ir3emu: some fixes

1 files changed, 55 insertions, 32 deletions

diff --git a/ir3emu.c b/ir3emu.c
index 2319714..c5bacfc 100644
--- a/ir3emu.c
+++ b/ir3emu.c
@@ -365,26 +365,22 @@ emu_run_cat1(struct ir3_emu_ctx *ctx, instr_cat1_t *cat1)
 		});
 	}
 
-	if (type_float(cat1->src_type) &&
-			type_uint(cat1->dst_type)) {
+	if (type_float(cat1->src_type) && type_uint(cat1->dst_type)) {
 		dst.u = src.f;
-	} else if (type_float(cat1->src_type) &&
-			type_sint(cat1->dst_type)) {
+	} else if (type_float(cat1->src_type) && type_sint(cat1->dst_type)) {
 		dst.i = src.f;
-	} else if (type_uint(cat1->src_type) &&
-			type_sint(cat1->dst_type)) {
+	} else if (type_uint(cat1->src_type) && type_sint(cat1->dst_type)) {
 		dst.i = src.u;
-	} else if (type_uint(cat1->src_type) &&
-			type_float(cat1->dst_type)) {
-		dst.i = src.f;
-	} else if (type_sint(cat1->src_type) &&
-			type_float(cat1->dst_type)) {
-		dst.f = src.i;
-	} else if (type_sint(cat1->src_type) &&
-			type_uint(cat1->dst_type)) {
+	} else if (type_uint(cat1->src_type) && type_float(cat1->dst_type)) {
 		dst.f = src.u;
+	} else if (type_sint(cat1->src_type) && type_float(cat1->dst_type)) {
+		dst.f = src.i;
+	} else if (type_sint(cat1->src_type) && type_uint(cat1->dst_type)) {
+		dst.u = src.i;
+	} else if (cat1->src_type == cat1->dst_type) {
+		dst.u = src.u;
 	} else {
-		dst.f = src.f;
+		assert(0);
 	}
 
 	write_dst(ctx, dst, (struct ir3_regdst){
@@ -399,6 +395,19 @@ emu_run_cat2(struct ir3_emu_ctx *ctx, instr_cat2_t *cat2)
 {
 	struct ir3_regsrc src2rs;
 	regval_t dst, src1, src2;
+	bool absnegf;
+
+	switch (cat2->opc) {
+	/* TODO the .b instructions do bitwise negate, rather than
+	 * float abs/neg, so they also need special handling..
+	 */
+	case OPC_ABSNEG_S:
+		absnegf = false;
+		break;
+	default:
+		absnegf = true;
+		break;
+	}
 
 	if (cat2->c1.src1_c) {
 		src1 = read_src(ctx, 0, (struct ir3_regsrc){
@@ -407,8 +416,8 @@ emu_run_cat2(struct ir3_emu_ctx *ctx, instr_cat2_t *cat2)
 			.r    = cat2->src1_r,
 			.c    = cat2->c1.src1_c,
 			.im   = cat2->src1_im,
-			.neg  = cat2->src1_neg,
-			.abs  = cat2->src1_abs,
+			.neg  = cat2->src1_neg && absnegf,
+			.abs  = cat2->src1_abs && absnegf,
 		});
 	} else if (cat2->rel1.src1_rel) {
 		src1 = read_src(ctx, 0, (struct ir3_regsrc){
@@ -417,8 +426,8 @@ emu_run_cat2(struct ir3_emu_ctx *ctx, instr_cat2_t *cat2)
 			.r    = cat2->src1_r,
 			.c    = cat2->rel1.src1_c,
 			.im   = cat2->src1_im,
-			.neg  = cat2->src1_neg,
-			.abs  = cat2->src1_abs,
+			.neg  = cat2->src1_neg && absnegf,
+			.abs  = cat2->src1_abs && absnegf,
 			.rel  = cat2->rel1.src1_rel,
 		});
 	} else {
@@ -427,8 +436,8 @@ emu_run_cat2(struct ir3_emu_ctx *ctx, instr_cat2_t *cat2)
 			.full = cat2->full,
 			.r    = cat2->src1_r,
 			.im   = cat2->src1_im,
-			.neg  = cat2->src1_neg,
-			.abs  = cat2->src1_abs,
+			.neg  = cat2->src1_neg && absnegf,
+			.abs  = cat2->src1_abs && absnegf,
 		});
 	}
 
