* bug#31474: logxor+ash trigger compilation bug?
@ 2018-05-16 18:16 Jan Nieuwenhuizen
2018-05-28 2:13 ` Mark H Weaver
0 siblings, 1 reply; 5+ messages in thread
From: Jan Nieuwenhuizen @ 2018-05-16 18:16 UTC (permalink / raw)
To: 31474
Hi!
Trying to implement bit-fields for MesCC, I stumble upon this. I was
looking for a bitwise left shift that introduces 1's instead of zeros.
This code
--8<---------------cut here---------------start------------->8---
;; foo.scm
(let* ((set-mask (pk 'set-mask (ash 3 3)))
(clear-mask (pk 'clear-mask (logxor set-mask -1))))
(pk 'expected (logxor 24 -1))
(display clear-mask)
(newline)
clear-mask)
--8<---------------cut here---------------end--------------->8---
behaves as I expect when compilation is turned off (compile or
auto-compile behave alike for me)
--8<---------------cut here---------------start------------->8---
19:50:43 janneke@dundal:~/src/mes
$ guile --no-auto-compile foo.scm
;;; (set-mask 24)
;;; (clear-mask -25)
;;; (expected -25)
-25
--8<---------------cut here---------------end--------------->8---
but when (auto)compiled, look:
--8<---------------cut here---------------start------------->8---
19:50:47 janneke@dundal:~/src/mes
$ guile foo.scm
;;; note: auto-compilation is enabled, set GUILE_AUTO_COMPILE=0
;;; or pass the --no-auto-compile argument to disable.
;;; compiling /home/janneke/src/mes/foo.scm
;;; compiled /home/janneke/.cache/guile/ccache/2.2-LE-8-3.A/home/janneke/src/mes/foo.scm.go
;;; (set-mask 24)
;;; (clear-mask -1)
;;; (expected -25)
-1
--8<---------------cut here---------------end--------------->8---
I'm using guile-2.2.3 from Guix master.
Is this a bug, can you suggest a workaround?
Greetings,
janneke
--
Jan Nieuwenhuizen <janneke@gnu.org> | GNU LilyPond http://lilypond.org
Freelance IT http://JoyofSource.com | Avatar® http://AvatarAcademy.com
^ permalink raw reply [flat|nested] 5+ messages in thread
* bug#31474: logxor+ash trigger compilation bug?
2018-05-16 18:16 bug#31474: logxor+ash trigger compilation bug? Jan Nieuwenhuizen
@ 2018-05-28 2:13 ` Mark H Weaver
2018-05-28 12:03 ` Mark H Weaver
0 siblings, 1 reply; 5+ messages in thread
From: Mark H Weaver @ 2018-05-28 2:13 UTC (permalink / raw)
To: Jan Nieuwenhuizen; +Cc: 31474
[-- Attachment #1: Type: text/plain, Size: 947 bytes --]
Hi Jan,
Jan Nieuwenhuizen <janneke@gnu.org> writes:
> ;; foo.scm
> (let* ((set-mask (pk 'set-mask (ash 3 3)))
> (clear-mask (pk 'clear-mask (logxor set-mask -1))))
> (pk 'expected (logxor 24 -1))
> (display clear-mask)
> (newline)
> clear-mask)
>
>
> behaves as I expect when compilation is turned off
[...]
> but when (auto)compiled, look:
[...]
> ;;; (set-mask 24)
>
> ;;; (clear-mask -1)
>
> ;;; (expected -25)
> -1
Indeed, thanks for the report. Guile 2.2's type inference pass
contained several bugs in the range analysis of bitwise logical
operators. I've attached below a preliminary (not fully tested) patch
that hopefully fixes these problems, and also makes some improvements.
> Is this a bug, can you suggest a workaround?
The specific workaround here would be to use (lognot x) instead of
(logxor x -1), which is a bit nicer anyway. They are equivalent.
Another equivalent formulation is (- -1 x).
