Opened 5 years ago

Closed 5 years ago

#1624 closed defect (fixed)

Flonum unboxing too agressive

Reported by: sjamaan Owned by:
Priority: critical Milestone: 5.2
Component: compiler Version: 5.1.0
Keywords: lfa2 Cc:
Estimated difficulty: medium

Description (last modified by sjamaan)

Found by Sven Hartrumpf. The following program will segfault:

(define (test-it)
  (let loop ((done #f))
    (if done
        (cons  (or (read) 0.0) (loop #t)))))

(print (test-it))

When run with a DEBUGBUILD CHICKEN, you'll see it fails with a nice panic due to a low-level assertion error triggered by a C_flonum_magnitude call when you enter a fixnum:

[panic] Low-level type assertion C_blockp((C_word)C_VAL1(C__PREV_TMPST.n1))=#t failed at test.c:176 - execution terminated

Generated code failing:

/* k150 in loop in test-it in k131 in k128 in k125 */
static void C_ccall f_152(C_word c,C_word *av){
C_word tmp;
C_word t0=av[0];
C_word t1=av[1];
C_word t2;
C_word t3;
C_word t4;
double f0;
C_word *a;
C_save_and_reclaim((void *)f_152,2,av);}
f0=C_flonum_magnitude((C_truep(t1)?t1:lf[1])); /// <--- offending code
C_trace(C_text("test.scm:5: loop"));

Most likely introduced by 79cf7427638eebf695f2be54731bb6bbf4d0fff2.

ALSO: In the generated code in this case, it unboxes and immediately boxes the flonum. There's no operation in between, so that's wasteful.

It only makes sense with >1 operations (though with exactly 1 it should be okay too but adds no perf improvement)

Change History (4)

comment:1 Changed 5 years ago by felix winkelmann

Estimated difficulty: hardmedium

The offending code seems to be lfa2.scm:349, the rule for walking if/##core#cond nodes. The result type is always taken from the 2nd branch, but should instead be the union of the result types of both branches.

comment:2 Changed 5 years ago by sjamaan

Description: modified (diff)

comment:3 Changed 5 years ago by felix winkelmann

AFAICT, the transformation in this pass doesn't create redundant boxing. I think, in this particular example the pointless boxing/unboxing is an artifact of the completely wrong type computed by walking the conditional.

comment:4 Changed 5 years ago by sjamaan

Resolution: fixed
Status: newclosed

Fixed by 9413c4bf1b4b0c95e60ba12e082f4e03b4c0f8ef

Note: See TracTickets for help on using tickets.