From c800fef2e2c65f7d81215cb316de6f73b15ba6c5 Mon Sep 17 00:00:00 2001 From: Eric Anholt Date: Tue, 18 Aug 2015 22:07:47 -0700 Subject: [PATCH] vc4: Allow unpack_8[abcd]_f's src to stay in r4. I had QPU emit code to do it, but forgot to flag the register class. total instructions in shared programs: 97974 -> 97590 (-0.39%) instructions in affected programs: 25291 -> 24907 (-1.52%) --- src/gallium/drivers/vc4/vc4_register_allocate.c | 16 +++++++++++++++- 1 file changed, 15 insertions(+), 1 deletion(-) diff --git a/src/gallium/drivers/vc4/vc4_register_allocate.c b/src/gallium/drivers/vc4/vc4_register_allocate.c index 58d812bce8a..93013147d49 100644 --- a/src/gallium/drivers/vc4/vc4_register_allocate.c +++ b/src/gallium/drivers/vc4/vc4_register_allocate.c @@ -279,7 +279,21 @@ vc4_register_allocate(struct vc4_context *vc4, struct vc4_compile *c) } if (qir_src_needs_a_file(inst)) { - class_bits[inst->src[0].index] &= CLASS_BIT_A; + switch (inst->op) { + case QOP_UNPACK_8A_F: + case QOP_UNPACK_8B_F: + case QOP_UNPACK_8C_F: + case QOP_UNPACK_8D_F: + /* Special case: these can be done as R4 + * unpacks, as well. + */ + class_bits[inst->src[0].index] &= (CLASS_BIT_A | + CLASS_BIT_R4); + break; + default: + class_bits[inst->src[0].index] &= CLASS_BIT_A; + break; + } } ip++; } -- 2.30.2