Skip to contentSkip to conversation

mem_address_not_aligned trap for unaligned PC

PATCH In review Aurelien Jarno 2 participants 4 messages Apr 1, 2007

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 mem_address_not_aligned trap for unaligned PC

Hi all,

According to the SPARCv8 and SPARCv9 manuals, the jmpl, rett and return instructions should generate a mem_address_not_aligned trap if either of the low-order two bits of the target address is nonzero.

The patch below implements that.

Bye,
Aurelien

Index: target-sparc/cpu.h
===================================================================
RCS file: /sources/qemu/qemu/target-sparc/cpu.h,v
retrieving revision 1.32

cpu.h +2 -0
diff -u -d -p -r1.32 cpu.h
@@ -35,6 +35,7 @@
 #define TT_NFPU_INSN 0x04
 #define TT_WIN_OVF  0x05
 #define TT_WIN_UNF  0x06 
+#define TT_UNALIGNED 0x07 
 #define TT_FP_EXCP  0x08
 #define TT_DFAULT   0x09
 #define TT_TOVF     0x0a
@@ -55,6 +56,7 @@
 #define TT_DFAULT   0x30
 #define TT_DMISS    0x31
 #define TT_DPROT    0x32
+#define TT_UNALIGNED 0x34
 #define TT_PRIV_ACT 0x37
 #define TT_EXTINT   0x40
 #define TT_SPILL    0x80

Index: target-sparc/op.c
===================================================================
RCS file: /sources/qemu/qemu/target-sparc/op.c,v
retrieving revision 1.27

op.c +4 -1
diff -u -d -p -r1.27 op.c
@@ -1486,7 +1486,10 @@ void OPPROTO op_movl_npc_im(void)
 
 void OPPROTO op_movl_npc_T0(void)
 {
-    env->npc = T0;
+    if (T0 & 0x3)
+	raise_exception(TT_UNALIGNED);
+    else
+	env->npc = T0;
 }
 
 void OPPROTO op_mov_pc_npc(void)
Signature · 4 lines
  .''`.  Aurelien Jarno	            | GPG: 1024D/F1BCDB73
 : :' :  Debian developer           | Electrical Engineer
 `. `'   [email protected]         | [email protected]
   `-    people.debian.org/~aurel32 | www.aurel32.net
#1 lore raw

RE: [Qemu-devel] [PATCH][SPARC] mem_address_not_aligned trap forunaligned PC

Quoted 3 lines

According to the SPARCv8 and SPARCv9 manuals, the jmpl, rett and return instructions should generate a mem_address_not_aligned trap if either of the low-order two bits of the target address is nonzero.

This is true, but in that case alignment should be enforced for loads ands stores as well. The checks also incur a performance penalty for little advantage. Maybe the checks should be enabled only with a compile/run-time option.

_________________________________________________________________
Express yourself instantly with MSN Messenger! Download today it's FREE! http://messenger.msn.click-url.com/go/onm00200471ave/direct/01/

#2 lore raw
Quoted 9 lines · Blue Swirl
Quoted 3 lines

According to the SPARCv8 and SPARCv9 manuals, the jmpl, rett and return instructions should generate a mem_address_not_aligned trap if either of the low-order two bits of the target address is nonzero.

This is true, but in that case alignment should be enforced for loads ands stores as well. The checks also incur a performance penalty for little advantage. Maybe the checks should be enabled only with a compile/run-time option.

I also have a patch for load/store, but openbios is doing a few unaligned memory accesses, so the patch can't be used until openbios is fixed.

I haven't made any benchmark, but the performance penality is probably very small. If you look at softmmu_template.h, you will see that it only adds a if test, while the whole function is already a few dozen of lines long.

I would really like to see the QEMU having the same behaviour as real hardware, this allow for example debugging SIGBUS problems in a program without having real hardware.

Signature · 4 lines
  .''`.  Aurelien Jarno	            | GPG: 1024D/F1BCDB73
 : :' :  Debian developer           | Electrical Engineer
 `. `'   [email protected]         | [email protected]
   `-    people.debian.org/~aurel32 | www.aurel32.net
#3 lore raw

PATCH 1/1 mem_address_not_aligned trap for unaligned PC and load/store

Quoted 11 lines · Aurelien Jarno
Quoted 9 lines · Blue Swirl
Quoted 3 lines

According to the SPARCv8 and SPARCv9 manuals, the jmpl, rett and return instructions should generate a mem_address_not_aligned trap if either of the low-order two bits of the target address is nonzero.

