Heavily improve the LLVM struct type generation to improve ABI

4 files changed, 150 insertions, 136 deletions

diff --git a/src/llvm_backend.hpp b/src/llvm_backend.hpp
index d0ba29964..3ae4e24b2 100644
--- a/src/llvm_backend.hpp
+++ b/src/llvm_backend.hpp
@@ -91,6 +91,8 @@ struct lbIncompleteDebugType {
 	LLVMMetadataRef metadata;
 };
 
+typedef Slice<i32> lbStructFieldRemapping;
+
 struct lbModule {
 	LLVMModuleRef mod;
 	LLVMContextRef ctx;
@@ -100,8 +102,9 @@ struct lbModule {
 	CheckerInfo *info;
 	AstPackage *pkg; // associated
 
-	Map<LLVMTypeRef> types; // Key: Type *
-	Map<Type *> llvm_types; // Key: LLVMTypeRef
+	Map<LLVMTypeRef> types;                             // Key: Type *
+	Map<lbStructFieldRemapping> struct_field_remapping; // Key: LLVMTypeRef or Type *
+	Map<Type *> llvm_types;                             // Key: LLVMTypeRef
 	i32 internal_type_level;
 
 	Map<lbValue>  values;           // Key: Entity *
@@ -447,7 +450,10 @@ lbCopyElisionHint lb_set_copy_elision_hint(lbProcedure *p, lbAddr const &addr, A
 void lb_reset_copy_elision_hint(lbProcedure *p, lbCopyElisionHint prev_hint);
 lbValue lb_consume_copy_elision_hint(lbProcedure *p);
 
+
 bool lb_struct_has_padding_prefix(Type *t);
+lbStructFieldRemapping lb_get_struct_remapping(lbModule *m, Type *t);
+LLVMTypeRef lb_type_padding_filler(lbModule *m, i64 padding, i64 padding_align);
 
 #define LB_STARTUP_RUNTIME_PROC_NAME   "__$startup_runtime"
 #define LB_STARTUP_TYPE_INFO_PROC_NAME "__$startup_type_info"
@@ -461,51 +467,51 @@ bool lb_struct_has_padding_prefix(Type *t);
 
