GNU bug report logs - #78606
31.0.50; [nativecomp] native-ice in emit_mvar_rval

Previous Next

To reply to this bug, email your comments to 78606 AT debbugs.gnu.org.

Toggle the display of automated, internal messages from the tracker.

Message #5 received at submit <at> debbugs.gnu.org (full text, mbox):

From: Pip Cet <pipcet <at> protonmail.com>
To: Andrea Corallo <acorallo <at> gnu.org>, bug-gnu-emacs <at> gnu.org
Subject: 31.0.50; [nativecomp] native-ice in emit_mvar_rval
Date: Tue, 27 May 2025 13:38:02 +0000

This reduced test case:

;; -*- lexical-binding: t; -*-

(defun f (x)
  (and (= x 1)
       (if (eql x 1)
           1
         x)))

causes a native-ice:

Error: native-ice (\"can't find data in relocation containers\")

when Emacs is run as:

./src/emacs -Q --batch --eval '(native-compile "native.el")'

The problem, as far as I can see, is that emit_mvar_rval appears to
assume that if the constraint part of an mvar allows only a single
value, that value can also be fetched from the reloc array.  In this
case, the constraints on x make it so only 1.0 is a legal value in one
basic block, but 1.0 isn't in the data array.

Note that it is not a valid fix to make obj_to_reloc fail gently if it
cannot find obj in either data container: if 1.0 is present in the
ephemeral vector but not in the default vector, referring to the
ephemeral vector after top_level_run has completed will fail.

The problem is this limple insn:

(set #(mvar 17592188503872 1 (member 1.0)) #(mvar 17592188503233 0 (member 1.0)))

The RHS mvar was not created with :constant 1.0 (so 1.0 was never added
to the data container), but it is comp-cstr-imm-vld-p.

The test case is reduced from mule-diag.el:

(defun list-block-of-chars (charset row min max)
  (let (i ch)
    (insert-char ?- (+ 7 (* 4 16)))
    (insert "\n     ")
    (setq i 0)
    (while (< i 16)
      (insert (format "%4X" i))
      (setq i (1+ i)))
    (setq i (* (/ min 16) 16))
    (while (<= i max)
      (if (= (% i 16) 0)
	  (insert (format "\n%6Xx" (/ (+ (* row 256) i) 16))))
      (setq ch (if (< i min)
		   32
		 (or (decode-char charset (+ (* row 256) i))
		     32)))		; gap in mapping
      ;; Don't insert control codes, non-Unicode characters.
      (if (or (< ch 32) (= ch 127))
	  (setq ch (single-key-description ch))
	(if (and (>= ch 128) (< ch 160))
	    (setq ch (format "%02Xh" ch))
	  (if (> ch #x10FFFF)
	      (setq ch 32))))
      (insert "\t" ch)
      (setq i (1+ i))))
  (insert "\n"))

But the original problem in mule-diag.el (generating a basic block which
would call (single-key-description 127.0) ) only showed up with other
modifications to the nativecomp compiler.

There are several ways to fix this, but I must note that adding a
constant that is highly likely to be already present in the register
file is usually wasteful: in this case, we don't want to create the
constant 1.0, we simply want to return the register which we now know to
contain 1.0.

Message #8 received at 78606 <at> debbugs.gnu.org (full text, mbox):

From: Eli Zaretskii <eliz <at> gnu.org>
To: acorallo <at> gnu.org, Pip Cet <pipcet <at> protonmail.com>
Cc: 78606 <at> debbugs.gnu.org
Subject: Re: bug#78606: 31.0.50; [nativecomp] native-ice in emit_mvar_rval
Date: Sat, 07 Jun 2025 11:42:02 +0300

Ping!  Andrea, any comments or suggestions?

> Date: Tue, 27 May 2025 13:38:02 +0000
> From:  Pip Cet via "Bug reports for GNU Emacs,
>  the Swiss army knife of text editors" <bug-gnu-emacs <at> gnu.org>
> 
> This reduced test case:
> 
> ;; -*- lexical-binding: t; -*-
> 
> (defun f (x)
>   (and (= x 1)
>        (if (eql x 1)
>            1
>          x)))
> 
> causes a native-ice:
> 
> Error: native-ice (\"can't find data in relocation containers\")
> 
> when Emacs is run as:
> 
> ./src/emacs -Q --batch --eval '(native-compile "native.el")'
> 
> The problem, as far as I can see, is that emit_mvar_rval appears to
> assume that if the constraint part of an mvar allows only a single
> value, that value can also be fetched from the reloc array.  In this
> case, the constraints on x make it so only 1.0 is a legal value in one
> basic block, but 1.0 isn't in the data array.
> 
> Note that it is not a valid fix to make obj_to_reloc fail gently if it
> cannot find obj in either data container: if 1.0 is present in the
> ephemeral vector but not in the default vector, referring to the
> ephemeral vector after top_level_run has completed will fail.
> 
> The problem is this limple insn:
> 
> (set #(mvar 17592188503872 1 (member 1.0)) #(mvar 17592188503233 0 (member 1.0)))
> 
> The RHS mvar was not created with :constant 1.0 (so 1.0 was never added
> to the data container), but it is comp-cstr-imm-vld-p.
> 
> The test case is reduced from mule-diag.el:
> 
> (defun list-block-of-chars (charset row min max)
>   (let (i ch)
>     (insert-char ?- (+ 7 (* 4 16)))
>     (insert "\n     ")
>     (setq i 0)
>     (while (< i 16)
>       (insert (format "%4X" i))
>       (setq i (1+ i)))
>     (setq i (* (/ min 16) 16))
>     (while (<= i max)
>       (if (= (% i 16) 0)
> 	  (insert (format "\n%6Xx" (/ (+ (* row 256) i) 16))))
>       (setq ch (if (< i min)
> 		   32
> 		 (or (decode-char charset (+ (* row 256) i))
> 		     32)))		; gap in mapping
>       ;; Don't insert control codes, non-Unicode characters.
>       (if (or (< ch 32) (= ch 127))
> 	  (setq ch (single-key-description ch))
> 	(if (and (>= ch 128) (< ch 160))
> 	    (setq ch (format "%02Xh" ch))
> 	  (if (> ch #x10FFFF)
> 	      (setq ch 32))))
>       (insert "\t" ch)
>       (setq i (1+ i))))
>   (insert "\n"))
> 
> But the original problem in mule-diag.el (generating a basic block which
> would call (single-key-description 127.0) ) only showed up with other
> modifications to the nativecomp compiler.
> 
> There are several ways to fix this, but I must note that adding a
> constant that is highly likely to be already present in the register
> file is usually wasteful: in this case, we don't want to create the
> constant 1.0, we simply want to return the register which we now know to
> contain 1.0.
> 
> 
> 
> 
> 

Previous Next

GNU bug tracking system
Copyright (C) 1999 Darren O. Benham, 1997,2003 nCipher Corporation Ltd, 1994-97 Ian Jackson.