This is true, but in that case alignment should be enforced for loads ands stores as well. The checks also incur a performance penalty for little advantage. Maybe the checks should be enabled only with a compile/run-time option.

Quoted 4 lines

I also have a patch for load/store, but openbios is doing a few unaligned memory accesses, so the patch can't be used until openbios is fixed.

I have finally been able to found and fix the bug in openbios. I have send a patch to the openbios mailing list. The patch and the fixed bios is available on http://temp.aurel32.net .

The patch below is for QEMU and implements mem_address_not_aligned for unaligned PC (jmpl, rett and return instructions) and for load/store operations.

I will try to do some benchmarks about the performance penality, but I think it is very small. The patch basically adds 2 or 3 host assembly instructions for each load/store, jmpl, rett or return instruction.

Index: target-sparc/cpu.h
===================================================================
RCS file: /sources/qemu/qemu/target-sparc/cpu.h,v
retrieving revision 1.32

cpu.h +2 -0
diff -u -d -p -r1.32 cpu.h
@@ -35,6 +35,7 @@
 #define TT_NFPU_INSN 0x04
 #define TT_WIN_OVF  0x05
 #define TT_WIN_UNF  0x06 
+#define TT_UNALIGNED 0x07 
 #define TT_FP_EXCP  0x08
 #define TT_DFAULT   0x09
 #define TT_TOVF     0x0a
@@ -55,6 +56,7 @@
 #define TT_DFAULT   0x30
 #define TT_DMISS    0x31
 #define TT_DPROT    0x32
+#define TT_UNALIGNED 0x34
 #define TT_PRIV_ACT 0x37
 #define TT_EXTINT   0x40
 #define TT_SPILL    0x80

Index: target-sparc/op.c
===================================================================
RCS file: /sources/qemu/qemu/target-sparc/op.c,v
retrieving revision 1.27

op.c +4 -1
diff -u -d -p -r1.27 op.c
@@ -1486,7 +1486,10 @@ void OPPROTO op_movl_npc_im(void)
 
 void OPPROTO op_movl_npc_T0(void)
 {
-    env->npc = T0;
+    if (T0 & 0x3)
+	raise_exception(TT_UNALIGNED);
+    else
+	env->npc = T0;
 }
 
 void OPPROTO op_mov_pc_npc(void)

Index: target-sparc/op_helper.c
===================================================================
RCS file: /sources/qemu/qemu/target-sparc/op_helper.c,v retrieving revision 1.23

op_helper.c +8 -0
diff -u -d -p -r1.23 op_helper.c
@@ -872,7 +872,10 @@ void do_interrupt(int intno)
 
 #if !defined(CONFIG_USER_ONLY) 
 
+static void do_unaligned_access (target_ulong addr, int is_write, int is_user, void *retaddr);
+
 #define MMUSUFFIX _mmu
+#define ALIGNED_ONLY
 #define GETPC() (__builtin_return_address(0))
 
 #define SHIFT 0
@@ -887,6 +890,11 @@ void do_interrupt(int intno)
 #define SHIFT 3
 #include "softmmu_template.h"
 
+static void do_unaligned_access (target_ulong addr, int is_write, int is_user, void *retaddr)
+{
+	/* Comment the following line to disable mem_address_not_aligned traps */
+	raise_exception(TT_UNALIGNED);
+}
 
 /* try to fill the TLB and return an exception if error. If retaddr is
    NULL, it means that the function was called in C code (i.e. not

Index: target-sparc/translate.c
===================================================================
RCS file: /sources/qemu/qemu/target-sparc/translate.c,v retrieving revision 1.44

translate.c +0 -1
diff -u -d -p -r1.44 translate.c
@@ -25,7 +25,6 @@
    Rest of V9 instructions, VIS instructions
    NPC/PC static optimisations (use JUMP_TB when possible)
    Optimize synthetic instructions
-   Optional alignment check
    128-bit float
 */
 
Signature · 4 lines
  .''`.  Aurelien Jarno	            | GPG: 1024D/F1BCDB73
 : :' :  Debian developer           | Electrical Engineer
 `. `'   [email protected]         | [email protected]
   `-    people.debian.org/~aurel32 | www.aurel32.net