Date:	Tue, 8 Nov 94 03:25:43 JST
From:	Stephen Turnbull <turnbull AT shako DOT sk DOT tsukuba DOT ac DOT jp>
To:	cspt AT ludens DOT elte DOT hu
Cc:	djgpp AT sun DOT soe DOT clarkson DOT edu
Subject:	2.4/3-2.4/3!=0

   A simple comparison changed the value of a variable (x) when I

No floating point operation is "simple."  Not even assignment (in an
environment of high precision floating point hardware).

   compiled the program with -O1. When I compiled a program with these

With optimization, the program uses common subexpression elimination
and does a FPU register to CPU register load for the value of x.  This
is faster than the memory reference to address &x, thus preferred to
the memory reference.  The compiler assumes that the value it just
stored into that memory location is unchanged.  This doesn't mean it
stored the same value as the source (can't be---FPU registers are 10
bytes, doubles are only 8 bytes); it means that the result of
repeating the store operation would be the same.

   operations without optimizations (or with optimizations, but
   declaring all double variables volatile), 1-x was always >=0. When

The compiler may not optimize away references to volatile memory locations.
The compiler may also not use associativity to change the order of
operations if the value of the operation is assigned to a variable,
even if that variable is later optimized into nonexistence.

   I optimized it, there was this bug, which caused the program to
   terminate with a floating point exception when it tried to
   calculate sqrt(1-x). I hardly found this bug, because when I
   compared 1-x to 0, (if(1-x<0) printf("akurvaeletbe\n"); or
   something like that before sqrt), the program did not hang (the
   comparison corrected the value of x, so it became 0).

	   So my question is:
   Is it possible to compile this program with -O1 (and -ffloat-store), but 
   without this bug? 
   If it is, which optimization flag must I switch off after -O1?

I forget what GCC calls it; look for "common subexpression" or "very
busy expression" in the man page.  Switch that one off, it might work.

- Raw text -