd64/isel.c
index c54b580..73e0860 100644
--- a/amd64/isel.c
+++ b/amd64/isel.c
@@ -429,25 +429,8 @@ sel(Ins i, Num *tn, Fn *fn)
 	case Oexts:
 	case Otruncd:
 	case Ocast:
-	case Oxselieq:
-	case Oxseline:
-	case Oxselisge:
-	case Oxselisgt:
-	case Oxselisle:
-	case Oxselislt:
-	case Oxseliuge:
-	case Oxseliugt:
-	case Oxseliule:
-	case Oxseliult:
-	case Oxselfeq:
-	case Oxselfge:
-	case Oxselfgt:
-	case Oxselfle:
-	case Oxselflt:
-	case Oxselfne:
-	case Oxselfo:
-	case Oxselfuo:
-	case_OExt:
+	case_Oxsel:
+	case_Oext:
 Emit:
 		emiti(i);
 		i1 = curi; /* fixarg() can change curi */
@@ -461,7 +444,9 @@ Emit:
 		break;
 	default:
 		if (isext(i.op))
-			goto case_OExt;
+			goto case_Oext;
+		if (isxsel(i.op))
+			goto case_Oxsel;
 		if (isload(i.op))
 			goto case_Oload;
 		if (iscmp(i.op, &kc, &x)) {
@@ -518,13 +503,13 @@ flagi(Ins *i0, Ins *i)
 static Ins*
 selsel(Fn *fn, Blk *b, Ins *i, Num *tn)
 {
-	Ref r, cr0, cr1;
+	Ref r, cr[2];
 	int c, k, swap, gencmp, gencpy;
 	Ins *isel0, *isel1, *fi;
 	Tmp *t;
 
 	assert(i->op == Osel1);
-	for (isel0 = i; b->ins < isel0; isel0--) {
+	for (isel0=i; b->ins<isel0; isel0--) {
 		if (isel0->op == Osel0)
 			break;
 		assert(isel0->op == Osel1);
@@ -534,29 +519,38 @@ selsel(Fn *fn, Blk *b, Ins *i, Num *tn)
 	assert(rtype(r) == RTmp);
 	t = &fn->tmp[r.val];
 	fi = flagi(b->ins, isel0);
-	cr0 = cr1 = R;
+	cr[0] = cr[1] = R;
 	gencmp = gencpy = swap = 0;
 	k = Kw;
 	c = Cine;
 	if (!fi || !req(fi->to, r)) {
 		gencmp = 1;
-		cr0 = r;
-		cr1 = CON_Z;
-	} else if (iscmp(fi->op, &k, &c)
-		   && c != NCmpI+Cfeq /* see sel() */
-		   && c != NCmpI+Cfne) {
+		cr[0] = r;
+		cr[1] = CON_Z;
+	}
+	else if (iscmp(fi->op, &k, &c)) {
+		if (c == NCmpI+Cfeq
+		|| c == NCmpI+Cfne) {
+			/* these are selected as 'and'
+			 * or 'or', so we check their
+			 * result with Cine
+			 */
+			c = Cine;
+			goto Other;
+		}
 		swap = cmpswap(fi->arg, c);
 		if (swap)
 			c = cmpop(c);
 		if (t->nuse == 1) {
 			gencmp = 1;
-			cr0 = fi->arg[0];
-			cr1 = fi->arg[1];
+			cr[0] = fi->arg[0];
+			cr[1] = fi->arg[1];
 			*fi = (Ins){.op = Onop};
 		}
-	} else if (fi->op == Oand && t->nuse == 1
-		&& (rtype(fi->arg[0]) == RTmp ||
-		    rtype(fi->arg[1]) == RTmp)) {
+	}
+	else if (fi->op == Oand && t->nuse == 1
+	     && (rtype(fi->arg[0]) == RTmp ||
+	         rtype(fi->arg[1]) == RTmp)) {
 		fi->op = Oxtest;
 		fi->to = R;
 		if (rtype(fi->arg[1]) == RCon) {
@@ -564,7 +558,9 @@ selsel(Fn *fn, Blk *b, Ins *i, Num *tn)
 			fi->arg[1] = fi->arg[0];
 			fi->arg[0] = r;
 		}
-	} else {
+	}
+	else {
+	Other:
 		/* since flags are not tracked in liveness,
 		 * the result of the flag-setting instruction
 		 * has to be marked as live
@@ -573,18 +569,15 @@ selsel(Fn *fn, Blk *b, Ins *i, Num *tn)
 			gencpy = 1;
 	}
 	/* generate conditional moves */
-	for (isel1 = i; isel0 < isel1; --isel1) {
-		isel1->op = Oxselieq+c;
+	for (isel1=i; isel0<isel1; --isel1) {
+		isel1->op = Oxsel+c;
 		sel(*isel1, tn, fn);
 	}
-	if (gencmp) {
-		assert(!gencpy);
-		selcmp((Ref[2]){cr0, cr1}, k, swap, fn);
-	}
-	if (gencpy) {
-		assert(!gencmp);
+	assert(!gencmp || !gencpy);
+	if (gencmp)
+		selcmp(cr, k, swap, fn);
+	if (gencpy)
 		emit(Ocopy, Kw, R, r, R);
-	}
 	*isel0 = (Ins){.op = Onop};
 	return isel0;
 }
@@ -618,7 +611,7 @@ seljmp(Blk *b, Fn *fn)
 		b->jmp.type = Jjf + Cine;
 	}
 	else if (iscmp(fi->op, &k, &c)
-	     && c != NCmpI+Cfeq /* see sel() */
+	     && c != NCmpI+Cfeq /* see sel(), selsel() */
 	     && c != NCmpI+Cfne) {
 		swap = cmpswap(fi->arg, c);
 		if (swap)
diff --git a/ops.h b/ops.h
index 2336d5f..f59e757 100644
--- a/ops.h
+++ b/ops.h
@@ -145,7 +145,7 @@ O(nop,     T(x,x,x,x, x,x,x,x), F(0,0,0,0,0,0,0,0,0,0)) X(0,0,1) V(0)
 O(addr,    T(m,m,e,e, x,x,e,e), F(0,0,0,0,0,0,0,0,0,0)) X(0,0,1) V(0)
 O(blit0,   T(m,e,e,e, m,e,e,e), F(0,0,0,0,0,0,0,0,0,1)) X(0,1,0) V(0)
 O(blit1,   T(w,e,e,e, x,e,e,e), F(0,0,0,0,0,0,0,0,0,1)) X(0,1,0) V(0)
-O(sel0,    T(w,e,e,e, x,x,x,x), F(0,0,0,0,0,0,0,0,0,1)) X(0,0,0) V(0)
+O(sel0,    T(w,e,e,e, x,e,e,e), F(0,0,0,0,0,0,0,0,0,1)) X(0,0,0) V(0)
 O(sel1,    T(w,l,e,e, w,l,e,e), F(0,0,0,0,0,0,0,0,0,1)) X(0,0,0) V(0)
 O(swap,    T(w,l,s,d, w,l,s,d), F(0,0,0,0,0,0,0,0,0,0)) X(1,0,0) V(0)
 O(sign,    T(w,l,e,e, x,x,e,e), F(0,0,0,0,0,0,0,0,0,0)) X(0,0,0) V(0)