@@ -439,8 +448,8 @@ emu_run_cat2(struct ir3_emu_ctx *ctx, instr_cat2_t *cat2)
 			.r    = cat2->src2_r,
 			.c    = cat2->c2.src2_c,
 			.im   = cat2->src2_im,
-			.neg  = cat2->src2_neg,
-			.abs  = cat2->src2_abs,
+			.neg  = cat2->src2_neg && absnegf,
+			.abs  = cat2->src2_abs && absnegf,
 		};
 	} else if (cat2->rel2.src2_rel) {
 		src2rs = (struct ir3_regsrc){
@@ -449,8 +458,8 @@ emu_run_cat2(struct ir3_emu_ctx *ctx, instr_cat2_t *cat2)
 			.r    = cat2->src2_r,
 			.c    = cat2->rel2.src2_c,
 			.im   = cat2->src2_im,
-			.neg  = cat2->src2_neg,
-			.abs  = cat2->src2_abs,
+			.neg  = cat2->src2_neg && absnegf,
+			.abs  = cat2->src2_abs && absnegf,
 			.rel  = cat2->rel2.src2_rel,
 		};
 	} else {
@@ -459,8 +468,8 @@ emu_run_cat2(struct ir3_emu_ctx *ctx, instr_cat2_t *cat2)
 			.full = cat2->full,
 			.r    = cat2->src2_r,
 			.im   = cat2->src2_im,
-			.neg  = cat2->src2_neg,
-			.abs  = cat2->src2_abs,
+			.neg  = cat2->src2_neg && absnegf,
+			.abs  = cat2->src2_abs && absnegf,
 		};
 	}
 
@@ -473,6 +482,22 @@ emu_run_cat2(struct ir3_emu_ctx *ctx, instr_cat2_t *cat2)
 		src2 = read_src(ctx, 1, src2rs);
 	}
 
+	if (!absnegf) {
+		switch (cat2->opc) {
+		/* TODO the .b instructions do bitwise negate, rather than
+		 * float abs/neg, so they also need special handling..
+		 */
+		case OPC_ABSNEG_S:
+			if (cat2->src1_neg)
+				src1.i = -src1.i;
+			if (cat2->src1_abs)
+				src1.i = abs(src1.i);
+			break;
+		default:
+			break;
+		}
+	}
+
 	switch (cat2->opc) {
 	case OPC_ADD_F:
 		dst.f = src1.f + src2.f;
@@ -490,8 +515,7 @@ emu_run_cat2(struct ir3_emu_ctx *ctx, instr_cat2_t *cat2)
 		dst.i = src1.i * src2.i;
 		break;
 	case OPC_MULL_U:
-		/* technically, I think, this should multiply just the low bits.. */
-		dst.u = src1.u * src2.u;
+		dst.u = (src1.u & 0xffff) * (src2.u & 0xffff);
 		break;
 	case OPC_CMPS_F:
 		switch (cat2->cond) {
@@ -660,14 +684,13 @@ emu_run_cat3(struct ir3_emu_ctx *ctx, instr_cat3_t *cat3)
 		dst.f = (src1.f * src2.f) + src3.f;
 		break;
 	case OPC_MADSH_M16:
-		/* technically this should mul-add shift high, ie. ah * bl << 16 */
-		dst.u = (src1.u * src2.u) + src3.u;
+		dst.u = (((src1.u >> 16) * (src2.u & 0xffff)) << 16) + src3.u;
 		break;
 	case OPC_SEL_F32:
 		dst.f = (src2.i >= 0) ? src3.f : src1.f;
 		break;
 	case OPC_SEL_B32:
-		dst.f = src2.u ? src3.f : src1.f;
+		dst.f = src2.u ? src1.f : src3.f;
 		break;
 	default:
 		emu_fatal(ctx, "unhandled cat3 opc: %u", cat3->opc);