ARM9: Implement SMLA<x><y>

include/mgba/internal/arm/arm.h

@@ -75,7 +75,8 @@ union PSR {
 		unsigned z : 1;
 		unsigned c : 1;
 		unsigned v : 1;
-		unsigned unused : 20;
+		unsigned q : 1;
+		unsigned unused : 19;
 		unsigned i : 1;
 		unsigned f : 1;
 		unsigned t : 1;
@@ -85,7 +86,8 @@ union PSR {
 		unsigned t : 1;
 		unsigned f : 1;
 		unsigned i : 1;
-		unsigned unused : 20;
+		unsigned unused : 19;
+		unsigned q : 1;
 		unsigned v : 1;
 		unsigned c : 1;
 		unsigned z : 1;

include/mgba/internal/arm/decoder.h

@@ -188,6 +188,10 @@ enum ARMMnemonic {
 	ARM_MN_RSB,
 	ARM_MN_RSC,
 	ARM_MN_SBC,
+	ARM_MN_SMLABB,
+	ARM_MN_SMLABT,
+	ARM_MN_SMLATB,
+	ARM_MN_SMLATT,
 	ARM_MN_SMLAL,
 	ARM_MN_SMULL,
 	ARM_MN_STC,

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

@@ -104,13 +104,13 @@
 	DECLARE_INSTRUCTION_ARM(EMITTER, ILL), \
 	DECLARE_INSTRUCTION_ARM(EMITTER, ILL), \
 	DECLARE_INSTRUCTION_ARM(EMITTER, ILL), \
-	DECLARE_INSTRUCTION_ARM(EMITTER, ILL), \
+	MIN_V(DECLARE_INSTRUCTION_ARM(EMITTER, SMLABB), DECLARE_INSTRUCTION_ARM(EMITTER, ILL), V >= 5), \
 	DECLARE_INSTRUCTION_ARM(EMITTER, SWP), \
-	DECLARE_INSTRUCTION_ARM(EMITTER, ILL), \
+	MIN_V(DECLARE_INSTRUCTION_ARM(EMITTER, SMLATB), DECLARE_INSTRUCTION_ARM(EMITTER, ILL), V >= 5), \
 	DECLARE_INSTRUCTION_ARM(EMITTER, STRHP), \
+	MIN_V(DECLARE_INSTRUCTION_ARM(EMITTER, SMLABT), DECLARE_INSTRUCTION_ARM(EMITTER, ILL), V >= 5), \
 	DECLARE_INSTRUCTION_ARM(EMITTER, ILL), \
-	DECLARE_INSTRUCTION_ARM(EMITTER, ILL), \
+	MIN_V(DECLARE_INSTRUCTION_ARM(EMITTER, SMLATT), DECLARE_INSTRUCTION_ARM(EMITTER, ILL), V >= 5), \
-	DECLARE_INSTRUCTION_ARM(EMITTER, ILL), \
 	DECLARE_INSTRUCTION_ARM(EMITTER, ILL), \
 	DECLARE_ARM_ALU_BLOCK(EMITTER, TST, ILL, LDRHP, LDRSBP, LDRSHP), \
 	DECLARE_INSTRUCTION_ARM(EMITTER, MSR), \

src/arm/decoder-arm.c

@@ -311,6 +311,11 @@ DEFINE_LONG_MULTIPLY_DECODER_ARM(SMULL)
 DEFINE_LONG_MULTIPLY_DECODER_ARM(UMLAL)
 DEFINE_LONG_MULTIPLY_DECODER_ARM(UMULL)
 
+DEFINE_MULTIPLY_DECODER_EX_ARM(SMLABB, SMLABB, 0, ARM_OPERAND_REGISTER_4)
+DEFINE_MULTIPLY_DECODER_EX_ARM(SMLABT, SMLABT, 0, ARM_OPERAND_REGISTER_4)
+DEFINE_MULTIPLY_DECODER_EX_ARM(SMLATB, SMLATB, 0, ARM_OPERAND_REGISTER_4)
+DEFINE_MULTIPLY_DECODER_EX_ARM(SMLATT, SMLATT, 0, ARM_OPERAND_REGISTER_4)
+
 // Begin load/store definitions
 
