idem.

4 files changed, 87 insertions, 51 deletions

diff --git a/avx.h b/avx.h
index eea76a1..5ae12ec 100644
--- a/avx.h
+++ b/avx.h
@@ -26,14 +26,14 @@
 #define VZEROALL	VEX3(0,0,0,VEX_m_0F,0,0,VEX_L_256,VEX_p_NO); BYTE $0x77
 
 /* VMOVUPD */
-#define VMOVUPD_128mr(off, s, d)	VEX3(0,0,0,VEX_m_0F,0,0,VEX_L_128,VEX_p_66);	\
-				VOPi(0x10, 0x1, (d), (s), (off))
-#define VMOVUPD_128rr(s, d)	VEX3(0,0,0,VEX_m_0F,0,0,VEX_L_128,VEX_p_66);		\
-			VOP(0x10, 0x3, (d), (s))
-#define VMOVUPD_256mr(off, s, d)	VEX3(0,0,0,VEX_m_0F,0,0,VEX_L_256,VEX_p_66);	\
-				VOPi(0x10, 0x1, (d), (s), (off))
-#define VMOVUPD_256rr(s, d)	VEX3(0,0,0,VEX_m_0F,0,0,VEX_L_256,VEX_p_66);		\
-			VOP(0x10, 0x3, (d), (s))
+#define VMOVUPD_128mr(s, d)	VEX3(0,0,0,VEX_m_0F,0,0,VEX_L_128,VEX_p_66);	\
+				VOP(0x10, 0x0, (d), (s))
+#define VMOVUPD_128rm(s, d)	VEX3(0,0,0,VEX_m_0F,0,0,VEX_L_128,VEX_p_66);		\
+			VOP(0x11, 0x0, (s), (d))
+#define VMOVUPD_256mr(s, d)	VEX3(0,0,0,VEX_m_0F,0,0,VEX_L_256,VEX_p_66);	\
+				VOP(0x10, 0x0, (d), (s))
+#define VMOVUPD_256rm(s, d)	VEX3(0,0,0,VEX_m_0F,0,0,VEX_L_256,VEX_p_66);		\
+			VOP(0x11, 0x0, (s), (d))
 
 /* VMOVAPD */
 #define VMOVAPD_128mr(off, s, d)	VEX3(0,0,0,VEX_m_0F,0,0,VEX_L_128,VEX_p_66);	\
diff --git a/dppd.s b/dppd.s
index db805ff..55ee3d9 100644
--- a/dppd.s
+++ b/dppd.s
@@ -6,40 +6,41 @@ DATA one(SB)/8,$1.0
 GLOBL one(SB), $8
 
 TEXT dotvec2_sse4(SB), 1, $0
-	MOVQ SP, AX
-	MOVDQU_mr(8, rAX, rX0)	/* MOVDQU a+0(FP), X0 */
-	MOVDQU_mr(32, rAX, rX1)	/* MOVDQU b+24(FP), X1 */
+	MOVUPD a+0(FP), X0
+	MOVUPD b+24(FP), X1
 	DPPD(rX1, rX0)		/* DPPD $0x31, X1, X0 */
 	RET
 
 TEXT dotvec2_avx(SB), 1, $0
 	MOVQ SP, AX
-	VMOVUPD_128mr(8, rAX, rX0)	/* VMOVUPD a+0(FP), X0 */
-	VMOVUPD_128mr(32, rAX, rX1)	/* VMOVUPD b+24(FP), X1 */
+	ADDQ $8, AX
+	VMOVUPD_128mr(rAX, rX0)
+	ADDQ $24, AX
+	VMOVUPD_128mr(rAX, rX1)
 	VDPPD(rX1, rX0, rX0)		/* VDPPD $0x31, X1, X0, X0 */
+	VZEROUPPER
 	RET
 
