arm: 0.826
device: 0.770
architecture: 0.706
graphic: 0.678
register: 0.628
debug: 0.578
performance: 0.520
semantic: 0.397
kernel: 0.376
ppc: 0.346
risc-v: 0.337
vnc: 0.309
mistranslation: 0.304
PID: 0.298
boot: 0.277
files: 0.271
user-level: 0.230
socket: 0.224
TCG: 0.205
permissions: 0.190
VMM: 0.107
virtual: 0.093
network: 0.087
hypervisor: 0.048
assembly: 0.033
i386: 0.020
peripherals: 0.012
KVM: 0.006
x86: 0.001

gdbstub XML generation for ARM is done for every CPU
Description of problem:
- As arm_cpu_register_gdb_regs_for_features is called from the device
   realize stage for each vCPU in user mode we end up uselessly
   regenerating the XML for every new thread. Once you get up to 100
   threads this starts exceeding the large maps done for QHT and PageDesc
Steps to reproduce:
See above command line, valgrind picks it up
Additional information:
See also #866, #967