diff --git a/.gitignore b/.gitignore
index 8cf92bc5..73f59c0d 100644
--- a/.gitignore
+++ b/.gitignore
@@ -12,6 +12,7 @@ Makefile.in
 TAGS
 tags
 *.o
+*.vect
 Vips-8.0.gir
 Vips-8.0.typelib
 .*.swp
diff --git a/ChangeLog b/ChangeLog
index 2c68dfc4..43f75ba7 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -25,6 +25,7 @@
 - added vips_gaussblur() convenience function
 - added --vips-profile, records and dumps thread timing info 
 - added vipsprofile, visualises --vips-profile output
+- auto-vectorization-friendly inner loops
 
 20/11/13	started 7.36.5
 - better cache sizing in unbuffered sequential mode
diff --git a/TODO b/TODO
index 3b339cb2..7b238beb 100644
--- a/TODO
+++ b/TODO
@@ -1,3 +1,18 @@
+- check vectorizer on linear.c
+
+  do some more packages, we've just done arithmetic so far
+
+  time add with auto vec and with orc, is orc worthwhile?
+
+  make sure __restrict__ turns off if the compiler does not support it, is
+  there a configure thing?
+
+  how much would alignment buy us? is there any way we can do this? probably
+  not, since we need to be able to generate any sub-area
+
+  how about avg? do we need -ffast-math to vec that?
+
+
 - seen some leaks from 
 
 	vips dzsave --layout google wtc.jpg x
diff --git a/libvips/arithmetic/abs.c b/libvips/arithmetic/abs.c
index 8b268463..1faf999f 100644
--- a/libvips/arithmetic/abs.c
+++ b/libvips/arithmetic/abs.c
@@ -1,4 +1,4 @@
-/* im_abs()
+/* absolute value
  *
  * Copyright: 1990, N. Dessipris, based on im_powtra()
  * Author: Nicos Dessipris
@@ -93,25 +93,19 @@ vips_abs_build( VipsObject *object )
 /* Integer abs operation: just test and negate.
  */
 #define ABS_INT( TYPE ) { \
-	TYPE *p = (TYPE *) in[0]; \
-	TYPE *q = (TYPE *) out; \
+	TYPE * __restrict__ p = (TYPE *) in[0]; \
+	TYPE * __restrict__ q = (TYPE *) out; \
 	int x; \
 	\
-	for( x = 0; x < sz; x++ ) { \
-		TYPE v = p[x]; \
-		\
-		if( v < 0 ) \
-			q[x] = 0 - v; \
-		else \
-			q[x] = v; \
-	} \
+	for( x = 0; x < sz; x++ ) \
+		q[x] = p[x] < 0 ? 0 - p[x] : p[x]; \
 }
 
 /* Float abs operation: call fabs().
  */
 #define ABS_FLOAT( TYPE ) { \
-	TYPE *p = (TYPE *) in[0]; \
-	TYPE *q = (TYPE *) out; \
+	TYPE * __restrict__ p = (TYPE *) in[0]; \
+	TYPE * __restrict__ q = (TYPE *) out; \
 	int x; \
 	\
 	for( x = 0; x < sz; x++ ) \
@@ -124,8 +118,8 @@ vips_abs_build( VipsObject *object )
 #ifdef HAVE_HYPOT
 
 #define ABS_COMPLEX( TYPE ) { \
-	TYPE *p = (TYPE *) in[0]; \
-	TYPE *q = (TYPE *) out; \
+	TYPE * __restrict__ p = (TYPE *) in[0]; \
+	TYPE * __restrict__ q = (TYPE *) out; \
 	int x; \
 	\
 	for( x = 0; x < sz; x++ ) { \
@@ -137,8 +131,8 @@ vips_abs_build( VipsObject *object )
 #else /*HAVE_HYPOT*/
 
 #define ABS_COMPLEX( TYPE ) { \
-	TYPE *p = (TYPE *) in[0]; \
-	TYPE *q = (TYPE *) out; \
+	TYPE * __restrict__ p = (TYPE *) in[0]; \
+	TYPE * __restrict__ q = (TYPE *) out; \
 	int x; \
 	\
 	for( x = 0; x < sz; x++ ) { \
diff --git a/libvips/arithmetic/add.c b/libvips/arithmetic/add.c
index 30a13ea4..fcee7eb5 100644
--- a/libvips/arithmetic/add.c
+++ b/libvips/arithmetic/add.c
@@ -88,9 +88,9 @@ typedef VipsBinaryClass VipsAddClass;
 G_DEFINE_TYPE( VipsAdd, vips_add, VIPS_TYPE_BINARY );
 
