work through loads and stores

3 files changed, 445 insertions, 5 deletions

diff --git a/src/common.h b/src/common.h
index 218b589..e9f0b69 100644
--- a/src/common.h
+++ b/src/common.h
@@ -31,6 +31,17 @@ enum ejit_opcode {
 	LDXIF,
 	LDXID,
 
+	LDXR8,
+	LDXR16,
+	LDXR32,
+	LDXR64,
+	LDXRU8,
+	LDXRU16,
+	LDXRU32,
+	LDXRU64,
+	LDXRF,
+	LDXRD,
+
 	STI8,
 	STI16,
 	STI32,
@@ -53,6 +64,13 @@ enum ejit_opcode {
 	STXIF,
 	STXID,
 
+	EXTR8,
+	EXTR16,
+	EXTR32,
+	EXTRU8,
+	EXTRU16,
+	EXTRU32,
+
 	ADDR,
 	ADDR_F,
 	ADDI,
@@ -73,6 +91,11 @@ enum ejit_opcode {
 	NEGR,
 	COMR,
 
+	LSHI,
+	LSHR,
+	RSHI,
+	RSHR,
+
 	ANDR,
 	ANDI,
 
@@ -117,6 +140,7 @@ enum ejit_opcode {
 	BLTGTR_F,
 
 	JMP,
+	JMPR,
 
 	BMCI,
 	BMCR,
diff --git a/src/ejit.c b/src/ejit.c
index c4e82b1..4f407aa 100644
--- a/src/ejit.c
+++ b/src/ejit.c
@@ -233,8 +233,58 @@ void ejit_stxi_d(struct ejit_func *s, struct ejit_fpr r0, struct ejit_gpr r1,
 	emit_insn_i(s, STXID, r0.f, r1.r, o);
 }
 
+void ejit_ldi_i8(struct ejit_func *s, struct ejit_gpr r0, void *p)
+{
+	emit_insn_p(s, LDI8, r0.r, 0, p);
+}
+
+void ejit_ldi_i16(struct ejit_func *s, struct ejit_gpr r0, void *p)
+{
+	emit_insn_p(s, LDI16, r0.r, 0, p);
+}
+
+void ejit_ldi_i32(struct ejit_func *s, struct ejit_gpr r0, void *p)
+{
+	emit_insn_p(s, LDI32, r0.r, 0, p);
+}
+
+void ejit_ldi_i64(struct ejit_func *s, struct ejit_gpr r0, void *p)
+{
+	emit_insn_p(s, LDI64, r0.r, 0, p);
+}
+
+void ejit_ldi_u8(struct ejit_func *s, struct ejit_gpr r0, void *p)
+{
+	emit_insn_p(s, LDIU8, r0.r, 0, p);
+}
+
+void ejit_ldi_u16(struct ejit_func *s, struct ejit_gpr r0, void *p)
+{
+	emit_insn_p(s, LDIU16, r0.r, 0, p);
+}
+
+void ejit_ldi_u32(struct ejit_func *s, struct ejit_gpr r0, void *p)
+{
+	emit_insn_p(s, LDIU32, r0.r, 0, p);
+}
+
+void ejit_ldi_u64(struct ejit_func *s, struct ejit_gpr r0, void *p)
+{
+	emit_insn_p(s, LDIU64, r0.r, 0, p);
+}
+
+void ejit_ldi_f(struct ejit_func *s, struct ejit_fpr r0, void *p)
+{
+	emit_insn_p(s, LDIF, r0.f, 0, p);
+}
+
+void ejit_ldi_d(struct ejit_func *s, struct ejit_fpr r0, void *p)
+{
+	emit_insn_p(s, LDID, r0.f, 0, p);
+}
+
 void ejit_ldxi_i8(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
-                   int64_t o)
+                  int64_t o)
 {
 	emit_insn_i(s, LDXI8, r0.r, r1.r, o);
 }
@@ -258,7 +308,7 @@ void ejit_ldxi_i64(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
 }
 
 void ejit_ldxi_u8(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
-                   int64_t o)
+                  int64_t o)
 {
 	emit_insn_i(s, LDXIU8, r0.r, r1.r, o);
 }
@@ -282,17 +332,77 @@ void ejit_ldxi_u64(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
 }
 
 void ejit_ldxi_f(struct ejit_func *s, struct ejit_fpr r0, struct ejit_gpr r1,
-                   int64_t o)
+                 int64_t o)
 {
 	emit_insn_i(s, LDXIF, r0.f, r1.r, o);
 }
 
 void ejit_ldxi_d(struct ejit_func *s, struct ejit_fpr r0, struct ejit_gpr r1,
-                   int64_t o)
+                 int64_t o)
 {
 	emit_insn_i(s, LDXID, r0.f, r1.r, o);
 }
 
