diff --git a/TODO b/TODO
index d1a555bf..3c5156ea 100644
--- a/TODO
+++ b/TODO
@@ -1,3 +1,28 @@
+- large and puzzling performance regression cf. 7.18 on the benchmark
+
+  7.18
+	$ time ./vips.py wtc_tiled_small.tif wtc2.tif
+	vips warning: im_conv: 6534 overflows and 13439 underflows detected
+	real	0m1.891s
+	user	0m3.110s
+	sys	0m0.320s
+
+  7.20 (and 7.22)
+
+	$ time ./vips.py wtc_tiled_small.tif wtc2.tif
+	vips warning: im_conv: 6502 overflows and 13429 underflows detected
+	real	0m2.962s
+	user	0m4.950s
+	sys	0m0.250s
+
+  and why are the conv numbers different?
+
+  seems to be a difference in the bilinear interpolator? we did increase
+  accuracy in 7.18 -> 7.20
+
+  I suppose conv must have changed too?
+
+
 
 - use D65 in cmsCreateLab4Profile() ? not sure
 
diff --git a/libvips/resample/interpolate.c b/libvips/resample/interpolate.c
index 9fe25d7e..adabae99 100644
--- a/libvips/resample/interpolate.c
+++ b/libvips/resample/interpolate.c
@@ -409,10 +409,11 @@ vips_interpolate_bilinear_interpolate( VipsInterpolate *interpolate,
 	const int tx = (six + 1) >> 1;
 	const int ty = (siy + 1) >> 1;
 
-	/* We know x/y are always positive, so we can just (int) them. 
+	/* We want ((int)x) ... save this double -> int conversion by just
+	 * shifting sx down.
 	 */
-	const int ix = (int) x;
-	const int iy = (int) y;
+	const int ix = sx >> (VIPS_TRANSFORM_SHIFT + 1);
+	const int iy = sy >> (VIPS_TRANSFORM_SHIFT + 1);
 
 	const PEL *p1 = (PEL *) IM_REGION_ADDR( in, ix, iy );
 	const PEL *p2 = p1 + ps;