* libc/machine/sh/memset.S: Avoid clobbering volatile

	objects following a tiny to-be-set array in the same quadword.

diff --git a/newlib/ChangeLog b/newlib/ChangeLog
index 6efbd70..a2266a6 100644 (file)
--- a/newlib/ChangeLog
+++ b/newlib/ChangeLog
@@ -1,3 +1,8 @@
+2003-04-09  J"orn Rennecke <joern.rennecke@superh.com>
+
+       * libc/machine/sh/memset.S: Avoid clobbering volatile
+       objects following a tiny to-be-set array in the same quadword.
+
 2001-04-09  Corinna Vinschen  <corinna@vinschen.de>
 
        * libc/include/wchar.h: Add definitions for wcswidth and wcwidth.  

diff --git a/newlib/libc/machine/sh/memset.S b/newlib/libc/machine/sh/memset.S
index ab71d95..2b45aff 100644 (file)
--- a/newlib/libc/machine/sh/memset.S
+++ b/newlib/libc/machine/sh/memset.S
@@ -12,9 +12,7 @@
 ENTRY(memset)
 #if __SHMEDIA__
        pta/l multiquad, tr0
-       mshflo.b r3,r3,r3
        ptabs r18, tr2
-       mperm.w r3, r63, r3     // Fill pattern now in every byte of r3
 
        andi r2, -8, r25
        add r2, r4, r5
@@ -23,16 +21,40 @@ ENTRY(memset)
        cmveq r4, r25, r20
        bne/u r25, r20, tr0 // multiquad
 
-       ldlo.q r2, 0, r7
-       shlli r4, 2, r4
-       movi -1, r8
-       SHHI r8, r4, r8
-       SHHI r8, r4, r8
-       mcmv r7, r8, r3
-       stlo.q r2, 0, r3
+!      This sequence could clobber volatile objects that are in the same
+!      quadword as a very short char array.
+!      ldlo.q r2, 0, r7
+!      shlli r4, 2, r4
+!      movi -1, r8
+!      SHHI r8, r4, r8
+!      SHHI r8, r4, r8
+!      mcmv r7, r8, r3
+!      stlo.q r2, 0, r3
+
+       pta/l setlongs, tr0
+       movi 4, r8
+       bgeu/u r4, r8, tr0
+       pta/l endset, tr0
+       beqi/u r4, 0, tr0
+       st.b r2, 0, r3
+       beqi/u r4, 1, tr0
+       nop
+       st.b r2, 1, r3
+       beqi/l r4, 2, tr0
+       st.b r2,2,r3
+endset: blink tr2, r63
+setlongs:
+       mshflo.b r3, r3, r3
+       mperm.w r3, r63, r3     // Fill pattern now in every byte of r3
+       stlo.l r2, 0, r3
+       nop
+       nop
+       sthi.l r5, -1, r3
        blink tr2, r63
 
 multiquad:
+       mshflo.b r3, r3, r3
+       mperm.w r3, r63, r3     // Fill pattern now in every byte of r3
        pta/l lastquad, tr0
        stlo.q r2, 0, r3
        sub r20, r25, r24