Mark
[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: [PATCH] Fix type inference for bitwise logical operators --]
[-- Type: text/x-patch, Size: 8832 bytes --]
From 25eee7be61f4e467a5ce83856fbf8a7770cf5dca Mon Sep 17 00:00:00 2001
From: Mark H Weaver <mhw@netris.org>
Date: Sun, 27 May 2018 21:58:48 -0400
Subject: [PATCH] Fix type inference for bitwise logical operators.
Fixes <https://bugs.gnu.org/31474> and related bugs.
Reported by Jan Nieuwenhuizen <janneke@gnu.org>.
* module/language/cps/types.scm (next-power-of-two): Remove procedure.
(non-negative?, saturate+, saturate-, lognot*, logand-bounds): New
procedures. Use them to improve and fix bugs in the range analysis of
the type inferrers for 'logand', 'logsub', 'logior', 'ulogior',
'logxor', 'ulogxor', and 'lognot'.
---
module/language/cps/types.scm | 158 +++++++++++++++++++++-------------
1 file changed, 97 insertions(+), 61 deletions(-)
diff --git a/module/language/cps/types.scm b/module/language/cps/types.scm
index c24f9b99d..80073966d 100644
--- a/module/language/cps/types.scm
+++ b/module/language/cps/types.scm
@@ -1,5 +1,5 @@
;;; Type analysis on CPS
-;;; Copyright (C) 2014, 2015 Free Software Foundation, Inc.
+;;; Copyright (C) 2014, 2015, 2018 Free Software Foundation, Inc.
;;;
;;; This library is free software: you can redistribute it and/or modify
;;; it under the terms of the GNU Lesser General Public License as
@@ -1273,32 +1273,49 @@ minimum, and maximum."
(define! result &u64 0 &u64-max)))
(define-type-aliases ulsh ulsh/immediate)
-(define (next-power-of-two n)
- (let lp ((out 1))
- (if (< n out)
- out
- (lp (ash out 1)))))
+(define-inlinable (non-negative? n)
+ (not (negative? n)))
+
+(define (saturate+ n)
+ (if (inf? n)
+ +inf.0
+ (1- (ash 1 (integer-length n)))))
+
+(define (saturate- n)
+ (if (inf? n)
+ -inf.0
+ (ash -1 (integer-length n))))
+
+(define (lognot* n)
+ (- -1 n))
+
+(define (logand-bounds a0 a1 b0 b1)
+ ;; (a0 <= a <= a1) and (b0 <= b <= b1)
+ (cond ((and (non-negative? a0) (non-negative? b0))
+ (values 0 (min a1 b1)))
+ ((non-negative? a0)
+ (values 0 a1))
+ ((non-negative? b0)
+ (values 0 b1))
+ (else
+ (values (saturate- (min a0 b0))
+ (cond ((and (negative? a1) (negative? b1))
+ (min a1 b1))
+ ((negative? a1)
+ b1)
+ ((negative? b1)
+ a1)
+ (else
+ (saturate+ (max a1 b1))))))))
(define-simple-type-checker (logand &exact-integer &exact-integer))
(define-type-inferrer (logand a b result)
- (define (logand-min a b)
- (if (and (negative? a) (negative? b))
- (let ((min (min a b)))
- (if (inf? min)
- -inf.0
- (- 1 (next-power-of-two (- min)))))
- 0))
- (define (logand-max a b)
- (cond
- ((or (and (positive? a) (positive? b))
- (and (negative? a) (negative? b)))
- (min a b))
- (else (max a b))))
(restrict! a &exact-integer -inf.0 +inf.0)
(restrict! b &exact-integer -inf.0 +inf.0)
- (define! result &exact-integer
- (logand-min (&min a) (&min b))
- (logand-max (&max a) (&max b))))
+ (call-with-values (lambda ()
+ (logand-bounds (&min a) (&max a) (&min b) (&max b)))
+ (lambda (min max)
+ (define! result &exact-integer min max))))
(define-simple-type-checker (ulogand &u64 &u64))
(define-type-inferrer (ulogand a b result)
@@ -1308,22 +1325,8 @@ minimum, and maximum."
