doc/guide/src/CombineConversions.adoc

   1 CombineConversions
   2 ==================
   3
   4 <:CombineConversions:> is an optimization pass for the <:SSA:>
   5 <:IntermediateLanguage:>, invoked from <:SSASimplify:>.
   6
   7 == Description ==
   8
   9 This pass looks for and simplifies nested calls to (signed)
  10 extension/truncation.
  11
  12 == Implementation ==
  13
  14 * <!ViewGitFile(mlton,master,mlton/ssa/combine-conversions.fun)>
  15
  16 == Details and Notes ==
  17
  18 It processes each block in dfs order (visiting definitions before uses):
  19
  20 * If the statement is not a `PrimApp` with `Word_extdToWord`, skip it.
  21 * After processing a conversion, it tags the `Var` for subsequent use.
  22 * When inspecting a conversion, check if the `Var` operand is also the
  23 result of a conversion. If it is, try to combine the two operations.
  24 Repeatedly simplify until hitting either a non-conversion `Var` or a
  25 case where the conversion cannot be simplified.
  26
  27 The optimization rules are very simple:
  28 ----
  29 x1 = ...
  30 x2 = Word_extdToWord (W1, W2, {signed=s1}) x1
  31 x3 = Word_extdToWord (W2, W3, {signed=s2}) x2
  32 ----
  33
  34 * If `W1 = W2`, then there is no conversions before `x_1`.
  35 +
  36 This is guaranteed because `W2 = W3` will always trigger optimization.
  37
  38 * Case `W1 <= W3 <= W2`:
  39 +
  40 ----
  41 x3 = Word_extdToWord (W1, W3, {signed=s1}) x1
  42 ----
  43
  44 * Case `W1 <  W2 <  W3  AND  ((NOT s1) OR s2)`:
  45 +
  46 ----
  47 x3 = Word_extdToWord (W1, W3, {signed=s1}) x1
  48 ----
  49
  50 * Case `W1 =  W2 <  W3`:
  51 +
  52 unoptimized, because there are no conversions past `W1` and `x2 = x1`
  53
  54 * Case `W3 <= W2 <= W1  OR  W3 <= W1 <= W2`:
  55 +
  56 ----
  57 x_3 = Word_extdToWord (W1, W3, {signed=_}) x1
  58 ----
  59 +
  60 because `W3 <= W1 && W3 <= W2`, just clip `x1`
  61
  62 * Case `W2 < W1 <= W3  OR  W2 < W3 <= W1`:
  63 +
  64 unoptimized, because `W2 < W1 && W2 < W3`, has truncation effect
  65
  66 * Case `W1 < W2 < W3  AND  (s1 AND (NOT s2))`:
  67 +
  68 unoptimized, because each conversion affects the result separately