- Improved the precision of (Deg)ree division.

2 files changed, 65 insertions, 12 deletions

diff --git a/luxc/src/lux/compiler/js/rt.clj b/luxc/src/lux/compiler/js/rt.clj
index f17998871..d5f503f01 100644
--- a/luxc/src/lux/compiler/js/rt.clj
+++ b/luxc/src/lux/compiler/js/rt.clj
@@ -618,8 +618,24 @@
                  
                  "return LuxRT$addI64(top,bottomAndMiddle);"
                  "})")
+   "countLeadingZeroes" (str "(function LuxRT$countLeadingZeroes(input) {"
+                             "var zeroes = 64;"
+                             (str "while(!LuxRT$eqI64(input,LuxRT$ZERO)) {"
+                                  "zeroes--;"
+                                  "input = LuxRT$ushrI64(input,1);"
+                                  "}")
+                             "return zeroes;"
+                             "})")
    "divD64" (str "(function LuxRT$divD64(l,r) {"
-                 "return LuxRT$shlI64(LuxRT$divI64(l,LuxRT$fromNumberI64(r.H)),32);"
+                 (str "if(LuxRT$eqI64(l,r)) {"
+                      "return LuxRT$negateI64(LuxRT$ONE);" ;; ~= 1.0 DEG
+                      "}"
+                      "else {"
+                      "var minShift = Math.min(LuxRT$countLeadingZeroes(l), LuxRT$countLeadingZeroes(r));"
+                      "l = LuxRT$shlI64(l,minShift);"
+                      "r = LuxRT$shlI64(r,minShift);"
+                      "return LuxRT$shlI64(LuxRT$divI64(l,LuxRT$fromNumberI64(r.H)),32);"
+                      "}")
                  "})")
    "degToReal" (str "(function LuxRT$degToReal(input) {"
                     "var two32 = Math.pow(2,32);"
diff --git a/luxc/src/lux/compiler/jvm/rt.clj b/luxc/src/lux/compiler/jvm/rt.clj
index a1b171a25..aa0c881d0 100644
--- a/luxc/src/lux/compiler/jvm/rt.clj
+++ b/luxc/src/lux/compiler/jvm/rt.clj
@@ -368,17 +368,54 @@
              (.visitInsn Opcodes/LRETURN)
              (.visitMaxs 0 0)
              (.visitEnd))
-         _ (doto (.visitMethod =class (+ Opcodes/ACC_PUBLIC Opcodes/ACC_STATIC) "div_deg" "(JJ)J" nil nil)
-             (.visitCode)
-             ;; Based on: http://stackoverflow.com/a/8510587/6823464
-             (.visitVarInsn Opcodes/LLOAD 0)
-             (.visitVarInsn Opcodes/LLOAD 2) high-4b
-             (.visitInsn Opcodes/LDIV)
-             (.visitLdcInsn (int 32))
-             (.visitInsn Opcodes/LSHL)
-             (.visitInsn Opcodes/LRETURN)
-             (.visitMaxs 0 0)
-             (.visitEnd))
+         _ (let [$loop-start (new Label)
+                 $done (new Label)]
+             (doto (.visitMethod =class (+ Opcodes/ACC_PUBLIC Opcodes/ACC_STATIC) "count_leading_zeros" "(J)I" nil nil)
+               (.visitCode)
+               (.visitLdcInsn (int 64))
+               (.visitLabel $loop-start)
+               (.visitVarInsn Opcodes/LLOAD 0)
+               (.visitLdcInsn (long 0))
+               (.visitInsn Opcodes/LCMP)
+               (.visitJumpInsn Opcodes/IFEQ $done)
+               (.visitVarInsn Opcodes/LLOAD 0)
+               (.visitLdcInsn (int 1))
+               (.visitInsn Opcodes/LUSHR)
+               (.visitVarInsn Opcodes/LSTORE 0)
+               (.visitLdcInsn (int 1))
+               (.visitInsn Opcodes/ISUB)
+               (.visitJumpInsn Opcodes/GOTO $loop-start)
+               (.visitLabel $done)
+               (.visitInsn Opcodes/IRETURN)
+               (.visitMaxs 0 0)
+               (.visitEnd)))
+         _ (let [$same (new Label)]
+             (doto (.visitMethod =class (+ Opcodes/ACC_PUBLIC Opcodes/ACC_STATIC) "div_deg" "(JJ)J" nil nil)
+               (.visitCode)
+               (.visitVarInsn Opcodes/LLOAD 0)
+               (.visitVarInsn Opcodes/LLOAD 2)
+               (.visitInsn Opcodes/LCMP)
+               (.visitJumpInsn Opcodes/IFEQ $same)
+               ;; Based on: http://stackoverflow.com/a/8510587/6823464
+               ;; Shifting the operands as much as possible can help
+               ;; avoid some loss of precision later.
+               (.visitVarInsn Opcodes/LLOAD 0)
+               (.visitMethodInsn Opcodes/INVOKESTATIC "lux/LuxRT" "count_leading_zeros" "(J)I")
+               (.visitVarInsn Opcodes/LLOAD 2)
+               (.visitMethodInsn Opcodes/INVOKESTATIC "lux/LuxRT" "count_leading_zeros" "(J)I")
+               (.visitMethodInsn Opcodes/INVOKESTATIC "java/lang/Math" "min" "(II)I")
+               (.visitVarInsn Opcodes/ISTORE 4)
+               (.visitVarInsn Opcodes/LLOAD 0) (.visitVarInsn Opcodes/ILOAD 4) (.visitInsn Opcodes/LSHL)
+               (.visitVarInsn Opcodes/LLOAD 2) (.visitVarInsn Opcodes/ILOAD 4) (.visitInsn Opcodes/LSHL) high-4b
+               (.visitInsn Opcodes/LDIV)
+               (.visitLdcInsn (int 32))
+               (.visitInsn Opcodes/LSHL)
+               (.visitInsn Opcodes/LRETURN)
+               (.visitLabel $same)
+               (.visitLdcInsn (long -1)) ;; ~= 1.0 DEG
+               (.visitInsn Opcodes/LRETURN)
+               (.visitMaxs 0 0)
+               (.visitEnd)))
          _ (doto (.visitMethod =class (+ Opcodes/ACC_PUBLIC Opcodes/ACC_STATIC) "deg-to-real" "(J)D" nil nil)
              (.visitCode)
              ;; Translate high bytes