summaryrefslogtreecommitdiffhomepage
path: root/src/3p/chibicc/codegen.c
diff options
context:
space:
mode:
Diffstat (limited to 'src/3p/chibicc/codegen.c')
-rw-r--r--src/3p/chibicc/codegen.c1595
1 files changed, 1595 insertions, 0 deletions
diff --git a/src/3p/chibicc/codegen.c b/src/3p/chibicc/codegen.c
new file mode 100644
index 0000000..da11fd7
--- /dev/null
+++ b/src/3p/chibicc/codegen.c
@@ -0,0 +1,1595 @@
+#include "chibicc.h"
+
+#define GP_MAX 6
+#define FP_MAX 8
+
+static FILE *output_file;
+static int depth;
+static char *argreg8[] = {"%dil", "%sil", "%dl", "%cl", "%r8b", "%r9b"};
+static char *argreg16[] = {"%di", "%si", "%dx", "%cx", "%r8w", "%r9w"};
+static char *argreg32[] = {"%edi", "%esi", "%edx", "%ecx", "%r8d", "%r9d"};
+static char *argreg64[] = {"%rdi", "%rsi", "%rdx", "%rcx", "%r8", "%r9"};
+static Obj *current_fn;
+
+static void gen_expr(Node *node);
+static void gen_stmt(Node *node);
+
+__attribute__((format(printf, 1, 2)))
+static void println(char *fmt, ...) {
+ va_list ap;
+ va_start(ap, fmt);
+ vfprintf(output_file, fmt, ap);
+ va_end(ap);
+ fprintf(output_file, "\n");
+}
+
+static int count(void) {
+ static int i = 1;
+ return i++;
+}
+
+static void push(void) {
+ println(" push %%rax");
+ depth++;
+}
+
+static void pop(char *arg) {
+ println(" pop %s", arg);
+ depth--;
+}
+
+static void pushf(void) {
+ println(" sub $8, %%rsp");
+ println(" movsd %%xmm0, (%%rsp)");
+ depth++;
+}
+
+static void popf(int reg) {
+ println(" movsd (%%rsp), %%xmm%d", reg);
+ println(" add $8, %%rsp");
+ depth--;
+}
+
+// Round up `n` to the nearest multiple of `align`. For instance,
+// align_to(5, 8) returns 8 and align_to(11, 8) returns 16.
+int align_to(int n, int align) {
+ return (n + align - 1) / align * align;
+}
+
+static char *reg_dx(int sz) {
+ switch (sz) {
+ case 1: return "%dl";
+ case 2: return "%dx";
+ case 4: return "%edx";
+ case 8: return "%rdx";
+ }
+ unreachable();
+}
+
+static char *reg_ax(int sz) {
+ switch (sz) {
+ case 1: return "%al";
+ case 2: return "%ax";
+ case 4: return "%eax";
+ case 8: return "%rax";
+ }
+ unreachable();
+}
+
+// Compute the absolute address of a given node.
+// It's an error if a given node does not reside in memory.
+static void gen_addr(Node *node) {
+ switch (node->kind) {
+ case ND_VAR:
+ // Variable-length array, which is always local.
+ if (node->var->ty->kind == TY_VLA) {
+ println(" mov %d(%%rbp), %%rax", node->var->offset);
+ return;
+ }
+
+ // Local variable
+ if (node->var->is_local) {
+ println(" lea %d(%%rbp), %%rax", node->var->offset);
+ return;
+ }
+
+ if (opt_fpic) {
+ // Thread-local variable
+ if (node->var->is_tls) {
+ println(" data16 lea %s@tlsgd(%%rip), %%rdi", node->var->name);
+ println(" .value 0x6666");
+ println(" rex64");
+ println(" call __tls_get_addr@PLT");
+ return;
+ }
+
+ // Function or global variable
+ println(" mov %s@GOTPCREL(%%rip), %%rax", node->var->name);
+ return;
+ }
+
+ // Thread-local variable
+ if (node->var->is_tls) {
+ println(" mov %%fs:0, %%rax");
+ println(" add $%s@tpoff, %%rax", node->var->name);
+ return;
+ }
+
+ // Here, we generate an absolute address of a function or a global
+ // variable. Even though they exist at a certain address at runtime,
+ // their addresses are not known at link-time for the following
+ // two reasons.
+ //
+ // - Address randomization: Executables are loaded to memory as a
+ // whole but it is not known what address they are loaded to.
+ // Therefore, at link-time, relative address in the same
+ // exectuable (i.e. the distance between two functions in the
+ // same executable) is known, but the absolute address is not
+ // known.
+ //
+ // - Dynamic linking: Dynamic shared objects (DSOs) or .so files
+ // are loaded to memory alongside an executable at runtime and
+ // linked by the runtime loader in memory. We know nothing
+ // about addresses of global stuff that may be defined by DSOs
+ // until the runtime relocation is complete.
+ //
+ // In order to deal with the former case, we use RIP-relative
+ // addressing, denoted by `(%rip)`. For the latter, we obtain an
+ // address of a stuff that may be in a shared object file from the
+ // Global Offset Table using `@GOTPCREL(%rip)` notation.
+
+ // Function
+ if (node->ty->kind == TY_FUNC) {
+ if (node->var->is_definition)
+ println(" lea %s(%%rip), %%rax", node->var->name);
+ else
+ println(" mov %s@GOTPCREL(%%rip), %%rax", node->var->name);
+ return;
+ }
+
+ // Global variable
+ println(" lea %s(%%rip), %%rax", node->var->name);
+ return;
+ case ND_DEREF:
+ gen_expr(node->lhs);
+ return;
+ case ND_COMMA:
+ gen_expr(node->lhs);
+ gen_addr(node->rhs);
+ return;
+ case ND_MEMBER:
+ gen_addr(node->lhs);
+ println(" add $%d, %%rax", node->member->offset);
+ return;
+ case ND_FUNCALL:
+ if (node->ret_buffer) {
+ gen_expr(node);
+ return;
+ }
+ break;
+ case ND_ASSIGN:
+ case ND_COND:
+ if (node->ty->kind == TY_STRUCT || node->ty->kind == TY_UNION) {
+ gen_expr(node);
+ return;
+ }
+ break;
+ case ND_VLA_PTR:
+ println(" lea %d(%%rbp), %%rax", node->var->offset);
+ return;
+ }
+
+ error_tok(node->tok, "not an lvalue");
+}
+
+// Load a value from where %rax is pointing to.
+static void load(Type *ty) {
+ switch (ty->kind) {
+ case TY_ARRAY:
+ case TY_STRUCT:
+ case TY_UNION:
+ case TY_FUNC:
+ case TY_VLA:
+ // If it is an array, do not attempt to load a value to the
+ // register because in general we can't load an entire array to a
+ // register. As a result, the result of an evaluation of an array
+ // becomes not the array itself but the address of the array.
+ // This is where "array is automatically converted to a pointer to
+ // the first element of the array in C" occurs.