 #define LOOP( IN, OUT ) { \
-	IN *left = (IN *) in[0]; \
-	IN *right = (IN *) in[1]; \
-	OUT *q = (OUT *) out; \
+	IN * __restrict__ left = (IN *) in[0]; \
+	IN * __restrict__ right = (IN *) in[1]; \
+	OUT * __restrict__ q = (OUT *) out; \
 	\
 	for( x = 0; x < sz; x++ ) \
 		q[x] = left[x] + right[x]; \
diff --git a/libvips/arithmetic/boolean.c b/libvips/arithmetic/boolean.c
index 9c2434cf..9f038ca5 100644
--- a/libvips/arithmetic/boolean.c
+++ b/libvips/arithmetic/boolean.c
@@ -108,18 +108,18 @@ vips_boolean_build( VipsObject *object )
 }
 
 #define LOOP( TYPE, OP ) { \
-	TYPE *left = (TYPE *) in[0]; \
-	TYPE *right = (TYPE *) in[1]; \
-	TYPE *q = (TYPE *) out; \
+	TYPE * __restrict__ left = (TYPE *) in[0]; \
+	TYPE * __restrict__ right = (TYPE *) in[1]; \
+	TYPE * __restrict__ q = (TYPE *) out; \
 	\
 	for( x = 0; x < sz; x++ ) \
 		q[x] = left[x] OP right[x]; \
 }
 
 #define FLOOP( TYPE, OP ) { \
-	TYPE *left = (TYPE *) in[0]; \
-	TYPE *right = (TYPE *) in[1]; \
-	int *q = (int *) out; \
+	TYPE * __restrict__ left = (TYPE *) in[0]; \
+	TYPE * __restrict__ right = (TYPE *) in[1]; \
+	int * __restrict__ q = (int *) out; \
 	\
 	for( x = 0; x < sz; x++ ) \
 		q[x] = ((int) left[x]) OP ((int) right[x]); \
diff --git a/libvips/arithmetic/complex.c b/libvips/arithmetic/complex.c
index d610ad68..93095cb6 100644
--- a/libvips/arithmetic/complex.c
+++ b/libvips/arithmetic/complex.c
@@ -80,8 +80,8 @@ typedef VipsUnaryClass VipsComplexClass;
 G_DEFINE_TYPE( VipsComplex, vips_complex, VIPS_TYPE_UNARY );
 
 #define LOOP( IN, OUT, OP ) { \
-	IN *p = (IN *) in[0]; \
-	OUT *q = (OUT *) out; \
+	IN * __restrict__ p = (IN *) in[0]; \
+	OUT * __restrict__ q = (OUT *) out; \
 	\
 	for( x = 0; x < sz; x++ ) { \
 		OP( q, p[x], 0.0 ); \
@@ -91,8 +91,8 @@ G_DEFINE_TYPE( VipsComplex, vips_complex, VIPS_TYPE_UNARY );
 }
 
 #define CLOOP( IN, OUT, OP ) { \
-	IN *p = (IN *) in[0]; \
-	OUT *q = (OUT *) out; \
+	IN * __restrict__ p = (IN *) in[0]; \
+	OUT * __restrict__ q = (OUT *) out; \
 	\
 	for( x = 0; x < sz; x++ ) { \
 		OP( q, p[0], p[1] ); \
diff --git a/libvips/arithmetic/divide.c b/libvips/arithmetic/divide.c
index a82c5718..cb9edb65 100644
--- a/libvips/arithmetic/divide.c
+++ b/libvips/arithmetic/divide.c
@@ -90,9 +90,9 @@ G_DEFINE_TYPE( VipsDivide, vips_divide, VIPS_TYPE_BINARY );
 
