23:47 <kcxt> hmm, so i'm poking around board_f, and I notice that for initf_dm we are walking the entire FDT with dcache disabled to bind/probe pre-reloc drivers

23:48 <kcxt> this takes an entire second on qc6490 qcom platforms (probably similar for all qcom platforms) since we have a HUGE DT

23:48 <kcxt> post-reloc we enable dcache and build the livetree before initr_dm

23:50 <kcxt> since we almost always already have UART configured, it's not necessary to have a device-model pre-reloc

23:50 <kcxt> if you disable SYS_MALLOC_F then initf_dm is skipped, but this configuration doesn't actually work since log_init calls calloc() unconditionally .... so unless you have no log drivers you always need some heap memory pre reloc

23:52 <kcxt> I tried simply calling dm_init(false) instead of the whole dm_init_and_scan(), this takes the time-to-uboot-shell from ~3100ms to ~2100ms on QCS6490

23:53 <kcxt> a bunch of qcom devices skip enabling UART in sbl1 unless you pull the rx pin up, so having u-boot init unconditionally pre-reloc is helpful for some bringup

23:54 <kcxt> I guess we could introduce a config option to skip setting up the device model pre-reloc and enable it by default in qcom_defconfig, but this feels non-ideal

23:56 <kcxt> better would be to enable dcache and build the livetree pre-reloc, but that populates a bunch of global variables which won't get copied during relocation

23:57 <kcxt> buuut we don't actually /need/ to relocate u-boot at all on qcom anyways (it seems this isn't even needed for efi runtime)