Add LLVM intrinsics for floor/ceil/trunc/abs.

Add tests for vector trunc, round, floor, ceil.
Add fast algorithm for round.

base/float.jl

@@ -168,8 +168,11 @@ round {T<:Integer}(::Type{T}, x::FloatingPoint) = trunc(T,round(x))
 
 
 # this is needed very early because it is used by Range and colon
-round(x::Float64) = ccall((:round, Base.libm_name), Float64, (Float64,), x)
-floor(x::Float64) = ccall((:floor, Base.libm_name), Float64, (Float64,), x)
+function round(x::Float64)
+ y = trunc(x)
+ ifelse(x==y,y,trunc(2.0*x-y))
+end
+floor(x::Float64) = box(Float64,floor_llvm(unbox(Float64,x)))
 
 ## floating point promotions ##
 promote_rule(::Type{Float32}, ::Type{Float16}) = Float32

base/math.jl

@@ -25,7 +25,7 @@ import Base: log, exp, sin, cos, tan, sinh, cosh, tanh, asin,
  max, min, minmax, ceil, floor, trunc, round, ^, exp2,
  exp10, expm1, log1p
 
-import Core.Intrinsics: nan_dom_err, sqrt_llvm, box, unbox, powi_llvm
+import Core.Intrinsics: nan_dom_err, ceil_llvm, floor_llvm, trunc_llvm, sqrt_llvm, box, unbox, powi_llvm
 
 # non-type specific math functions
 
@@ -132,18 +132,29 @@ sqrt(x::Float32) = box(Float32,sqrt_llvm(unbox(Float32,x)))
 sqrt(x::Real) = sqrt(float(x))
 @vectorize_1arg Number sqrt
 
-for f in (:ceil, :trunc, :significand, :rint) # :nearbyint
+ceil(x::Float64) = box(Float64,ceil_llvm(unbox(Float64,x)))
+ceil(x::Float32) = box(Float32,ceil_llvm(unbox(Float32,x)))
+@vectorize_1arg Real ceil
+
+trunc(x::Float64) = box(Float64,trunc_llvm(unbox(Float64,x)))
+trunc(x::Float32) = box(Float32,trunc_llvm(unbox(Float32,x)))
+@vectorize_1arg Real trunc
+
+for f in (:significand, :rint) # :nearbyint
  @eval begin
  ($f)(x::Float64) = ccall(($(string(f)),libm), Float64, (Float64,), x)
  ($f)(x::Float32) = ccall(($(string(f,"f")),libm), Float32, (Float32,), x)
  @vectorize_1arg Real $f
  end
 end
 
-round(x::Float32) = ccall((:roundf, libm), Float32, (Float32,), x)
+function round(x::Float32)
+ y = trunc(x)
+ ifelse(x==y,y,trunc(2.f0*x-y))
+end
 @vectorize_1arg Real round
 
-floor(x::Float32) = ccall((:floorf, libm), Float32, (Float32,), x)
+floor(x::Float32) = box(Float32,floor_llvm(unbox(Float32,x)))
 @vectorize_1arg Real floor
 
 hypot(x::Real, y::Real) = hypot(promote(float(x), float(y))...)

src/intrinsics.cpp

@@ -29,7 +29,7 @@ namespace JL_I {
  nan_dom_err,
  // functions
  abs_float, copysign_float, flipsign_int, select_value,
- sqrt_llvm, powi_llvm,
+ ceil_llvm, floor_llvm, trunc_llvm, sqrt_llvm, powi_llvm,
  // pointer access
  pointerref, pointerset, pointertoref,
  // c interface
@@ -1099,12 +1099,18 @@ static Value *emit_intrinsic(intrinsic f, jl_value_t **args, size_t nargs,
  HANDLE(abs_float,1)
  {
  x = FP(x);
+#ifdef LLVM34
+ return builder.CreateCall(Intrinsic::getDeclaration(jl_Module, Intrinsic::fabs,
+ ArrayRef<Type*>(x->getType())),
+ x);
+#else
  Type *intt = JL_INTT(x->getType());
  Value *bits = builder.CreateBitCast(FP(x), intt);
  Value *absbits =
  builder.CreateAnd(bits,
  ConstantInt::get(intt, APInt::getSignedMaxValue(((IntegerType*)intt)->getBitWidth())));
  return builder.CreateBitCast(absbits, x->getType());
+#endif
  }
  HANDLE(copysign_float,2)
  {
@@ -1147,6 +1153,24 @@ static Value *emit_intrinsic(intrinsic f, jl_value_t **args, size_t nargs,
  HANDLE(jl_alloca,1) {
  return builder.CreateAlloca(IntegerType::get(jl_LLVMContext, 8),JL_INT(x));
  }
+ HANDLE(ceil_llvm,1) {
+ x = FP(x);
+ return builder.CreateCall(Intrinsic::getDeclaration(jl_Module, Intrinsic::ceil,
+ ArrayRef<Type*>(x->getType())),
+ x);
+ }
+ HANDLE(floor_llvm,1) {
+ x = FP(x);
+ return builder.CreateCall(Intrinsic::getDeclaration(jl_Module, Intrinsic::floor,
+ ArrayRef<Type*>(x->getType())),
+ x);
+ }
+ HANDLE(trunc_llvm,1) {
+ x = FP(x);
+ return builder.CreateCall(Intrinsic::getDeclaration(jl_Module, Intrinsic::trunc,
+ ArrayRef<Type*>(x->getType())),
+ x);
+ }
  HANDLE(sqrt_llvm,1) {
  x = FP(x);
  raise_exception_unless(builder.CreateFCmpUGE(x, ConstantFP::get(x->getType(),0.0)),
@@ -1245,7 +1269,7 @@ extern "C" void jl_init_intrinsic_functions(void)
  ADD_I(uitofp); ADD_I(sitofp);
  ADD_I(fptrunc); ADD_I(fpext);
  ADD_I(abs_float); ADD_I(copysign_float);
- ADD_I(flipsign_int); ADD_I(select_value); ADD_I(sqrt_llvm);
+ ADD_I(flipsign_int); ADD_I(select_value); ADD_I(ceil_llvm); ADD_I(floor_llvm); ADD_I(trunc_llvm); ADD_I(sqrt_llvm);
  ADD_I(powi_llvm);
  ADD_I(pointerref); ADD_I(pointerset); ADD_I(pointertoref);
  ADD_I(checked_sadd); ADD_I(checked_uadd);

test/numbers.jl

@@ -1330,6 +1330,21 @@ for x = 2^24-10:2^24+10
  @test ceil(Int,y) == i
 end
 
+# rounding vectors
+let ≈(x,y) = x==y && typeof(x)==typeof(y)
+ for t in [Float32,Float64]
+ # try different vector lengths
+ for n in [0,3,255,256]
+ r = (1:n)-div(n,2)
+ y = t[x/4 for x in r]
+ @test trunc(y) ≈ t[div(i,4) for i in r]
+ @test round(y) ≈ t[(i+1+(i>=0))>>2 for i in r]
+ @test floor(y) ≈ t[i>>2 for i in r]
+ @test ceil(y) ≈ t[(i+3)>>2 for i in r]
+ end
+ end
+end
+
 @test_throws InexactError round(Int,Inf)
 @test_throws InexactError round(Int,NaN)
 @test round(Int,2.5) == 3