 /* This is going to be much slower */
 #define CLOOP( TYPE ) { \
-	TYPE *left = (TYPE *) in[0]; \
-	TYPE *right = (TYPE *) in[1]; \
-	TYPE *q = (TYPE *) out; \
+	TYPE * __restrict__ left = (TYPE *) in[0]; \
+	TYPE * __restrict__ right = (TYPE *) in[1]; \
+	TYPE * __restrict__ q = (TYPE *) out; \
 	int i; \
         \
 	for( i = 0; i < sz; i++ ) { \
@@ -120,9 +120,9 @@ G_DEFINE_TYPE( VipsDivide, vips_divide, VIPS_TYPE_BINARY );
 #else /* USE_MODARG_DIV */
 
 #define CLOOP( TYPE ) {                                     \
-	TYPE *left = (TYPE *) in[0]; \
-	TYPE *right = (TYPE *) in[1]; \
-	TYPE *q = (TYPE *) out; \
+	TYPE * __restrict__ left = (TYPE *) in[0]; \
+	TYPE * __restrict__ right = (TYPE *) in[1]; \
+	TYPE * __restrict__ q = (TYPE *) out; \
 	int i; \
         \
 	for( i = 0; i < sz; i++ ) { \
@@ -157,15 +157,12 @@ G_DEFINE_TYPE( VipsDivide, vips_divide, VIPS_TYPE_BINARY );
 /* Real divide. Cast in to OUT before divide so we work for float output.
  */
 #define RLOOP( IN, OUT ) { \
-	IN *left = (IN *) in[0]; \
-	IN *right = (IN *) in[1]; \
-	OUT *q = (OUT *) out; \
+	IN * __restrict__ left = (IN *) in[0]; \
+	IN * __restrict__ right = (IN *) in[1]; \
+	OUT * __restrict__ q = (OUT *) out; \
 	\
 	for( x = 0; x < sz; x++ ) \
-		if( right[x] == 0 ) \
-			q[x] = 0; \
-		else \
-			q[x] = (OUT) left[x] / (OUT) right[x]; \
+		q[x] = right[x] == 0 ? q[x] : (OUT) left[x] / (OUT) right[x]; \
 }
 
 static void
diff --git a/libvips/arithmetic/invert.c b/libvips/arithmetic/invert.c
index 87a66397..d9bac71d 100644
--- a/libvips/arithmetic/invert.c
+++ b/libvips/arithmetic/invert.c
@@ -69,24 +69,24 @@ typedef VipsUnaryClass VipsInvertClass;
 G_DEFINE_TYPE( VipsInvert, vips_invert, VIPS_TYPE_UNARY );
 
 #define LOOP( TYPE, L ) { \
-	TYPE *p = (TYPE *) in[0]; \
-	TYPE *q = (TYPE *) out; \
+	TYPE * __restrict__ p = (TYPE *) in[0]; \
+	TYPE * __restrict__ q = (TYPE *) out; \
 	\
 	for( x = 0; x < sz; x++ ) \
 		q[x] = (L) - p[x]; \
 }
 
 #define LOOPN( TYPE ) { \
-	TYPE *p = (TYPE *) in[0]; \
-	TYPE *q = (TYPE *) out; \
+	TYPE * __restrict__ p = (TYPE *) in[0]; \
+	TYPE * __restrict__ q = (TYPE *) out; \
 	\
 	for( x = 0; x < sz; x++ ) \
 		q[x] = -1 * p[x]; \
 }
 
 #define LOOPC( TYPE ) { \
-	TYPE *p = (TYPE *) in[0]; \
-	TYPE *q = (TYPE *) out; \
+	TYPE * __restrict__ p = (TYPE *) in[0]; \
+	TYPE * __restrict__ q = (TYPE *) out; \
 	\
 	for( x = 0; x < sz; x++ ) { \
 		q[0] = -1 * p[0]; \
diff --git a/libvips/arithmetic/linear.c b/libvips/arithmetic/linear.c
index 98408b51..061f6120 100644
--- a/libvips/arithmetic/linear.c
+++ b/libvips/arithmetic/linear.c
@@ -163,8 +163,8 @@ vips_linear_build( VipsObject *object )
 /* Non-complex input, any output.
  */
 #define LOOPN( IN, OUT ) { \
-	IN *p = (IN *) in[0]; \
-	OUT *q = (OUT *) out; \
+	IN * __restrict__ p = (IN *) in[0]; \
+	OUT * __restrict__ q = (OUT *) out; \
 	\
 	for( i = 0, x = 0; x < width; x++ ) \
 		for( k = 0; k < nb; k++, i++ ) \
@@ -174,8 +174,8 @@ vips_linear_build( VipsObject *object )
 /* Complex input, complex output. 
  */
 #define LOOPCMPLXN( IN, OUT ) { \
-	IN *p = (IN *) in[0]; \
-	OUT *q = (OUT *) out; \
+	IN * __restrict__ p = (IN *) in[0]; \
+	OUT * __restrict__ q = (OUT *) out; \
 	\
 	for( x = 0; x < width; x++ ) \
 		for( k = 0; k < nb; k++ ) { \
diff --git a/libvips/arithmetic/math.c b/libvips/arithmetic/math.c
index e9182d46..1b649802 100644
--- a/libvips/arithmetic/math.c
+++ b/libvips/arithmetic/math.c
@@ -97,8 +97,8 @@ vips_math_build( VipsObject *object )
 }
 