(define-simple-type-checker (logsub &exact-integer &exact-integer))
(define-type-inferrer (logsub a b result)
- (define (logsub-bounds min-a max-a min-b max-b)
- (cond
- ((negative? max-b)
- ;; Sign bit always set on B, so result will never be negative.
- ;; If A might be negative (all leftmost bits 1), we don't know
- ;; how positive the result might be.
- (values 0 (if (negative? min-a) +inf.0 max-a)))
- ((negative? min-b)
- ;; Sign bit might be set on B.
- (values min-a (if (negative? min-a) +inf.0 max-a)))
- ((negative? min-a)
- ;; Sign bit never set on B -- result will have the sign of A.
- (values -inf.0 max-a))
- (else
- ;; Sign bit never set on A and never set on B -- the nice case.
- (values 0 max-a))))
+ (define (logsub-bounds a0 a1 b0 b1)
+ (logand-bounds a0 a1 (lognot* b1) (lognot* b0)))
(restrict! a &exact-integer -inf.0 +inf.0)
(restrict! b &exact-integer -inf.0 +inf.0)
(call-with-values (lambda ()
@@ -1339,24 +1342,30 @@ minimum, and maximum."
(define-simple-type-checker (logior &exact-integer &exact-integer))
(define-type-inferrer (logior a b result)
- ;; Saturate all bits of val.
- (define (saturate val)
- (1- (next-power-of-two val)))
- (define (logior-min a b)
- (cond ((and (< a 0) (<= 0 b)) a)
- ((and (< b 0) (<= 0 a)) b)
- (else (max a b))))
- (define (logior-max a b)
- ;; If either operand is negative, just assume the max is -1.
- (cond
- ((or (< a 0) (< b 0)) -1)
- ((or (inf? a) (inf? b)) +inf.0)
- (else (saturate (logior a b)))))
+ (define (logior-bounds a0 a1 b0 b1)
+ ;; (a0 <= a <= a1) and (b0 <= b <= b1)
+ (cond ((and (negative? a1) (negative? b1))
+ (values (max a0 b0) -1))
+ ((negative? a1)
+ (values a0 -1))
+ ((negative? b1)
+ (values b0 -1))
+ (else
+ (values (cond ((and (non-negative? a0) (non-negative? b0))
+ (max a0 b0))
+ ((non-negative? a0)
+ b0)
+ ((non-negative? b0)
+ a0)
+ (else
+ (saturate- (min a0 b0))))
+ (saturate+ (max a1 b1))))))
(restrict! a &exact-integer -inf.0 +inf.0)
(restrict! b &exact-integer -inf.0 +inf.0)
- (define! result &exact-integer
- (logior-min (&min a) (&min b))
- (logior-max (&max a) (&max b))))
+ (call-with-values (lambda ()
+ (logior-bounds (&min a) (&max a) (&min b) (&max b)))
+ (lambda (min max)
+ (define! result &exact-integer min max))))
(define-simple-type-checker (ulogior &u64 &u64))
(define-type-inferrer (ulogior a b result)
@@ -1364,23 +1373,50 @@ minimum, and maximum."
(restrict! b &u64 0 &u64-max)
(define! result &u64
(max (&min/0 a) (&min/0 b))
- (1- (next-power-of-two (logior (&max/u64 a) (&max/u64 b))))))
-
-;; For our purposes, treat logxor the same as logior.
