adding intrinsic to compute iround efficiently

adding itrunc for explicit truncating conversion to integer preparation for fixing issue #170
JuliaLang · Aug 12, 2011 · 2e826e4 · 2e826e4
1 parent 07b1273
commit 2e826e4
Show file tree

Hide file tree

Showing 5 changed files with 45 additions and 1 deletion.
diff --git a/Makefile b/Makefile
@@ -21,6 +21,7 @@ sys.ji: VERSION j/sysimg.j j/start_image.j src/boot.j src/dump.c j/*.j
 clean:
 	rm -f julia
 	rm -f *~ *#
+	rm -f sys.ji
 	$(MAKE) -C j clean
 	$(MAKE) -C src clean
 	$(MAKE) -C ui clean

diff --git a/j/float.j b/j/float.j
@@ -1,5 +1,11 @@
 ## floating point conversions ##
 
+iround(x::Float32) = boxsi32(fpiround32(unbox32(x)))
+iround(x::Float64) = boxsi64(fpiround64(unbox64(x)))
+
+itrunc(x::Float32) = boxsi32(fptosi32(unbox32(x)))
+itrunc(x::Float64) = boxsi64(fptosi64(unbox64(x)))
+
 convert(::Type{Float32}, x::Bool)    = boxf32(sitofp32(unbox8(x)))
 convert(::Type{Float32}, x::Int8)    = boxf32(sitofp32(unbox8(x)))
 convert(::Type{Float32}, x::Int16)   = boxf32(sitofp32(unbox16(x)))

diff --git a/j/floatfuncs.j b/j/floatfuncs.j
@@ -43,3 +43,6 @@ function hex2num(s)
     end
     return boxf64(unbox64(parse_int(Int64, s, 16)))
 end
+
+@vectorize_1arg Real iround
+@vectorize_1arg Real itrunc
diff --git a/j/math_libm.j b/j/math_libm.j
@@ -72,7 +72,7 @@ end
 @libmfunc_1arg_float Real   round
 
 #@libmfunc_1arg_int Real lrint
-@libmfunc_1arg_int Real lround iround
+#@libmfunc_1arg_int Real lround iround
 @libmfunc_1arg_int Real ilogb
 
 @libfdmfunc_2arg Number atan2

diff --git a/src/intrinsics.cpp b/src/intrinsics.cpp
@@ -25,6 +25,7 @@ namespace JL_I {
         trunc8, trunc16, trunc32, trunc64, trunc_int,
         fptoui8, fptosi8, fptoui16, fptosi16, fptoui32, fptosi32,
         fptoui64, fptosi64, 
+        fpiround32, fpiround64,
         uitofp32, sitofp32, uitofp64, sitofp64,
         fptrunc32, fpext64,
         // functions
@@ -957,6 +958,38 @@ static Value *emit_intrinsic(intrinsic f, jl_value_t **args, size_t nargs,
         return builder.CreateFPToUI(FP(x), T_int64);
     HANDLE(fptosi64,1)
         return builder.CreateFPToSI(FP(x), T_int64);
+    HANDLE(fpiround32,1)
+    {
+        Value *bits = builder.CreateBitCast(x, T_int32);
+        Value *half = builder.CreateBitCast(ConstantFP::get(T_float32, 0.5),
+                                            T_int32);
+        Value *signedhalf =
+            builder.CreateOr(half,
+                             builder.CreateAnd(bits,
+                                               ConstantInt::get(T_int32,
+                                                                BIT31)));
+        return builder.
+            CreateFPToSI(builder.CreateFAdd(x,
+                                            builder.CreateBitCast(signedhalf,
+                                                                  T_float32)),
+                         T_int32);
+    }
+    HANDLE(fpiround64,1)
+    {
+        Value *bits = builder.CreateBitCast(x, T_int64);
+        Value *half = builder.CreateBitCast(ConstantFP::get(T_float64, 0.5),
+                                            T_int64);
+        Value *signedhalf =
+            builder.CreateOr(half,
+                             builder.CreateAnd(bits,
+                                               ConstantInt::get(T_int64,
+                                                                BIT63)));
+        return builder.
+            CreateFPToSI(builder.CreateFAdd(x,
+                                            builder.CreateBitCast(signedhalf,
+                                                                  T_float64)),
+                         T_int64);
+    }
     HANDLE(uitofp32,1)
         return builder.CreateUIToFP(x, T_float32);
     HANDLE(sitofp32,1)
@@ -1081,6 +1114,7 @@ extern "C" void jl_init_intrinsic_functions()
     ADD_I(fptoui8); ADD_I(fptosi8);
     ADD_I(fptoui16); ADD_I(fptosi16); ADD_I(fptoui32); ADD_I(fptosi32);
     ADD_I(fptoui64); ADD_I(fptosi64);
+    ADD_I(fpiround32); ADD_I(fpiround64);
     ADD_I(uitofp32); ADD_I(sitofp32); ADD_I(uitofp64); ADD_I(sitofp64);
     ADD_I(fptrunc32); ADD_I(fpext64);
     ADD_I(sqrt_float); ADD_I(powi_float); ADD_I(pow_float);