 #define LOOP( IN, OUT, OP ) { \
-	IN *p = (IN *) in[0]; \
-	OUT *q = (OUT *) out; \
+	IN * __restrict__ p = (IN *) in[0]; \
+	OUT * __restrict__ q = (OUT *) out; \
 	\
 	for( x = 0; x < sz; x++ ) \
 		q[x] = OP( p[x] ); \
diff --git a/libvips/arithmetic/math2.c b/libvips/arithmetic/math2.c
index 16d57af3..36c8410a 100644
--- a/libvips/arithmetic/math2.c
+++ b/libvips/arithmetic/math2.c
@@ -102,9 +102,9 @@ vips_math2_build( VipsObject *object )
 }
 
 #define LOOP( IN, OUT, OP ) { \
-	IN *p1 = (IN *) in[0]; \
-	IN *p2 = (IN *) in[1]; \
-	OUT *q = (OUT *) out; \
+	IN * __restrict__ p1 = (IN *) in[0]; \
+	IN * __restrict__ p2 = (IN *) in[1]; \
+	OUT * __restrict__ q = (OUT *) out; \
 	\
 	for( x = 0; x < sz; x++ ) \
 		OP( q[x], p1[x], p2[x] ); \
@@ -137,12 +137,9 @@ vips_math2_build( VipsObject *object )
 	double left = (double) (X); \
 	double right = (double) (E); \
 	\
-	if( left == 0.0 && right < 0.0 ) \
-		/* Division by zero! Difficult to report tho' \
-		 */ \
-		(Y) = 0.0; \
-	else \
-		(Y) = pow( left, right ); \
+	/* Division by zero! Difficult to report tho' \
+	 */ \
+	(Y) = (left == 0.0 && right < 0.0) ? 0.0 : pow( left, right ); \
 }
 
 #define WOP( Y, X, E ) POW( Y, E, X )
@@ -355,9 +352,9 @@ vips_math2_const_build( VipsObject *object )
 }
 
 #define LOOPC( IN, OUT, OP ) { \
-	IN *p = (IN *) in[0]; \
-	OUT *q = (OUT *) out; \
-	double *c = (double *) uconst->c_ready; \
+	IN * __restrict__ p = (IN *) in[0]; \
+	OUT * __restrict__ q = (OUT *) out; \
+	double * __restrict__ c = (double *) uconst->c_ready; \
 	\
 	for( i = 0, x = 0; x < width; x++ ) \
 		for( b = 0; b < bands; b++, i++ ) \
diff --git a/libvips/arithmetic/multiply.c b/libvips/arithmetic/multiply.c
index 5e2fd8f3..7e40d86d 100644
--- a/libvips/arithmetic/multiply.c
+++ b/libvips/arithmetic/multiply.c
@@ -81,9 +81,9 @@ G_DEFINE_TYPE( VipsMultiply, vips_multiply, VIPS_TYPE_BINARY );
 /* Complex multiply.
  */
 #define CLOOP( TYPE ) { \
-	TYPE *left = (TYPE *) in[0]; \
-	TYPE *right = (TYPE *) in[1]; \
-	TYPE *q = (TYPE *) out; \
+	TYPE * __restrict__ left = (TYPE *) in[0]; \
+	TYPE * __restrict__ right = (TYPE *) in[1]; \
+	TYPE * __restrict__ q = (TYPE *) out; \
 	\
 	for( x = 0; x < sz; x++ ) { \
 		double x1 = left[0]; \
@@ -104,9 +104,9 @@ G_DEFINE_TYPE( VipsMultiply, vips_multiply, VIPS_TYPE_BINARY );
 /* Real multiply.
  */
 #define RLOOP( IN, OUT ) { \
-	IN *left = (IN *) in[0]; \
-	IN *right = (IN *) in[1]; \
-	OUT *q = (OUT *) out; \
+	IN * __restrict__ left = (IN *) in[0]; \
+	IN * __restrict__ right = (IN *) in[1]; \
+	OUT * __restrict__ q = (OUT *) out; \
 	\
 	for( x = 0; x < sz; x++ ) \
 		q[x] = left[x] * right[x]; \
diff --git a/libvips/arithmetic/relational.c b/libvips/arithmetic/relational.c
index a6064f2a..80b0d46a 100644
--- a/libvips/arithmetic/relational.c
+++ b/libvips/arithmetic/relational.c
@@ -110,18 +110,18 @@ vips_relational_build( VipsObject *object )
 }
 
