While looking at issue chrome-os-partner:56243 (ChromeOS/ARM slowness), I found sometimes malloc is a hotspot, and there is a toolchain bug for handling TLS variable threadlocal_heap_ used in tcmalloc making the performance worse. That bug is filed in chromium:650137. 

I also found an additional optimization opportunity: changing the TLS model for that variable from initial-exec to local-exec if the code is in the executable, not in a shared library. Below is the code of reading the TLS variable on ARM:

(initial-exec)
  36:   4bc3            ldr     r3, [pc, #780]  ; (344 <_ZN12_GLOBAL__N_19do_mallocEj+0x344>)
  38:   ee1d 2f70       mrc     15, 0, r2, cr13, cr0, {3}
  3c:   447b            add     r3, pc
  3e:   681b            ldr     r3, [r3, #0]
  40:   58d6            ldr     r6, [r2, r3]

(local-exec)

  36:   ee1d 3f70       mrc     15, 0, r3, cr13, cr0, {3}
  3a:   4ac1            ldr     r2, [pc, #772]  ; (340 <_ZN12_GLOBAL__N_19do_mallocEj+0x340>)
  3c:   589e            ldr     r6, [r3, r2]

Changing it from initial-exec to local-exec may also sidestep the toolchain bug (not sure as it hasn't been root-caused).