+ return;
+ case TY_FLOAT:
+ println(" movss (%%rax), %%xmm0");
+ return;
+ case TY_DOUBLE:
+ println(" movsd (%%rax), %%xmm0");
+ return;
+ case TY_LDOUBLE:
+ println(" fldt (%%rax)");
+ return;
+ }
+
+ char *insn = ty->is_unsigned ? "movz" : "movs";
+
+ // When we load a char or a short value to a register, we always
+ // extend them to the size of int, so we can assume the lower half of
+ // a register always contains a valid value. The upper half of a
+ // register for char, short and int may contain garbage. When we load
+ // a long value to a register, it simply occupies the entire register.
+ if (ty->size == 1)
+ println(" %sbl (%%rax), %%eax", insn);
+ else if (ty->size == 2)
+ println(" %swl (%%rax), %%eax", insn);
+ else if (ty->size == 4)
+ println(" movsxd (%%rax), %%rax");
+ else
+ println(" mov (%%rax), %%rax");
+}
+
+// Store %rax to an address that the stack top is pointing to.
+static void store(Type *ty) {
+ pop("%rdi");
+
+ switch (ty->kind) {
+ case TY_STRUCT:
+ case TY_UNION:
+ for (int i = 0; i < ty->size; i++) {
+ println(" mov %d(%%rax), %%r8b", i);
+ println(" mov %%r8b, %d(%%rdi)", i);
+ }
+ return;
+ case TY_FLOAT:
+ println(" movss %%xmm0, (%%rdi)");
+ return;
+ case TY_DOUBLE:
+ println(" movsd %%xmm0, (%%rdi)");
+ return;
+ case TY_LDOUBLE:
+ println(" fstpt (%%rdi)");
+ return;
+ }
+
+ if (ty->size == 1)
+ println(" mov %%al, (%%rdi)");
+ else if (ty->size == 2)
+ println(" mov %%ax, (%%rdi)");
+ else if (ty->size == 4)
+ println(" mov %%eax, (%%rdi)");
+ else
+ println(" mov %%rax, (%%rdi)");
+}
+
+static void cmp_zero(Type *ty) {
+ switch (ty->kind) {
+ case TY_FLOAT:
+ println(" xorps %%xmm1, %%xmm1");
+ println(" ucomiss %%xmm1, %%xmm0");
+ return;
+ case TY_DOUBLE:
+ println(" xorpd %%xmm1, %%xmm1");
+ println(" ucomisd %%xmm1, %%xmm0");
+ return;
+ case TY_LDOUBLE:
+ println(" fldz");
+ println(" fucomip");
+ println(" fstp %%st(0)");
+ return;
+ }
+
+ if (is_integer(ty) && ty->size <= 4)
+ println(" cmp $0, %%eax");
+ else
+ println(" cmp $0, %%rax");
+}
+
+enum { I8, I16, I32, I64, U8, U16, U32, U64, F32, F64, F80 };
+
+static int getTypeId(Type *ty) {
+ switch (ty->kind) {
+ case TY_CHAR:
+ return ty->is_unsigned ? U8 : I8;
+ case TY_SHORT:
+ return ty->is_unsigned ? U16 : I16;
+ case TY_INT:
+ return ty->is_unsigned ? U32 : I32;
+ case TY_LONG:
+ return ty->is_unsigned ? U64 : I64;
+ case TY_FLOAT:
+ return F32;
+ case TY_DOUBLE:
+ return F64;
+ case TY_LDOUBLE:
+ return F80;
+ }
+ return U64;
+}
+
+// The table for type casts
+static char i32i8[] = "movsbl %al, %eax";
+static char i32u8[] = "movzbl %al, %eax";
+static char i32i16[] = "movswl %ax, %eax";
+static char i32u16[] = "movzwl %ax, %eax";
+static char i32f32[] = "cvtsi2ssl %eax, %xmm0";
+static char i32i64[] = "movsxd %eax, %rax";
+static char i32f64[] = "cvtsi2sdl %eax, %xmm0";
+static char i32f80[] = "mov %eax, -4(%rsp); fildl -4(%rsp)";
+
+static char u32f32[] = "mov %eax, %eax; cvtsi2ssq %rax, %xmm0";
+static char u32i64[] = "mov %eax, %eax";
+static char u32f64[] = "mov %eax, %eax; cvtsi2sdq %rax, %xmm0";
+static char u32f80[] = "mov %eax, %eax; mov %rax, -8(%rsp); fildll -8(%rsp)";
+
+static char i64f32[] = "cvtsi2ssq %rax, %xmm0";
+static char i64f64[] = "cvtsi2sdq %rax, %xmm0";
+static char i64f80[] = "movq %rax, -8(%rsp); fildll -8(%rsp)";
+
+static char u64f32[] = "cvtsi2ssq %rax, %xmm0";
+static char u64f64[] =
+ "test %rax,%rax; js 1f; pxor %xmm0,%xmm0; cvtsi2sd %rax,%xmm0; jmp 2f; "
+ "1: mov %rax,%rdi; and $1,%eax; pxor %xmm0,%xmm0; shr %rdi; "
+ "or %rax,%rdi; cvtsi2sd %rdi,%xmm0; addsd %xmm0,%xmm0; 2:";
+static char u64f80[] =
+ "mov %rax, -8(%rsp); fildq -8(%rsp); test %rax, %rax; jns 1f;"
+ "mov $1602224128, %eax; mov %eax, -4(%rsp); fadds -4(%rsp); 1:";
+
+static char f32i8[] = "cvttss2sil %xmm0, %eax; movsbl %al, %eax";
+static char f32u8[] = "cvttss2sil %xmm0, %eax; movzbl %al, %eax";
+static char f32i16[] = "cvttss2sil %xmm0, %eax; movswl %ax, %eax";
+static char f32u16[] = "cvttss2sil %xmm0, %eax; movzwl %ax, %eax";
+static char f32i32[] = "cvttss2sil %xmm0, %eax";
+static char f32u32[] = "cvttss2siq %xmm0, %rax";
+static char f32i64[] = "cvttss2siq %xmm0, %rax";
+static char f32u64[] = "cvttss2siq %xmm0, %rax";
+static char f32f64[] = "cvtss2sd %xmm0, %xmm0";
+static char f32f80[] = "movss %xmm0, -4(%rsp); flds -4(%rsp)";
+
+static char f64i8[] = "cvttsd2sil %xmm0, %eax; movsbl %al, %eax";
+static char f64u8[] = "cvttsd2sil %xmm0, %eax; movzbl %al, %eax";
+static char f64i16[] = "cvttsd2sil %xmm0, %eax; movswl %ax, %eax";
+static char f64u16[] = "cvttsd2sil %xmm0, %eax; movzwl %ax, %eax";
+static char f64i32[] = "cvttsd2sil %xmm0, %eax";
+static char f64u32[] = "cvttsd2siq %xmm0, %rax";
+static char f64i64[] = "cvttsd2siq %xmm0, %rax";
+static char f64u64[] = "cvttsd2siq %xmm0, %rax";
+static char f64f32[] = "cvtsd2ss %xmm0, %xmm0";
+static char f64f80[] = "movsd %xmm0, -8(%rsp); fldl -8(%rsp)";
+
+#define FROM_F80_1 \
+ "fnstcw -10(%rsp); movzwl -10(%rsp), %eax; or $12, %ah; " \
+ "mov %ax, -12(%rsp); fldcw -12(%rsp); "
+
+#define FROM_F80_2 " -24(%rsp); fldcw -10(%rsp); "
+
+static char f80i8[] = FROM_F80_1 "fistps" FROM_F80_2 "movsbl -24(%rsp), %eax";