-(define-type-aliases logior logxor)
+ (saturate+ (max (&max/u64 a) (&max/u64 b)))))
+
+(define-simple-type-checker (logxor &exact-integer &exact-integer))
+(define-type-inferrer (logxor a b result)
+ (define (logxor-bounds a0 a1 b0 b1)
+ ;; (a0 <= a <= a1) and (b0 <= b <= b1)
+ (cond ((and (non-negative? a0) (non-negative? b0))
+ (values 0 (saturate+ (max a1 b1))))
+ ((and (negative? a1) (negative? b1))
+ (values 0 (saturate+ (min a0 b0))))
+ ((and (non-negative? a0) (negative? b1))
+ (values (saturate- (max a1 (lognot* b0))) -1))
+ ((and (negative? a1) (non-negative? b0))
+ (values (saturate- (max b1 (lognot* a0))) -1))
+ ((and (negative? a0) (non-negative? a1)
+ (negative? b0) (non-negative? b1))
+ (values (saturate- (max a1 b1 (lognot* a0) (lognot* b0)))
+ (saturate+ (max a1 b1 (lognot* a0) (lognot* b0)))))
+ (else
+ (values (if (and (non-negative? a1) (negative? b0))
+ (saturate- (max a1 (lognot* b0)))
+ (saturate- (max b1 (lognot* a0))))
+ (if (and (non-negative? a1) (non-negative? b1))
+ (saturate+ (max a1 b1))
+ (saturate+ (min a0 b0)))))))
+ (restrict! a &exact-integer -inf.0 +inf.0)
+ (restrict! b &exact-integer -inf.0 +inf.0)
+ (call-with-values (lambda ()
+ (logxor-bounds (&min a) (&max a) (&min b) (&max b)))
+ (lambda (min max)
+ (define! result &exact-integer min max))))
(define-simple-type-checker (ulogxor &u64 &u64))
(define-type-inferrer (ulogxor a b result)
(restrict! a &u64 0 &u64-max)
(restrict! b &u64 0 &u64-max)
- (define! result &u64 0 &u64-max))
+ (define! result &u64 0 (saturate+ (max (&max/u64 a) (&max/u64 b)))))
(define-simple-type-checker (lognot &exact-integer))
(define-type-inferrer (lognot a result)
(restrict! a &exact-integer -inf.0 +inf.0)
(define! result &exact-integer
- (- -1 (&max a))
- (- -1 (&min a))))
+ (lognot* (&max a))
+ (lognot* (&min a))))
(define-simple-type-checker (logtest &exact-integer &exact-integer))
(define-predicate-inferrer (logtest a b true?)
--
2.17.0
^ permalink raw reply related [flat|nested] 5+ messages in thread
* bug#31474: logxor+ash trigger compilation bug?
2018-05-28 2:13 ` Mark H Weaver
@ 2018-05-28 12:03 ` Mark H Weaver
2018-05-28 21:17 ` Jan Nieuwenhuizen
0 siblings, 1 reply; 5+ messages in thread
From: Mark H Weaver @ 2018-05-28 12:03 UTC (permalink / raw)
To: Jan Nieuwenhuizen; +Cc: 31474
[-- Attachment #1: Type: text/plain, Size: 229 bytes --]
Here's an improved version of my patch. It's functionally equivalent
but with more comprehensible code and more comments. I think this is
ready to push to the stable-2.2 branch. Comments and suggestions
welcome.
Mark
[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: [PATCH] Fix type inference for bitwise logical operators (v2) --]
[-- Type: text/x-patch, Size: 13480 bytes --]
From aefb4c3627596335a2ef2cf6f721f9e04b49ae7e Mon Sep 17 00:00:00 2001
From: Mark H Weaver <mhw@netris.org>
Date: Sun, 27 May 2018 21:58:48 -0400
Subject: [PATCH] Fix type inference for bitwise logical operators.
Fixes <https://bugs.gnu.org/31474> and related bugs.
Reported by Jan Nieuwenhuizen <janneke@gnu.org>.
* module/language/cps/types.scm (next-power-of-two): Remove procedure.
(non-negative?, lognot*, saturate+, saturate-, logand-bounds)
(logsub-bounds, logior-bounds, logxor-bounds): New procedures. Use them
to improve and fix bugs in the range analysis of the type inferrers for
'logand', 'logsub', 'logior', 'ulogior', 'logxor', 'ulogxor', and
'lognot'.
---
module/language/cps/types.scm | 230 +++++++++++++++++++++++++---------
1 file changed, 169 insertions(+), 61 deletions(-)
diff --git a/module/language/cps/types.scm b/module/language/cps/types.scm
index c24f9b99d..4326a8d37 100644
--- a/module/language/cps/types.scm
+++ b/module/language/cps/types.scm
@@ -1,5 +1,5 @@
;;; Type analysis on CPS
-;;; Copyright (C) 2014, 2015 Free Software Foundation, Inc.