 #define RLOOP( TYPE, ROP ) { \
-	TYPE *left = (TYPE *) in[0]; \
-	TYPE *right = (TYPE *) in[1]; \
-	VipsPel *q = (VipsPel *) out; \
+	TYPE * __restrict__ left = (TYPE *) in[0]; \
+	TYPE * __restrict__ right = (TYPE *) in[1]; \
+	VipsPel * __restrict__ q = (VipsPel *) out; \
 	\
 	for( x = 0; x < sz; x++ ) \
 		q[x] = (left[x] ROP right[x]) ? 255 : 0; \
 }
 
 #define CLOOP( TYPE, COP ) { \
-	TYPE *left = (TYPE *) in[0]; \
-	TYPE *right = (TYPE *) in[1]; \
-	VipsPel *q = (VipsPel *) out; \
+	TYPE * __restrict__ left = (TYPE *) in[0]; \
+	TYPE * __restrict__ right = (TYPE *) in[1]; \
+	VipsPel * __restrict__ q = (VipsPel *) out; \
 	\
 	for( x = 0; x < sz; x++ ) { \
 		q[x] = COP( left[0], left[1], right[0], right[1]) ? 255 : 0; \
@@ -478,8 +478,8 @@ vips_relational_const_build( VipsObject *object )
 }
 
 #define RLOOPC( TYPE, OP ) { \
-	TYPE *p = (TYPE *) in[0]; \
-	TYPE *c = (TYPE *) uconst->c_ready; \
+	TYPE * __restrict__ p = (TYPE *) in[0]; \
+	TYPE * __restrict__ c = (TYPE *) uconst->c_ready; \
  	\
 	for( i = 0, x = 0; x < width; x++ ) \
 		for( b = 0; b < bands; b++, i++ ) \
@@ -487,10 +487,10 @@ vips_relational_const_build( VipsObject *object )
 }
 
 #define CLOOPC( TYPE, OP ) { \
-	TYPE *p = (TYPE *) in[0]; \
+	TYPE * __restrict__ p = (TYPE *) in[0]; \
  	\
 	for( i = 0, x = 0; x < width; x++ ) { \
-		TYPE *c = (TYPE *) uconst->c_ready; \
+		TYPE * __restrict__ c = (TYPE *) uconst->c_ready; \
 		\
 		for( b = 0; b < bands; b++, i++ ) { \
 			out[i] = OP( p[0], p[1], c[0], c[1]) ? 255 : 0; \
diff --git a/libvips/arithmetic/remainder.c b/libvips/arithmetic/remainder.c
index f5ad0cfa..b53881fb 100644
--- a/libvips/arithmetic/remainder.c
+++ b/libvips/arithmetic/remainder.c
@@ -92,32 +92,26 @@ vips_remainder_build( VipsObject *object )
 /* Integer remainder-after-division.
  */
 #define IREMAINDER( TYPE ) { \
-	TYPE *p1 = (TYPE *) in[0]; \
-	TYPE *p2 = (TYPE *) in[1]; \
-	TYPE *q = (TYPE *) out; \
+	TYPE * __restrict__ p1 = (TYPE *) in[0]; \
+	TYPE * __restrict__ p2 = (TYPE *) in[1]; \
+	TYPE * __restrict__ q = (TYPE *) out; \
 	\
 	for( x = 0; x < sz; x++ ) \
-		if( p2[x] ) \
-			q[x] = p1[x] % p2[x]; \
-		else \
-			q[x] = -1; \
+		q[x] = p2[x] ? p1[x] % p2[x] : -1; \
 }
 
