ARM9: Implemented SMLAW<y> and SMULW<y> (#609)

CHANGES

@@ -12,6 +12,7 @@ Bugfixes:
 Misc:
  - DS: Set boot complete bit in RAM on boot (fixes mgba.io/i/576, mgba.io/i/580, mgba.io/i/586)
  - DS Memory: Ensure DS9 I/O is 8-byte aligned
+ - ARM9: Implement SMLAW and SMULW
  - Qt: Add .nds files to the extension list in Info.plist
 
 0.6.0: (Future)

include/mgba/internal/arm/decoder.h

@@ -190,14 +190,18 @@ enum ARMMnemonic {
 	ARM_MN_SBC,
 	ARM_MN_SMLABB,
 	ARM_MN_SMLABT,
+	ARM_MN_SMLAL,
 	ARM_MN_SMLATB,
 	ARM_MN_SMLATT,
+	ARM_MN_SMLAWB,
+	ARM_MN_SMLAWT,
 	ARM_MN_SMULBB,
 	ARM_MN_SMULBT,
+	ARM_MN_SMULL,
 	ARM_MN_SMULTB,
 	ARM_MN_SMULTT,
-	ARM_MN_SMLAL,
-	ARM_MN_SMULL,
+	ARM_MN_SMULWB,
+	ARM_MN_SMULWT,
 	ARM_MN_STC,
 	ARM_MN_STM,
 	ARM_MN_STR,

include/mgba/internal/arm/emitter-arm.h

@@ -143,13 +143,13 @@
 	DECLARE_INSTRUCTION_ARM(EMITTER, ILL), \
 	DECLARE_INSTRUCTION_ARM(EMITTER, ILL), \
 	DECLARE_INSTRUCTION_ARM(EMITTER, BKPT), \
+	MIN_V(DECLARE_INSTRUCTION_ARM(EMITTER, SMLAWB), DECLARE_INSTRUCTION_ARM(EMITTER, ILL), V >= 5), \
 	DECLARE_INSTRUCTION_ARM(EMITTER, ILL), \
-	DECLARE_INSTRUCTION_ARM(EMITTER, ILL), \
-	DECLARE_INSTRUCTION_ARM(EMITTER, ILL), \
+	MIN_V(DECLARE_INSTRUCTION_ARM(EMITTER, SMULWB), DECLARE_INSTRUCTION_ARM(EMITTER, ILL), V >= 5), \
 	DECLARE_INSTRUCTION_ARM(EMITTER, STRHPW), \
+	MIN_V(DECLARE_INSTRUCTION_ARM(EMITTER, SMLAWT), DECLARE_INSTRUCTION_ARM(EMITTER, ILL), V >= 5), \
 	DECLARE_INSTRUCTION_ARM(EMITTER, ILL), \
-	DECLARE_INSTRUCTION_ARM(EMITTER, ILL), \
-	DECLARE_INSTRUCTION_ARM(EMITTER, ILL), \
+	MIN_V(DECLARE_INSTRUCTION_ARM(EMITTER, SMULWT), DECLARE_INSTRUCTION_ARM(EMITTER, ILL), V >= 5), \
 	DECLARE_INSTRUCTION_ARM(EMITTER, ILL), \
 	DECLARE_ARM_ALU_BLOCK(EMITTER, TEQ, ILL, LDRHPW, LDRSBPW, LDRSHPW), \
 	DECLARE_INSTRUCTION_ARM(EMITTER, MRSR), \

src/arm/decoder-arm.c

@@ -320,6 +320,11 @@ DEFINE_MULTIPLY_DECODER_EX_ARM(SMULBT, SMULBT, 0, 0)
 DEFINE_MULTIPLY_DECODER_EX_ARM(SMULTB, SMULTB, 0, 0)
 DEFINE_MULTIPLY_DECODER_EX_ARM(SMULTT, SMULTT, 0, 0)
 
+DEFINE_MULTIPLY_DECODER_EX_ARM(SMLAWB, SMLAWB, 0, ARM_OPERAND_REGISTER_4)
+DEFINE_MULTIPLY_DECODER_EX_ARM(SMLAWT, SMLAWT, 0, ARM_OPERAND_REGISTER_4)
+DEFINE_MULTIPLY_DECODER_EX_ARM(SMULWB, SMULWB, 0, 0)
+DEFINE_MULTIPLY_DECODER_EX_ARM(SMULWT, SMULWT, 0, 0)
+
 // Begin load/store definitions
 
 DEFINE_LOAD_STORE_MODE_2_DECODER_ARM(LDR, LDR, LOAD_CYCLES, ARM_ACCESS_WORD)

src/arm/decoder.c

@@ -278,14 +278,18 @@ static const char* _armMnemonicStrings[] = {
 	"sbc",
 	"smlabb",
 	"smlabt",
+	"smlal",
 	"smlatb",
 	"smlatt",
+	"smlawb",
+	"smlawt",
 	"smulbb",
 	"smulbt",
+	"smull",
 	"smultb",
 	"smultt",
-	"smlal",
-	"smull",
+	"smulwb",
+	"smulwt",
 	"stc",
 	"stm",
 	"str",

src/arm/isa-arm.c

@@ -401,6 +401,14 @@ static inline void _immediate(struct ARMCore* cpu, uint32_t opcode) {
 		y = ARM_SXT_16(cpu->gprs[rs] >> 16); \
 		BODY)
 
+#define DEFINE_MULTIPLY_INSTRUCTION_WY_ARM(NAME, BODY) \
+	DEFINE_MULTIPLY_INSTRUCTION_3_ARM(NAME ## B, \
+		y = ARM_SXT_16(cpu->gprs[rs]); \
+		BODY) \
+	DEFINE_MULTIPLY_INSTRUCTION_3_ARM(NAME ## T, \
+		y = ARM_SXT_16(cpu->gprs[rs] >> 16); \
+		BODY) \
+
 #define DEFINE_LOAD_STORE_INSTRUCTION_EX_ARM(NAME, ADDRESS, WRITEBACK, BODY) \
 	DEFINE_INSTRUCTION_ARM(NAME, \
 		uint32_t address; \
@@ -576,6 +584,14 @@ DEFINE_MULTIPLY_INSTRUCTION_XY_ARM(SMLA,
 
 DEFINE_MULTIPLY_INSTRUCTION_XY_ARM(SMUL, cpu->gprs[rd] = x * y;)
 
+DEFINE_MULTIPLY_INSTRUCTION_WY_ARM(SMLAW,
+	int32_t dn = cpu->gprs[rn]; \
+	int32_t d = (((int64_t) cpu->gprs[rm]) * ((int64_t) y)) >> 16; \
+	cpu->gprs[rd] = d + dn; \
+	cpu->cpsr.q = cpu->cpsr.q || ARM_V_ADDITION(d, dn, cpu->gprs[rd]);)
+
+DEFINE_MULTIPLY_INSTRUCTION_WY_ARM(SMULW, cpu->gprs[rd] = (((int64_t) cpu->gprs[rm]) * ((int64_t) y)) >> 16;)
+
 DEFINE_MULTIPLY_INSTRUCTION_2_ARM(SMULL,
 	int64_t d = ((int64_t) cpu->gprs[rm]) * ((int64_t) cpu->gprs[rs]);
 	cpu->gprs[rd] = d;