+static char f80u8[] = FROM_F80_1 "fistps" FROM_F80_2 "movzbl -24(%rsp), %eax";
+static char f80i16[] = FROM_F80_1 "fistps" FROM_F80_2 "movzbl -24(%rsp), %eax";
+static char f80u16[] = FROM_F80_1 "fistpl" FROM_F80_2 "movswl -24(%rsp), %eax";
+static char f80i32[] = FROM_F80_1 "fistpl" FROM_F80_2 "mov -24(%rsp), %eax";
+static char f80u32[] = FROM_F80_1 "fistpl" FROM_F80_2 "mov -24(%rsp), %eax";
+static char f80i64[] = FROM_F80_1 "fistpq" FROM_F80_2 "mov -24(%rsp), %rax";
+static char f80u64[] = FROM_F80_1 "fistpq" FROM_F80_2 "mov -24(%rsp), %rax";
+static char f80f32[] = "fstps -8(%rsp); movss -8(%rsp), %xmm0";
+static char f80f64[] = "fstpl -8(%rsp); movsd -8(%rsp), %xmm0";
+
+static char *cast_table[][11] = {
+ // i8 i16 i32 i64 u8 u16 u32 u64 f32 f64 f80
+ {NULL, NULL, NULL, i32i64, i32u8, i32u16, NULL, i32i64, i32f32, i32f64, i32f80}, // i8
+ {i32i8, NULL, NULL, i32i64, i32u8, i32u16, NULL, i32i64, i32f32, i32f64, i32f80}, // i16
+ {i32i8, i32i16, NULL, i32i64, i32u8, i32u16, NULL, i32i64, i32f32, i32f64, i32f80}, // i32
+ {i32i8, i32i16, NULL, NULL, i32u8, i32u16, NULL, NULL, i64f32, i64f64, i64f80}, // i64
+
+ {i32i8, NULL, NULL, i32i64, NULL, NULL, NULL, i32i64, i32f32, i32f64, i32f80}, // u8
+ {i32i8, i32i16, NULL, i32i64, i32u8, NULL, NULL, i32i64, i32f32, i32f64, i32f80}, // u16
+ {i32i8, i32i16, NULL, u32i64, i32u8, i32u16, NULL, u32i64, u32f32, u32f64, u32f80}, // u32
+ {i32i8, i32i16, NULL, NULL, i32u8, i32u16, NULL, NULL, u64f32, u64f64, u64f80}, // u64
+
+ {f32i8, f32i16, f32i32, f32i64, f32u8, f32u16, f32u32, f32u64, NULL, f32f64, f32f80}, // f32
+ {f64i8, f64i16, f64i32, f64i64, f64u8, f64u16, f64u32, f64u64, f64f32, NULL, f64f80}, // f64
+ {f80i8, f80i16, f80i32, f80i64, f80u8, f80u16, f80u32, f80u64, f80f32, f80f64, NULL}, // f80
+};
+
+static void cast(Type *from, Type *to) {
+ if (to->kind == TY_VOID)
+ return;
+
+ if (to->kind == TY_BOOL) {
+ cmp_zero(from);
+ println(" setne %%al");
+ println(" movzx %%al, %%eax");
+ return;
+ }
+
+ int t1 = getTypeId(from);
+ int t2 = getTypeId(to);
+ if (cast_table[t1][t2])
+ println(" %s", cast_table[t1][t2]);
+}
+
+// Structs or unions equal or smaller than 16 bytes are passed
+// using up to two registers.
+//
+// If the first 8 bytes contains only floating-point type members,
+// they are passed in an XMM register. Otherwise, they are passed
+// in a general-purpose register.
+//
+// If a struct/union is larger than 8 bytes, the same rule is
+// applied to the the next 8 byte chunk.
+//
+// This function returns true if `ty` has only floating-point
+// members in its byte range [lo, hi).
+static bool has_flonum(Type *ty, int lo, int hi, int offset) {
+ if (ty->kind == TY_STRUCT || ty->kind == TY_UNION) {
+ for (Member *mem = ty->members; mem; mem = mem->next)
+ if (!has_flonum(mem->ty, lo, hi, offset + mem->offset))
+ return false;
+ return true;
+ }
+
+ if (ty->kind == TY_ARRAY) {
+ for (int i = 0; i < ty->array_len; i++)
+ if (!has_flonum(ty->base, lo, hi, offset + ty->base->size * i))
+ return false;
+ return true;
+ }
+
+ return offset < lo || hi <= offset || ty->kind == TY_FLOAT || ty->kind == TY_DOUBLE;
+}
+
+static bool has_flonum1(Type *ty) {
+ return has_flonum(ty, 0, 8, 0);
+}
+
+static bool has_flonum2(Type *ty) {
+ return has_flonum(ty, 8, 16, 0);
+}
+
+static void push_struct(Type *ty) {
+ int sz = align_to(ty->size, 8);
+ println(" sub $%d, %%rsp", sz);
+ depth += sz / 8;
+
+ for (int i = 0; i < ty->size; i++) {
+ println(" mov %d(%%rax), %%r10b", i);
+ println(" mov %%r10b, %d(%%rsp)", i);
+ }
+}
+
+static void push_args2(Node *args, bool first_pass) {
+ if (!args)
+ return;
+ push_args2(args->next, first_pass);
+
+ if ((first_pass && !args->pass_by_stack) || (!first_pass && args->pass_by_stack))
+ return;
+
+ gen_expr(args);
+
+ switch (args->ty->kind) {
+ case TY_STRUCT:
+ case TY_UNION:
+ push_struct(args->ty);
+ break;
+ case TY_FLOAT:
+ case TY_DOUBLE:
+ pushf();
+ break;
+ case TY_LDOUBLE:
+ println(" sub $16, %%rsp");
+ println(" fstpt (%%rsp)");
+ depth += 2;
+ break;
+ default:
+ push();
+ }
+}
+
+// Load function call arguments. Arguments are already evaluated and
+// stored to the stack as local variables. What we need to do in this
+// function is to load them to registers or push them to the stack as
+// specified by the x86-64 psABI. Here is what the spec says:
+//
+// - Up to 6 arguments of integral type are passed using RDI, RSI,
+// RDX, RCX, R8 and R9.
+//
+// - Up to 8 arguments of floating-point type are passed using XMM0 to
+// XMM7.
+//
+// - If all registers of an appropriate type are already used, push an
+// argument to the stack in the right-to-left order.
+//
+// - Each argument passed on the stack takes 8 bytes, and the end of
+// the argument area must be aligned to a 16 byte boundary.
+//
+// - If a function is variadic, set the number of floating-point type
+// arguments to RAX.
+static int push_args(Node *node) {
+ int stack = 0, gp = 0, fp = 0;
+
+ // If the return type is a large struct/union, the caller passes
+ // a pointer to a buffer as if it were the first argument.
+ if (node->ret_buffer && node->ty->size > 16)
+ gp++;
+
+ // Load as many arguments to the registers as possible.