 TEXT dotvec3_sse4(SB), 1, $0
-	MOVQ SP, AX
-	MOVLPD(8, rAX, rX0)	/* MOVLPD a+0(FP), X0 */
-	MOVHPD(16, rAX, rX0)	/* MOVHPD a+8(FP), X0 */
-	MOVLPD(40, rAX, rX1)	/* MOVLPD b+32(FP), X1 */
-	MOVHPD(48, rAX, rX1)	/* MOVHPD b+40(FP), X1 */
-	DPPD(rX1, rX0)		/* DPPD $0x31, X1, X0 */
-	MOVSD one(SB), X1
-	MOVHPD(24, rAX, rX0)	/* MOVHPD a+16(FP), X0 */
-	MOVHPD(56, rAX, rX1)	/* MOVHPD b+48(FP), X1 */
+	MOVUPD a+0(FP), X0
+	MOVUPD b+32(FP), X1
 	DPPD(rX1, rX0)		/* DPPD $0x31, X1, X0 */
+	MOVSD a+16(FP), X1
+	MULSD b+48(FP), X1
+	ADDSD X1, X0
 	RET
 
 TEXT dotvec3_avx(SB), 1, $0
 	MOVQ SP, AX
-	VMOVUPD_128mr(8, rAX, rX0)	/* VMOVUPD a+0(FP), X0 */
-	VMOVUPD_128mr(40, rAX, rX1)	/* VMOVUPD b+32(FP), X1 */
+	ADDQ $8, AX
+	VMOVUPD_128mr(rAX, rX0)
+	ADDQ $32, AX
+	VMOVUPD_128mr(rAX, rX1)
 	VDPPD(rX1, rX0, rX0)		/* VDPPD $0x31, X1, X0, X0 */
 	MOVSD a+16(FP), X1
 	MOVSD b+48(FP), X2
 	VFMADD231SD(rX1, rX2, rX0)
+	VZEROUPPER
 	RET
 
 TEXT Pt2b(SB), 1, $0
@@ -53,21 +54,18 @@ TEXT Pt2b(SB), 1, $0
 	RET
 
 TEXT hsubpd(SB), 1, $0
-	MOVQ SP, AX
-	MOVLPD(8, rAX, rX0)	/* MOVLPD a+0(FP), X0 */
-	MOVHPD(16, rAX, rX0)	/* MOVHPD b+8(FP), X0 */
+	MOVLPD a+0(FP), X0
+	MOVHPD b+8(FP), X0
 	HSUBPD(rX0, rX0)	/* HSUBPD X0, X0 */
 	RET
 
 TEXT crossvec3_sse(SB), 1, $0
-	MOVQ SP, AX
-	ADDQ $8, AX
-	MOVLPD(40, rAX, rX0)	/* MOVLPD b+32(FP), X0 */
-	MOVHPD(8, rAX, rX0)	/* MOVHPD a+0(FP), X0 */
-	MOVLPD(16, rAX, rX1)	/* MOVLPD a+8(FP), X1 */
-	MOVHPD(48, rAX, rX1)	/* MOVHPD b+40(FP), X1 */
-	MOVLPD(56, rAX, rX2)	/* MOVLPD b+48(FP), X2 */
-	MOVHPD(24, rAX, rX2)	/* MOVHPD a+16(FP), X2 */
+	MOVLPD b+40(FP), X0
+	MOVHPD a+8(FP), X0	/* X0 := [a.x][b.x] */
+	MOVLPD a+16(FP), X1
+	MOVHPD b+48(FP), X1	/* X1 := [b.y][a.y] */
+	MOVLPD b+56(FP), X2
+	MOVHPD a+24(FP), X2	/* X2 := [a.z][b.z] */
 	MOVAPD X1, X3
 	MULPD X2, X3
 	HSUBPD(rX3, rX3)	/* x */
@@ -99,21 +97,38 @@ TEXT fma(SB), 1, $0
 	VFMADD231SD(rX1, rX2, rX0)
 	RET
 
