[Mesa-dev] [PATCH v2 1/3] ac: add support for more types with struct/raw LLVM intrinsics

[Samuel Pitoiset]

LLVM 9+ now supports 8-bit and 16-bit types.

v2: - rebase on master (without the vec3 change)

Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
---
 src/amd/common/ac_llvm_build.c | 46 +++++++++++++++++++---------------
 1 file changed, 26 insertions(+), 20 deletions(-)

diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index 8f40ddda0b5..92dd4104592 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -1135,6 +1135,7 @@ ac_build_llvm8_buffer_store_common(struct ac_llvm_context *ctx,
 				   LLVMValueRef voffset,
 				   LLVMValueRef soffset,
 				   unsigned num_channels,
+				   LLVMTypeRef base_type,
 				   bool glc,
 				   bool slc,
 				   bool writeonly_memory,
@@ -1150,18 +1151,19 @@ ac_build_llvm8_buffer_store_common(struct ac_llvm_context *ctx,
 	args[idx++] = voffset ? voffset : ctx->i32_0;
 	args[idx++] = soffset ? soffset : ctx->i32_0;
 	args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0);
-	unsigned func = CLAMP(num_channels, 1, 3) - 1;
-
-	const char *type_names[] = {"f32", "v2f32", "v4f32"};
+	unsigned func = num_channels == 3 ? 4 : num_channels;
 	const char *indexing_kind = structurized ? "struct" : "raw";
-	char name[256];
+	char name[256], type_name[8];
+
+	LLVMTypeRef type = func > 1 ? LLVMVectorType(base_type, func) : base_type;
+	ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
 
 	if (use_format) {
 		snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.format.%s",
-			 indexing_kind, type_names[func]);
+			 indexing_kind, type_name);
 	} else {
 		snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.%s",
-			 indexing_kind, type_names[func]);
+			 indexing_kind, type_name);
 	}
 
 	ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
@@ -1181,8 +1183,8 @@ ac_build_buffer_store_format(struct ac_llvm_context *ctx,
 	if (HAVE_LLVM >= 0x800) {
 		ac_build_llvm8_buffer_store_common(ctx, rsrc, data, vindex,
 						   voffset, NULL, num_channels,
-						   glc, false, writeonly_memory,
-						   true, true);
+						   ctx->f32, glc, false,
+						   writeonly_memory, true, true);
 	} else {
 		ac_build_buffer_store_common(ctx, rsrc, data, vindex, voffset,
 					     num_channels, glc, false,
@@ -1245,6 +1247,7 @@ ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
 							   ctx->i32_0,
 							   voffset, offset,
 							   num_channels,
+							   ctx->f32,
 							   glc, slc,
 							   writeonly_memory,
 							   false, false);
@@ -1320,6 +1323,7 @@ ac_build_llvm8_buffer_load_common(struct ac_llvm_context *ctx,
 				  LLVMValueRef voffset,
 				  LLVMValueRef soffset,
 				  unsigned num_channels,
+				  LLVMTypeRef base_type,
 				  bool glc,
 				  bool slc,
 				  bool can_speculate,
@@ -1334,23 +1338,22 @@ ac_build_llvm8_buffer_load_common(struct ac_llvm_context *ctx,
 	args[idx++] = voffset ? voffset : ctx->i32_0;
 	args[idx++] = soffset ? soffset : ctx->i32_0;
 	args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0);
-	unsigned func = CLAMP(num_channels, 1, 3) - 1;
-
-	LLVMTypeRef types[] = {ctx->f32, ctx->v2f32, ctx->v4f32};
-	const char *type_names[] = {"f32", "v2f32", "v4f32"};
+	unsigned func = num_channels == 3 ? 4 : num_channels;
 	const char *indexing_kind = structurized ? "struct" : "raw";
-	char name[256];
+	char name[256], type_name[8];
+
+	LLVMTypeRef type = func > 1 ? LLVMVectorType(base_type, func) : base_type;
+	ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
 
 	if (use_format) {
 		snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.format.%s",
-			 indexing_kind, type_names[func]);
+			 indexing_kind, type_name);
 	} else {
 		snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.%s",
-			 indexing_kind, type_names[func]);
+			 indexing_kind, type_name);
 	}
 
-	return ac_build_intrinsic(ctx, name, types[func], args,
-				  idx,
+	return ac_build_intrinsic(ctx, name, type, args, idx,
 				  ac_get_load_intr_attribs(can_speculate));
 }
 
@@ -1409,7 +1412,8 @@ ac_build_buffer_load(struct ac_llvm_context *ctx,
 	if (HAVE_LLVM >= 0x0800) {
 		return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex,
 							 offset, ctx->i32_0,
-							 num_channels, glc, slc,
+							 num_channels, ctx->f32,
+							 glc, slc,
 							 can_speculate, false,
 							 false);
 	}
@@ -1429,7 +1433,8 @@ LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
 {
 	if (HAVE_LLVM >= 0x800) {
 		return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
-							 num_channels, glc, false,
+							 num_channels, ctx->f32,
+							 glc, false,
 							 can_speculate, true, true);
 	}
 	return ac_build_buffer_load_common(ctx, rsrc, vindex, voffset,
@@ -1447,7 +1452,8 @@ LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
 {
 	if (HAVE_LLVM >= 0x800) {
 		return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
-							 num_channels, glc, false,
+							 num_channels, ctx->f32,
+							 glc, false,
 							 can_speculate, true, true);
 	}
 
-- 
2.21.0