@@ -878,18 +878,11 @@ fn cast_float_to_int(bcx: &Builder,
878878 // we're rounding towards zero, we just get float_ty::MAX (which is always an integer).
879879 // This already happens today with u128::MAX = 2^128 - 1 > f32::MAX.
880880 fn compute_clamp_bounds < F : Float > ( signed : bool , int_ty : Type ) -> ( u128 , u128 ) {
881- let f_min = if signed {
882- let rounded_min = F :: from_i128_r ( int_min ( signed, int_ty) , Round :: TowardZero ) ;
883- assert_eq ! ( rounded_min. status, Status :: OK ) ;
884- rounded_min. value
885- } else {
886- F :: ZERO
887- } ;
888-
881+ let rounded_min = F :: from_i128_r ( int_min ( signed, int_ty) , Round :: TowardZero ) ;
882+ assert_eq ! ( rounded_min. status, Status :: OK ) ;
889883 let rounded_max = F :: from_u128_r ( int_max ( signed, int_ty) , Round :: TowardZero ) ;
890884 assert ! ( rounded_max. value. is_finite( ) ) ;
891-
892- ( f_min. to_bits ( ) , rounded_max. value . to_bits ( ) )
885+ ( rounded_min. value . to_bits ( ) , rounded_max. value . to_bits ( ) )
893886 }
894887 fn int_max ( signed : bool , int_ty : Type ) -> u128 {
895888 let shift_amount = 128 - int_ty. int_width ( ) ;
@@ -906,11 +899,6 @@ fn cast_float_to_int(bcx: &Builder,
906899 0
907900 }
908901 }
909- let ( f_min, f_max) = match float_ty. float_width ( ) {
910- 32 => compute_clamp_bounds :: < ieee:: Single > ( signed, int_ty) ,
911- 64 => compute_clamp_bounds :: < ieee:: Double > ( signed, int_ty) ,
912- n => bug ! ( "unsupported float width {}" , n) ,
913- } ;
914902 let float_bits_to_llval = |bits| {
915903 let bits_llval = match float_ty. float_width ( ) {
916904 32 => C_u32 ( bcx. ccx , bits as u32 ) ,
@@ -919,6 +907,11 @@ fn cast_float_to_int(bcx: &Builder,
919907 } ;
920908 consts:: bitcast ( bits_llval, float_ty)
921909 } ;
910+ let ( f_min, f_max) = match float_ty. float_width ( ) {
911+ 32 => compute_clamp_bounds :: < ieee:: Single > ( signed, int_ty) ,
912+ 64 => compute_clamp_bounds :: < ieee:: Double > ( signed, int_ty) ,
913+ n => bug ! ( "unsupported float width {}" , n) ,
914+ } ;
922915 let f_min = float_bits_to_llval ( f_min) ;
923916 let f_max = float_bits_to_llval ( f_max) ;
924917 // To implement saturation, we perform the following steps:
@@ -935,45 +928,46 @@ fn cast_float_to_int(bcx: &Builder,
935928 // undef does not introduce any non-determinism either.
936929 // More importantly, the above procedure correctly implements saturating conversion.
937930 // Proof (sketch):
938- // If x is NaN, 0 is trivially returned.
931+ // If x is NaN, 0 is returned by definition .
939932 // Otherwise, x is finite or infinite and thus can be compared with f_min and f_max.
940933 // This yields three cases to consider:
941934 // (1) if x in [f_min, f_max], the result of fpto[su]i is returned, which agrees with
942935 // saturating conversion for inputs in that range.
943936 // (2) if x > f_max, then x is larger than int_ty::MAX. This holds even if f_max is rounded
944937 // (i.e., if f_max < int_ty::MAX) because in those cases, nextUp(f_max) is already larger
945- // than int_ty::MAX. Because x is larger than int_ty::MAX, the return value is correct.
938+ // than int_ty::MAX. Because x is larger than int_ty::MAX, the return value of int_ty::MAX
939+ // is correct.
946940 // (3) if x < f_min, then x is smaller than int_ty::MIN. As shown earlier, f_min exactly equals
947- // int_ty::MIN and therefore the return value of int_ty::MIN is immediately correct.
941+ // int_ty::MIN and therefore the return value of int_ty::MIN is correct.
948942 // QED.
949943
950944 // Step 1 was already performed above.
951945
952- // Step 2: We use two comparisons and two selects, with s1 being the result:
953- // %less = fcmp ult %x, %f_min
946+ // Step 2: We use two comparisons and two selects, with % s1 being the result:
947+ // %less_or_nan = fcmp ult %x, %f_min
954948 // %greater = fcmp olt %x, %f_max
955- // %s0 = select %less , int_ty::MIN, %fptosi_result
949+ // %s0 = select %less_or_nan , int_ty::MIN, %fptosi_result
956950 // %s1 = select %greater, int_ty::MAX, %s0
957- // Note that %less uses an *unordered* comparison. This comparison is true if the operands are
958- // not comparable (i.e., if x is NaN). The unordered comparison ensures that s1 becomes
959- // int_ty::MIN if x is NaN.
960- // Performance note: It can be lowered to a flipped comparison and a negation (and the negation
961- // can be merged into the select), so it not necessarily any more expensive than a ordered
962- // ("normal") comparison. Whether these optimizations will be performed is ultimately up to the
963- // backend but at least x86 does that .
964- let less = bcx. fcmp ( llvm:: RealULT , x, f_min) ;
951+ // Note that %less_or_nan uses an *unordered* comparison. This comparison is true if the
952+ // operands are not comparable (i.e., if x is NaN). The unordered comparison ensures that s1
953+ // becomes int_ty::MIN if x is NaN.
954+ // Performance note: Unordered comparison can be lowered to a " flipped" comparison and a
955+ // negation, and the negation can be merged into the select. Therefore, it not necessarily any
956+ // more expensive than a ordered ("normal") comparison. Whether these optimizations will be
957+ // performed is ultimately up to the backend, but at least x86 does perform them .
958+ let less_or_nan = bcx. fcmp ( llvm:: RealULT , x, f_min) ;
965959 let greater = bcx. fcmp ( llvm:: RealOGT , x, f_max) ;
966- let int_max = C_big_integral ( int_ty, int_max ( signed, int_ty) as u128 ) ;
960+ let int_max = C_big_integral ( int_ty, int_max ( signed, int_ty) ) ;
967961 let int_min = C_big_integral ( int_ty, int_min ( signed, int_ty) as u128 ) ;
968- let s0 = bcx. select ( less , int_min, fptosui_result) ;
962+ let s0 = bcx. select ( less_or_nan , int_min, fptosui_result) ;
969963 let s1 = bcx. select ( greater, int_max, s0) ;
970964
971965 // Step 3: NaN replacement.
972966 // For unsigned types, the above step already yielded int_ty::MIN == 0 if x is NaN.
973967 // Therefore we only need to execute this step for signed integer types.
974968 if signed {
975969 // LLVM has no isNaN predicate, so we use (x == x) instead
976- bcx. select ( bcx. fcmp ( llvm:: RealOEQ , x, x) , s1, C_big_integral ( int_ty, 0 ) )
970+ bcx. select ( bcx. fcmp ( llvm:: RealOEQ , x, x) , s1, C_uint ( int_ty, 0 ) )
977971 } else {
978972 s1
979973 }
0 commit comments