ATI Technologies Inc Radeon HD 4770 [RV740]
drm-radeon-testing kernel (git tip)
libdrm, mesa and xf86-video-intel git master
mesa git master = 99b9019716d0a5cfc7438677d2e11090d676c054
this is sort of copy&paste from bug #34674. The problem (fallback to software rendering) described there also happens on r600g, and probably (judging from the state tracker code) on all gallium drivers.
Anyway, from profiling I found out that pipe_put_tile_z and pipe_get_tile_z consume like 90% of the CPU time. These two get called by fallback_copy_texsubimage, which is called from st_copy_texsubimage.
Still happens with mesa git master (6bb230970cb2da5e3fe9bc63991b4b5f92445e92)
Can no longer reproduce this with up-to-date mesa git master.