+ for (Node *arg = node->args; arg; arg = arg->next) {
+ Type *ty = arg->ty;
+
+ switch (ty->kind) {
+ case TY_STRUCT:
+ case TY_UNION:
+ if (ty->size > 16) {
+ arg->pass_by_stack = true;
+ stack += align_to(ty->size, 8) / 8;
+ } else {
+ bool fp1 = has_flonum1(ty);
+ bool fp2 = has_flonum2(ty);
+
+ if (fp + fp1 + fp2 < FP_MAX && gp + !fp1 + !fp2 < GP_MAX) {
+ fp = fp + fp1 + fp2;
+ gp = gp + !fp1 + !fp2;
+ } else {
+ arg->pass_by_stack = true;
+ stack += align_to(ty->size, 8) / 8;
+ }
+ }
+ break;
+ case TY_FLOAT:
+ case TY_DOUBLE:
+ if (fp++ >= FP_MAX) {
+ arg->pass_by_stack = true;
+ stack++;
+ }
+ break;
+ case TY_LDOUBLE:
+ arg->pass_by_stack = true;
+ stack += 2;
+ break;
+ default:
+ if (gp++ >= GP_MAX) {
+ arg->pass_by_stack = true;
+ stack++;
+ }
+ }
+ }
+
+ if ((depth + stack) % 2 == 1) {
+ println(" sub $8, %%rsp");
+ depth++;
+ stack++;
+ }
+
+ push_args2(node->args, true);
+ push_args2(node->args, false);
+
+ // If the return type is a large struct/union, the caller passes
+ // a pointer to a buffer as if it were the first argument.
+ if (node->ret_buffer && node->ty->size > 16) {
+ println(" lea %d(%%rbp), %%rax", node->ret_buffer->offset);
+ push();
+ }
+
+ return stack;
+}
+
+static void copy_ret_buffer(Obj *var) {
+ Type *ty = var->ty;
+ int gp = 0, fp = 0;
+
+ if (has_flonum1(ty)) {
+ assert(ty->size == 4 || 8 <= ty->size);
+ if (ty->size == 4)
+ println(" movss %%xmm0, %d(%%rbp)", var->offset);
+ else
+ println(" movsd %%xmm0, %d(%%rbp)", var->offset);
+ fp++;
+ } else {
+ for (int i = 0; i < MIN(8, ty->size); i++) {
+ println(" mov %%al, %d(%%rbp)", var->offset + i);
+ println(" shr $8, %%rax");
+ }
+ gp++;
+ }
+
+ if (ty->size > 8) {
+ if (has_flonum2(ty)) {
+ assert(ty->size == 12 || ty->size == 16);
+ if (ty->size == 12)
+ println(" movss %%xmm%d, %d(%%rbp)", fp, var->offset + 8);
+ else
+ println(" movsd %%xmm%d, %d(%%rbp)", fp, var->offset + 8);
+ } else {
+ char *reg1 = (gp == 0) ? "%al" : "%dl";
+ char *reg2 = (gp == 0) ? "%rax" : "%rdx";
+ for (int i = 8; i < MIN(16, ty->size); i++) {
+ println(" mov %s, %d(%%rbp)", reg1, var->offset + i);
+ println(" shr $8, %s", reg2);
+ }
+ }
+ }
+}
+
+static void copy_struct_reg(void) {
+ Type *ty = current_fn->ty->return_ty;
+ int gp = 0, fp = 0;
+
+ println(" mov %%rax, %%rdi");
+
+ if (has_flonum(ty, 0, 8, 0)) {
+ assert(ty->size == 4 || 8 <= ty->size);
+ if (ty->size == 4)
+ println(" movss (%%rdi), %%xmm0");
+ else
+ println(" movsd (%%rdi), %%xmm0");
+ fp++;
+ } else {
+ println(" mov $0, %%rax");
+ for (int i = MIN(8, ty->size) - 1; i >= 0; i--) {
+ println(" shl $8, %%rax");
+ println(" mov %d(%%rdi), %%al", i);
+ }
+ gp++;
+ }
+
+ if (ty->size > 8) {
+ if (has_flonum(ty, 8, 16, 0)) {
+ assert(ty->size == 12 || ty->size == 16);
+ if (ty->size == 4)
+ println(" movss 8(%%rdi), %%xmm%d", fp);
+ else
+ println(" movsd 8(%%rdi), %%xmm%d", fp);
+ } else {
+ char *reg1 = (gp == 0) ? "%al" : "%dl";
+ char *reg2 = (gp == 0) ? "%rax" : "%rdx";
+ println(" mov $0, %s", reg2);
+ for (int i = MIN(16, ty->size) - 1; i >= 8; i--) {
+ println(" shl $8, %s", reg2);
+ println(" mov %d(%%rdi), %s", i, reg1);
+ }
+ }
+ }
+}
+
+static void copy_struct_mem(void) {
+ Type *ty = current_fn->ty->return_ty;
+ Obj *var = current_fn->params;
+
+ println(" mov %d(%%rbp), %%rdi", var->offset);
+
+ for (int i = 0; i < ty->size; i++) {
+ println(" mov %d(%%rax), %%dl", i);
+ println(" mov %%dl, %d(%%rdi)", i);
+ }
+}
+
+static void builtin_alloca(void) {
+ // Align size to 16 bytes.
+ println(" add $15, %%rdi");
+ println(" and $0xfffffff0, %%edi");
+
+ // Shift the temporary area by %rdi.
+ println(" mov %d(%%rbp), %%rcx", current_fn->alloca_bottom->offset);
+ println(" sub %%rsp, %%rcx");
+ println(" mov %%rsp, %%rax");
+ println(" sub %%rdi, %%rsp");
+ println(" mov %%rsp, %%rdx");
+ println("1:");
+ println(" cmp $0, %%rcx");
+ println(" je 2f");
+ println(" mov (%%rax), %%r8b");
+ println(" mov %%r8b, (%%rdx)");
+ println(" inc %%rdx");
+ println(" inc %%rax");
+ println(" dec %%rcx");
+ println(" jmp 1b");
+ println("2:");
+
+ // Move alloca_bottom pointer.
+ println(" mov %d(%%rbp), %%rax", current_fn->alloca_bottom->offset);
+ println(" sub %%rdi, %%rax");
+ println(" mov %%rax, %d(%%rbp)", current_fn->alloca_bottom->offset);
+}
+
+// Generate code for a given node.