 
 enum lbCallingConventionKind {
-    lbCallingConvention_C = 0,
-    lbCallingConvention_Fast = 8,
-    lbCallingConvention_Cold = 9,
-    lbCallingConvention_GHC = 10,
-    lbCallingConvention_HiPE = 11,
-    lbCallingConvention_WebKit_JS = 12,
-    lbCallingConvention_AnyReg = 13,
-    lbCallingConvention_PreserveMost = 14,
-    lbCallingConvention_PreserveAll = 15,
-    lbCallingConvention_Swift = 16,
-    lbCallingConvention_CXX_FAST_TLS = 17,
-    lbCallingConvention_FirstTargetCC = 64,
-    lbCallingConvention_X86_StdCall = 64,
-    lbCallingConvention_X86_FastCall = 65,
-    lbCallingConvention_ARM_APCS = 66,
-    lbCallingConvention_ARM_AAPCS = 67,
-    lbCallingConvention_ARM_AAPCS_VFP = 68,
-    lbCallingConvention_MSP430_INTR = 69,
-    lbCallingConvention_X86_ThisCall = 70,
-    lbCallingConvention_PTX_Kernel = 71,
-    lbCallingConvention_PTX_Device = 72,
-    lbCallingConvention_SPIR_FUNC = 75,
-    lbCallingConvention_SPIR_KERNEL = 76,
-    lbCallingConvention_Intel_OCL_BI = 77,
-    lbCallingConvention_X86_64_SysV = 78,
-    lbCallingConvention_Win64 = 79,
-    lbCallingConvention_X86_VectorCall = 80,
-    lbCallingConvention_HHVM = 81,
-    lbCallingConvention_HHVM_C = 82,
-    lbCallingConvention_X86_INTR = 83,
-    lbCallingConvention_AVR_INTR = 84,
-    lbCallingConvention_AVR_SIGNAL = 85,
-    lbCallingConvention_AVR_BUILTIN = 86,
-    lbCallingConvention_AMDGPU_VS = 87,
-    lbCallingConvention_AMDGPU_GS = 88,
-    lbCallingConvention_AMDGPU_PS = 89,
-    lbCallingConvention_AMDGPU_CS = 90,
-    lbCallingConvention_AMDGPU_KERNEL = 91,
-    lbCallingConvention_X86_RegCall = 92,
-    lbCallingConvention_AMDGPU_HS = 93,
-    lbCallingConvention_MSP430_BUILTIN = 94,
-    lbCallingConvention_AMDGPU_LS = 95,
-    lbCallingConvention_AMDGPU_ES = 96,
-    lbCallingConvention_AArch64_VectorCall = 97,
-    lbCallingConvention_MaxID = 1023,
+	lbCallingConvention_C = 0,
+	lbCallingConvention_Fast = 8,
+	lbCallingConvention_Cold = 9,
+	lbCallingConvention_GHC = 10,
+	lbCallingConvention_HiPE = 11,
+	lbCallingConvention_WebKit_JS = 12,
+	lbCallingConvention_AnyReg = 13,
+	lbCallingConvention_PreserveMost = 14,
+	lbCallingConvention_PreserveAll = 15,
+	lbCallingConvention_Swift = 16,
+	lbCallingConvention_CXX_FAST_TLS = 17,
+	lbCallingConvention_FirstTargetCC = 64,
+	lbCallingConvention_X86_StdCall = 64,
+	lbCallingConvention_X86_FastCall = 65,
+	lbCallingConvention_ARM_APCS = 66,
+	lbCallingConvention_ARM_AAPCS = 67,
+	lbCallingConvention_ARM_AAPCS_VFP = 68,
+	lbCallingConvention_MSP430_INTR = 69,
+	lbCallingConvention_X86_ThisCall = 70,
+	lbCallingConvention_PTX_Kernel = 71,
+	lbCallingConvention_PTX_Device = 72,
+	lbCallingConvention_SPIR_FUNC = 75,
+	lbCallingConvention_SPIR_KERNEL = 76,
+	lbCallingConvention_Intel_OCL_BI = 77,
+	lbCallingConvention_X86_64_SysV = 78,
+	lbCallingConvention_Win64 = 79,
+	lbCallingConvention_X86_VectorCall = 80,
+	lbCallingConvention_HHVM = 81,
+	lbCallingConvention_HHVM_C = 82,
+	lbCallingConvention_X86_INTR = 83,
+	lbCallingConvention_AVR_INTR = 84,
+	lbCallingConvention_AVR_SIGNAL = 85,
+	lbCallingConvention_AVR_BUILTIN = 86,
+	lbCallingConvention_AMDGPU_VS = 87,
+	lbCallingConvention_AMDGPU_GS = 88,
+	lbCallingConvention_AMDGPU_PS = 89,
+	lbCallingConvention_AMDGPU_CS = 90,
+	lbCallingConvention_AMDGPU_KERNEL = 91,
+	lbCallingConvention_X86_RegCall = 92,
+	lbCallingConvention_AMDGPU_HS = 93,
+	lbCallingConvention_MSP430_BUILTIN = 94,
+	lbCallingConvention_AMDGPU_LS = 95,
+	lbCallingConvention_AMDGPU_ES = 96,
+	lbCallingConvention_AArch64_VectorCall = 97,
+	lbCallingConvention_MaxID = 1023,
 };
 
 lbCallingConventionKind const lb_calling_convention_map[ProcCC_MAX] = {
diff --git a/src/llvm_backend_const.cpp b/src/llvm_backend_const.cpp
index cb9369c72..2797f7317 100644
--- a/src/llvm_backend_const.cpp
+++ b/src/llvm_backend_const.cpp
@@ -139,15 +139,12 @@ LLVMValueRef llvm_const_named_struct(lbModule *m, Type *t, LLVMValueRef *values,
 	
 	GB_ASSERT(value_count_ == bt->Struct.fields.count);
 	