 /* Float remainder-after-division.
  */
 #define FREMAINDER( TYPE ) { \
-	TYPE *p1 = (TYPE *) in[0]; \
-	TYPE *p2 = (TYPE *) in[1]; \
-	TYPE *q = (TYPE *) out; \
+	TYPE * __restrict__ p1 = (TYPE *) in[0]; \
+	TYPE * __restrict__ p2 = (TYPE *) in[1]; \
+	TYPE * __restrict__ q = (TYPE *) out; \
 	\
 	for( x = 0; x < sz; x++ ) { \
 		double a = p1[x]; \
 		double b = p2[x]; \
 		\
-		if( b ) \
-			q[x] = a - b * floor (a / b); \
-		else \
-			q[x] = -1; \
+		q[x] = b ? a - b * floor (a / b) : -1; \
 	} \
 }
 
@@ -262,9 +256,9 @@ vips_remainder_const_build( VipsObject *object )
 /* Integer remainder-after-divide, per-band constant.
  */
 #define IREMAINDERCONST( TYPE ) { \
-	TYPE *p = (TYPE *) in[0]; \
-	TYPE *q = (TYPE *) out; \
-	TYPE *c = (TYPE *) uconst->c_ready; \
+	TYPE * __restrict__ p = (TYPE *) in[0]; \
+	TYPE * __restrict__ q = (TYPE *) out; \
+	TYPE * __restrict__ c = (TYPE *) uconst->c_ready; \
 	\
 	for( i = 0, x = 0; x < width; x++ ) \
 		for( b = 0; b < bands; b++, i++ ) \
@@ -274,19 +268,18 @@ vips_remainder_const_build( VipsObject *object )
 /* Float remainder-after-divide, per-band constant.
  */
 #define FREMAINDERCONST( TYPE ) { \
-	TYPE *p = (TYPE *) in[0]; \
-	TYPE *q = (TYPE *) out; \
-	TYPE *c = (TYPE *) uconst->c_ready; \
+	TYPE * __restrict__ p = (TYPE *) in[0]; \
+	TYPE * __restrict__ q = (TYPE *) out; \
+	TYPE * __restrict__ c = (TYPE *) uconst->c_ready; \
 	\
 	for( i = 0, x = 0; x < width; x++ ) \
 		for( b = 0; b < bands; b++, i++ ) { \
 			double left = p[i]; \
 			double right = c[b]; \
 			\
-			if( right ) \
-				q[i] = left - right * floor( left / right ); \
-			else \
-				q[i] = -1; \
+			q[i] = right ? \
+				left - right * floor( left / right ) : \
+				-1; \
 		} \
 }
 
diff --git a/libvips/arithmetic/round.c b/libvips/arithmetic/round.c
index 5a951f4a..602c19e2 100644
--- a/libvips/arithmetic/round.c
+++ b/libvips/arithmetic/round.c
@@ -85,8 +85,8 @@ vips_round_build( VipsObject *object )
 }
 
 #define LOOP( TYPE, OP ) { \
-	TYPE *p = (TYPE *) in[0]; \
-	TYPE *q = (TYPE *) out; \
+	TYPE * __restrict__ p = (TYPE *) in[0]; \
+	TYPE * __restrict__ q = (TYPE *) out; \
 	\
 	for( x = 0; x < sz; x++ ) \
 		q[x] = OP( p[x] ); \
diff --git a/libvips/arithmetic/sign.c b/libvips/arithmetic/sign.c
index 1476dff8..69f11e8e 100644
--- a/libvips/arithmetic/sign.c
+++ b/libvips/arithmetic/sign.c
@@ -58,8 +58,8 @@ typedef VipsUnaryClass VipsSignClass;
 G_DEFINE_TYPE( VipsSign, vips_sign, VIPS_TYPE_UNARY );
 
