broadcom: fix GS load_input locations
This fixes an underneath bug exposed by !9050 (merged).
VS (and VS_BIN specifically) has a lowering pass for output stores that converts from driver locations to VPM offsets. GS (and GS_BIN specifically) needs to deal with offsets when loading inputs.
The bug is due the way GS computes this offsets. Roughly speaking, VS can store VARYING_SLOT_VAR0.xyz
and VARYING_SLOT_VAR1.xyzw
, but not VARYING_SLOT_VAR0.w
. This "hole" is not included in the VPM offsets, so the offsets for VARYING_SLOT_VAR1.xyzw
are not the same as if VARYING_SLOT_VAR0.w
were also stored.