+;;; Copyright (C) 2014, 2015, 2018 Free Software Foundation, Inc.
;;;
;;; This library is free software: you can redistribute it and/or modify
;;; it under the terms of the GNU Lesser General Public License as
@@ -1273,32 +1273,79 @@ minimum, and maximum."
(define! result &u64 0 &u64-max)))
(define-type-aliases ulsh ulsh/immediate)
-(define (next-power-of-two n)
- (let lp ((out 1))
- (if (< n out)
- out
- (lp (ash out 1)))))
+(define-inlinable (non-negative? n)
+ "Return true if N is non-negative, otherwise return false."
+ (not (negative? n)))
+
+;; Like 'lognot', but handles infinities.
+(define-inlinable (lognot* n)
+ "Return the bitwise complement of N. If N is infinite, return -N."
+ (- -1 n))
+
+(define saturate+
+ (case-lambda
+ "Let N be the least upper bound of the integer lengths of the
+arguments. Return the greatest integer whose integer length is N.
+If any of the arguments are infinite, return positive infinity."
+ ((a b)
+ (if (or (inf? a) (inf? b))
+ +inf.0
+ (1- (ash 1 (max (integer-length a)
+ (integer-length b))))))
+ ((a b c)
+ (saturate+ (saturate+ a b) c))
+ ((a b c d)
+ (saturate+ (saturate+ a b) c d))))
+
+(define saturate-
+ (case-lambda
+ "Let N be the least upper bound of the integer lengths of the
+arguments. Return the least integer whose integer length is N.
+If any of the arguments are infinite, return negative infinity."
+ ((a b) (lognot* (saturate+ a b)))
+ ((a b c) (lognot* (saturate+ a b c)))
+ ((a b c d) (lognot* (saturate+ a b c d)))))
+
+(define (logand-bounds a0 a1 b0 b1)
+ "Return two values: lower and upper bounds for (logand A B)
+where (A0 <= A <= A1) and (B0 <= B <= B1)."
+ ;; For each argument, we consider three cases: (1) the argument is
+ ;; non-negative, (2) its sign is unknown, or (3) it is negative.
+ ;; To handle both arguments, we must consider a total of 9 cases:
+ ;;
+ ;; -----------------------------------------------------------------------
+ ;; LOGAND | non-negative B | unknown-sign B | negative B
+ ;; -----------------------------------------------------------------------
+ ;; non-negative A | 0 .. (min A1 B1) | 0 .. A1 | 0 .. A1
+ ;; -----------------------------------------------------------------------
+ ;; unknown-sign A | 0 .. B1 | (sat- A0 B0) | (sat- A0 B0)
+ ;; | | .. | .. A1
+ ;; | | (sat+ A1 B1) |
+ ;; -----------------------------------------------------------------------
+ ;; negative A | 0 .. B1 | (sat- A0 B0) | (sat- A0 B0)
+ ;; | | .. B1 | .. (min A1 B1)
+ ;; -----------------------------------------------------------------------
+ (values (if (or (non-negative? a0) (non-negative? b0))
+ 0
+ (saturate- a0 b0))
+ (cond ((or (and (non-negative? a0) (non-negative? b0))
+ (and (negative? a1) (negative? b1)))
+ (min a1 b1))
+ ((or (non-negative? a0) (negative? b1))
+ a1)
+ ((or (non-negative? b0) (negative? a1))
+ b1)
+ (else
+ (saturate+ a1 b1)))))
(define-simple-type-checker (logand &exact-integer &exact-integer))
(define-type-inferrer (logand a b result)
- (define (logand-min a b)
- (if (and (negative? a) (negative? b))
- (let ((min (min a b)))
- (if (inf? min)
- -inf.0
- (- 1 (next-power-of-two (- min)))))
- 0))
- (define (logand-max a b)
- (cond
- ((or (and (positive? a) (positive? b))
- (and (negative? a) (negative? b)))
- (min a b))
- (else (max a b))))
(restrict! a &exact-integer -inf.0 +inf.0)
(restrict! b &exact-integer -inf.0 +inf.0)
- (define! result &exact-integer
- (logand-min (&min a) (&min b))
- (logand-max (&max a) (&max b))))
+ (call-with-values (lambda ()
+ (logand-bounds (&min a) (&max a) (&min b) (&max b)))
+ (lambda (min max)
+ (define! result &exact-integer min max))))
(define-simple-type-checker (ulogand &u64 &u64))
(define-type-inferrer (ulogand a b result)
@@ -1306,24 +1353,17 @@ minimum, and maximum."
