Mesa (master): ac/nir: avoid loading unused VS input components

[Samuel Pitoiset]

Module: Mesa
Branch: master
Commit: 2f54d7382db3e3a23fb1b3e61c246931828309e5
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=2f54d7382db3e3a23fb1b3e61c246931828309e5

Author: Samuel Pitoiset <samuel.pitoiset at gmail.com>
Date:   Thu Feb  1 11:35:06 2018 +0100

ac/nir: avoid loading unused VS input components

Polaris10:
Totals from affected shaders:
SGPRS: 122840 -> 120984 (-1.51 %)
VGPRS: 78812 -> 78440 (-0.47 %)
Spilled SGPRs: 177 -> 129 (-27.12 %)
Code Size: 2950028 -> 2941276 (-0.30 %) bytes
Max Waves: 17899 -> 17976 (0.43 %)

Vega10:
Totals from affected shaders:
SGPRS: 117144 -> 115776 (-1.17 %)
VGPRS: 77580 -> 77532 (-0.06 %)
Spilled SGPRs: 0 -> 152 (0.00 %)
Code Size: 3352656 -> 3347860 (-0.14 %) bytes
Max Waves: 19756 -> 19866 (0.56 %)

This increases SGPRs spilling a bit with Talos, but I have
some other ideas that might reduce it.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas at basnieuwenhuizen.nl>

---

 src/amd/common/ac_nir_to_llvm.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index e44afbb9b2..15d6759ced 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -5317,6 +5317,9 @@ handle_vs_input_decl(struct nir_to_llvm_context *ctx,
 	int index = variable->data.location - VERT_ATTRIB_GENERIC0;
 	int idx = variable->data.location;
 	unsigned attrib_count = glsl_count_attribute_slots(variable->type, true);
+	uint8_t input_usage_mask =
+		ctx->shader_info->info.vs.input_usage_mask[variable->data.location];
+	unsigned num_channels = util_last_bit(input_usage_mask);
 
 	variable->data.driver_location = idx * 4;
 
@@ -5341,7 +5344,9 @@ handle_vs_input_decl(struct nir_to_llvm_context *ctx,
 		input = ac_build_buffer_load_format(&ctx->ac, t_list,
 						    buffer_index,
 						    ctx->ac.i32_0,
-						    4, false, true);
+						    num_channels, false, true);
+
+		input = ac_build_expand_to_vec4(&ctx->ac, input, num_channels);
 
 		for (unsigned chan = 0; chan < 4; chan++) {
 			LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, chan, false);