intel/compiler: Don't copy prop source mods into PICK_HIGH_32BIT
authorKenneth Graunke <kenneth@whitecape.org>
Wed, 22 Apr 2020 22:03:03 +0000 (15:03 -0700)
committerKenneth Graunke <kenneth@whitecape.org>
Thu, 23 Apr 2020 00:03:18 +0000 (17:03 -0700)
commit220f0e10d8c27684b1c91c5291bdf8a8012e2a1a
tree5920d76098d4dedeae87aa34cf912983949398c8
parentbe33cf8ad04855d37a7f756ba46c1ee49461b908
intel/compiler: Don't copy prop source mods into PICK_HIGH_32BIT

VEC4_OPCODE_PICK_HIGH_32BIT performs 32-bit UD access on a 64-bit DF
value.  abs and negate make sense on DF, but break entirely when
trying to access pieces of the value as unsigned integer dwords.

Fixes an fsign Piglit test on Ivybridge:
tests/spec/arb_gpu_shader_fp64/execution/built-in-functions/vs-sign-neg-abs

It had regressed when I removed nir_lower_to_source_modifiers, as that
caused us to start generating different code which provoked this bug.

Fixes: b7c47c4f7cf ("intel/compiler: Drop nir_lower_to_source_mods() and related handling.")
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/2817
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4691>
src/intel/compiler/brw_vec4_copy_propagation.cpp