+TEXT addpt2_sse(SB), 1, $0
+	MOVUPD a+8(FP), X0
+	MOVUPD b+32(FP), X1
+	ADDPD X1, X0
+	MOVSD a+24(FP), X2
+	ADDSD b+48(FP), X2
+	MOVQ BP, DI
+	MOVUPD X0, (DI)
+	MOVSD X2, 16(DI)
+	RET
+
 /* TODO: write only 24 bytes */
 TEXT addpt2_avx(SB), 1, $0
 	MOVQ SP, AX
-	ADDQ $8, AX
-	VMOVDQU_256mr(8, rAX, rX0)
-	VMOVDQU_256mr(32, rAX, rX1)
+	ADDQ $16, AX
+	VMOVUPD_256mr(rAX, rX0)
+	ADDQ $24, AX
+	VMOVUPD_256mr(rAX, rX1)
 	VADDPD_256rr(rX1, rX0, rX0)
-	VMOVDQU_256rm(rX0, rAX)
+	MOVQ BP, DI
+	VMOVUPD_256rm(rX0, rDI)
+	VZEROUPPER
 	RET
 
 TEXT addpt3_avx(SB), 1, $0
 	MOVQ SP, AX
-	ADDQ $8, AX
-	VMOVDQU_256mr(8, rAX, rX0)
-	VMOVDQU_256mr(40, rAX, rX1)
+	ADDQ $16, AX
+	VMOVUPD_256mr(rAX, rX0)
+	ADDQ $32, AX
+	VMOVUPD_256mr(rAX, rX1)
 	VADDPD_256rr(rX1, rX0, rX0)
-	VMOVDQU_256rm(rX0, rAX)
+	MOVQ BP, DI
+	VMOVUPD_256rm(rX0, rDI)
+	VZEROUPPER
 	RET
diff --git a/main.c b/main.c
index 02fc11c..51a2f89 100644
--- a/main.c
+++ b/main.c
@@ -11,6 +11,7 @@ Point2 Pt2b(double, double, double);
 Point3 crossvec3_sse(Point3, Point3);
 double hsubpd(double, double);
 double fma(double, double, double);
+Point2 addpt2_sse(Point2, Point2);
 Point2 addpt2_avx(Point2, Point2);
 Point3 addpt3_avx(Point3, Point3);
 
@@ -40,8 +41,10 @@ main(int argc, char *argv[])
 	a = strtod(argv[0], nil);
 	b = strtod(argv[1], nil);
 
+	r = 0;
 	r = fmin(a, b);
 	print("fmin(%g, %g) = %g\n", a, b, r);
+	r = 0;
 	r = min(a, b);
 	print("min(%g, %g) = %g\n", a, b, r);
 
@@ -49,10 +52,13 @@ main(int argc, char *argv[])
 
 	p0 = Pt2b(a, 1, 1);
 	p1 = Pt2b(b, 3, 1);
-	r = dotvec2_sse4(p0, p1);
-	print("dotvec2_sse4(%v, %v) = %g\n", p0, p1, r);
+	r = 0;
 	r = dotvec2(p0, p1);
 	print("dotvec2(%v, %v) = %g\n", p0, p1, r);
+	r = 0;
+	r = dotvec2_sse4(p0, p1);
+	print("dotvec2_sse4(%v, %v) = %g\n", p0, p1, r);
+	r = 0;
 	r = dotvec2_avx(p0, p1);
 	print("dotvec2_avx(%v, %v) = %g\n", p0, p1, r);
 