+void ejit_ldxr_i8(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+                  struct ejit_gpr r2)
+{
+	emit_insn_r(s, LDXR8, r0.r, r1.r, r2.r);
+}
+
+void ejit_ldxr_i16(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+                   struct ejit_gpr r2)
+{
+	emit_insn_r(s, LDXR16, r0.r, r1.r, r2.r);
+}
+
+void ejit_ldxr_i32(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+                   struct ejit_gpr r2)
+{
+	emit_insn_r(s, LDXR32, r0.r, r1.r, r2.r);
+}
+
+void ejit_ldxr_i64(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+                   struct ejit_gpr r2)
+{
+	emit_insn_r(s, LDXR64, r0.r, r1.r, r2.r);
+}
+
+void ejit_ldxr_u8(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+                  struct ejit_gpr r2)
+{
+	emit_insn_r(s, LDXRU8, r0.r, r1.r, r2.r);
+}
+
+void ejit_ldxr_u16(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+                   struct ejit_gpr r2)
+{
+	emit_insn_r(s, LDXRU16, r0.r, r1.r, r2.r);
+}
+
+void ejit_ldxr_u32(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+                   struct ejit_gpr r2)
+{
+	emit_insn_r(s, LDXRU32, r0.r, r1.r, r2.r);
+}
+
+void ejit_ldxr_u64(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+                   struct ejit_gpr r2)
+{
+	emit_insn_r(s, LDXIU64, r0.r, r1.r, r2.r);
+}
+
+void ejit_ldxr_f(struct ejit_func *s, struct ejit_fpr r0, struct ejit_gpr r1,
+                 struct ejit_gpr r2)
+{
+	emit_insn_r(s, LDXRF, r0.f, r1.r, r2.r);
+}
+
+void ejit_ldxr_d(struct ejit_func *s, struct ejit_fpr r0, struct ejit_gpr r1,
+                 struct ejit_gpr r2)
+{
+	emit_insn_r(s, LDXRD, r0.f, r1.r, r2.r);
+}
+
 void ejit_retr(struct ejit_func *s, struct ejit_gpr r0)
 {
 	emit_insn_r(s, RETR, r0.r, 0, 0);
@@ -313,6 +423,36 @@ void ejit_reti_f(struct ejit_func *s, double f)
 	emit_insn_f(s, RETI_F, 0, 0, f);
 }
 
+void ejit_extr_8(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1)
+{
+	emit_insn_i(s, EXTR8, r0.r, r1.r, 0);
+}
+
+void ejit_extr_16(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1)
+{
+	emit_insn_i(s, EXTR16, r0.r, r1.r, 0);
+}
+
+void ejit_extr_32(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1)
+{
+	emit_insn_i(s, EXTR32, r0.r, r1.r, 0);
+}
+
+void ejit_extr_u8(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1)
+{
+	emit_insn_i(s, EXTRU8, r0.r, r1.r, 0);
+}
+
+void ejit_extr_u16(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1)
+{
+	emit_insn_i(s, EXTRU16, r0.r, r1.r, 0);
+}
+
+void ejit_extr_u32(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1)
+{
+	emit_insn_i(s, EXTRU32, r0.r, r1.r, 0);
+}
+
 void ejit_addr(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
                struct ejit_gpr r2)
 {
@@ -378,6 +518,30 @@ void ejit_divr_f(struct ejit_func *s, struct ejit_fpr r0, struct ejit_fpr r1,
 	emit_insn_r(s, DIVR_F, r0.f, r1.f, r2.f);
 }
 
+void ejit_lshi(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+               int64_t o)
+{
+	emit_insn_i(s, LSHI, r0.r, r1.r, o);
+}
+
+void ejit_lshr(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+               struct ejit_gpr r2)
+{
+	emit_insn_r(s, LSHR, r0.r, r1.r, r2.r);
+}
+
+void ejit_rshi(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+               int64_t o)
+{
+	emit_insn_i(s, RSHI, r0.r, r1.r, o);
+}
+
+void ejit_rshr(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+               struct ejit_gpr r2)
+{
+	emit_insn_r(s, RSHR, r0.r, r1.r, r2.r);
+}
+
 void ejit_andr(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
                struct ejit_gpr r2)
 {
@@ -420,17 +584,83 @@ void ejit_eqr(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
 }
 
 void ejit_eqr_f(struct ejit_func *s, struct ejit_gpr r0, struct ejit_fpr r1,
-		struct ejit_fpr r2)
+                struct ejit_fpr r2)
 {
 	emit_insn_r(s, EQR_F, r0.r, r1.f, r2.f);
 }
 
