Mockout inline asm for riscv

core/rexcode/isa/riscv/encoder.odin

@@ -455,6 +455,10 @@ resolve_relocation_inline :: #force_inline proc(
 	return true
 }
 
+is_implicit_op_inline :: #force_inline proc "contextless" (op: Operand_Type) -> bool {
+	return op == .GPR_SP
+}
+
 // =============================================================================
 // Scatter / gather helpers for B-type and J-type immediates
 // =============================================================================

core/rexcode/isa/riscv/encoding_types.odin

@@ -52,11 +52,12 @@ Feature :: enum u8 {
 }
 
 Encoding_Flags :: bit_field u8 {
-	rv32_only: bool | 1,   // RV32 base only
-	rv64_only: bool | 1,   // RV64 base only (e.g. LD/SD/ADDIW/...)
-	branch:    bool | 1,   // changes PC
-	fp_round:  bool | 1,   // funct3 doubles as FP rounding-mode field
-	_:         u8   | 4,
+	rv32_only:      bool | 1, // RV32 base only
+	rv64_only:      bool | 1, // RV64 base only (e.g. LD/SD/ADDIW/...)
+	branch:         bool | 1, // changes PC
+	fp_round:       bool | 1, // funct3 doubles as FP rounding-mode field
+	explicit_count: u8   | 3,
+	has_implicit:   bool | 1,
 }
 
 // What the user passes in.
@@ -149,7 +150,7 @@ Encoding :: struct #packed {
 	enc:      [4]Operand_Encoding, // 4
 	bits:     u32,                 // 4 -- static bit pattern
 	mask:     u32,                 // 4 -- which bits are static
-	feature: Feature,                 // 1
+	feature:  Feature,             // 1
 	flags:    Encoding_Flags,      // 1
 }
 #assert(size_of(Encoding) == 20)

core/rexcode/isa/riscv/tablegen/gen.odin

@@ -113,7 +113,22 @@ emit_encode_tables :: proc() -> (total: int) {
 		fmt.sbprintfln(&sb, "\t// .%v", m)
 		for &form in forms {
 			f := &form.encoding
-			write_row(&sb, f.mnemonic, f.ops, f.enc, f.bits, f.mask, f.feature, f.flags)
+			flags := f.flags
+
+			encoding_operand_count: u8 = 0
+			has_implicit := false
+			for op_type in f.ops {
+				if op_type == .NONE { break }
+				if lib.is_implicit_op_inline(op_type) {
+					has_implicit = true
+				} else {
+					encoding_operand_count += 1
+				}
+			}
+			flags.explicit_count = encoding_operand_count
+			flags.has_implicit   = has_implicit
+
+			write_row(&sb, f.mnemonic, f.ops, f.enc, f.bits, f.mask, f.feature, flags)
 		}
 	}
 	strings.write_string(&sb, "}\n\n")

core/rexcode/isa/x86/tablegen/cpp-compiler/cpp-gen.odin

@@ -309,6 +309,20 @@ main :: proc() {
 					// NOTE: SideEffectFlag_HINT deliberately excluded — inert, may be DCE'd.
 				return ((side_effects & VOLATILE_SE) != 0);
 			}
+
+			u8 is_call_or_mem() const {
+				return (cast(u16)side_effects & SideEffectFlag_CONTROL) != 0 ||
+					(cast(u16)implicit_wr & ClobberReg_RSP) != 0;
+			}
+			bool has_control() const {
+				return (cast(u16)side_effects & SideEffectFlag_CONTROL) != 0;
+			}
+			bool has_halt() const {
+				return (cast(u16)side_effects & SideEffectFlag_HALT) != 0;
+			}
+			bool is_conditional() const {
+				return has_control() && (cast(u16)flags_rd != 0);
+			}
 		};
 
 		void clobber_implicit_regs(StringSet *clobber_registers_set, u16 implicit_regs) {
@@ -381,13 +395,6 @@ main :: proc() {
 			fmt.sbprintf(&sb, "return cast(u8)((flags>>%du)&((1u<<%d)-1));", bit_offset, bit_size)
 			strings.write_string(&sb, " }\n")
 		}
-		{
-			bit_offset := intrinsics.type_field_bit_offset(Encoding_Flags, "op_count")
-			bit_size   := intrinsics.type_field_bit_size(Encoding_Flags, "op_count")
-			strings.write_string(&sb, "\t\tu8   op_count      () const { ")
-			fmt.sbprintf(&sb, "return cast(u8)((flags>>%du)&((1u<<%d)-1));", bit_offset, bit_size)
-			strings.write_string(&sb, " }\n")
-		}
 		{
 			bit_offset := intrinsics.type_field_bit_offset(Encoding_Flags, "lock_ok")
 			strings.write_string(&sb, "\t\tbool lock_ok       () const { ")
@@ -421,7 +428,8 @@ main :: proc() {
 
 	strings.write_string(&sb, """
 
-		bool init() {
+		bool init(i64 word_size) {
+			gb_unused(word_size);
 			string_map_init(&mnemonic_map, MNEMONIC_COUNT*2);
 			for (u16 m = M_INVALID+1; m < MNEMONIC_COUNT; m++) {
 				string_map_set(&mnemonic_map, mnemonic_strings[m], cast(Mnemonic)m);

src/asm_tables.cpp

@@ -54,7 +54,9 @@ gb_global String const asm_operand_kind_expected_strings[AsmOperand_COUNT] = {
 };
 
