Question

There is a part of code:

*det_ptr++ = (float)(dx*dy - 0.81*dxy*dxy);

where dx, dy, and dxy are floats.

Apple LLVM 3.0 compiler makes the following assembly for it:

+0x250  vldr.32                        s0, [r13, #+140]
+0x254  vldr.32                        s1, [r13, #+136]
+0x258  vmul.f32                       s0, s0, s1
+0x25c  vcvt.f64.f32                   d16, s0 <-------------- cast from float to double
+0x260  vldr.32                        s0, [r13, #+132]
+0x264  vcvt.f64.f32                   d17, s0 <-------------- cast from float to double
+0x268  vldr.64                        d18, [r13, #+16]
+0x26c  vmul.f64                       d17, d18, d17
+0x270  vldr.32                        s0, [r13, #+132]
+0x274  vcvt.f64.f32                   d19, s0 <-------------- cast from float to double
+0x278  vmul.f64                       d17, d17, d19
+0x27c  vsub.f64                       d16, d16, d17
+0x280  vcvt.f32.f64                   s0, d16
+0x284  ldr                            r0, [sp, #+104]
+0x286  adds                           r1, r0, #4  ; 0x4
+0x288  str                            r1, [sp, #+104]
+0x28a  vstr.32                        s0, [r0]

Is there any way to forbid these casts?

Answer 1

The way in which you wrote your program requires those casts. 0.81 is a double-precision literal, so dxy must be promoted to double before the multiplication takes place, and dx*dy must be promoted before the subtraction. The fact that you cast the final result back to float doesn’t matter–the C standard is perfectly clear that those terms are evaluated in double-precision regardless.

To prevent the promotion to double, use a single-precision literal instead (by adding the f suffix):

*det_ptr++ = dx*dy - 0.81f*dxy*dxy;