amd64 asm: extend family emits mnemonic that doesnt exist

This commit is contained in:
kalsprite
2026-08-19 14:03:56 -07:00
parent 43d44e5942
commit b9983105bd
2 changed files with 96 additions and 2 deletions

View File

@@ -185,6 +185,10 @@ struct lbAsmGenerate {
}
virtual char instruction_size_suffix(AstAsmInstruction *instr) = 0;
// Some AT&T mnemonics encode BOTH operand widths and so cannot be spelled as a name plus one
// suffix: `movsx` from i8 to i32 is `movsbl`. Returns the complete mnemonic, or {} when the
// name-plus-suffix spelling is the right one
virtual String instruction_att_mnemonic(AstAsmInstruction *instr) = 0;
virtual char size_suffix_for_operand(Ast *op) = 0;
virtual gbString write_memory_operand(gbString asm_string, Slice<i32> const &op_number, AstAsmMemoryOperand *mem_op, u32 flags) = 0;
virtual lbValue emit_call(lbProcedure *p, Array<lbValue> const &args) = 0;
@@ -221,6 +225,50 @@ struct lbAsmGenerate_amd64 : lbAsmGenerate {
}
// The sign/zero-extend family is the only one whose two operands differ in width. Its AT&T
// mnemonic names both: `movsx` i8 -> i32 is `movsbl`, never `movsx` plus a suffix. `movsxd` is
// the same rule, movs + l + q
String instruction_att_mnemonic(AstAsmInstruction *instr) override {
bool sign_extend;
switch (instr->mnemonic) {
case Asm_amd64::M_MOVSX:
case Asm_amd64::M_MOVSXD:
sign_extend = true;
break;
case Asm_amd64::M_MOVZX:
sign_extend = false;
break;
default:
return {};
}
auto forms = g_asm_amd64.encoding_forms(instr->mnemonic);
if (instr->valid_form_index < 0 || instr->valid_form_index >= forms.count) {
return {};
}
auto const &form = forms[instr->valid_form_index];
// Intel operand order: dst first
i32 dst = g_asm_amd64.operand_type_bit_width(form.ops[0]);
i32 src = g_asm_amd64.operand_type_bit_width(form.ops[1]);
if (sign_extend) {
if (src == 8 && dst == 16) { return str_lit("movsbw"); }
if (src == 8 && dst == 32) { return str_lit("movsbl"); }
if (src == 8 && dst == 64) { return str_lit("movsbq"); }
if (src == 16 && dst == 32) { return str_lit("movswl"); }
if (src == 16 && dst == 64) { return str_lit("movswq"); }
if (src == 32 && dst == 64) { return str_lit("movslq"); }
} else {
if (src == 8 && dst == 16) { return str_lit("movzbw"); }
if (src == 8 && dst == 32) { return str_lit("movzbl"); }
if (src == 8 && dst == 64) { return str_lit("movzbq"); }
if (src == 16 && dst == 32) { return str_lit("movzwl"); }
if (src == 16 && dst == 64) { return str_lit("movzwq"); }
}
return {};
}
// Scan an instruction's operands for an annotated memory operand and return its
// size suffix, or 0 if none. The checker has already verified the annotation
// agrees with the matched encoding form, so a suffix here can never conflict.
@@ -508,6 +556,10 @@ struct lbAsmGenerate_amd64 : lbAsmGenerate {
case_ast_node(instr, AsmInstruction, instr_);
asm_string = gb_string_appendc(asm_string, "\t");
String name = instr->name->Ident.token.string;
String att = this->instruction_att_mnemonic(instr);
if (att.len != 0) {
name = att;
}
asm_string = gb_string_append_length(asm_string, name.text, name.len);
// If a memory operand carries an explicit size annotation ([p]:u8) and
@@ -515,8 +567,10 @@ struct lbAsmGenerate_amd64 : lbAsmGenerate {
// encoded as a mnemonic suffix (crc32 -> crc32b). The checker has already
// verified the annotation agrees with the matched form, so an emitted
// suffix can never conflict with a register operand's implied width.
if (char suffix = this->instruction_size_suffix(instr)) {
asm_string = gb_string_append_length(asm_string, &suffix, 1);
if (att.len == 0) {
if (char suffix = this->instruction_size_suffix(instr)) {
asm_string = gb_string_append_length(asm_string, &suffix, 1);
}
}
asm_string = gb_string_appendc(asm_string, " ");

View File

@@ -0,0 +1,40 @@
package test_internal
import "core:testing"
// The sign/zero-extend family is the only one whose two operands differ in width, so its AT&T
// mnemonic names both -- `movsx` from i8 to i32 is `movsbl`. Emitting the name plus a single width
// suffix produced `movsxl`, which no assembler has, so none of these forms could be built
@(test)
asm_extend_mnemonics :: proc(t: ^testing.T) {
when ODIN_ARCH == .amd64 {
sx_8_16 :: asm(a: i8) -> (r: i16) { movsx r, a; }
sx_8_32 :: asm(a: i8) -> (r: i32) { movsx r, a; }
sx_8_64 :: asm(a: i8) -> (r: i64) { movsx r, a; }
sx_16_32 :: asm(a: i16) -> (r: i32) { movsx r, a; }
sx_16_64 :: asm(a: i16) -> (r: i64) { movsx r, a; }
zx_8_16 :: asm(a: u8) -> (r: u16) { movzx r, a; }
zx_8_32 :: asm(a: u8) -> (r: u32) { movzx r, a; }
zx_8_64 :: asm(a: u8) -> (r: u64) { movzx r, a; }
zx_16_32 :: asm(a: u16) -> (r: u32) { movzx r, a; }
zx_16_64 :: asm(a: u16) -> (r: u64) { movzx r, a; }
sxd :: asm(a: i32) -> (r: i64) { movsxd r, a; }
testing.expect_value(t, sx_8_16(-1), i16(-1))
testing.expect_value(t, sx_8_32(-1), i32(-1))
testing.expect_value(t, sx_8_64(-1), i64(-1))
testing.expect_value(t, sx_16_32(-300), i32(-300))
testing.expect_value(t, sx_16_64(-300), i64(-300))
testing.expect_value(t, sxd(-123456), i64(-123456))
testing.expect_value(t, zx_8_16(0xFF), u16(255))
testing.expect_value(t, zx_8_32(0xFF), u32(255))
testing.expect_value(t, zx_8_64(0xFF), u64(255))
testing.expect_value(t, zx_16_32(0xFFFF), u32(65535))
testing.expect_value(t, zx_16_64(0xFFFF), u64(65535))
// positive values too, so a mnemonic that merely truncates cannot pass
testing.expect_value(t, sx_8_32(127), i32(127))
testing.expect_value(t, zx_8_32(1), u32(1))
}
}