[mono][jit] Reenable branch optimizations when running with llvm. (#97189)

* [mono][jit] Reenable branch optimizations when running with llvm.

* [mono][llvm] Fix the return type of OP_SSE2_PACKUS.

This is supposed to return a 16xi8 so there is no need to cast it.
This only worked by accident because the phi handling code casted it
to the right type most of the time.

* Disable moving out of line bblocks to the end of code when using llvm.

* Fix a crash in OP_SETRET.

Author: vargaz

src/mono/mono/mini/branch-opts.c

@@ -776,7 +776,7 @@ mono_if_conversion (MonoCompile *cfg)
 }
 
 void
-mono_nullify_basic_block (MonoBasicBlock *bb)
+mono_nullify_basic_block (MonoCompile *cfg, MonoBasicBlock *bb)
 {
 	bb->in_count = 0;
 	bb->out_count = 0;
@@ -960,7 +960,7 @@ remove_block_if_useless (MonoCompile *cfg, MonoBasicBlock *bb, MonoBasicBlock *p
 		}
 
 		previous_bb->next_bb = bb->next_bb;
-		mono_nullify_basic_block (bb);
+		mono_nullify_basic_block (cfg, bb);
 
 		return TRUE;
 	} else {
@@ -1046,7 +1046,7 @@ mono_merge_basic_blocks (MonoCompile *cfg, MonoBasicBlock *bb, MonoBasicBlock *b
 		if (bb->next_bb == bbn)
 			bb->next_bb = bbn->next_bb;
 	}
-	mono_nullify_basic_block (bbn);
+	mono_nullify_basic_block (cfg, bbn);
 
 	/*
 	 * If bbn fell through to its next bblock, have to add a branch, since bb
@@ -1291,7 +1291,7 @@ mono_optimize_branches (MonoCompile *cfg)
 				for (i = 0; i < bbn->out_count; i++)
 					replace_in_block (bbn->out_bb [i], bbn, NULL);
 
-				mono_nullify_basic_block (bbn);
+				mono_nullify_basic_block (cfg, bbn);
 				changed = TRUE;
 			}
 
@@ -1341,7 +1341,7 @@ mono_optimize_branches (MonoCompile *cfg)
 				for (i = 0; i < bbn->out_count; i++)
 					replace_in_block (bbn->out_bb [i], bbn, NULL);
 
-				mono_nullify_basic_block (bbn);
+				mono_nullify_basic_block (cfg, bbn);
 				changed = TRUE;
 				continue;
 			}
@@ -1463,7 +1463,7 @@ mono_optimize_branches (MonoCompile *cfg)
 				}
 
 				if (bb->last_ins && MONO_IS_COND_BRANCH_NOFP (bb->last_ins)) {
-					if (bb->last_ins->inst_false_bb && bb->last_ins->inst_false_bb->out_of_line && (bb->region == bb->last_ins->inst_false_bb->region) && !cfg->disable_out_of_line_bblocks) {
+					if (!COMPILE_LLVM (cfg) && bb->last_ins->inst_false_bb && bb->last_ins->inst_false_bb->out_of_line && (bb->region == bb->last_ins->inst_false_bb->region) && !cfg->disable_out_of_line_bblocks) {
 						/* Reverse the branch */
 						bb->last_ins->opcode = GUINT32_TO_OPCODE (mono_reverse_branch_op (bb->last_ins->opcode));
 						bbn = bb->last_ins->inst_false_bb;

src/mono/mono/mini/mini-llvm.c

@@ -6147,7 +6147,7 @@ process_bb (EmitContext *ctx, MonoBasicBlock *bb)
 					break;
 				case LLVMArgVtypeAsScalar:
 					if (is_simd) {
-						retval = LLVMBuildBitCast (builder, values [ins->sreg1], ret_type, "setret_simd_vtype_as_scalar");
+						retval = LLVMBuildBitCast (builder, lhs, ret_type, "setret_simd_vtype_as_scalar");
 					} else {
 						g_assert (addresses [ins->sreg1]);
 						retval = LLVMBuildLoad2 (builder, ret_type, build_ptr_cast (builder, addresses [ins->sreg1]->value, pointer_type (ret_type)), "");
@@ -9579,9 +9579,7 @@ MONO_RESTORE_WARNING
 			LLVMValueRef args [2];
 			args [0] = convert (ctx, lhs, sse_i2_t);
 			args [1] = convert (ctx, rhs, sse_i2_t);
-			values [ins->dreg] = convert (ctx,
-				call_intrins (ctx, INTRINS_SSE_PACKUSWB, args, dname),
-				type_to_sse_type (GTMREG_TO_INT (ins->inst_c1)));
+			values [ins->dreg] = call_intrins (ctx, INTRINS_SSE_PACKUSWB, args, dname);
 			break;
 		}
 

src/mono/mono/mini/mini.c

@@ -3549,8 +3549,6 @@ mini_method_compile (MonoMethod *method, guint32 opts, JitFlags flags, int parts
 		}
 
 		cfg->opt &= ~MONO_OPT_LINEARS;
-
-		cfg->opt &= ~MONO_OPT_BRANCH;
 	}
 
 	cfg->after_method_to_ir = TRUE;
@@ -3975,7 +3973,7 @@ mini_method_compile (MonoMethod *method, guint32 opts, JitFlags flags, int parts
 	if (!cfg->compile_aot)
 		mono_lldb_save_method_info (cfg);
 
-	if (cfg->verbose_level >= 2) {
+	if (cfg->verbose_level >= 2 && !cfg->llvm_only) {
 		char *id =  mono_method_full_name (cfg->method, TRUE);
 		g_print ("\n*** ASM for %s ***\n", id);
 		mono_disassemble_code (cfg, cfg->native_code, cfg->code_len, id + 3);

src/mono/mono/mini/mini.h

@@ -2184,7 +2184,7 @@ void      mono_link_bblock                  (MonoCompile *cfg, MonoBasicBlock *f
 void      mono_unlink_bblock                (MonoCompile *cfg, MonoBasicBlock *from, MonoBasicBlock* to);
 gboolean  mono_bblocks_linked               (MonoBasicBlock *bb1, MonoBasicBlock *bb2);
 void      mono_remove_bblock                (MonoCompile *cfg, MonoBasicBlock *bb);
-void      mono_nullify_basic_block          (MonoBasicBlock *bb);
+void      mono_nullify_basic_block          (MonoCompile *cfg, MonoBasicBlock *bb);
 void      mono_merge_basic_blocks           (MonoCompile *cfg, MonoBasicBlock *bb, MonoBasicBlock *bbn);
 void      mono_optimize_branches            (MonoCompile *cfg);