(restrict! b &u64 0 &u64-max)
(define! result &u64 0 (min (&max/u64 a) (&max/u64 b))))
+(define (logsub-bounds a0 a1 b0 b1)
+ "Return two values: lower and upper bounds for (logsub A B),
+i.e. (logand A (lognot B)), where (A0 <= A <= A1) and (B0 <= B <= B1)."
+ ;; Here we use 'logand-bounds' to compute the bounds, after
+ ;; computing the bounds of (lognot B) from the bounds of B.
+ ;; From (B0 <= B <= B1) it follows that (~B1 <= ~B <= ~B0),
+ ;; where ~X means (lognot X).
+ (logand-bounds a0 a1 (lognot* b1) (lognot* b0)))
+
(define-simple-type-checker (logsub &exact-integer &exact-integer))
(define-type-inferrer (logsub a b result)
- (define (logsub-bounds min-a max-a min-b max-b)
- (cond
- ((negative? max-b)
- ;; Sign bit always set on B, so result will never be negative.
- ;; If A might be negative (all leftmost bits 1), we don't know
- ;; how positive the result might be.
- (values 0 (if (negative? min-a) +inf.0 max-a)))
- ((negative? min-b)
- ;; Sign bit might be set on B.
- (values min-a (if (negative? min-a) +inf.0 max-a)))
- ((negative? min-a)
- ;; Sign bit never set on B -- result will have the sign of A.
- (values -inf.0 max-a))
- (else
- ;; Sign bit never set on A and never set on B -- the nice case.
- (values 0 max-a))))
(restrict! a &exact-integer -inf.0 +inf.0)
(restrict! b &exact-integer -inf.0 +inf.0)
(call-with-values (lambda ()
@@ -1337,26 +1377,47 @@ minimum, and maximum."
(restrict! b &u64 0 &u64-max)
(define! result &u64 0 (&max/u64 a)))
+(define (logior-bounds a0 a1 b0 b1)
+ "Return two values: lower and upper bounds for (logior A B)
+where (A0 <= A <= A1) and (B0 <= B <= B1)."
+ ;; For each argument, we consider three cases: (1) the argument is
+ ;; non-negative, (2) its sign is unknown, or (3) it is negative.
+ ;; To handle both arguments, we must consider a total of 9 cases.
+ ;;
+ ;; ---------------------------------------------------------------------
+ ;; LOGIOR | non-negative B | unknown-sign B | negative B
+ ;; ---------------------------------------------------------------------
+ ;; non-negative A | (max A0 B0) | B0 | B0 .. -1
+ ;; | .. | .. |
+ ;; | (sat+ A1 B1) | (sat+ A1 B1) |
+ ;; ---------------------------------------------------------------------
+ ;; unknown-sign A | A0 | (sat- A0 B0) | B0 .. -1
+ ;; | .. | .. |
+ ;; | (sat+ A1 B1) | (sat+ A1 B1) |
+ ;; ---------------------------------------------------------------------
+ ;; negative A | A0 .. -1 | A0 .. -1 | (max A0 B0) .. -1
+ ;; ---------------------------------------------------------------------
+ (values (cond ((or (and (non-negative? a0) (non-negative? b0))
+ (and (negative? a1) (negative? b1)))
+ (max a0 b0))
+ ((or (non-negative? a0) (negative? b1))
+ b0)
+ ((or (non-negative? b0) (negative? a1))
+ a0)
+ (else
+ (saturate- a0 b0)))
+ (if (or (negative? a1) (negative? b1))
+ -1
+ (saturate+ a1 b1))))
+
(define-simple-type-checker (logior &exact-integer &exact-integer))
(define-type-inferrer (logior a b result)
- ;; Saturate all bits of val.
