[Mesa-dev] [PATCH 4/8] radeonsi: set the calling convention for inlined function calls

[Marek Olšák]

From: Marek Olšák <marek.olsak at amd.com>

otherwise the behavior is undefined
---
 src/amd/common/ac_llvm_build.c                         | 8 ++++++++
 src/amd/common/ac_llvm_build.h                         | 3 +++
 src/gallium/drivers/radeonsi/si_compute_prim_discard.c | 2 +-
 src/gallium/drivers/radeonsi/si_shader.c               | 2 +-
 4 files changed, 13 insertions(+), 2 deletions(-)

diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index 1e6247ad72e..cdd4c36f5da 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -4427,10 +4427,18 @@ ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij)
 
 LLVMValueRef
 ac_build_load_helper_invocation(struct ac_llvm_context *ctx)
 {
 	LLVMValueRef result = ac_build_intrinsic(ctx, "llvm.amdgcn.ps.live",
 						 ctx->i1, NULL, 0,
 						 AC_FUNC_ATTR_READNONE);
 	result = LLVMBuildNot(ctx->builder, result, "");
 	return LLVMBuildSExt(ctx->builder, result, ctx->i32, "");
 }
+
+LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMValueRef func,
+			   LLVMValueRef *args, unsigned num_args)
+{
+	LLVMValueRef ret = LLVMBuildCall(ctx->builder, func, args, num_args, "");
+	LLVMSetInstructionCallConv(ret, LLVMGetFunctionCallConv(func));
+	return ret;
+}
diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
index a1654d2b2c4..1928843c78c 100644
--- a/src/amd/common/ac_llvm_build.h
+++ b/src/amd/common/ac_llvm_build.h
@@ -713,20 +713,23 @@ ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
 LLVMValueRef
 ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
 		    unsigned bitsize);
 
 LLVMValueRef
 ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij);
 
 LLVMValueRef
 ac_build_load_helper_invocation(struct ac_llvm_context *ctx);
 
+LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMValueRef func,
+			   LLVMValueRef *args, unsigned num_args);
+
 LLVMValueRef ac_build_atomic_rmw(struct ac_llvm_context *ctx, LLVMAtomicRMWBinOp op,
 				 LLVMValueRef ptr, LLVMValueRef val,
 				 const char *sync_scope);
 
 LLVMValueRef ac_build_atomic_cmp_xchg(struct ac_llvm_context *ctx, LLVMValueRef ptr,
 				      LLVMValueRef cmp, LLVMValueRef val,
 				      const char *sync_scope);
 
 #ifdef __cplusplus
 }
diff --git a/src/gallium/drivers/radeonsi/si_compute_prim_discard.c b/src/gallium/drivers/radeonsi/si_compute_prim_discard.c
index 0f2934243a1..28da7b92250 100644
--- a/src/gallium/drivers/radeonsi/si_compute_prim_discard.c
+++ b/src/gallium/drivers/radeonsi/si_compute_prim_discard.c
@@ -660,21 +660,21 @@ void si_build_prim_discard_compute_shader(struct si_shader_context *ctx)
 			index[1] = LLVMBuildSelect(builder, prim_is_odd, index0, index1, "");
 		}
 	}
 
 	/* Execute the vertex shader for each vertex to get vertex positions. */
 	LLVMValueRef pos[3][4];
 	for (unsigned i = 0; i < vertices_per_prim; i++) {
 		vs_params[param_vertex_id] = index[i];
 		vs_params[param_instance_id] = instance_id;
 
-		LLVMValueRef ret = LLVMBuildCall(builder, vs, vs_params, num_vs_params, "");
+		LLVMValueRef ret = ac_build_call(&ctx->ac, vs, vs_params, num_vs_params);
 		for (unsigned chan = 0; chan < 4; chan++)
 			pos[i][chan] = LLVMBuildExtractValue(builder, ret, chan, "");
 	}
 
 	/* Divide XYZ by W. */
 	for (unsigned i = 0; i < vertices_per_prim; i++) {
 		for (unsigned chan = 0; chan < 3; chan++)
 			pos[i][chan] = ac_build_fdiv(&ctx->ac, pos[i][chan], pos[i][3]);
 	}
 
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index eb75e2a77a4..54b29d0ae01 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -6584,21 +6584,21 @@ static void si_build_wrapper_function(struct si_shader_context *ctx,
 					}
 				} else {
 					arg = LLVMBuildBitCast(builder, arg, param_type, "");
 				}
 			}
 
 			in[param_idx] = arg;
 			out_idx += param_size;
 		}
 
-		ret = LLVMBuildCall(builder, parts[part], in, num_params, "");
+		ret = ac_build_call(&ctx->ac, parts[part], in, num_params);
 
 		if (is_merged_shader(ctx) &&
 		    part + 1 == next_shader_first_part) {
 			lp_build_endif(&if_state);
 
 			/* The second half of the merged shader should use
 			 * the inputs from the toplevel (wrapper) function,
 			 * not the return value from the last call.
 			 *
 			 * That's because the last call was executed condi-
-- 
2.17.1