@@ -60,15 +66,19 @@ main(int argc, char *argv[])
 
 	p0t = Pt3(a, 1, 9, 1);
 	p1t = Pt3(b, 3, 4, 1);
-	r = dotvec3_sse4(p0t, p1t);
-	print("dotvec3_sse4(%V, %V) = %g\n", p0t, p1t, r);
+	r = 0;
 	r = dotvec3(p0t, p1t);
 	print("dotvec3(%V, %V) = %g\n", p0t, p1t, r);
+	r = 0;
+	r = dotvec3_sse4(p0t, p1t);
+	print("dotvec3_sse4(%V, %V) = %g\n", p0t, p1t, r);
+	r = 0;
 	r = dotvec3_avx(p0t, p1t);
 	print("dotvec3_avx(%V, %V) = %g\n", p0t, p1t, r);
 
 	print("\n");
 
+	r = 0;
 	r = hsubpd(a, b);
 	print("hsubpd(%g, %g) = %g\n", a, b, r);
 
@@ -76,15 +86,19 @@ main(int argc, char *argv[])
 
 	p0t = Pt3(a, 1, 9, 1);
 	p1t = Pt3(b, 3, 4, 1);
-	prt = crossvec3_sse(p0t, p1t);
-	print("crossvec3_sse(%V, %V) = %V\n", p0t, p1t, prt);
+	prt = Vec3(0,0,0);
 	prt = crossvec3(p0t, p1t);
 	print("crossvec3(%V, %V) = %V\n", p0t, p1t, prt);
+	prt = Vec3(0,0,0);
+	prt = crossvec3_sse(p0t, p1t);
+	print("crossvec3_sse(%V, %V) = %V\n", p0t, p1t, prt);
 
 	print("\n");
 
+	r = 0;
 	r = madd(a, b, 21);
 	print("madd(%g, %g, 21) = %g\n", a, b, r);
+	r = 0;
 	r = fma(a, b, 21);
 	print("fma(%g, %g, 21) = %g\n", a, b, r);
 
@@ -92,8 +106,13 @@ main(int argc, char *argv[])
 
 	p0 = Pt2b(a, 1, 1);
 	p1 = Pt2b(b, 3, 1);
+	pr = Vec2(0,0);
 	pr = addpt2(p0, p1);
 	print("addpt2(%v, %v) = %v\n", p0, p1, pr);
+	pr = Vec2(0,0);
+	pr = addpt2_sse(p0, p1);
+	print("addpt2_sse(%v, %v) = %v\n", p0, p1, pr);
+	pr = Vec2(0,0);
 	pr = addpt2_avx(p0, p1);
 	print("addpt2_avx(%v, %v) = %v\n", p0, p1, pr);
 
@@ -101,8 +120,10 @@ main(int argc, char *argv[])
 
 	p0t = Pt3(a, 1, 1, b);
 	p1t = Pt3(b, 3, 1, a);
+	prt = Vec3(0,0,0);
 	prt = addpt3(p0t, p1t);
 	print("addpt3(%V, %V) = %V\n", p0t, p1t, prt);
+	prt = Vec3(0,0,0);
 	prt = addpt3_avx(p0t, p1t);
 	print("addpt3_avx(%V, %V) = %V\n", p0t, p1t, prt);
 
diff --git a/sse.h b/sse.h
index cd98f86..0a711d5 100644
--- a/sse.h
+++ b/sse.h
@@ -23,13 +23,13 @@
 //opcode = 660F12
 //modrm  = 01 000 000 [AX → X0] / 01 001 000 [AX → X1]
 //disp8 = 8 / 32
-#define MOVLPD(off, s, d) OPi(0x12, 0x1, (d), (s), (off))
+//#define MOVLPD(off, s, d) OPi(0x12, 0x1, (d), (s), (off))
 
 /* MOVHPD */
 //opcode = 660F16
 //modrm  = 01 000 000 [AX → X0] / 01 001 000 [AX → X1]
 //disp8 = 16 / 40
-#define MOVHPD(off, s, d) OPi(0x16, 0x1, (d), (s), (off))
+//#define MOVHPD(off, s, d) OPi(0x16, 0x1, (d), (s), (off))
 
 /* HSUBPD */
 //opcode = 660F7D = 01100110 00001111 01111101