- (define (saturate val)
- (1- (next-power-of-two val)))
- (define (logior-min a b)
- (cond ((and (< a 0) (<= 0 b)) a)
- ((and (< b 0) (<= 0 a)) b)
- (else (max a b))))
- (define (logior-max a b)
- ;; If either operand is negative, just assume the max is -1.
- (cond
- ((or (< a 0) (< b 0)) -1)
- ((or (inf? a) (inf? b)) +inf.0)
- (else (saturate (logior a b)))))
(restrict! a &exact-integer -inf.0 +inf.0)
(restrict! b &exact-integer -inf.0 +inf.0)
- (define! result &exact-integer
- (logior-min (&min a) (&min b))
- (logior-max (&max a) (&max b))))
+ (call-with-values (lambda ()
+ (logior-bounds (&min a) (&max a) (&min b) (&max b)))
+ (lambda (min max)
+ (define! result &exact-integer min max))))
(define-simple-type-checker (ulogior &u64 &u64))
(define-type-inferrer (ulogior a b result)
@@ -1364,23 +1425,70 @@ minimum, and maximum."
(restrict! b &u64 0 &u64-max)
(define! result &u64
(max (&min/0 a) (&min/0 b))
- (1- (next-power-of-two (logior (&max/u64 a) (&max/u64 b))))))
-
-;; For our purposes, treat logxor the same as logior.
-(define-type-aliases logior logxor)
+ (saturate+ (&max/u64 a) (&max/u64 b))))
+
+(define (logxor-bounds a0 a1 b0 b1)
+ "Return two values: lower and upper bounds for (logxor A B)
+where (A0 <= A <= A1) and (B0 <= B <= B1)."
+ ;; For each argument, we consider three cases: (1) the argument is
+ ;; non-negative, (2) its sign is unknown, or (3) it is negative.
+ ;; To handle both arguments, we must consider a total of 9 cases.
+ ;;
+ ;; --------------------------------------------------------------------
+ ;; LOGXOR | non-negative B | unknown-sign B | negative B
+ ;; --------------------------------------------------------------------
+ ;; non-negative A | 0 | (sat- A1 B0) | (sat- A1 B0)
+ ;; | .. | .. | ..
+ ;; | (sat+ A1 B1) | (sat+ A1 B1) | -1
+ ;; --------------------------------------------------------------------
+ ;; unknown-sign A | (sat- A0 B1) | (sat- A0 B1 A1 B0) | (sat- A1 B0)
+ ;; | .. | .. | ..
+ ;; | (sat+ A1 B1) | (sat+ A1 B1 A0 B0) | (sat+ A0 B0)
+ ;; --------------------------------------------------------------------
+ ;; negative A | (sat- A0 B1) | (sat- A0 B1) | 0
+ ;; | .. | .. | ..