 #include "asm_tables_amd64.cpp"
+#include "asm_tables_riscv.cpp"
 
-void init_asm_tables() {
-	g_asm_amd64.init();
+void init_asm_tables(i64 word_size) {
+	g_asm_amd64.init(word_size);
+	g_asm_riscv.init(word_size);
 }

src/asm_tables_amd64.cpp

@@ -312,6 +312,20 @@ struct Asm_amd64 {
 				// NOTE: SideEffectFlag_HINT deliberately excluded — inert, may be DCE'd.
 			return ((side_effects & VOLATILE_SE) != 0);
 		}
+
+		u8 is_call_or_mem() const {
+			return (cast(u16)side_effects & SideEffectFlag_CONTROL) != 0 ||
+				(cast(u16)implicit_wr & ClobberReg_RSP) != 0;
+		}
+		bool has_control() const {
+			return (cast(u16)side_effects & SideEffectFlag_CONTROL) != 0;
+		}
+		bool has_halt() const {
+			return (cast(u16)side_effects & SideEffectFlag_HALT) != 0;
+		}
+		bool is_conditional() const {
+			return has_control() && (cast(u16)flags_rd != 0);
+		}
 	};
 
 	void clobber_implicit_regs(StringSet *clobber_registers_set, u16 implicit_regs) {
@@ -424,7 +438,6 @@ struct Asm_amd64 {
 
 		bool has_implicit  () const { return ((flags>>21u)&1) != 0; }
 		u8   explicit_count() const { return cast(u8)((flags>>18u)&((1u<<3)-1)); }
-		u8   op_count      () const { return cast(u8)((flags>>22u)&((1u<<3)-1)); }
 		bool lock_ok       () const { return ((flags>>14u)&1) != 0; }
 		bool rep_ok        () const { return ((flags>>15u)&1) != 0; }
 	};
@@ -447,7 +460,8 @@ struct Asm_amd64 {
 	StringMap<Prefix>   prefix_map;
 	StringMap<Register> register_map;
 
-	bool init() {
+	bool init(i64 word_size) {
+		gb_unused(word_size);
 		string_map_init(&mnemonic_map, MNEMONIC_COUNT*2);
 		for (u16 m = M_INVALID+1; m < MNEMONIC_COUNT; m++) {
 			string_map_set(&mnemonic_map, mnemonic_strings[m], cast(Mnemonic)m);

src/check_asm.cpp

@@ -1080,8 +1080,7 @@ gb_internal void check_mnemonic(AsmCtx *asm_ctx, CheckerContext *ctx, Entity *tm
 		// aligned at the call boundary) or manipulates RSP directly. Plain memory access
 		// through a parameter pointer does NOT require stack realignment, so
 		// implies_clobber_memory() is intentionally NOT used here.
-		if ((cast(u16)clobber.side_effects & asm_ctx->SideEffectFlag_CONTROL) != 0 ||
-		    (cast(u16)clobber.implicit_wr & asm_ctx->ClobberReg_RSP) != 0) {
+		if (clobber.is_call_or_mem()) {
 			asm_acc->saw_call_or_mem = true;
 		}
 
@@ -1159,17 +1158,16 @@ gb_internal void check_mnemonic(AsmCtx *asm_ctx, CheckerContext *ctx, Entity *tm
 		// is not merely CONTROL-with-fallthrough. We approximate "unconditional" as
 		// CONTROL|HALT with no explicit label/operand fallthrough below.
 		{
-			u16 se = cast(u16)clobber.side_effects;
-			bool control = (se & asm_ctx->SideEffectFlag_CONTROL) != 0;
-			bool halt    = (se & asm_ctx->SideEffectFlag_HALT)    != 0;
+			bool control = clobber.has_control();
+			bool halt    = clobber.has_halt();
 			// A conditional branch reads a flag and can fall through -> not terminal.
-			bool conditional = control && (cast(u16)clobber.flags_rd != 0);
+			bool conditional = clobber.is_conditional();
 			asm_acc->last_is_terminal = halt || (control && !conditional);
 		}
 
 		// A branch/call inside the template means subsequent instructions may be reached
 		// out of textual order; stop trusting the linear def model past this point.
-		if (cast(u16)clobber.side_effects & asm_ctx->SideEffectFlag_CONTROL) {
+		if (clobber.has_control()) {
 			asm_acc->straight_line = false;
 		}
 		asm_ctx->clobber_implicit_regs(&tmpl_entity->AsmTemplate.clobber_registers_set, produced);
@@ -2126,6 +2124,8 @@ gb_internal void check_asm_template(AsmCtx *asm_ctx, CheckerContext *ctx, Entity
 gb_internal void check_asm_template_from_entity(CheckerContext *c, Entity *e, DeclInfo *d) {
 	if (build_context.metrics.arch == TargetArch_amd64) {
 		check_asm_template(&g_asm_amd64, c, e, d);
+	} else if (build_context.metrics.arch == TargetArch_riscv64) {
+		check_asm_template(&g_asm_riscv, c, e, d);
 	} else {
 		error(e->token, "asm templates are not currently supported for this target");
 	}

src/main.cpp

@@ -4321,7 +4321,7 @@ int main(int arg_count, char const **arg_ptr) {
 	bool failed_to_cache_parsing = false;
 
 	TIME_SECTION("init asm tables");
-	init_asm_tables();
+	init_asm_tables(build_context.metrics.ptr_size);
 
 	MAIN_TIME_SECTION("parse files");