+static void gen_expr(Node *node) {
+ println(" .loc %d %d", node->tok->file->file_no, node->tok->line_no);
+
+ switch (node->kind) {
+ case ND_NULL_EXPR:
+ return;
+ case ND_NUM: {
+ switch (node->ty->kind) {
+ case TY_FLOAT: {
+ union { float f32; uint32_t u32; } u = { node->fval };
+ println(" mov $%u, %%eax # float %Lf", u.u32, node->fval);
+ println(" movq %%rax, %%xmm0");
+ return;
+ }
+ case TY_DOUBLE: {
+ union { double f64; uint64_t u64; } u = { node->fval };
+ println(" mov $%lu, %%rax # double %Lf", u.u64, node->fval);
+ println(" movq %%rax, %%xmm0");
+ return;
+ }
+ case TY_LDOUBLE: {
+ union { long double f80; uint64_t u64[2]; } u;
+ memset(&u, 0, sizeof(u));
+ u.f80 = node->fval;
+ println(" mov $%lu, %%rax # long double %Lf", u.u64[0], node->fval);
+ println(" mov %%rax, -16(%%rsp)");
+ println(" mov $%lu, %%rax", u.u64[1]);
+ println(" mov %%rax, -8(%%rsp)");
+ println(" fldt -16(%%rsp)");
+ return;
+ }
+ }
+
+ println(" mov $%ld, %%rax", node->val);
+ return;
+ }
+ case ND_NEG:
+ gen_expr(node->lhs);
+
+ switch (node->ty->kind) {
+ case TY_FLOAT:
+ println(" mov $1, %%rax");
+ println(" shl $31, %%rax");
+ println(" movq %%rax, %%xmm1");
+ println(" xorps %%xmm1, %%xmm0");
+ return;
+ case TY_DOUBLE:
+ println(" mov $1, %%rax");
+ println(" shl $63, %%rax");
+ println(" movq %%rax, %%xmm1");
+ println(" xorpd %%xmm1, %%xmm0");
+ return;
+ case TY_LDOUBLE:
+ println(" fchs");
+ return;
+ }
+
+ println(" neg %%rax");
+ return;
+ case ND_VAR:
+ gen_addr(node);
+ load(node->ty);
+ return;
+ case ND_MEMBER: {
+ gen_addr(node);
+ load(node->ty);
+
+ Member *mem = node->member;
+ if (mem->is_bitfield) {
+ println(" shl $%d, %%rax", 64 - mem->bit_width - mem->bit_offset);
+ if (mem->ty->is_unsigned)
+ println(" shr $%d, %%rax", 64 - mem->bit_width);
+ else
+ println(" sar $%d, %%rax", 64 - mem->bit_width);
+ }
+ return;
+ }
+ case ND_DEREF:
+ gen_expr(node->lhs);
+ load(node->ty);
+ return;
+ case ND_ADDR:
+ gen_addr(node->lhs);
+ return;
+ case ND_ASSIGN:
+ gen_addr(node->lhs);
+ push();
+ gen_expr(node->rhs);
+
+ if (node->lhs->kind == ND_MEMBER && node->lhs->member->is_bitfield) {
+ println(" mov %%rax, %%r8");
+
+ // If the lhs is a bitfield, we need to read the current value
+ // from memory and merge it with a new value.
+ Member *mem = node->lhs->member;
+ println(" mov %%rax, %%rdi");
+ println(" and $%ld, %%rdi", (1L << mem->bit_width) - 1);
+ println(" shl $%d, %%rdi", mem->bit_offset);
+
+ println(" mov (%%rsp), %%rax");
+ load(mem->ty);
+
+ long mask = ((1L << mem->bit_width) - 1) << mem->bit_offset;
+ println(" mov $%ld, %%r9", ~mask);
+ println(" and %%r9, %%rax");
+ println(" or %%rdi, %%rax");
+ store(node->ty);
+ println(" mov %%r8, %%rax");
+ return;
+ }
+
+ store(node->ty);
+ return;
+ case ND_STMT_EXPR:
+ for (Node *n = node->body; n; n = n->next)
+ gen_stmt(n);
+ return;
+ case ND_COMMA:
+ gen_expr(node->lhs);
+ gen_expr(node->rhs);
+ return;
+ case ND_CAST:
+ gen_expr(node->lhs);
+ cast(node->lhs->ty, node->ty);
+ return;
+ case ND_MEMZERO:
+ // `rep stosb` is equivalent to `memset(%rdi, %al, %rcx)`.
+ println(" mov $%d, %%rcx", node->var->ty->size);
+ println(" lea %d(%%rbp), %%rdi", node->var->offset);
+ println(" mov $0, %%al");
+ println(" rep stosb");
+ return;
+ case ND_COND: {
+ int c = count();
+ gen_expr(node->cond);
+ cmp_zero(node->cond->ty);
+ println(" je .L.else.%d", c);
+ gen_expr(node->then);
+ println(" jmp .L.end.%d", c);
+ println(".L.else.%d:", c);
+ gen_expr(node->els);
+ println(".L.end.%d:", c);
+ return;
+ }
+ case ND_NOT:
+ gen_expr(node->lhs);
+ cmp_zero(node->lhs->ty);
+ println(" sete %%al");
+ println(" movzx %%al, %%rax");
+ return;
+ case ND_BITNOT:
+ gen_expr(node->lhs);
+ println(" not %%rax");
+ return;
+ case ND_LOGAND: {
+ int c = count();
+ gen_expr(node->lhs);
+ cmp_zero(node->lhs->ty);
+ println(" je .L.false.%d", c);
+ gen_expr(node->rhs);
+ cmp_zero(node->rhs->ty);
+ println(" je .L.false.%d", c);
+ println(" mov $1, %%rax");
+ println(" jmp .L.end.%d", c);
+ println(".L.false.%d:", c);
+ println(" mov $0, %%rax");
+ println(".L.end.%d:", c);
+ return;
+ }
+ case ND_LOGOR: {
+ int c = count();
+ gen_expr(node->lhs);
+ cmp_zero(node->lhs->ty);
+ println(" jne .L.true.%d", c);
+ gen_expr(node->rhs);
+ cmp_zero(node->rhs->ty);
+ println(" jne .L.true.%d", c);
+ println(" mov $0, %%rax");
+ println(" jmp .L.end.%d", c);
+ println(".L.true.%d:", c);
+ println(" mov $1, %%rax");
+ println(".L.end.%d:", c);
+ return;
+ }
+ case ND_FUNCALL: {
+ if (node->lhs->kind == ND_VAR && !strcmp(node->lhs->var->name, "alloca")) {
+ gen_expr(node->args);
+ println(" mov %%rax, %%rdi");
+ builtin_alloca();
+ return;
+ }
+
+ int stack_args = push_args(node);
+ gen_expr(node->lhs);
+
+ int gp = 0, fp = 0;
+
+ // If the return type is a large struct/union, the caller passes
+ // a pointer to a buffer as if it were the first argument.
+ if (node->ret_buffer && node->ty->size > 16)
+ pop(argreg64[gp++]);
+
+ for (Node *arg = node->args; arg; arg = arg->next) {
+ Type *ty = arg->ty;
+
+ switch (ty->kind) {
+ case TY_STRUCT:
+ case TY_UNION:
+ if (ty->size > 16)
+ continue;
+
+ bool fp1 = has_flonum1(ty);
+ bool fp2 = has_flonum2(ty);
+
+ if (fp + fp1 + fp2 < FP_MAX && gp + !fp1 + !fp2 < GP_MAX) {
+ if (fp1)
+ popf(fp++);
+ else
+ pop(argreg64[gp++]);
+
+ if (ty->size > 8) {
+ if (fp2)
+ popf(fp++);
+ else
+ pop(argreg64[gp++]);
+ }
+ }
+ break;
+ case TY_FLOAT:
+ case TY_DOUBLE:
+ if (fp < FP_MAX)
+ popf(fp++);
+ break;
+ case TY_LDOUBLE:
+ break;
+ default:
+ if (gp < GP_MAX)
+ pop(argreg64[gp++]);
+ }
+ }
+
+ println(" mov %%rax, %%r10");
+ println(" mov $%d, %%rax", fp);
+ println(" call *%%r10");
+ println(" add $%d, %%rsp", stack_args * 8);
+
+ depth -= stack_args;
+
+ // It looks like the most significant 48 or 56 bits in RAX may
+ // contain garbage if a function return type is short or bool/char,
+ // respectively. We clear the upper bits here.
