From c37347a4630fda601b4c40585ca25fff42f756c6 Mon Sep 17 00:00:00 2001
From: Quentin Carbonneaux <quentin@c9x.me>
Date: Tue, 12 Mar 2019 17:44:14 +0100
Subject: emit valid code for mem->mem copies

---
 amd64/emit.c       | 14 ++++++++---
 test/_phispill.ssa | 68 ------------------------------------------------------
 test/spill1.ssa    | 68 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 79 insertions(+), 71 deletions(-)
 delete mode 100644 test/_phispill.ssa
 create mode 100644 test/spill1.ssa

diff --git a/amd64/emit.c b/amd64/emit.c
index 653c548..ac7a1f3 100644
--- a/amd64/emit.c
+++ b/amd64/emit.c
@@ -360,7 +360,7 @@ emitins(Ins i, Fn *fn, FILE *f)
 {
 	Ref r;
 	int64_t val;
-	int o;
+	int o, t0;
 
 	switch (i.op) {
 	default:
@@ -434,19 +434,27 @@ emitins(Ins i, Fn *fn, FILE *f)
 		 * also, we can use a trick to load 64-bits
 		 * registers, it's detailed in my note below
 		 * http://c9x.me/art/notes.html?09/19/2015 */
+		t0 = rtype(i.arg[0]);
 		if (req(i.to, R) || req(i.arg[0], R))
 			break;
 		if (isreg(i.to)
-		&& rtype(i.arg[0]) == RCon
+		&& t0 == RCon
 		&& i.cls == Kl
 		&& fn->con[i.arg[0].val].type == CBits
 		&& (val = fn->con[i.arg[0].val].bits.i) >= 0
 		&& val <= UINT32_MAX) {
 			emitf("movl %W0, %W=", &i, fn, f);
 		} else if (isreg(i.to)
-		&& rtype(i.arg[0]) == RCon
+		&& t0 == RCon
 		&& fn->con[i.arg[0].val].type == CAddr) {
 			emitf("lea%k %M0, %=", &i, fn, f);
+		} else if (rtype(i.to) == RSlot
+		&& (t0 == RSlot || t0 == RMem)) {
+			i.cls = KWIDE(i.cls) ? Kd : Ks;
+			i.arg[1] = TMP(XMM0+15);
+			emitf("mov%k %0, %1", &i, fn, f);
+			emitf("mov%k %1, %=", &i, fn, f);
+
 		} else if (!req(i.arg[0], i.to))
 			emitf("mov%k %0, %=", &i, fn, f);
 		break;
diff --git a/test/_phispill.ssa b/test/_phispill.ssa
deleted file mode 100644
index 21e98c2..0000000
--- a/test/_phispill.ssa
+++ /dev/null
@@ -1,68 +0,0 @@
-export
-function w $f(w %n0) {
-@start
-@loop
-	%n1 =w phi @start %n0, @loop %n2
-
-	%p0 =w phi @start 0, @loop %p01
-	%p1 =w phi @start 0, @loop %p11
-	%p2 =w phi @start 0, @loop %p21
-	%p3 =w phi @start 0, @loop %p31
-	%p4 =w phi @start 0, @loop %p41
-	%p5 =w phi @start 0, @loop %p51
-	%p6 =w phi @start 0, @loop %p61
-	%p7 =w phi @start 0, @loop %p71
-	%p8 =w phi @start 0, @loop %p81
-	%p9 =w phi @start 0, @loop %p91
-	%pa =w phi @start 0, @loop %pa1
-	%pb =w phi @start 0, @loop %pb1
-	%pc =w phi @start 0, @loop %pc1
-	%pd =w phi @start 0, @loop %pd1
-	%pe =w phi @start 0, @loop %pe1
-	%pf =w phi @start 0, @loop %pf1
