[Mesa-dev] [PATCH 1/4] radeonsi: make emit_waitcnt externally accessible

[Nicolai Hähnle]

From: Nicolai Hähnle <nicolai.haehnle at amd.com>

---
 src/gallium/drivers/radeonsi/si_shader.c          | 15 +++++----------
 src/gallium/drivers/radeonsi/si_shader_internal.h |  7 +++++++
 2 files changed, 12 insertions(+), 10 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 837cc1c..39ce080 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -3400,26 +3400,21 @@ static void emit_optimization_barrier(struct si_shader_context *ctx,
 		vgpr = LLVMBuildBitCast(builder, vgpr, LLVMVectorType(ctx->i32, vgpr_size / 4), "");
 		vgpr0 = LLVMBuildExtractElement(builder, vgpr, ctx->i32_0, "");
 		vgpr0 = LLVMBuildCall(builder, inlineasm, &vgpr0, 1, "");
 		vgpr = LLVMBuildInsertElement(builder, vgpr, vgpr0, ctx->i32_0, "");
 		vgpr = LLVMBuildBitCast(builder, vgpr, vgpr_type, "");
 
 		*pvgpr = vgpr;
 	}
 }
 
-/* Combine these with & instead of |. */
-#define NOOP_WAITCNT 0xf7f
-#define LGKM_CNT 0x07f
-#define VM_CNT 0xf70
-
-static void emit_waitcnt(struct si_shader_context *ctx, unsigned simm16)
+void si_emit_waitcnt(struct si_shader_context *ctx, unsigned simm16)
 {
 	struct gallivm_state *gallivm = &ctx->gallivm;
 	LLVMBuilderRef builder = gallivm->builder;
 	LLVMValueRef args[1] = {
 		LLVMConstInt(ctx->i32, simm16, 0)
 	};
 	lp_build_intrinsic(builder, "llvm.amdgcn.s.waitcnt",
 			   ctx->voidt, args, 1, 0);
 }
 
@@ -3438,21 +3433,21 @@ static void membar_emit(
 
 	if (flags & (TGSI_MEMBAR_ATOMIC_BUFFER |
 		     TGSI_MEMBAR_SHADER_BUFFER |
 		     TGSI_MEMBAR_SHADER_IMAGE))
 		waitcnt &= VM_CNT;
 
 	if (flags & TGSI_MEMBAR_SHARED)
 		waitcnt &= LGKM_CNT;
 
 	if (waitcnt != NOOP_WAITCNT)
-		emit_waitcnt(ctx, waitcnt);
+		si_emit_waitcnt(ctx, waitcnt);
 }
 
 static void clock_emit(
 		const struct lp_build_tgsi_action *action,
 		struct lp_build_tgsi_context *bld_base,
 		struct lp_build_emit_data *emit_data)
 {
 	struct si_shader_context *ctx = si_shader_context(bld_base);
 	struct gallivm_state *gallivm = &ctx->gallivm;
 	LLVMValueRef tmp;
@@ -3948,21 +3943,21 @@ static void load_emit(
 	const struct tgsi_shader_info *info = &ctx->shader->selector->info;
 	char intrinsic_name[64];
 	bool readonly_memory = false;
 
 	if (inst->Src[0].Register.File == TGSI_FILE_MEMORY) {
 		load_emit_memory(ctx, emit_data);
 		return;
 	}
 
 	if (inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE)
-		emit_waitcnt(ctx, VM_CNT);
+		si_emit_waitcnt(ctx, VM_CNT);
 
 	readonly_memory = !(inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE) &&
 			  is_oneway_access_only(inst, info,
 						info->shader_buffers_store |
 						info->shader_buffers_atomic,
 						info->images_store |
 						info->images_atomic);
 
 	if (inst->Src[0].Register.File == TGSI_FILE_BUFFER) {
 		load_emit_buffer(ctx, emit_data, readonly_memory);
@@ -4165,21 +4160,21 @@ static void store_emit(
 	unsigned target = inst->Memory.Texture;
 	char intrinsic_name[64];
 	bool writeonly_memory = false;
 
 	if (inst->Dst[0].Register.File == TGSI_FILE_MEMORY) {
 		store_emit_memory(ctx, emit_data);
 		return;
 	}
 
 	if (inst->Memory.Qualifier & TGSI_MEMORY_VOLATILE)
-		emit_waitcnt(ctx, VM_CNT);
+		si_emit_waitcnt(ctx, VM_CNT);
 
 	writeonly_memory = is_oneway_access_only(inst, info,
 						 info->shader_buffers_load |
 						 info->shader_buffers_atomic,
 						 info->images_load |
 						 info->images_atomic);
 
 	if (inst->Dst[0].Register.File == TGSI_FILE_BUFFER) {
 		store_emit_buffer(ctx, emit_data, writeonly_memory);
 		return;
@@ -5682,21 +5677,21 @@ static void si_llvm_emit_barrier(const struct lp_build_tgsi_action *action,
 {
 	struct si_shader_context *ctx = si_shader_context(bld_base);
 	struct gallivm_state *gallivm = &ctx->gallivm;
 
 	/* SI only (thanks to a hw bug workaround):
 	 * The real barrier instruction isn’t needed, because an entire patch
 	 * always fits into a single wave.
 	 */
 	if (ctx->screen->b.chip_class == SI &&
 	    ctx->type == PIPE_SHADER_TESS_CTRL) {
-		emit_waitcnt(ctx, LGKM_CNT & VM_CNT);
+		si_emit_waitcnt(ctx, LGKM_CNT & VM_CNT);
 		return;
 	}
 
 	lp_build_intrinsic(gallivm->builder,
 			   "llvm.amdgcn.s.barrier",
 			   ctx->voidt, NULL, 0, LP_FUNC_ATTR_CONVERGENT);
 }
 
 static const struct lp_build_tgsi_action tex_action = {
 	.fetch_args = tex_fetch_args,
diff --git a/src/gallium/drivers/radeonsi/si_shader_internal.h b/src/gallium/drivers/radeonsi/si_shader_internal.h
index 35315ca..c3913de 100644
--- a/src/gallium/drivers/radeonsi/si_shader_internal.h
+++ b/src/gallium/drivers/radeonsi/si_shader_internal.h
@@ -280,13 +280,20 @@ LLVMValueRef si_llvm_emit_fetch_64bit(struct lp_build_tgsi_context *bld_base,
 LLVMValueRef si_llvm_emit_fetch(struct lp_build_tgsi_context *bld_base,
 				const struct tgsi_full_src_register *reg,
 				enum tgsi_opcode_type type,
 				unsigned swizzle);
 
 void si_llvm_emit_store(struct lp_build_tgsi_context *bld_base,
 			const struct tgsi_full_instruction *inst,
 			const struct tgsi_opcode_info *info,
 			LLVMValueRef dst[4]);
 
+/* Combine these with & instead of |. */
+#define NOOP_WAITCNT 0xf7f
+#define LGKM_CNT 0x07f
+#define VM_CNT 0xf70
+
+void si_emit_waitcnt(struct si_shader_context *ctx, unsigned simm16);
+
 void si_shader_context_init_alu(struct lp_build_tgsi_context *bld_base);
 
 #endif
-- 
2.9.3