+ switch (node->ty->kind) {
+ case TY_BOOL:
+ println(" movzx %%al, %%eax");
+ return;
+ case TY_CHAR:
+ if (node->ty->is_unsigned)
+ println(" movzbl %%al, %%eax");
+ else
+ println(" movsbl %%al, %%eax");
+ return;
+ case TY_SHORT:
+ if (node->ty->is_unsigned)
+ println(" movzwl %%ax, %%eax");
+ else
+ println(" movswl %%ax, %%eax");
+ return;
+ }
+
+ // If the return type is a small struct, a value is returned
+ // using up to two registers.
+ if (node->ret_buffer && node->ty->size <= 16) {
+ copy_ret_buffer(node->ret_buffer);
+ println(" lea %d(%%rbp), %%rax", node->ret_buffer->offset);
+ }
+
+ return;
+ }
+ case ND_LABEL_VAL:
+ println(" lea %s(%%rip), %%rax", node->unique_label);
+ return;
+ case ND_CAS: {
+ gen_expr(node->cas_addr);
+ push();
+ gen_expr(node->cas_new);
+ push();
+ gen_expr(node->cas_old);
+ println(" mov %%rax, %%r8");
+ load(node->cas_old->ty->base);
+ pop("%rdx"); // new
+ pop("%rdi"); // addr
+
+ int sz = node->cas_addr->ty->base->size;
+ println(" lock cmpxchg %s, (%%rdi)", reg_dx(sz));
+ println(" sete %%cl");
+ println(" je 1f");
+ println(" mov %s, (%%r8)", reg_ax(sz));
+ println("1:");
+ println(" movzbl %%cl, %%eax");
+ return;
+ }
+ case ND_EXCH: {
+ gen_expr(node->lhs);
+ push();
+ gen_expr(node->rhs);
+ pop("%rdi");
+
+ int sz = node->lhs->ty->base->size;
+ println(" xchg %s, (%%rdi)", reg_ax(sz));
+ return;
+ }
+ }
+
+ switch (node->lhs->ty->kind) {
+ case TY_FLOAT:
+ case TY_DOUBLE: {
+ gen_expr(node->rhs);
+ pushf();
+ gen_expr(node->lhs);
+ popf(1);
+
+ char *sz = (node->lhs->ty->kind == TY_FLOAT) ? "ss" : "sd";
+
+ switch (node->kind) {
+ case ND_ADD:
+ println(" add%s %%xmm1, %%xmm0", sz);
+ return;
+ case ND_SUB:
+ println(" sub%s %%xmm1, %%xmm0", sz);
+ return;
+ case ND_MUL:
+ println(" mul%s %%xmm1, %%xmm0", sz);
+ return;
+ case ND_DIV:
+ println(" div%s %%xmm1, %%xmm0", sz);
+ return;
+ case ND_EQ:
+ case ND_NE:
+ case ND_LT:
+ case ND_LE:
+ println(" ucomi%s %%xmm0, %%xmm1", sz);
+
+ if (node->kind == ND_EQ) {
+ println(" sete %%al");
+ println(" setnp %%dl");
+ println(" and %%dl, %%al");
+ } else if (node->kind == ND_NE) {
+ println(" setne %%al");
+ println(" setp %%dl");
+ println(" or %%dl, %%al");
+ } else if (node->kind == ND_LT) {
+ println(" seta %%al");
+ } else {
+ println(" setae %%al");
+ }
+
+ println(" and $1, %%al");
+ println(" movzb %%al, %%rax");
+ return;
+ }
+
+ error_tok(node->tok, "invalid expression");
+ }
+ case TY_LDOUBLE: {
+ gen_expr(node->lhs);
+ gen_expr(node->rhs);
+
+ switch (node->kind) {
+ case ND_ADD:
+ println(" faddp");
+ return;
+ case ND_SUB:
+ println(" fsubrp");
+ return;
+ case ND_MUL:
+ println(" fmulp");
+ return;
+ case ND_DIV:
+ println(" fdivrp");
+ return;
+ case ND_EQ:
+ case ND_NE:
+ case ND_LT:
+ case ND_LE:
+ println(" fcomip");
+ println(" fstp %%st(0)");
+
+ if (node->kind == ND_EQ)
+ println(" sete %%al");
+ else if (node->kind == ND_NE)
+ println(" setne %%al");
+ else if (node->kind == ND_LT)
+ println(" seta %%al");
+ else
+ println(" setae %%al");
+
+ println(" movzb %%al, %%rax");
+ return;
+ }
+
+ error_tok(node->tok, "invalid expression");
+ }
+ }
+
+ gen_expr(node->rhs);
+ push();
+ gen_expr(node->lhs);
+ pop("%rdi");
+
+ char *ax, *di, *dx;
+
+ if (node->lhs->ty->kind == TY_LONG || node->lhs->ty->base) {
+ ax = "%rax";
+ di = "%rdi";
+ dx = "%rdx";
+ } else {
+ ax = "%eax";
+ di = "%edi";
+ dx = "%edx";
+ }
+
+ switch (node->kind) {
+ case ND_ADD:
+ println(" add %s, %s", di, ax);
+ return;
+ case ND_SUB:
+ println(" sub %s, %s", di, ax);
+ return;
+ case ND_MUL:
+ println(" imul %s, %s", di, ax);
+ return;
+ case ND_DIV:
+ case ND_MOD:
+ if (node->ty->is_unsigned) {
+ println(" mov $0, %s", dx);
+ println(" div %s", di);
+ } else {
+ if (node->lhs->ty->size == 8)
+ println(" cqo");
+ else
+ println(" cdq");
+ println(" idiv %s", di);
+ }
+
+ if (node->kind == ND_MOD)
+ println(" mov %%rdx, %%rax");
+ return;
+ case ND_BITAND:
+ println(" and %s, %s", di, ax);
+ return;
+ case ND_BITOR:
+ println(" or %s, %s", di, ax);
+ return;
+ case ND_BITXOR:
+ println(" xor %s, %s", di, ax);
+ return;
+ case ND_EQ:
+ case ND_NE:
+ case ND_LT:
+ case ND_LE:
+ println(" cmp %s, %s", di, ax);
+
+ if (node->kind == ND_EQ) {
+ println(" sete %%al");
+ } else if (node->kind == ND_NE) {
+ println(" setne %%al");
+ } else if (node->kind == ND_LT) {
+ if (node->lhs->ty->is_unsigned)
+ println(" setb %%al");
+ else
+ println(" setl %%al");
+ } else if (node->kind == ND_LE) {
+ if (node->lhs->ty->is_unsigned)
+ println(" setbe %%al");
+ else