+void ejit_gtr(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+              struct ejit_gpr r2)
+{
+	emit_insn_r(s, GTR, r0.r, r1.r, r2.r);
+}
+
+void ejit_gtr_u(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+                struct ejit_gpr r2)
+{
+	emit_insn_r(s, GTR_U, r0.r, r1.r, r2.r);
+}
+
+void ejit_gtr_f(struct ejit_func *s, struct ejit_gpr r0, struct ejit_fpr r1,
+                struct ejit_fpr r2)
+{
+	emit_insn_r(s, GTR_F, r0.r, r1.f, r2.f);
+}
+
 void ejit_ltr(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
               struct ejit_gpr r2)
 {
 	emit_insn_r(s, GER, r0.r, r2.r, r1.r);
 }
 
+void ejit_ltr_u(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+                struct ejit_gpr r2)
+{
+	emit_insn_r(s, GER_U, r0.r, r2.r, r1.r);
+}
+
+void ejit_ltr_f(struct ejit_func *s, struct ejit_gpr r0, struct ejit_fpr r1,
+                struct ejit_fpr r2)
+{
+	emit_insn_r(s, GER_F, r0.r, r2.f, r1.f);
+}
+
+void ejit_ger(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+              struct ejit_gpr r2)
+{
+	emit_insn_r(s, GER, r0.r, r1.r, r2.r);
+}
+
+void ejit_ger_u(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+                struct ejit_gpr r2)
+{
+	emit_insn_r(s, GER_U, r0.r, r1.r, r2.r);
+}
+
+void ejit_ger_f(struct ejit_func *s, struct ejit_gpr r0, struct ejit_fpr r1,
+                struct ejit_fpr r2)
+{
+	emit_insn_r(s, GER_F, r0.r, r1.f, r2.f);
+}
+
+void ejit_ler(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+              struct ejit_gpr r2)
+{
+	emit_insn_r(s, GTR, r0.r, r2.r, r1.r);
+}
+
+void ejit_ler_u(struct ejit_func *s, struct ejit_gpr r0, struct ejit_gpr r1,
+                struct ejit_gpr r2)
+{
+	emit_insn_r(s, GTR_U, r0.r, r2.r, r1.r);
+}
+
+void ejit_ler_f(struct ejit_func *s, struct ejit_gpr r0, struct ejit_fpr r1,
+                struct ejit_fpr r2)
+{
+	emit_insn_r(s, GTR_F, r0.r, r2.f, r1.f);
+}
+
 struct ejit_reloc ejit_bner(struct ejit_func *s, struct ejit_gpr r0,
                             struct ejit_gpr r1)
 {
@@ -648,6 +878,13 @@ struct ejit_reloc ejit_jmp(struct ejit_func *s)
 	return (struct ejit_reloc){.insn = addr};
 }
 
+struct ejit_reloc ejit_jmpr(struct ejit_func *s, struct ejit_gpr r0)
+{
+	size_t addr = vec_len(&s->insns);
+	emit_insn_i(s, JMPR, 0, r0.r, 0);
+	return (struct ejit_reloc){.insn = addr};
+}
+
 struct ejit_reloc ejit_bmci(struct ejit_func *s, struct ejit_gpr r0, int64_t o)
 {
 	size_t addr = vec_len(&s->insns);
diff --git a/src/interp.c b/src/interp.c
index 383599e..30a7672 100644
--- a/src/interp.c
+++ b/src/interp.c
@@ -13,6 +13,13 @@ union interp_ret ejit_interp(struct ejit_func *f, size_t argc,
 		[MOVR] = &&MOVR,
 		[MOVR_F] = &&MOVR_F,
 
+		[EXTR8] = &&EXTR8,
+		[EXTR16] = &&EXTR16,
+		[EXTR32] = &&EXTR32,
+		[EXTRU8] = &&EXTRU8,
+		[EXTRU16] = &&EXTRU16,
+		[EXTRU32] = &&EXTRU32,
+
 		[ADDR] = &&ADDR,
 		[ADDR_F] = &&ADDR_F,
 		[ADDI] = &&ADDI,
@@ -29,6 +36,11 @@ union interp_ret ejit_interp(struct ejit_func *f, size_t argc,
 		[DIVR_U] = &&DIVR_U,
 		[DIVR_F] = &&DIVR_F,
 
