Re: Patch to allow Ada to work with tree-ssa

[Andrew Haley <aph at redhat dot com>]

Richard Kenner writes:
 >      > I somebody could point out what tree transformation is going wrong, I'll
 >      > be glad to fix it.
 > 
 >     It seems that something has changed the depth-first, left-to-right
 >     parsing of expressions.  Have you any idea what might have caused this?
 > 
 > "parsing"?  Did you mean gimplifying?

Yes.

 > Indeed a number of things I did might have had the effect of
 > changing the order in which subtrees are gimplified.  Since I
 > didn't think that anything depended on that, I didn't pay that much
 > attention to the issue.  If there are places that do depend on
 > that, can you identify them so that comments can be placed in
 > gimplify.c saying what's order-dependent.

Java assumes that all expressions are fully evaluated in strict
left-to-right order.  

For example, in Java "a += a++ + a" is well-defined.

 > The changes would mostly relate to the order of gimplification of MODIFY_EXPR
 > and certain compound values, I'd guess.
 > 
 > Can you give an example of an expression that's handled differently?

Ranjit Mathew has reported that this breaks:

public class err6 {
  public static void main(String[] args) {
    int[] x = { 10, 11, 12, 1, 14 };
    int[] y = { 1, 2, 3, 4, 5, 6 };

    if ( x[(x=y)[2]] == 1 ) {
      System.out.println("OK");
    } else {
      System.out.println("NG:[1]-->[" +x[(x=y)[2]]+ "]");
    }
  }
}

The rule we're checking is that each dimension expression is fully
evaluated before any part of any dimension expression to its right.

So, this

  x[(x=y)[2]]

is gimplified as

  x1 = x
  x = y
  tmp = x1[2]
  x1[tmp]

I'll provide a more detailed diagnosis tomorrow.  It's quite possible,
of course, that you've revealed a bug in the Java gimplification code.

Andrew.