optimization bug with -fexpensive-optimizations

Godmar Back gback@cs.utah.edu
Sun Dec 5 22:07:00 GMT 1999


The following program demonstrates a bug where "-O -fexpensive-optimization"
wrongly removes a needed zero-extending "movzbl" instruction after a 
byte-setting "setne" instruction.  I suspect that the instruction is
removed in combine.c:try_combine(), but I cannot say whether the problem
is there or in the Java front-end.  Maybe the Java front-end is
feeding wrong information about the sizes of its basic types to
the optimizer (?).

To reproduce the problem, simply compile and run the attached program.
I've included the relevant assembly output.  This is with a current CVS
on a Linux box.

	- Godmar

// bug.java
public class bug {
    public static void main(String av[]) {

    /* This is miscompiled under "-O -fexpensive-optimizations" on x86;
     * both from source and from class file.

  0: new #30=<Class java.lang.Integer>
  3: dup
  4: iload_0
  5: ifeq 12
  8: iconst_1
  9: goto 13
 12: iconst_0
 13: invokespecial #33=<Method java.lang.Integer.<init> (int)void>
 16: areturn

        pushl   %ebp
        movl    %esp, %ebp
        subl    $16, %esp
        pushl   %esi
        pushl   %ebx
        movb    8(%ebp), %bl
        subl    $12, %esp
        pushl   $_CL_3bug
        call    _Jv_InitClass
        subl    $8, %esp
        pushl   $12
        pushl   $_CL_Q34java4lang7Integer
        call    _Jv_AllocObject
        movl    %eax, %esi
        addl    $32, %esp
        testb   %bl, %bl
        setne   %al
### 	higher-order bits of %eax are non-zero
### 	we need a movzbl  %al, %eax instruction
###	but expensive-optimization removes it.  
###	Bug in combine.c?
        subl    $8, %esp
        pushl   %eax
        pushl   %esi
        call    __Q34java4lang7Integeri
        movl    %esi, %eax
        leal    -24(%ebp), %esp
        popl    %ebx
        popl    %esi
        movl    %ebp, %esp
        popl    %ebp

    static Integer get(boolean f) { return new Integer(f ? 1 : 0); }

More information about the Gcc-bugs mailing list