[mono][interp] Avoid running super instruction pass more than once

We will avoid running it if we know will do another iteration later (especially useful when first iteration is ssa disabled). It can still happen for it to run multiple times (when bblock optimization requests retry), but this should be very rare. This will save compilation time and should also remove the necessity of handling super instructions in constant folding pass.

Reverts dotnet#99055 since it is no longer necessary.

src/mono/mono/mini/interp/transform-opt.c

@@ -2130,12 +2130,6 @@ interp_get_mt_for_ldind (int ldind_op)
 		result.field = op val->field; \
 		break;
 
-#define INTERP_FOLD_SHIFTOP_IMM(opcode,local_type,field,shift_op,cast_type) \
-	case opcode: \
-		result.type = local_type; \
-		result.field = (cast_type)val->field shift_op ins->data [0]; \
-		break;
-
 #define INTERP_FOLD_CONV(opcode,val_type_dst,field_dst,val_type_src,field_src,cast_type) \
 	case opcode: \
 		result.type = val_type_dst; \
@@ -2175,19 +2169,6 @@ interp_fold_unop (TransformData *td, InterpInst *ins)
 		INTERP_FOLD_UNOP (MINT_NOT_I8, VAR_VALUE_I8, l, ~);
 		INTERP_FOLD_UNOP (MINT_CEQ0_I4, VAR_VALUE_I4, i, 0 ==);
 
-		INTERP_FOLD_UNOP (MINT_ADD_I4_IMM, VAR_VALUE_I4, i, ((gint32)(gint16)ins->data [0])+);
-		INTERP_FOLD_UNOP (MINT_ADD_I8_IMM, VAR_VALUE_I8, l, ((gint64)(gint16)ins->data [0])+);
-
-		INTERP_FOLD_UNOP (MINT_MUL_I4_IMM, VAR_VALUE_I4, i, ((gint32)(gint16)ins->data [0])*);
-		INTERP_FOLD_UNOP (MINT_MUL_I8_IMM, VAR_VALUE_I8, l, ((gint64)(gint16)ins->data [0])*);
-
-		INTERP_FOLD_SHIFTOP_IMM (MINT_SHR_UN_I4_IMM, VAR_VALUE_I4, i, >>, guint32);
-		INTERP_FOLD_SHIFTOP_IMM (MINT_SHR_UN_I8_IMM, VAR_VALUE_I8, l, >>, guint64);
-		INTERP_FOLD_SHIFTOP_IMM (MINT_SHL_I4_IMM, VAR_VALUE_I4, i, <<, gint32);
-		INTERP_FOLD_SHIFTOP_IMM (MINT_SHL_I8_IMM, VAR_VALUE_I8, l, <<, gint64);
-		INTERP_FOLD_SHIFTOP_IMM (MINT_SHR_I4_IMM, VAR_VALUE_I4, i, >>, gint32);
-		INTERP_FOLD_SHIFTOP_IMM (MINT_SHR_I8_IMM, VAR_VALUE_I8, l, >>, gint64);
-
 		INTERP_FOLD_CONV (MINT_CONV_I1_I4, VAR_VALUE_I4, i, VAR_VALUE_I4, i, gint8);
 		INTERP_FOLD_CONV (MINT_CONV_I1_I8, VAR_VALUE_I4, i, VAR_VALUE_I8, l, gint8);
 		INTERP_FOLD_CONV (MINT_CONV_U1_I4, VAR_VALUE_I4, i, VAR_VALUE_I4, i, guint8);
@@ -2921,7 +2902,7 @@ interp_cprop (TransformData *td)
 				td->var_values [dreg].type = VAR_VALUE_I4;
 				td->var_values [dreg].i = (gint32)td->data_items [ins->data [0]];
 #endif
-			} else if (MINT_IS_UNOP (opcode) || MINT_IS_BINOP_IMM (opcode)) {
+			} else if (MINT_IS_UNOP (opcode)) {
 				ins = interp_fold_unop (td, ins);
 			} else if (MINT_IS_UNOP_CONDITIONAL_BRANCH (opcode)) {
 				ins = interp_fold_unop_cond_br (td, bb, ins);
@@ -3925,8 +3906,15 @@ interp_optimize_code (TransformData *td)
 
 	// We run this after var deadce to detect more single use vars. This pass will clear
 	// unnecessary instruction on the fly so deadce is no longer needed to run.
-	if (mono_interp_opt & INTERP_OPT_SUPER_INSTRUCTIONS)
-		MONO_TIME_TRACK (mono_interp_stats.super_instructions_time, interp_super_instructions (td));
+	if (mono_interp_opt & INTERP_OPT_SUPER_INSTRUCTIONS) {
+		// This pass is enough to be called only once, after all cprop and other optimizations
+		// are done. The problem is that currently it needs to run over code in SSA form, so we
+		// can't just run it at the very end of optimization cycles. Also bblock optimization can
+		// lead to another optimization iteration, so we can still end up running it multiple times.
+		// Basic block optimization currently needs to run after we exited SSA.
+		if (!ssa_enabled_retry && !td->need_optimization_retry)
+			MONO_TIME_TRACK (mono_interp_stats.super_instructions_time, interp_super_instructions (td));
+	}
 
 	if (!td->disable_ssa)
 		interp_exit_ssa (td);