Missing PPC TRACE exception on a branch

[Christophe PLE]

I am using qemu-1.1.1-1 to emulate a PPC PREP machine on intel host linux ubuntu 11-10

In the following example, i will show that the POWERPC TRACE exception at vector 0xD00 is not taken when executing a specific kind of branch.

I have the following stand-alone assembly source code for powerpc in file SE_TEST.s

 .text
 .global _start

_start:

/* copy vector handler at address 0xD00 */
 addis 3,0,(0xD00)@ha
 addi 3,3,(0xD00)@l
 addis 4,0,(vector_handler)@ha
 addi 4,4,(vector_handler)@l
 addi 3,3,-4
 addi 4,4,-4
 lwzu 7,4(4)
 stwu 7,4(3)
 lwzu 7,4(4)
 stwu 7,4(3)
 lwzu 7,4(4)
 stwu 7,4(3)
 lwzu 7,4(4)
 stwu 7,4(3)

/* set branch address in SRR0 register */
 addis 3,0,(branch)@ha
 addi 3,3,(branch)@l
 mtspr 26,3

/* Read MSR */
/* Set SE bit and clear IP bit then set value in SSR1 */

 mfmsr 4
 ori 4,4,0x0400
 andi. 4,4,0xFFBF
 mtspr 27,4
/* Set CR condition to execute not taken branch after rfi */
 addi 5,0,0
 cmpi 0,5,0
 rfi
 nop
 nop
 nop

branch:
 bne down <- branch where the error is.
branch_plus_1:
 nop
branch_plus_2:
 nop
down: nop
 nop
 nop
 nop

vector_handler:
 mfspr 6,26
 nop
 nop
 nop

It compiles with powerpc-eabi-gcc SE_TEST.s -o SE_TEST.elf

Then I run Qemu with the command : "./qemu-system-ppc -M prep -s -S"

Then i run a cross gdb with the command: "powerpc-eabi-gdb --nx"

On gdb prompt i execute the following gdb command :

file SE_TEST.elf
target remote :1234
load SE_TEST.elf
set $pc =_start
b *0xD08
c

echo "srr0 value in trace handler "
p/x $r6
echo "address of branch_plus_1 "
p/x &branch_plus_1
echo "address of branch_plus_2 "
p/x &branch_plus_2

The gdb command windows display the result:

(gdb) so co
0xfffffffc in ?? ()
Loading section .init, size 0x24 lma 0x1800074
Loading section .text, size 0x26c lma 0x1800098
Loading section .fini, size 0x20 lma 0x1800304
Loading section .eh_frame, size 0x8 lma 0x1810324
Loading section .ctors, size 0x8 lma 0x181032c
Loading section .dtors, size 0x8 lma 0x1810334
Loading section .jcr, size 0x4 lma 0x181033c
Loading section .data, size 0x4 lma 0x1810340
Start address 0x1800200, load size 720
Transfer rate: 5760 bits in <1 sec, 90 bytes/write.
Breakpoint 1 at 0xd08

Breakpoint 1, 0x00000d08 in ?? ()
"srr0 value in trace handler "$1 = 0x1800274
"address of branch_plus_1 "$2 = 0x1800270
"address of branch_plus_2 "$3 = 0x1800274

In trace exception handler, the SRR0 register value should be the value of address just following the branch id 0x1800270.
But it is not the case, the SRR0 value is the address of the next next instruction after the branch instruction.

It seems that the singlestep exception was not taken after executing the "bne down" instruction but was taken after executing the first following nop instruction.

I have make some other test with other kind of branch, the behaviour is correct and only for a conditionnal branch when it is not taken with issue appears.

[Michael Tokarev (mjt+launchpad-tls)]

This is fixed by qemu commit f0cc4aa8450376ca2aee3ebb09db71f9f2ff333b which went into qemu-1.3 release.