From 3f0631116ad0934dadce53b238bd26a7ee36f143 Mon Sep 17 00:00:00 2001
From: Matthew Flatt <mflatt@racket-lang.org>
Date: Thu, 3 Sep 2020 06:42:22 -0600
Subject: [PATCH] Chez Scheme: improve repair for arm32 return-address
 reference

Since the offset is always a multiple of 4, expand the range of
compactly encoded offsets.
---
 racket/src/ChezScheme/s/arm32.ss      | 14 +++++++++-----
 racket/src/ChezScheme/s/cpnanopass.ss | 15 +++++++++++----
 2 files changed, 20 insertions(+), 9 deletions(-)

diff --git a/racket/src/ChezScheme/s/arm32.ss b/racket/src/ChezScheme/s/arm32.ss
index dcc36d1811..2498e5a5d4 100644
--- a/racket/src/ChezScheme/s/arm32.ss
+++ b/racket/src/ChezScheme/s/arm32.ss
@@ -1674,10 +1674,14 @@
                          (bitwise-arithmetic-shift-left (logand n #xffffff) 8)
                          (bitwise-arithmetic-shift-right n 24)))))))))
 
-  ;; A region of funky12 where there's no number that fits when a smaller number doesn't
-  (define connected-funky12
+  ;; restrict funky12 so that an code offset n will not fit
+  ;; if a smaller offset wouldn't fit, which prevents bouncing
+  ;; in the loop that computes label offsets
+  (define code-offset-funky12
     (lambda (n)
-      (and (fixnum? n) (#%$fxu< n #x100)
+      (safe-assert (and (fixnum? n) (fx= 0 (fxand n 3))))
+      (and (fixnum? n)
+           (#%$fxu< n #x400)
            (funky12 n))))
 
   (define shift-count?
@@ -2192,11 +2196,11 @@
                  (let ([incr-offset (adjust-return-point-offset incr-offset l)])
                    (let ([disp (fx- next-addr (fx- offset incr-offset) 4)])
                      (cond
-                       [(connected-funky12 disp)
+                       [(code-offset-funky12 disp)
                         (Trivit (dest)
                           ; aka adr, encoding A1
                           (emit addi #f dest `(reg . ,%pc) disp '()))]
-                       [(connected-funky12 (- disp))
+                       [(code-offset-funky12 (- disp))
                         (Trivit (dest)
                           ; aka adr, encoding A2
                           (emit subi #f dest `(reg . ,%pc) (- disp) '()))]
diff --git a/racket/src/ChezScheme/s/cpnanopass.ss b/racket/src/ChezScheme/s/cpnanopass.ss
index ba0bcd7313..27c0c34626 100644
--- a/racket/src/ChezScheme/s/cpnanopass.ss
+++ b/racket/src/ChezScheme/s/cpnanopass.ss
@@ -15925,10 +15925,17 @@
                                      (dump (cdr trace-list) addr))))
                         (car trace-list)))))
                 (fprintf p "~d:~9t<end~@[ ~a~]>\n" size name))))
-          ; munge gets the code in forward order, but really wants to process it
-          ; backwards to find the label offsets.  Maybe the size would be better
-          ; tracked by doing it more like cp2 does right now and then patching in
-          ; the foward jumps and tightening up the code.
+          ;; munge gets the code in forward order, but really wants to process it
+          ;; backwards to find the label offsets.  Maybe the size would be better
+          ;; tracked by doing it more like cp2 does right now and then patching in
+          ;; the foward jumps and tightening up the code.
+          ;;
+          ;; If label addresses computed this time are not the same as last time,
+          ;; then `munge-recur?` is set, and some loop will try `munge` again.
+          ;; For that loop to converge, the instruction encoding for a larger label
+          ;; offset must not get smaller; otherwise, code might get further away,
+          ;; leading to some smaller encoding, which pulls code back closer, leading
+          ;; to a larger encoding, and so on.
           (define-who munge
             (lambda (c* size)
               (define (munge-pass c* iteration)