-	unsigned field_offset = 0;
-	if (lb_struct_has_padding_prefix(bt)) {
-		field_offset = 1;
-	}
-
-	unsigned values_with_padding_count = field_offset + cast(unsigned)(bt->Struct.fields.count*2 + 1);
+	auto field_remapping = lb_get_struct_remapping(m, t);
+	unsigned values_with_padding_count = LLVMCountStructElementTypes(struct_type);
+	
 	LLVMValueRef *values_with_padding = gb_alloc_array(permanent_allocator(), LLVMValueRef, values_with_padding_count);
 	for (unsigned i = 0; i < value_count; i++) {
-		values_with_padding[field_offset + i*2 + 1] = values[i];
+		values_with_padding[field_remapping[i]] = values[i];
 	}
 	for (unsigned i = 0; i < values_with_padding_count; i++) {
 		if (values_with_padding[i] == nullptr) {
@@ -832,16 +829,11 @@ lbValue lb_const_value(lbModule *m, Type *type, ExactValue value, bool allow_loc
 			if (is_type_raw_union(type)) {
 				return lb_const_nil(m, original_type);
 			}
-
-			isize offset = 0;
-			if (lb_struct_has_padding_prefix(type)) {
-				offset = 1;
-			}
 			
 			LLVMTypeRef struct_type = lb_type(m, original_type);
 
-			unsigned value_count = cast(unsigned)(offset + type->Struct.fields.count*2 + 1);
-			GB_ASSERT(LLVMCountStructElementTypes(struct_type) == value_count);
+			auto field_remapping = lb_get_struct_remapping(m, type);
+			unsigned value_count = LLVMCountStructElementTypes(struct_type);
 			
 			LLVMValueRef *values = gb_alloc_array(temporary_allocator(), LLVMValueRef, value_count);
 			bool *visited = gb_alloc_array(temporary_allocator(), bool, value_count);
@@ -859,7 +851,7 @@ lbValue lb_const_value(lbModule *m, Type *type, ExactValue value, bool allow_loc
 						Selection sel = lookup_field(type, name, false);
 						Entity *f = type->Struct.fields[sel.index[0]];
 							
-						isize index = offset + f->Variable.field_index*2 + 1;
+						i32 index = field_remapping[f->Variable.field_index];
 						if (elem_type_can_be_constant(f->type)) {
 							values[index] = lb_const_value(m, f->type, tav.value, allow_local).value;
 							visited[index] = true;
@@ -874,7 +866,7 @@ lbValue lb_const_value(lbModule *m, Type *type, ExactValue value, bool allow_loc
 							val = tav.value;
 						}
 						
-						isize index = offset + f->Variable.field_index*2 + 1;
+						i32 index = field_remapping[f->Variable.field_index];
 						if (elem_type_can_be_constant(f->type)) {
 							values[index]  = lb_const_value(m, f->type, val, allow_local).value;
 							visited[index] = true;
diff --git a/src/llvm_backend_general.cpp b/src/llvm_backend_general.cpp
index 8366e0b82..01699c978 100644
--- a/src/llvm_backend_general.cpp
+++ b/src/llvm_backend_general.cpp
@@ -56,6 +56,7 @@ void lb_init_module(lbModule *m, Checker *c) {
 
 	gbAllocator a = heap_allocator();
 	map_init(&m->types, a);
+	map_init(&m->struct_field_remapping, a);
 	map_init(&m->llvm_types, a);
 	map_init(&m->values, a);
 	map_init(&m->soa_values, a);
@@ -1109,10 +1110,10 @@ lbValue lb_emit_union_tag_ptr(lbProcedure *p, lbValue u) {
 
 	LLVMTypeRef uvt = LLVMGetElementType(LLVMTypeOf(u.value));
 	unsigned element_count = LLVMCountStructElementTypes(uvt);
-	GB_ASSERT_MSG(element_count == 3, "element_count=%u (%s) != (%s)", element_count, type_to_string(ut), LLVMPrintTypeToString(uvt));
+	GB_ASSERT_MSG(element_count == 2, "element_count=%u (%s) != (%s)", element_count, type_to_string(ut), LLVMPrintTypeToString(uvt));
 
 	lbValue tag_ptr = {};
-	tag_ptr.value = LLVMBuildStructGEP(p->builder, u.value, 2, "");
+	tag_ptr.value = LLVMBuildStructGEP(p->builder, u.value, 1, "");
 	tag_ptr.type = alloc_type_pointer(tag_type);
 	return tag_ptr;
 }
@@ -1646,22 +1647,25 @@ LLVMTypeRef lb_type_internal(lbModule *m, Type *type) {
 			GB_ASSERT(field_count == 2);
 			LLVMTypeRef *fields = gb_alloc_array(temporary_allocator(), LLVMTypeRef, field_count);
 
