From 4716083f3f2ac66db2d1457daae252000adf7c60 Mon Sep 17 00:00:00 2001
From: Greg King <gregdk@users.sf.net>
Date: Sun, 13 Dec 2015 07:17:41 -0500
Subject: [PATCH] Fixed a signed char shift optimization so that it won't be
 used on signed int also.

(It would lose significant bits from the high byte.)
---
 src/cc65/coptshift.c | 54 +++++++++++++++++++++++++++-----------------
 src/cc65/coptshift.h | 13 ++++++++---
 2 files changed, 43 insertions(+), 24 deletions(-)

diff --git a/src/cc65/coptshift.c b/src/cc65/coptshift.c
index a4b993073..2f8c2f833 100644
--- a/src/cc65/coptshift.c
+++ b/src/cc65/coptshift.c
@@ -307,51 +307,63 @@ NextEntry:
 
 
 
-unsigned OptShift2(CodeSeg* S)
-/* A call to the asrax1 routines may get replaced by something simpler, if
-** X is not used later:
+unsigned OptShift2 (CodeSeg* S)
+/* The sequence
+**
+**      bpl     L
+**      dex
+** L:   jsr     asraxN
+**
+** might be replaced by N copies of
 **
 **      cmp     #$80
 **      ror     a
+**
+** if X is not used later (X is assumed to be zero on entry).
 */
 {
     unsigned Changes = 0;
-    unsigned I;
+    unsigned I = 0;
 
     /* Walk over the entries */
-    I = 0;
     while (I < CS_GetEntryCount (S)) {
 
         unsigned Shift;
         unsigned Count;
+        CodeEntry* L[3];
 
         /* Get next entry */
-        CodeEntry* E = CS_GetEntry (S, I);
+        L[0] = CS_GetEntry (S, I);
 
         /* Check for the sequence */
-        if (E->OPC == OP65_JSR                          &&
-            (Shift = GetShift (E->Arg)) != SHIFT_NONE   &&
-            SHIFT_TYPE (Shift) == SHIFT_TYPE_ASR        &&
-            (Count = SHIFT_COUNT (Shift)) > 0           &&
-            Count * 100 <= S->CodeSizeFactor    &&
-            !RegXUsed (S, I+1)) {
+        if (L[0]->OPC == OP65_BPL                                       &&
+            L[0]->JumpTo != 0                                           &&
+            CS_GetEntries (S, L+1, I+1, 2)                              &&
+            L[1]->OPC == OP65_DEX                                       &&
+            L[0]->JumpTo->Owner == L[2]                                 &&
+            !CS_RangeHasLabel (S, I, 2)                                 &&
+            L[2]->OPC == OP65_JSR                                       &&
+            SHIFT_TYPE (Shift = GetShift (L[2]->Arg)) == SHIFT_TYPE_ASR &&
+            (Count = SHIFT_COUNT (Shift)) > 0                           &&
+            Count * 100 <= S->CodeSizeFactor                            &&
+            !RegXUsed (S, I+3)) {
 
             CodeEntry* X;
-            unsigned J = I+1;
+            unsigned J = I+2;
 
             /* Generate the replacement sequence */
             while (Count--) {
                 /* cmp #$80 */
-                X = NewCodeEntry (OP65_CMP, AM65_IMM, "$80", 0, E->LI);
-                CS_InsertEntry (S, X, J++);
+                X = NewCodeEntry (OP65_CMP, AM65_IMM, "$80", 0, L[2]->LI);
+                CS_InsertEntry (S, X, ++J);
 
                 /* ror a */
-                X = NewCodeEntry (OP65_ROR, AM65_ACC, "a", 0, E->LI);
-                CS_InsertEntry (S, X, J++);
+                X = NewCodeEntry (OP65_ROR, AM65_ACC, "a", 0, L[2]->LI);
+                CS_InsertEntry (S, X, ++J);
             }
 
-            /* Delete the call to asrax */
-            CS_DelEntry (S, I);
+            /* Remove the bpl/dex/jsr */
+            CS_DelEntries (S, I, 3);
 
             /* Remember, we had changes */
             ++Changes;
@@ -412,7 +424,7 @@ unsigned OptShift3 (CodeSeg* S)
             (Shift = GetShift (L[2]->Arg)) != SHIFT_NONE        &&
             SHIFT_DIR (Shift) == SHIFT_DIR_RIGHT                &&
             (Count = SHIFT_COUNT (Shift)) > 0) {
-                                                
+
             /* Add the replacement insn instead */
             CodeEntry* X = NewCodeEntry (OP65_ROR, AM65_ACC, "a", 0, L[2]->LI);
             CS_InsertEntry (S, X, I+3);
@@ -421,7 +433,7 @@ unsigned OptShift3 (CodeSeg* S)
                 CS_InsertEntry (S, X, I+4);
             }
 
-            /* Remove the bcs/dex/jsr */
+            /* Remove the bcc/inx/jsr */
             CS_DelEntries (S, I, 3);
 
             /* Remember, we had changes */
diff --git a/src/cc65/coptshift.h b/src/cc65/coptshift.h
index 0410652a1..e7f9cc359 100644
--- a/src/cc65/coptshift.h
+++ b/src/cc65/coptshift.h
@@ -60,12 +60,19 @@ unsigned OptShift1 (CodeSeg* S);
 **  L1:
 */
 
-unsigned OptShift2(CodeSeg* S);
-/* A call to the asrax1 routines may get replaced by something simpler, if
-** X is not used later:
+unsigned OptShift2 (CodeSeg* S);
+/* The sequence
+**
+**      bpl     L
+**      dex
+** L:   jsr     asraxN
+**
+** might be replaced by N copies of
 **
 **      cmp     #$80
 **      ror     a
+**
+** if X is not used later (X is assumed to be zero on entry).
 */
 
 unsigned OptShift3 (CodeSeg* S);