 DEFINE_LOAD_STORE_MODE_2_DECODER_ARM(LDR, LDR, LOAD_CYCLES, ARM_ACCESS_WORD)

src/arm/decoder.c

@@ -276,6 +276,10 @@ static const char* _armMnemonicStrings[] = {
 	"rsb",
 	"rsc",
 	"sbc",
+	"smlabb",
+	"smlabt",
+	"smlatb",
+	"smlatt",
 	"smlal",
 	"smull",
 	"stc",

src/arm/isa-arm.c

@@ -355,6 +355,25 @@ static inline void _immediate(struct ARMCore* cpu, uint32_t opcode) {
 	DEFINE_MULTIPLY_INSTRUCTION_2_EX_ARM(NAME, BODY, , WAIT) \
 	DEFINE_MULTIPLY_INSTRUCTION_2_EX_ARM(NAME ## S, BODY, S_BODY, WAIT)
 
+#define DEFINE_MULTIPLY_INSTRUCTION_3_ARM(NAME, BODY) \
+	DEFINE_INSTRUCTION_ARM(NAME, \
+		int rd = (opcode >> 16) & 0xF; \
+		int rs = (opcode >> 8) & 0xF; \
+		int rn = (opcode >> 12) & 0xF; \
+		int rm = opcode & 0xF; \
+		if (rd == ARM_PC) { \
+			return; \
+		} \
+		/* TODO: Timing */ \
+		int32_t x; \
+		int32_t y; \
+		BODY; \
+		int32_t dn = cpu->gprs[rn]; \
+		int32_t d = x * y; \
+		cpu->gprs[rd] = d + dn; \
+		cpu->cpsr.q = cpu->cpsr.q || ARM_V_ADDITION(d, dn, cpu->gprs[rd]); \
+		currentCycles += cpu->memory.activeNonseqCycles32 - cpu->memory.activeSeqCycles32)
+
 #define DEFINE_LOAD_STORE_INSTRUCTION_EX_ARM(NAME, ADDRESS, WRITEBACK, BODY) \
 	DEFINE_INSTRUCTION_ARM(NAME, \
 		uint32_t address; \
@@ -522,6 +541,22 @@ DEFINE_MULTIPLY_INSTRUCTION_2_ARM(SMLAL,
 	cpu->gprs[rdHi] = cpu->gprs[rdHi] + (d >> 32) + ARM_CARRY_FROM(dm, dn, cpu->gprs[rd]);,
 	ARM_NEUTRAL_HI_S(cpu->gprs[rd], cpu->gprs[rdHi]), 3)
 
+DEFINE_MULTIPLY_INSTRUCTION_3_ARM(SMLABB,
+	x = ARM_SXT_16(cpu->gprs[rm]);
+	y = ARM_SXT_16(cpu->gprs[rs]);)
+
+DEFINE_MULTIPLY_INSTRUCTION_3_ARM(SMLABT,
+	x = ARM_SXT_16(cpu->gprs[rm]);
+	y = ARM_SXT_16(cpu->gprs[rs] >> 16);)
+
+DEFINE_MULTIPLY_INSTRUCTION_3_ARM(SMLATB,
+	x = ARM_SXT_16(cpu->gprs[rm] >> 16);
+	y = ARM_SXT_16(cpu->gprs[rs]);)
+
+DEFINE_MULTIPLY_INSTRUCTION_3_ARM(SMLATT,
+	x = ARM_SXT_16(cpu->gprs[rm] >> 16);
+	y = ARM_SXT_16(cpu->gprs[rs] >> 16);)
+
 DEFINE_MULTIPLY_INSTRUCTION_2_ARM(SMULL,
 	int64_t d = ((int64_t) cpu->gprs[rm]) * ((int64_t) cpu->gprs[rs]);
 	cpu->gprs[rd] = d;