-			LLVMTypeRef padding_type = LLVMArrayType(lb_type(m, t_uintptr), 0);
 			LLVMTypeRef entries_fields[] = {
-				padding_type,
 				lb_type(m, t_rawptr), // data
-				padding_type,
 				lb_type(m, t_int), // len
-				padding_type,
 				lb_type(m, t_int), // cap
-				padding_type,
 				lb_type(m, t_allocator), // allocator
-				padding_type,
 			};
 
 			fields[0] = lb_type(m, internal_type->Struct.fields[0]->type);
 			fields[1] = LLVMStructTypeInContext(ctx, entries_fields, gb_count_of(entries_fields), false);
-
+			
+			{ // Add this to simplify things
+				lbStructFieldRemapping entries_field_remapping = {};
+				slice_init(&entries_field_remapping, permanent_allocator(), gb_count_of(entries_fields));
+				for_array(i, entries_field_remapping) {
+					entries_field_remapping[i] = cast(i32)i;
+				}
+				map_set(&m->struct_field_remapping, hash_pointer(fields[1]), entries_field_remapping);
+			}
+			
 			return LLVMStructTypeInContext(ctx, fields, field_count, false);
 		}
 
@@ -1678,23 +1682,20 @@ LLVMTypeRef lb_type_internal(lbModule *m, Type *type) {
 				fields[1] = LLVMArrayType(lb_type(m, t_u8), size_of_union);
 				return LLVMStructTypeInContext(ctx, fields, field_count, false);
 			}
-
-			isize offset = 0;
-			if (lb_struct_has_padding_prefix(type)) {
-				offset = 1;
-			}
+			
+			lbStructFieldRemapping field_remapping = {};
+			slice_init(&field_remapping, permanent_allocator(), type->Struct.fields.count);
 
 			m->internal_type_level += 1;
 			defer (m->internal_type_level -= 1);
 
-			unsigned field_count = cast(unsigned)(offset + type->Struct.fields.count*2 + 1);
-			LLVMTypeRef *fields = gb_alloc_array(temporary_allocator(), LLVMTypeRef, field_count);
-
-			LLVMTypeRef type_u8 = lb_type(m, t_u8);
-			LLVMTypeRef type_u16 = lb_type(m, t_u16);
-			LLVMTypeRef type_u32 = lb_type(m, t_u32);
-			LLVMTypeRef type_u64 = lb_type(m, t_u64);
+			auto fields = array_make<LLVMTypeRef>(temporary_allocator(), 0, type->Struct.fields.count*2 + 2);
 			
+			if (lb_struct_has_padding_prefix(type)) {
+				LLVMTypeRef padding_offset = lb_alignment_prefix_type_hack(m, type->Struct.custom_align);
+				array_add(&fields, padding_offset);
+			}
+
 			i64 padding_offset = 0;
 			for_array(i, type->Struct.fields) {
 				GB_ASSERT(type->Struct.offsets != nullptr);
@@ -1702,27 +1703,14 @@ LLVMTypeRef lb_type_internal(lbModule *m, Type *type) {
 				Entity *field = type->Struct.fields[i];
 				i64 padding = type->Struct.offsets[i] - padding_offset;
 
-				LLVMTypeRef padding_type = nullptr;
-				if (padding_offset == 0) {
-					padding_type = lb_alignment_prefix_type_hack(m, type_align_of(type));
-				} else {
-					i64 alignment = type_align_of(field->type);
-					// NOTE(bill): limit to `[N x u64]` to prevent ABI issues
-					alignment = gb_min(alignment, 8); 
-					if (padding % alignment == 0) {
-						isize len = padding/alignment;
-						switch (alignment) {
-						case 1: padding_type = LLVMArrayType(type_u8,  cast(unsigned)len); break;
-						case 2: padding_type = LLVMArrayType(type_u16, cast(unsigned)len); break;
-						case 4: padding_type = LLVMArrayType(type_u32, cast(unsigned)len); break;
-						case 8: padding_type = LLVMArrayType(type_u64, cast(unsigned)len); break;
-						}
-					} else {
-						padding_type = LLVMArrayType(type_u8, cast(unsigned)padding);
-					}
+				if (padding != 0) {
+					LLVMTypeRef padding_type = lb_type_padding_filler(m, padding, type_align_of(field->type));
+					array_add(&fields, padding_type);					
 				}
-				fields[offset + i*2 + 0] = padding_type;
-				fields[offset + i*2 + 1] = lb_type(m, field->type);
+				
+				field_remapping[i] = cast(i32)fields.count;
+				array_add(&fields, lb_type(m, field->type));
+				
 				if (!type->Struct.is_packed) {
 					padding_offset = align_formula(padding_offset, type_align_of(field->type));
 				}
@@ -1730,17 +1718,18 @@ LLVMTypeRef lb_type_internal(lbModule *m, Type *type) {
 			}
 			