+ ;; | -1 | (sat+ A0 B0) | (sat+ A0 B0)
+ ;; --------------------------------------------------------------------
+ (values (cond ((or (and (non-negative? a0) (non-negative? b0))
+ (and (negative? a1) (negative? b1)))
+ 0)
+ ((or (non-negative? a0) (negative? b1))
+ (saturate- a1 b0))
+ ((or (non-negative? b0) (negative? a1))
+ (saturate- a0 b1))
+ (else
+ (saturate- a0 b1 a1 b0)))
+ (cond ((or (and (non-negative? a0) (negative? b1))
+ (and (non-negative? b0) (negative? a1)))
+ -1)
+ ((or (non-negative? a0) (non-negative? b0))
+ (saturate+ a1 b1))
+ ((or (negative? a1) (negative? b1))
+ (saturate+ a0 b0))
+ (else
+ (saturate+ a1 b1 a0 b0)))))
+
+(define-simple-type-checker (logxor &exact-integer &exact-integer))
+(define-type-inferrer (logxor a b result)
+ (restrict! a &exact-integer -inf.0 +inf.0)
+ (restrict! b &exact-integer -inf.0 +inf.0)
+ (call-with-values (lambda ()
+ (logxor-bounds (&min a) (&max a) (&min b) (&max b)))
+ (lambda (min max)
+ (define! result &exact-integer min max))))
(define-simple-type-checker (ulogxor &u64 &u64))
(define-type-inferrer (ulogxor a b result)
(restrict! a &u64 0 &u64-max)
(restrict! b &u64 0 &u64-max)
- (define! result &u64 0 &u64-max))
+ (define! result &u64 0 (saturate+ (&max/u64 a) (&max/u64 b))))
(define-simple-type-checker (lognot &exact-integer))
(define-type-inferrer (lognot a result)
(restrict! a &exact-integer -inf.0 +inf.0)
(define! result &exact-integer
- (- -1 (&max a))
- (- -1 (&min a))))
+ (lognot* (&max a))
+ (lognot* (&min a))))
(define-simple-type-checker (logtest &exact-integer &exact-integer))
(define-predicate-inferrer (logtest a b true?)
--
2.17.0
^ permalink raw reply related [flat|nested] 5+ messages in thread
* bug#31474: logxor+ash trigger compilation bug?
2018-05-28 12:03 ` Mark H Weaver
@ 2018-05-28 21:17 ` Jan Nieuwenhuizen
2018-06-11 14:28 ` Mark H Weaver
0 siblings, 1 reply; 5+ messages in thread
From: Jan Nieuwenhuizen @ 2018-05-28 21:17 UTC (permalink / raw)
To: Mark H Weaver; +Cc: 31474
Mark H Weaver writes:
> Here's an improved version of my patch. It's functionally equivalent
> but with more comprehensible code and more comments. I think this is
> ready to push to the stable-2.2 branch. Comments and suggestions
> welcome.
I tried it and it works for me. Thanks a lot!
Also, thanks for the workarounds you suggested, they indeed work without
this patch.
Greetings,
janneke
--
Jan Nieuwenhuizen <janneke@gnu.org> | GNU LilyPond http://lilypond.org
Freelance IT http://JoyofSource.com | Avatar® http://AvatarAcademy.com
^ permalink raw reply [flat|nested] 5+ messages in thread
* bug#31474: logxor+ash trigger compilation bug?
2018-05-28 21:17 ` Jan Nieuwenhuizen
@ 2018-06-11 14:28 ` Mark H Weaver
0 siblings, 0 replies; 5+ messages in thread
From: Mark H Weaver @ 2018-06-11 14:28 UTC (permalink / raw)
To: Jan Nieuwenhuizen; +Cc: 31474-done
Jan Nieuwenhuizen <janneke@gnu.org> writes:
> Mark H Weaver writes:
>
>> Here's an improved version of my patch. It's functionally equivalent
>> but with more comprehensible code and more comments. I think this is
>> ready to push to the stable-2.2 branch. Comments and suggestions
>> welcome.
>
> I tried it and it works for me. Thanks a lot!
>
> Also, thanks for the workarounds you suggested, they indeed work without
> this patch.
I pushed the patch as commit 2733e97395db30c6233f79f341959e722b4bd4ff to
the stable-2.2 branch. I'm closing this bug now, but feel free to
reopen if needed.
Thanks,
Mark
^ permalink raw reply [flat|nested] 5+ messages in thread
end of thread, other threads:[~2018-06-11 14:28 UTC | newest]
Thread overview: 5+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2018-05-16 18:16 bug#31474: logxor+ash trigger compilation bug? Jan Nieuwenhuizen
2018-05-28 2:13 ` Mark H Weaver
2018-05-28 12:03 ` Mark H Weaver
2018-05-28 21:17 ` Jan Nieuwenhuizen
2018-06-11 14:28 ` Mark H Weaver
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).