[Mesa-dev] [PATCH] radv: Always lower indirect derefs after nir_lower_global_vars_to_local.

[Bas Nieuwenhuizen]

Otherwise new local variables can cause hangs on vega.

CC: <mesa-stable at lists.freedesktop.org>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105098
---
 src/amd/vulkan/radv_pipeline.c | 11 +++++--
 src/amd/vulkan/radv_shader.c   | 74 +++++++++++++++++++++++-------------------
 src/amd/vulkan/radv_shader.h   |  4 +++
 3 files changed, 53 insertions(+), 36 deletions(-)

diff --git a/src/amd/vulkan/radv_pipeline.c b/src/amd/vulkan/radv_pipeline.c
index 8f872e7c14..0d1958fc93 100644
--- a/src/amd/vulkan/radv_pipeline.c
+++ b/src/amd/vulkan/radv_pipeline.c
@@ -1525,9 +1525,16 @@ radv_link_shaders(struct radv_pipeline *pipeline, nir_shader **shaders)
 							   ordered_shaders[i - 1]);
 
 		if (progress) {
-			nir_lower_global_vars_to_local(ordered_shaders[i]);
+			if (nir_lower_global_vars_to_local(ordered_shaders[i])) {
+				radv_lower_indirect_derefs(ordered_shaders[i],
+				                           pipeline->device->physical_device);
+			}
 			radv_optimize_nir(ordered_shaders[i]);
-			nir_lower_global_vars_to_local(ordered_shaders[i - 1]);
+
+			if (nir_lower_global_vars_to_local(ordered_shaders[i - 1])) {
+				radv_lower_indirect_derefs(ordered_shaders[i - 1],
+				                           pipeline->device->physical_device);
+			}
 			radv_optimize_nir(ordered_shaders[i - 1]);
 		}
 	}
diff --git a/src/amd/vulkan/radv_shader.c b/src/amd/vulkan/radv_shader.c
index 42f83bb335..d9b8e209a9 100644
--- a/src/amd/vulkan/radv_shader.c
+++ b/src/amd/vulkan/radv_shader.c
@@ -114,6 +114,45 @@ void radv_DestroyShaderModule(
 	vk_free2(&device->alloc, pAllocator, module);
 }
 
+bool
+radv_lower_indirect_derefs(struct nir_shader *nir,
+                           struct radv_physical_device *device)
+{
+	/* While it would be nice not to have this flag, we are constrained
+	 * by the reality that LLVM 5.0 doesn't have working VGPR indexing
+	 * on GFX9.
+	 */
+	bool llvm_has_working_vgpr_indexing =
+		device->rad_info.chip_class <= VI;
+
+	/* TODO: Indirect indexing of GS inputs is unimplemented.
+	 *
+	 * TCS and TES load inputs directly from LDS or offchip memory, so
+	 * indirect indexing is trivial.
+	 */
+	nir_variable_mode indirect_mask = 0;
+	if (nir->info.stage == MESA_SHADER_GEOMETRY ||
+	    (nir->info.stage != MESA_SHADER_TESS_CTRL &&
+	     nir->info.stage != MESA_SHADER_TESS_EVAL &&
+	     !llvm_has_working_vgpr_indexing)) {
+		indirect_mask |= nir_var_shader_in;
+	}
+	if (!llvm_has_working_vgpr_indexing &&
+	    nir->info.stage != MESA_SHADER_TESS_CTRL)
+		indirect_mask |= nir_var_shader_out;
+
+	/* TODO: We shouldn't need to do this, however LLVM isn't currently
+	 * smart enough to handle indirects without causing excess spilling
+	 * causing the gpu to hang.
+	 *
+	 * See the following thread for more details of the problem:
+	 * https://lists.freedesktop.org/archives/mesa-dev/2017-July/162106.html
+	 */
+	indirect_mask |= nir_var_local;
+
+	return nir_lower_indirect_derefs(nir, indirect_mask);
+}
+
 void
 radv_optimize_nir(struct nir_shader *shader)
 {
@@ -254,40 +293,6 @@ radv_shader_compile_to_nir(struct radv_device *device,
 
 	nir_shader_gather_info(nir, entry_point->impl);
 
-	/* While it would be nice not to have this flag, we are constrained
-	 * by the reality that LLVM 5.0 doesn't have working VGPR indexing
-	 * on GFX9.
-	 */
-	bool llvm_has_working_vgpr_indexing =
-		device->physical_device->rad_info.chip_class <= VI;
-
-	/* TODO: Indirect indexing of GS inputs is unimplemented.
-	 *
-	 * TCS and TES load inputs directly from LDS or offchip memory, so
-	 * indirect indexing is trivial.
-	 */
-	nir_variable_mode indirect_mask = 0;
-	if (nir->info.stage == MESA_SHADER_GEOMETRY ||
-	    (nir->info.stage != MESA_SHADER_TESS_CTRL &&
-	     nir->info.stage != MESA_SHADER_TESS_EVAL &&
-	     !llvm_has_working_vgpr_indexing)) {
-		indirect_mask |= nir_var_shader_in;
-	}
-	if (!llvm_has_working_vgpr_indexing &&
-	    nir->info.stage != MESA_SHADER_TESS_CTRL)
-		indirect_mask |= nir_var_shader_out;
-
-	/* TODO: We shouldn't need to do this, however LLVM isn't currently
-	 * smart enough to handle indirects without causing excess spilling
-	 * causing the gpu to hang.
-	 *
-	 * See the following thread for more details of the problem:
-	 * https://lists.freedesktop.org/archives/mesa-dev/2017-July/162106.html
-	 */
-	indirect_mask |= nir_var_local;
-
-	nir_lower_indirect_derefs(nir, indirect_mask);
-
 	static const nir_lower_tex_options tex_options = {
 	  .lower_txp = ~0,
 	};
@@ -298,6 +303,7 @@ radv_shader_compile_to_nir(struct radv_device *device,
 	nir_lower_var_copies(nir);
 	nir_lower_global_vars_to_local(nir);
 	nir_remove_dead_variables(nir, nir_var_local);
+	radv_lower_indirect_derefs(nir, device->physical_device);
 	radv_optimize_nir(nir);
 
 	return nir;
diff --git a/src/amd/vulkan/radv_shader.h b/src/amd/vulkan/radv_shader.h
index b07f8a89e7..31fe739771 100644
--- a/src/amd/vulkan/radv_shader.h
+++ b/src/amd/vulkan/radv_shader.h
@@ -104,6 +104,10 @@ void
 radv_shader_variant_destroy(struct radv_device *device,
 			    struct radv_shader_variant *variant);
 
+bool
+radv_lower_indirect_derefs(struct nir_shader *nir,
+                           struct radv_physical_device *device);
+
 const char *
 radv_get_shader_name(struct radv_shader_variant *var, gl_shader_stage stage);
 
-- 
2.16.1