 			i64 end_padding = type_size_of(type)-padding_offset;
-			fields[field_count-1] = LLVMArrayType(type_u8, cast(unsigned)end_padding);
-
-			if (offset != 0) {
-				GB_ASSERT(offset == 1);
-				fields[0] = lb_alignment_prefix_type_hack(m, type->Struct.custom_align);
+			if (end_padding > 0) {
+				array_add(&fields, lb_type_padding_filler(m, end_padding, 1));
 			}
-			for (unsigned i = 0; i < field_count; i++) {
+
+			for_array(i, fields) {
 				GB_ASSERT(fields[i] != nullptr);
 			}
-
-			return LLVMStructTypeInContext(ctx, fields, field_count, type->Struct.is_packed);
+			
+			LLVMTypeRef struct_type = LLVMStructTypeInContext(ctx, fields.data, cast(unsigned)fields.count, type->Struct.is_packed);
+			map_set(&m->struct_field_remapping, hash_pointer(struct_type), field_remapping);
+			map_set(&m->struct_field_remapping, hash_pointer(type), field_remapping);
+			return struct_type;
 		}
 		break;
 
@@ -1755,29 +1744,25 @@ LLVMTypeRef lb_type_internal(lbModule *m, Type *type) {
 			gb_unused(size);
 
 			if (is_type_union_maybe_pointer_original_alignment(type)) {
-				LLVMTypeRef fields[1] = {lb_type(m, type->Union.variants[0])};
-				return LLVMStructTypeInContext(ctx, fields, 1, false);
+				LLVMTypeRef fields[] = {lb_type(m, type->Union.variants[0])};
+				return LLVMStructTypeInContext(ctx, fields, gb_count_of(fields), false);
 			}
 
 			unsigned block_size = cast(unsigned)type->Union.variant_block_size;
 
-			LLVMTypeRef fields[3] = {};
-			unsigned field_count = 1;
-			fields[0] = lb_alignment_prefix_type_hack(m, align);
+			auto fields = array_make<LLVMTypeRef>(temporary_allocator(), 0, 2);
 			if (is_type_union_maybe_pointer(type)) {
-				field_count += 1;
-				fields[1] = lb_type(m, type->Union.variants[0]);
+				LLVMTypeRef variant = lb_type(m, type->Union.variants[0]);
+				array_add(&fields, variant);
 			} else {
-				field_count += 2;
-				if (block_size == align) {
-					fields[1] = LLVMIntTypeInContext(m->ctx, 8*block_size);
-				} else {
-					fields[1] = LLVMArrayType(lb_type(m, t_u8), block_size);
-				}
-				fields[2] = lb_type(m, union_tag_type(type));
+				LLVMTypeRef block_type = lb_type_padding_filler(m, block_size, align);
+				LLVMTypeRef tag_type   = lb_type(m, union_tag_type(type));
+				array_add(&fields, block_type);
+				array_add(&fields, tag_type);
+				
 			}
-
-			return LLVMStructTypeInContext(ctx, fields, field_count, false);
+			
+			return LLVMStructTypeInContext(ctx, fields.data, cast(unsigned)fields.count, false);
 		}
 		break;
 
diff --git a/src/llvm_backend_utility.cpp b/src/llvm_backend_utility.cpp
index d1613a354..470347e82 100644
--- a/src/llvm_backend_utility.cpp
+++ b/src/llvm_backend_utility.cpp
@@ -807,25 +807,56 @@ lbValue lb_address_from_load(lbProcedure *p, lbValue value) {
 	return {};
 }
 