+ println(" setle %%al");
+ }
+
+ println(" movzb %%al, %%rax");
+ return;
+ case ND_SHL:
+ println(" mov %%rdi, %%rcx");
+ println(" shl %%cl, %s", ax);
+ return;
+ case ND_SHR:
+ println(" mov %%rdi, %%rcx");
+ if (node->lhs->ty->is_unsigned)
+ println(" shr %%cl, %s", ax);
+ else
+ println(" sar %%cl, %s", ax);
+ return;
+ }
+
+ error_tok(node->tok, "invalid expression");
+}
+
+static void gen_stmt(Node *node) {
+ println(" .loc %d %d", node->tok->file->file_no, node->tok->line_no);
+
+ switch (node->kind) {
+ case ND_IF: {
+ int c = count();
+ gen_expr(node->cond);
+ cmp_zero(node->cond->ty);
+ println(" je .L.else.%d", c);
+ gen_stmt(node->then);
+ println(" jmp .L.end.%d", c);
+ println(".L.else.%d:", c);
+ if (node->els)
+ gen_stmt(node->els);
+ println(".L.end.%d:", c);
+ return;
+ }
+ case ND_FOR: {
+ int c = count();
+ if (node->init)
+ gen_stmt(node->init);
+ println(".L.begin.%d:", c);
+ if (node->cond) {
+ gen_expr(node->cond);
+ cmp_zero(node->cond->ty);
+ println(" je %s", node->brk_label);
+ }
+ gen_stmt(node->then);
+ println("%s:", node->cont_label);
+ if (node->inc)
+ gen_expr(node->inc);
+ println(" jmp .L.begin.%d", c);
+ println("%s:", node->brk_label);
+ return;
+ }
+ case ND_DO: {
+ int c = count();
+ println(".L.begin.%d:", c);
+ gen_stmt(node->then);
+ println("%s:", node->cont_label);
+ gen_expr(node->cond);
+ cmp_zero(node->cond->ty);
+ println(" jne .L.begin.%d", c);
+ println("%s:", node->brk_label);
+ return;
+ }
+ case ND_SWITCH:
+ gen_expr(node->cond);
+
+ for (Node *n = node->case_next; n; n = n->case_next) {
+ char *ax = (node->cond->ty->size == 8) ? "%rax" : "%eax";
+ char *di = (node->cond->ty->size == 8) ? "%rdi" : "%edi";
+
+ if (n->begin == n->end) {
+ println(" cmp $%ld, %s", n->begin, ax);
+ println(" je %s", n->label);
+ continue;
+ }
+
+ // [GNU] Case ranges
+ println(" mov %s, %s", ax, di);
+ println(" sub $%ld, %s", n->begin, di);
+ println(" cmp $%ld, %s", n->end - n->begin, di);
+ println(" jbe %s", n->label);
+ }
+
+ if (node->default_case)
+ println(" jmp %s", node->default_case->label);
+
+ println(" jmp %s", node->brk_label);
+ gen_stmt(node->then);
+ println("%s:", node->brk_label);
+ return;
+ case ND_CASE:
+ println("%s:", node->label);
+ gen_stmt(node->lhs);
+ return;
+ case ND_BLOCK:
+ for (Node *n = node->body; n; n = n->next)
+ gen_stmt(n);
+ return;
+ case ND_GOTO:
+ println(" jmp %s", node->unique_label);
+ return;
+ case ND_GOTO_EXPR:
+ gen_expr(node->lhs);
+ println(" jmp *%%rax");
+ return;
+ case ND_LABEL:
+ println("%s:", node->unique_label);
+ gen_stmt(node->lhs);
+ return;
+ case ND_RETURN:
+ if (node->lhs) {
+ gen_expr(node->lhs);
+ Type *ty = node->lhs->ty;
+
+ switch (ty->kind) {
+ case TY_STRUCT:
+ case TY_UNION:
+ if (ty->size <= 16)
+ copy_struct_reg();
+ else
+ copy_struct_mem();
+ break;
+ }
+ }
+
+ println(" jmp .L.return.%s", current_fn->name);
+ return;
+ case ND_EXPR_STMT:
+ gen_expr(node->lhs);
+ return;
+ case ND_ASM:
+ println(" %s", node->asm_str);
+ return;
+ }
+
+ error_tok(node->tok, "invalid statement");
+}
+
+// Assign offsets to local variables.
+static void assign_lvar_offsets(Obj *prog) {
+ for (Obj *fn = prog; fn; fn = fn->next) {
+ if (!fn->is_function)
+ continue;
+
+ // If a function has many parameters, some parameters are
+ // inevitably passed by stack rather than by register.
+ // The first passed-by-stack parameter resides at RBP+16.
+ int top = 16;
+ int bottom = 0;
+
+ int gp = 0, fp = 0;
+
+ // Assign offsets to pass-by-stack parameters.
+ for (Obj *var = fn->params; var; var = var->next) {
+ Type *ty = var->ty;
+
+ switch (ty->kind) {
+ case TY_STRUCT:
+ case TY_UNION:
+ if (ty->size <= 16) {
+ bool fp1 = has_flonum(ty, 0, 8, 0);
+ bool fp2 = has_flonum(ty, 8, 16, 8);
+ if (fp + fp1 + fp2 < FP_MAX && gp + !fp1 + !fp2 < GP_MAX) {
+ fp = fp + fp1 + fp2;
+ gp = gp + !fp1 + !fp2;
+ continue;
+ }
+ }
+ break;
+ case TY_FLOAT:
+ case TY_DOUBLE:
+ if (fp++ < FP_MAX)
+ continue;
+ break;
+ case TY_LDOUBLE:
+ break;
+ default:
+ if (gp++ < GP_MAX)
+ continue;
+ }
+
+ top = align_to(top, 8);
+ var->offset = top;
+ top += var->ty->size;
+ }
+
+ // Assign offsets to pass-by-register parameters and local variables.
+ for (Obj *var = fn->locals; var; var = var->next) {
+ if (var->offset)
+ continue;
+
+ // AMD64 System V ABI has a special alignment rule for an array of
+ // length at least 16 bytes. We need to align such array to at least
+ // 16-byte boundaries. See p.14 of
+ // https://github.com/hjl-tools/x86-psABI/wiki/x86-64-psABI-draft.pdf.