+		[LSHI] = &&LSHI,
+		[LSHR] = &&LSHR,
+		[RSHI] = &&RSHI,
+		[RSHR] = &&RSHR,
+
 		[ANDR] = &&ANDR,
 		[ANDI] = &&ANDI,
 
@@ -56,6 +68,17 @@ union interp_ret ejit_interp(struct ejit_func *f, size_t argc,
 		[STXIF] = &&STXIF,
 		[STXID] = &&STXID,
 
+		[LDI8] = &&LDI8,
+		[LDI16] = &&LDI16,
+		[LDI32] = &&LDI32,
+		[LDI64] = &&LDI64,
+		[LDIU8] = &&LDIU8,
+		[LDIU16] = &&LDIU16,
+		[LDIU32] = &&LDIU32,
+		[LDIU64] = &&LDIU64,
+		[LDIF] = &&LDIF,
+		[LDID] = &&LDID,
+
 		[LDXI8] = &&LDXI8,
 		[LDXI16] = &&LDXI16,
 		[LDXI32] = &&LDXI32,
@@ -67,6 +90,17 @@ union interp_ret ejit_interp(struct ejit_func *f, size_t argc,
 		[LDXIF] = &&LDXIF,
 		[LDXID] = &&LDXID,
 
+		[LDXR8] = &&LDXR8,
+		[LDXR16] = &&LDXR16,
+		[LDXR32] = &&LDXR32,
+		[LDXR64] = &&LDXR64,
+		[LDXRU8] = &&LDXRU8,
+		[LDXRU16] = &&LDXRU16,
+		[LDXRU32] = &&LDXRU32,
+		[LDXRU64] = &&LDXRU64,
+		[LDXRF] = &&LDXRF,
+		[LDXRD] = &&LDXRD,
+
 		[BNER] = &&BNER,
 		[BNEI] = &&BNEI,
 		[BNER_F] = &&BNER_F,
@@ -96,6 +130,7 @@ union interp_ret ejit_interp(struct ejit_func *f, size_t argc,
 		[BLTGTR_F] = &&BLTGTR_F,
 
 		[JMP] = &&JMP,
+		[JMPR] = &&JMPR,
 
 		[BMCI] = &&BMCI,
 		[BMCR] = &&BMCR,
@@ -175,6 +210,30 @@ union interp_ret ejit_interp(struct ejit_func *f, size_t argc,
 	fpr[i.r0] = fpr[i.r1];
 	DISPATCH();
 
+	DO(EXTR8);
+	gpr[i.r0] = (int8_t)gpr[i.r1];
+	DISPATCH();
+
+	DO(EXTR16);
+	gpr[i.r0] = (int16_t)gpr[i.r1];
+	DISPATCH();
+
+	DO(EXTR32);
+	gpr[i.r0] = (int32_t)gpr[i.r1];
+	DISPATCH();
+
+	DO(EXTRU8);
+	gpr[i.r0] = (uint8_t)gpr[i.r1];
+	DISPATCH();
+
+	DO(EXTRU16);
+	gpr[i.r0] = (uint16_t)gpr[i.r1];
+	DISPATCH();
+
+	DO(EXTRU32);
+	gpr[i.r0] = (uint32_t)gpr[i.r1];
+	DISPATCH();
+
 	DO(ADDR);
 	gpr[i.r0] = gpr[i.r1] + gpr[i.r2];
 	DISPATCH();
@@ -219,6 +278,22 @@ union interp_ret ejit_interp(struct ejit_func *f, size_t argc,
 	fpr[i.r0] = fpr[i.r1] / fpr[i.r2];
 	DISPATCH();
 
+	DO(LSHI);
+	gpr[i.r0] = gpr[i.r1] << i.o;
+	DISPATCH();
+
+	DO(LSHR);
+	gpr[i.r0] = gpr[i.r1] << gpr[i.r2];
+	DISPATCH();
+
+	DO(RSHI);
+	gpr[i.r0] = gpr[i.r1] >> i.o;
+	DISPATCH();
+
+	DO(RSHR);
+	gpr[i.r0] = gpr[i.r1] >> gpr[i.r2];
+	DISPATCH();
+
 	DO(ANDR);
 	gpr[i.r0] = gpr[i.r1] & gpr[i.r2];
 	DISPATCH();
@@ -305,6 +380,56 @@ union interp_ret ejit_interp(struct ejit_func *f, size_t argc,
 	*addr = fpr[i.r0];
 	DISPATCH();
 