+
 bool lb_struct_has_padding_prefix(Type *t) {
 	Type *bt = base_type(t);
 	GB_ASSERT(bt->kind == Type_Struct);
 	return bt->Struct.custom_align != 0 && bt->Struct.fields.count == 0;
 }
+lbStructFieldRemapping lb_get_struct_remapping(lbModule *m, Type *t) {
+	t = base_type(t);
+	LLVMTypeRef struct_type = lb_type(m, t);
+	auto *field_remapping = map_get(&m->struct_field_remapping, hash_pointer(struct_type));
+	if (field_remapping == nullptr) {
+		field_remapping = map_get(&m->struct_field_remapping, hash_pointer(t));
+	}
+	GB_ASSERT(field_remapping != nullptr);
+	return *field_remapping;
+}
 
 i32 lb_convert_struct_index(lbModule *m, Type *t, i32 index) {
 	if (t->kind == Type_Struct) {
-		index = index*2 + 1;
-		if (lb_struct_has_padding_prefix(t)) {
-			index += 1;
+		auto field_remapping = lb_get_struct_remapping(m, t);
+		index = field_remapping[index];
+	}
+	return index;
+}
+
+LLVMTypeRef lb_type_padding_filler(lbModule *m, i64 padding, i64 padding_align) {
+	// NOTE(bill): limit to `[N x u64]` to prevent ABI issues
+	padding_align = gb_clamp(padding_align, 1, 8);
+	if (padding % padding_align == 0) {
+		LLVMTypeRef elem = nullptr;
+		isize len = padding/padding_align;
+		switch (padding_align) {
+		case 1: elem = lb_type(m, t_u8);  break;
+		case 2: elem = lb_type(m, t_u16); break;
+		case 4: elem = lb_type(m, t_u32); break;
+		case 8: elem = lb_type(m, t_u64); break;
 		}
 		
-		unsigned count = LLVMCountStructElementTypes(lb_type(m, t));
-		GB_ASSERT(count >= cast(unsigned)index);
+		GB_ASSERT_MSG(elem != nullptr, "Invalid lb_type_padding_filler padding and padding_align: %lld", padding_align);
+		if (len != 1) {
+			return LLVMArrayType(elem, cast(unsigned)len);
+		} else {
+			return elem;
+		}
+	} else {
+		return LLVMArrayType(lb_type(m, t_u8), cast(unsigned)padding);
 	}
-	return index;
 }
 
+
 char const *llvm_type_kinds[] = {
 	"LLVMVoidTypeKind",
 	"LLVMHalfTypeKind",
@@ -912,7 +943,6 @@ lbValue lb_emit_struct_ep(lbProcedure *p, lbValue s, i32 index) {
 		case 0: result_type = get_struct_field_type(gst, 0); break;
 		case 1: result_type = get_struct_field_type(gst, 1); break;
 		}
-		index = index*2 + 1;
 	} else if (is_type_array(t)) {
 		return lb_emit_array_epi(p, s, index);
 	} else if (is_type_relative_slice(t)) {
@@ -926,6 +956,7 @@ lbValue lb_emit_struct_ep(lbProcedure *p, lbValue s, i32 index) {
 
 	GB_ASSERT_MSG(result_type != nullptr, "%s %d", type_to_string(t), index);
 	
+	i32 original_index = index;
 	index = lb_convert_struct_index(p->module, t, index);
 	
 	if (lb_is_const(s)) {
@@ -943,7 +974,7 @@ lbValue lb_emit_struct_ep(lbProcedure *p, lbValue s, i32 index) {
 		// gb_printf_err("%d\n", index);
 		GB_ASSERT_MSG(LLVMGetTypeKind(st) == LLVMStructTypeKind, "%s", llvm_type_kinds[LLVMGetTypeKind(st)]);
 		unsigned count = LLVMCountStructElementTypes(st);
-		GB_ASSERT(count >= cast(unsigned)index);
+		GB_ASSERT_MSG(count >= cast(unsigned)index, "%u %d %d", count, index, original_index);
 		
 		res.value = LLVMBuildStructGEP(p->builder, s.value, cast(unsigned)index, "");
 		res.type = alloc_type_pointer(result_type);