With gdb build from trunk on s390x-linux with target board unix/-m31, I run into: ... (gdb) PASS: gdb.guile/scm-lazy-string.exp: bad length print ptr^M $1 = 0x804006b0 <error: Cannot access memory at address 0x804006b0>^M (gdb) FAIL: gdb.guile/scm-lazy-string.exp: ptr: print ptr ... A minimal example is: ... $ gdb -q -batch -ex "set trace-commands on" -x gdb.in +file scm-lazy-string/scm-lazy-string +break main Breakpoint 1 at 0x4005d2: file scm-lazy-string.c, line 23. +run Breakpoint 1, main () at scm-lazy-string.c:23 23 const char *ptr = "pointer"; +step 24 const char array[] = "array"; +print ptr $1 = 0x804006b0 <error: Cannot access memory at address 0x804006b0> ... If we delete the breakpoint after running to it, we have instead the expected: ... $ gdb -q -batch -ex "set trace-commands on" -x gdb.in +file scm-lazy-string +break main Breakpoint 1 at 0x4005d2: file scm-lazy-string.c, line 23. +run Breakpoint 1, main () at scm-lazy-string.c:23 23 const char *ptr = "pointer"; +delete +step 24 const char array[] = "array"; +print ptr $1 = 0x4006b0 "pointer" ...
The problem is in displaced stepping over insn: ... │B+> 0x4005d2 <main+10> larl %r1,0x4006b0 │... In the failing case we have: ... (gdb) si (gdb) p /x $r1 $1 = 0x3ff804006b0 (gdb) ... In the passing case we have: ... (gdb) si (gdb) p /x $r1 $2 = 0x3ff004006b0 ... The difference comes from this code in s390_displaced_step_fixup: ... /* Handle LOAD ADDRESS RELATIVE LONG. */ else if (is_ril (insn, op1_larl, op2_larl, &r1, &i2)) { /* Update PC. */ regcache_write_pc (regs, from + insnlen); /* Recompute output address in R1. */ regcache_cooked_write_unsigned (regs, S390_R0_REGNUM + r1, amode | (from + i2 * 2)); } ... where the "amode |" adds the 0x80000000. Using this patch: ... diff --git a/gdb/s390-tdep.c b/gdb/s390-tdep.c index d0dba7654bb..11b309dc817 100644 --- a/gdb/s390-tdep.c +++ b/gdb/s390-tdep.c @@ -557,7 +557,7 @@ s390_displaced_step_fixup (struct gdbarch *gdbarch, regcache_write_pc (regs, from + insnlen); /* Recompute output address in R1. */ regcache_cooked_write_unsigned (regs, S390_R0_REGNUM + r1, - amode | (from + i2 * 2)); + (from + i2 * 2)); } /* If we executed a breakpoint instruction, point PC right back at it. */ ... the test-case passes. I don't understand the architecture well enough to decide what the root cause is and whether this patch addresses it.
Ah right. LARL does *not* set the amode bit in 31-bit mode (in contrast to BASR / BRASL). The patch looks good to me.
https://sourceware.org/git/?p=binutils-gdb.git;a=search;s=Tom+de+Vries;st=author