summaryrefslogtreecommitdiffstats
path: root/gdb/Convex.notes
blob: 28d336bfed85e403a7400ac29a4edd003ecee417 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163

@node Convex,,, Top
@appendix Convex-specific info
@cindex Convex notes

Scalar registers are 64 bits long, which is a pain since
left half of an S register frequently contains noise.
Therefore there are two ways to obtain the value of an S register.

@table @kbd
@item $s0
returns the low half of the register as an int

@item $S0
returns the whole register as a long long
@end table

You can print the value in floating point by using @samp{p/f $s0} or @samp{p/f $S0}
to print a single or double precision value.

@cindex vector registers
Vector registers are handled similarly, with @samp{$V0} denoting the whole
64-bit register and @kbd{$v0} denoting the 32-bit low half; @samp{p/f $v0}
or @samp{p/f $V0} can be used to examine the register in floating point.
The length of the vector registers is taken from @samp{$vl}.  

Individual elements of a vector register are denoted in the obvious way;
@samp{print $v3[9]} prints the tenth element of register @kbd{v3}, and
@samp{set $v3[9] = 1234} alters it.

@kbd{$vl} and @kbd{$vs} are int, and @kbd{$vm} is an int vector.
Elements of @kbd{$vm} can't be assigned to.

@cindex communication registers
@kindex info comm-registers
Communication registers have names @kbd{$C0 .. $C63}, with @kbd{$c0 .. $c63}
denoting the low-order halves.  @samp{info comm-registers} will print them
all out, and tell which are locked.  (A communication register is
locked when a value is sent to it, and unlocked when the value is
received.)  Communication registers are, of course, global to all
threads, so it does not matter what the currently selected thread is.
@samp{info comm-reg @var{name}} prints just that one communication
register; @samp{name} may also be a communication register number
@samp{nn} or @samp{0xnn}.
@samp{info comm-reg @var{address}} prints the contents of the resource
structure at that address.

@kindex info psw
The command @samp{info psw} prints the processor status word @kbd{$ps}
bit by bit.

@kindex set base
GDB normally prints all integers in base 10, but the leading
@kbd{0x80000000} of pointers is intolerable in decimal, so the default
output radix has been changed to try to print addresses appropriately.
The @samp{set base} command can be used to change this.

@table @code
@item set base 10
Integer values always print in decimal.

@item set base 16
Integer values always print in hex.

@item set base
Go back to the initial state, which prints integer values in hex if they
look like pointers (specifically, if they start with 0x8 or 0xf in the
stack), otherwise in decimal.
@end table

@kindex set pipeline
When an exception such as a bus error or overflow happens, usually the PC
is several instructions ahead by the time the exception is detected.
The @samp{set pipe} command will disable this.

@table @code
@item set pipeline off
Forces serial execution of instructions; no vector chaining and no 
scalar instruction overlap.  With this, exceptions are detected with 
the PC pointing to the instruction after the one in error.

@item set pipeline on
Returns to normal, fast, execution.  This is the default.
@end table

@cindex parallel
In a parallel program, multiple threads may be executing, each
with its own registers, stack, and local memory.  When one of them
hits a breakpoint, that thread is selected.  Other threads do
not run while the thread is in the breakpoint.

@kindex 1cont
The selected thread can be single-stepped, given signals, and so
on.  Any other threads remain stopped.  When a @samp{cont} command is given,
all threads are resumed.  To resume just the selected thread, use
the command @samp{1cont}.

@kindex thread
The @samp{thread} command will show the active threads and the
instruction they are about to execute.  The selected thread is marked
with an asterisk.  The command @samp{thread @var{n}} will select thread @var{n},
shifting the debugger's attention to it for single-stepping,
registers, local memory, and so on.

@kindex info threads
The @samp{info threads} command will show what threads, if any, have
invisibly hit breakpoints or signals and are waiting to be noticed.

@kindex set parallel
The @samp{set parallel} command controls how many threads can be active.

@table @code
@item set parallel off
One thread.  Requests by the program that other threads join in
(spawn and pfork instructions) do not cause other threads to start up.
This does the same thing as the @samp{limit concurrency 1} command.

@item set parallel fixed
All CPUs are assigned to your program whenever it runs.  When it
executes a pfork or spawn instruction, it begins parallel execution
immediately.  This does the same thing as the @samp{mpa -f} command.

@item set parallel on
One or more threads.  Spawn and pfork cause CPUs to join in when and if
they are free.  This is the default.  It is very good for system
throughput, but not very good for finding bugs in parallel code.  If you
suspect a bug in parallel code, you probably want @samp{set parallel fixed.}
@end table

@subsection Limitations

WARNING: Convex GDB evaluates expressions in long long, because S
registers are 64 bits long.  However, GDB expression semantics are not
exactly C semantics.  This is a bug, strictly speaking, but it's not one I
know how to fix.  If @samp{x} is a program variable of type int, then it
is also type int to GDB, but @samp{x + 1} is long long, as is @samp{x + y}
or any other expression requiring computation.  So is the expression
@samp{1}, or any other constant.  You only really have to watch out for
calls.  The innocuous expression @samp{list_node (0x80001234)} has an
argument of type long long.  You must explicitly cast it to int.

It is not possible to continue after an uncaught fatal signal by using
@samp{signal 0}, @samp{return}, @samp{jump}, or anything else.  The difficulty is with
Unix, not GDB.

I have made no big effort to make such things as single-stepping a
@kbd{join} instruction do something reasonable.  If the program seems to
hang when doing this, type @kbd{ctrl-c} and @samp{cont}, or use
@samp{thread} to shift to a live thread.  Single-stepping a @kbd{spawn}
instruction apparently causes new threads to be born with their T bit set;
this is not handled gracefully.  When a thread has hit a breakpoint, other
threads may have invisibly hit the breakpoint in the background; if you
clear the breakpoint gdb will be surprised when threads seem to continue
to stop at it.  All of these situations produce spurious signal 5 traps;
if this happens, just type @samp{cont}.  If it becomes a nuisance, use
@samp{handle 5 nostop}.  (It will ask if you are sure.  You are.)

There is no way in GDB to store a float in a register, as with
@kbd{set $s0 = 3.1416}.  The identifier @kbd{$s0} denotes an integer,
and like any C expression which assigns to an integer variable, the
right-hand side is casted to type int.  If you should need to do
something like this, you can assign the value to @kbd{@{float@} ($sp-4)}
and then do @kbd{set $s0 = $sp[-4]}.  Same deal with @kbd{set $v0[69] = 6.9}.
OpenPOWER on IntegriCloud