+	DO(LDI8);
+	int8_t *addr = (int8_t *)i.p;
+	gpr[i.r0] = *addr;
+	DISPATCH();
+
+	DO(LDI16);
+	int16_t *addr = (int16_t *)i.p;
+	gpr[i.r0] = *addr;
+	DISPATCH();
+
+	DO(LDI32);
+	int32_t *addr = (int32_t *)i.p;
+	gpr[i.r0] = *addr;
+	DISPATCH();
+
+	DO(LDI64);
+	int64_t *addr = (int64_t *)i.p;
+	gpr[i.r0] = *addr;
+	DISPATCH();
+
+	DO(LDIU8);
+	uint8_t *addr = (uint8_t *)i.p;
+	gpr[i.r0] = *addr;
+	DISPATCH();
+
+	DO(LDIU16);
+	uint16_t *addr = (uint16_t *)i.p;
+	gpr[i.r0] = *addr;
+	DISPATCH();
+
+	DO(LDIU32);
+	uint32_t *addr = (uint32_t *)i.p;
+	gpr[i.r0] = *addr;
+	DISPATCH();
+
+	DO(LDIU64);
+	uint64_t *addr = (uint64_t *)i.p;
+	gpr[i.r0] = *addr;
+	DISPATCH();
+
+	DO(LDIF);
+	float *addr = (float *)i.p;
+	fpr[i.r0] = *addr;
+	DISPATCH();
+
+	DO(LDID);
+	double *addr = (double *)i.p;
+	fpr[i.r0] = *addr;
+	DISPATCH();
+
 	DO(LDXI8);
 	int8_t *addr = (int8_t *)(gpr[i.r1] + i.o);
 	gpr[i.r0] = *addr;
@@ -355,6 +480,56 @@ union interp_ret ejit_interp(struct ejit_func *f, size_t argc,
 	fpr[i.r0] = *addr;
 	DISPATCH();
 
+	DO(LDXR8);
+	int8_t *addr = (int8_t *)(gpr[i.r1] + gpr[i.r2]);
+	gpr[i.r0] = *addr;
+	DISPATCH();
+
+	DO(LDXR16);
+	int16_t *addr = (int16_t *)(gpr[i.r1] + gpr[i.r2]);
+	gpr[i.r0] = *addr;
+	DISPATCH();
+
+	DO(LDXR32);
+	int32_t *addr = (int32_t *)(gpr[i.r1] + gpr[i.r2]);
+	gpr[i.r0] = *addr;
+	DISPATCH();
+
+	DO(LDXR64);
+	int64_t *addr = (int64_t *)(gpr[i.r1] + gpr[i.r2]);
+	gpr[i.r0] = *addr;
+	DISPATCH();
+
+	DO(LDXRU8);
+	uint8_t *addr = (uint8_t *)(gpr[i.r1] + gpr[i.r2]);
+	gpr[i.r0] = *addr;
+	DISPATCH();
+
+	DO(LDXRU16);
+	uint16_t *addr = (uint16_t *)(gpr[i.r1] + gpr[i.r2]);
+	gpr[i.r0] = *addr;
+	DISPATCH();
+
+	DO(LDXRU32);
+	uint32_t *addr = (uint32_t *)(gpr[i.r1] + gpr[i.r2]);
+	gpr[i.r0] = *addr;
+	DISPATCH();
+
+	DO(LDXRU64);
+	uint64_t *addr = (uint64_t *)(gpr[i.r1] + gpr[i.r2]);
+	gpr[i.r0] = *addr;
+	DISPATCH();
+
+	DO(LDXRF);
+	float *addr = (float *)(gpr[i.r1] + gpr[i.r2]);
+	fpr[i.r0] = *addr;
+	DISPATCH();
+
+	DO(LDXRD);
+	double *addr = (double *)(gpr[i.r1] + gpr[i.r2]);
+	fpr[i.r0] = *addr;
+	DISPATCH();
+
 	DO(BNER);
 	if (gpr[i.r1] != gpr[i.r2])
 		JUMP(i.r0);
@@ -491,6 +666,10 @@ union interp_ret ejit_interp(struct ejit_func *f, size_t argc,
 	JUMP(i.r0);
 	DISPATCH();
 
+	DO(JMPR);
+	JUMP(gpr[i.r1]);
+	DISPATCH();
+
 	DO(BMSR);
 	if (gpr[i.r1] & gpr[i.r2])
 		JUMP(i.r0);