-
-	%p01 =w add 1, %p0
-	%p11 =w add 2, %p1
-	%p21 =w add 3, %p2
-	%p31 =w add 4, %p3
-	%p41 =w add 5, %p4
-	%p51 =w add 6, %p5
-	%p61 =w add 7, %p6
-	%p71 =w add 8, %p7
-	%p81 =w add 9, %p8
-	%p91 =w add 10, %p9
-	%pa1 =w add 11, %pa
-	%pb1 =w add 12, %pb
-	%pc1 =w add 13, %pc
-	%pd1 =w add 14, %pd
-	%pe1 =w add 15, %pe
-	%pf1 =w add 16, %pf
-
-	%n2 =w sub %n1, 1
-	jnz %n2, @loop, @end
-
-@end
-	%a =w sub 0, 0
-	%a =w add %p01, %a
-	%a =w add %p11, %a
-	%a =w add %p21, %a
-	%a =w add %p31, %a
-	%a =w add %p41, %a
-	%a =w add %p51, %a
-	%a =w add %p61, %a
-	%a =w add %p71, %a
-	%a =w add %p81, %a
-	%a =w add %p91, %a
-	%a =w add %pa1, %a
-	%a =w add %pb1, %a
-	%a =w add %pc1, %a
-	%a =w add %pd1, %a
-	%a =w add %pe1, %a
-	%a =w add %pf1, %a
-	ret %a
-}
-
-# >>> driver
-# extern int f(int);
-# int main() { return !(f(1) == 136); }
-# <<<
diff --git a/test/spill1.ssa b/test/spill1.ssa
new file mode 100644
index 0000000..21e98c2
--- /dev/null
+++ b/test/spill1.ssa
@@ -0,0 +1,68 @@
+export
+function w $f(w %n0) {
+@start
+@loop
+	%n1 =w phi @start %n0, @loop %n2
+
+	%p0 =w phi @start 0, @loop %p01
+	%p1 =w phi @start 0, @loop %p11
+	%p2 =w phi @start 0, @loop %p21
+	%p3 =w phi @start 0, @loop %p31
+	%p4 =w phi @start 0, @loop %p41
+	%p5 =w phi @start 0, @loop %p51
+	%p6 =w phi @start 0, @loop %p61
+	%p7 =w phi @start 0, @loop %p71
+	%p8 =w phi @start 0, @loop %p81
+	%p9 =w phi @start 0, @loop %p91
+	%pa =w phi @start 0, @loop %pa1
+	%pb =w phi @start 0, @loop %pb1
+	%pc =w phi @start 0, @loop %pc1
+	%pd =w phi @start 0, @loop %pd1
+	%pe =w phi @start 0, @loop %pe1
+	%pf =w phi @start 0, @loop %pf1
+
+	%p01 =w add 1, %p0
+	%p11 =w add 2, %p1
+	%p21 =w add 3, %p2
+	%p31 =w add 4, %p3
+	%p41 =w add 5, %p4
+	%p51 =w add 6, %p5
+	%p61 =w add 7, %p6
+	%p71 =w add 8, %p7
+	%p81 =w add 9, %p8
+	%p91 =w add 10, %p9
+	%pa1 =w add 11, %pa
+	%pb1 =w add 12, %pb
+	%pc1 =w add 13, %pc
+	%pd1 =w add 14, %pd
+	%pe1 =w add 15, %pe
+	%pf1 =w add 16, %pf
+
+	%n2 =w sub %n1, 1
+	jnz %n2, @loop, @end
+
+@end
+	%a =w sub 0, 0
+	%a =w add %p01, %a
+	%a =w add %p11, %a
+	%a =w add %p21, %a
+	%a =w add %p31, %a
+	%a =w add %p41, %a
+	%a =w add %p51, %a
+	%a =w add %p61, %a
+	%a =w add %p71, %a
+	%a =w add %p81, %a
+	%a =w add %p91, %a
+	%a =w add %pa1, %a
+	%a =w add %pb1, %a
+	%a =w add %pc1, %a
+	%a =w add %pd1, %a
+	%a =w add %pe1, %a
+	%a =w add %pf1, %a
+	ret %a
+}
+
+# >>> driver
+# extern int f(int);
+# int main() { return !(f(1) == 136); }
+# <<<
-- 
cgit 1.4.1