 #define CSIGN( TYPE ) { \
-	TYPE *p = (TYPE *) in[0]; \
-	TYPE *q = (TYPE *) out; \
+	TYPE * __restrict__ p = (TYPE *) in[0]; \
+	TYPE * __restrict__ q = (TYPE *) out; \
 	int x; \
 	\
 	for( x = 0; x < sz; x++ ) { \
@@ -83,8 +83,8 @@ G_DEFINE_TYPE( VipsSign, vips_sign, VIPS_TYPE_UNARY );
 }
 
 #define SIGN( TYPE ) { \
-	TYPE *p = (TYPE *) in[0]; \
-	signed char *q = (signed char *) out; \
+	TYPE * __restrict__ p = (TYPE *) in[0]; \
+	signed char * __restrict__ q = (signed char *) out; \
 	int x; \
 	\
 	for( x = 0; x < sz; x++ ) { \
diff --git a/libvips/arithmetic/subtract.c b/libvips/arithmetic/subtract.c
index a5f6223d..d7fece09 100644
--- a/libvips/arithmetic/subtract.c
+++ b/libvips/arithmetic/subtract.c
@@ -83,9 +83,9 @@ typedef VipsBinaryClass VipsSubtractClass;
 G_DEFINE_TYPE( VipsSubtract, vips_subtract, VIPS_TYPE_BINARY );
 
 #define LOOP( IN, OUT ) { \
-	IN *left = (IN *) in[0]; \
-	IN *right = (IN *) in[1]; \
-	OUT *q = (OUT *) out; \
+	IN * __restrict__ left = (IN *) in[0]; \
+	IN * __restrict__ right = (IN *) in[1]; \
+	OUT * __restrict__ q = (OUT *) out; \
 	\
 	for( x = 0; x < sz; x++ ) \
 		q[x] = left[x] - right[x]; \
diff --git a/libvips/arithmetic/unaryconst.c b/libvips/arithmetic/unaryconst.c
index 89a1f672..4740d374 100644
--- a/libvips/arithmetic/unaryconst.c
+++ b/libvips/arithmetic/unaryconst.c
@@ -52,7 +52,7 @@ G_DEFINE_ABSTRACT_TYPE( VipsUnaryConst, vips_unary_const, VIPS_TYPE_UNARY );
 /* Cast a vector of double to a vector of TYPE, clipping to a range.
  */
 #define CAST_CLIP( TYPE, N, X ) { \
-	TYPE *tq = (TYPE *) q; \
+	TYPE * __restrict__ tq = (TYPE *) q; \
 	\
 	for( i = 0; i < m; i++ ) { \
 		double v = p[VIPS_MIN( n - 1, i )]; \
@@ -64,7 +64,7 @@ G_DEFINE_ABSTRACT_TYPE( VipsUnaryConst, vips_unary_const, VIPS_TYPE_UNARY );
 /* Cast a vector of double to a vector of TYPE.
  */
 #define CAST( TYPE ) { \
-	TYPE *tq = (TYPE *) q; \
+	TYPE * __restrict__ tq = (TYPE *) q; \
 	\
 	for( i = 0; i < m; i++ ) \
 		tq[i] = (TYPE) p[VIPS_MIN( n - 1, i )]; \
@@ -73,7 +73,7 @@ G_DEFINE_ABSTRACT_TYPE( VipsUnaryConst, vips_unary_const, VIPS_TYPE_UNARY );
 /* Cast a vector of double to a complex vector of TYPE.
  */
 #define CASTC( TYPE ) { \
-	TYPE *tq = (TYPE *) q; \
+	TYPE * __restrict__ tq = (TYPE *) q; \
 	\
 	for( i = 0; i < m; i++ ) { \
 		tq[0] = (TYPE) p[VIPS_MIN( n - 1, i )]; \
@@ -86,7 +86,8 @@ G_DEFINE_ABSTRACT_TYPE( VipsUnaryConst, vips_unary_const, VIPS_TYPE_UNARY );
 /* Cast a n-band vector of double to a m-band vector in another format.
  */
 static VipsPel *
-make_pixel( VipsObject *obj, int m, VipsBandFmt fmt, int n, double *p )
+make_pixel( VipsObject *obj, 
+	int m, VipsBandFmt fmt, int n, double * __restrict__ p )
 {
 	VipsPel *q;
 	int i;