Ludovic Courtès wrote 5 years ago
(address . bug-guile@gnu.org)
87zh8pncst.fsf@inria.fr
Hello,
In https://issues.guix.gnu.org/42047, janneke found that
cross-compiling an ‘eqv?’ form from x86_64 to i686 leads to an incorrect
‘eqv?’ reduction (this was not the case in 3.0.2). Reproducer:
(lambda (i)
(eqv? i #x6478696f))
Then cross-compile from x86_64 like this:
guild compile -O3 --target=i586-pc-gnu -o bad.go reproducer.scm
and compile natively on a 32-bit platform with:
guild compile -O3 -o good.go reproducer.scm
If you disassemble good.go and bad.go, the diff goes like this:
--- /tmp/good.asm 2020-06-26 15:56:35.671024432 +0200
+++ /tmp/bad.asm 2020-06-26 15:56:35.651024548 +0200
@@ -1,55 +1,39 @@
Disassembly of <unnamed function> at #x98:
- 0 (instrument-entry 16348) at eqv-reduction-bug.scm:1:0
+ 0 (instrument-entry 16348) at /home/ludo/src/guile-debugging/eqv-reduction-bug.scm:1:0
2 (assert-nargs-ee/locals 1 0) ;; 1 slot (0 args)
- 3 (make-non-immediate 0 16343) ;; #<procedure a4f000 at eqv-reduction-bug.scm:1:0 (i)>
+ 3 (make-non-immediate 0 16343) ;; #<procedure a4f000 at /home/ludo/src/guile-debugging/eqv-reduction-bug.scm:1:?>
5 (handle-interrupts)
6 (return-values)
Disassembly of <unnamed function> at #xb4:
- 0 (instrument-entry 16351) at eqv-reduction-bug.scm:1:0
+ 0 (instrument-entry 16351) at /home/ludo/src/guile-debugging/eqv-reduction-bug.scm:1:0
2 (assert-nargs-ee/locals 2 0) ;; 2 slots (1 arg)
3 (static-ref 1 16346) ;; 1685612911
- 5 (eq? 0 1) at eqv-reduction-bug.scm:2:2
- 6 (je 21) ;; -> L3
- 7 (immediate-tag=? 0 7 0) ;; heap-object?
- 9 (jne 14) ;; -> L2
- 10 (heap-tag=? 0 127 23) ;; heap-number?
- 12 (jne 11) ;; -> L2
- 13 (heap-numbers-equal? 0 1)
- 14 (jne 5) ;; -> L1
- 15 (make-short-immediate 1 1028) ;; #t
- 16 (reset-frame 1) ;; 1 slot
- 17 (handle-interrupts)
- 18 (return-values)
+ 5 (eq? 0 1) at /home/ludo/src/guile-debugging/eqv-reduction-bug.scm:2:2
+ 6 (jne 5) ;; -> L1
+ 7 (make-short-immediate 1 1028) ;; #t
+ 8 (reset-frame 1) ;; 1 slot
+ 9 (handle-interrupts)
+ 10 (return-values)
L1:
- 19 (make-short-immediate 1 4) ;; #f
- 20 (reset-frame 1) ;; 1 slot
- 21 (handle-interrupts)
- 22 (return-values)
-L2:
- 23 (make-short-immediate 1 4) ;; #f
- 24 (reset-frame 1) ;; 1 slot
- 25 (handle-interrupts)
- 26 (return-values)
-L3:
- 27 (make-short-immediate 1 1028) ;; #t
- 28 (reset-frame 1) ;; 1 slot
- 29 (handle-interrupts)
- 30 (return-values)
+ 11 (make-short-immediate 1 4) ;; #f
+ 12 (reset-frame 1) ;; 1 slot
+ 13 (handle-interrupts)
+ 14 (return-values)
IOW, bad.go reduced ‘eqv?’ to ‘eq?’.
This might come from 7df3f3414bfb19a1bd7fbe29bc30a1ab28bf4319, perhaps
with an issue of the time at which ‘target-most-positive-fixnum’ is
called, similar to what was fixed in
a0b9d866380b04aff27dcbcf1e13051f3d9685ad.
To be continued…
Ludo’.