+ int align = (var->ty->kind == TY_ARRAY && var->ty->size >= 16)
+ ? MAX(16, var->align) : var->align;
+
+ bottom += var->ty->size;
+ bottom = align_to(bottom, align);
+ var->offset = -bottom;
+ }
+
+ fn->stack_size = align_to(bottom, 16);
+ }
+}
+
+static void emit_data(Obj *prog) {
+ for (Obj *var = prog; var; var = var->next) {
+ if (var->is_function || !var->is_definition)
+ continue;
+
+ if (var->is_static)
+ println(" .local %s", var->name);
+ else
+ println(" .globl %s", var->name);
+
+ int align = (var->ty->kind == TY_ARRAY && var->ty->size >= 16)
+ ? MAX(16, var->align) : var->align;
+
+ // Common symbol
+ if (opt_fcommon && var->is_tentative) {
+ println(" .comm %s, %d, %d", var->name, var->ty->size, align);
+ continue;
+ }
+
+ // .data or .tdata
+ if (var->init_data) {
+ if (var->is_tls)
+ println(" .section .tdata,\"awT\",@progbits");
+ else
+ println(" .data");
+
+ println(" .type %s, @object", var->name);
+ println(" .size %s, %d", var->name, var->ty->size);
+ println(" .align %d", align);
+ println("%s:", var->name);
+
+ Relocation *rel = var->rel;
+ int pos = 0;
+ while (pos < var->ty->size) {
+ if (rel && rel->offset == pos) {
+ println(" .quad %s%+ld", *rel->label, rel->addend);
+ rel = rel->next;
+ pos += 8;
+ } else {
+ println(" .byte %d", var->init_data[pos++]);
+ }
+ }
+ continue;
+ }
+
+ // .bss or .tbss
+ if (var->is_tls)
+ println(" .section .tbss,\"awT\",@nobits");
+ else
+ println(" .bss");
+
+ println(" .align %d", align);
+ println("%s:", var->name);
+ println(" .zero %d", var->ty->size);
+ }
+}
+
+static void store_fp(int r, int offset, int sz) {
+ switch (sz) {
+ case 4:
+ println(" movss %%xmm%d, %d(%%rbp)", r, offset);
+ return;
+ case 8:
+ println(" movsd %%xmm%d, %d(%%rbp)", r, offset);
+ return;
+ }
+ unreachable();
+}
+
+static void store_gp(int r, int offset, int sz) {
+ switch (sz) {
+ case 1:
+ println(" mov %s, %d(%%rbp)", argreg8[r], offset);
+ return;
+ case 2:
+ println(" mov %s, %d(%%rbp)", argreg16[r], offset);
+ return;
+ case 4:
+ println(" mov %s, %d(%%rbp)", argreg32[r], offset);
+ return;
+ case 8:
+ println(" mov %s, %d(%%rbp)", argreg64[r], offset);
+ return;
+ default:
+ for (int i = 0; i < sz; i++) {
+ println(" mov %s, %d(%%rbp)", argreg8[r], offset + i);
+ println(" shr $8, %s", argreg64[r]);
+ }
+ return;
+ }
+}
+
+static void emit_text(Obj *prog) {
+ for (Obj *fn = prog; fn; fn = fn->next) {
+ if (!fn->is_function || !fn->is_definition)
+ continue;
+
+ // No code is emitted for "static inline" functions
+ // if no one is referencing them.
+ if (!fn->is_live)
+ continue;
+
+ if (fn->is_static)
+ println(" .local %s", fn->name);
+ else
+ println(" .globl %s", fn->name);
+
+ println(" .text");
+ println(" .type %s, @function", fn->name);
+ println("%s:", fn->name);
+ current_fn = fn;
+
+ // Prologue
+ println(" push %%rbp");
+ println(" mov %%rsp, %%rbp");
+ println(" sub $%d, %%rsp", fn->stack_size);
+ println(" mov %%rsp, %d(%%rbp)", fn->alloca_bottom->offset);
+
+ // Save arg registers if function is variadic
+ if (fn->va_area) {
+ int gp = 0, fp = 0;
+ for (Obj *var = fn->params; var; var = var->next) {
+ if (is_flonum(var->ty))
+ fp++;
+ else
+ gp++;
+ }
+
+ int off = fn->va_area->offset;
+
+ // va_elem
+ println(" movl $%d, %d(%%rbp)", gp * 8, off); // gp_offset
+ println(" movl $%d, %d(%%rbp)", fp * 8 + 48, off + 4); // fp_offset
+ println(" movq %%rbp, %d(%%rbp)", off + 8); // overflow_arg_area
+ println(" addq $16, %d(%%rbp)", off + 8);
+ println(" movq %%rbp, %d(%%rbp)", off + 16); // reg_save_area
+ println(" addq $%d, %d(%%rbp)", off + 24, off + 16);
+
+ // __reg_save_area__
+ println(" movq %%rdi, %d(%%rbp)", off + 24);
+ println(" movq %%rsi, %d(%%rbp)", off + 32);
+ println(" movq %%rdx, %d(%%rbp)", off + 40);
+ println(" movq %%rcx, %d(%%rbp)", off + 48);
+ println(" movq %%r8, %d(%%rbp)", off + 56);
+ println(" movq %%r9, %d(%%rbp)", off + 64);
+ println(" movsd %%xmm0, %d(%%rbp)", off + 72);
+ println(" movsd %%xmm1, %d(%%rbp)", off + 80);
+ println(" movsd %%xmm2, %d(%%rbp)", off + 88);
+ println(" movsd %%xmm3, %d(%%rbp)", off + 96);
+ println(" movsd %%xmm4, %d(%%rbp)", off + 104);
+ println(" movsd %%xmm5, %d(%%rbp)", off + 112);
+ println(" movsd %%xmm6, %d(%%rbp)", off + 120);
+ println(" movsd %%xmm7, %d(%%rbp)", off + 128);
+ }
+
+ // Save passed-by-register arguments to the stack
+ int gp = 0, fp = 0;
+ for (Obj *var = fn->params; var; var = var->next) {
+ if (var->offset > 0)
+ continue;
+
+ Type *ty = var->ty;
+
+ switch (ty->kind) {
+ case TY_STRUCT:
+ case TY_UNION:
+ assert(ty->size <= 16);
+ if (has_flonum(ty, 0, 8, 0))
+ store_fp(fp++, var->offset, MIN(8, ty->size));
+ else
+ store_gp(gp++, var->offset, MIN(8, ty->size));
+
+ if (ty->size > 8) {
+ if (has_flonum(ty, 8, 16, 0))
+ store_fp(fp++, var->offset + 8, ty->size - 8);
+ else
+ store_gp(gp++, var->offset + 8, ty->size - 8);
+ }
+ break;
+ case TY_FLOAT:
+ case TY_DOUBLE:
+ store_fp(fp++, var->offset, ty->size);
+ break;
+ default:
+ store_gp(gp++, var->offset, ty->size);
+ }
+ }
+
+ // Emit code
+ gen_stmt(fn->body);
+ assert(depth == 0);
+
+ // [https://www.sigbus.info/n1570#5.1.2.2.3p1] The C spec defines
+ // a special rule for the main function. Reaching the end of the
+ // main function is equivalent to returning 0, even though the
+ // behavior is undefined for the other functions.
+ if (strcmp(fn->name, "main") == 0)
+ println(" mov $0, %%rax");
+
+ // Epilogue
+ println(".L.return.%s:", fn->name);
+ println(" mov %%rbp, %%rsp");
+ println(" pop %%rbp");
+ println(" ret");
+ }
+}
+
+void codegen(Obj *prog, FILE *out) {
+ output_file = out;
+
+ File **files = get_input_files();
+ for (int i = 0; files[i]; i++)
+ println(" .file %d \"%s\"", files[i]->file_no, files[i]->name);
+
+ assign_lvar_offsets(prog);
+ emit_data(prog);
+ emit_text(prog);
+}
generated by cgit v1.2.3 (git 2.25.1) at 2024-11-14 23:47:58 +0000