Skip to contentSkip to conversation

Adding type information to instructions

PATCH David Given 3 participants 3 messages Jul 4, 2008 – Jul 7, 2008 Sparse checker

lore
Share & tools
lore
#0 lore raw

I've been having trouble finding the types of pseudos; the approach I've been using, which is to follow the chain of pseudo's definers until I find an instruction with enough type information to use, turns out to fail utterly on OP_LOAD instructions; I can't find any way of getting enough information from the opcode's arguments to determine the type of the target.

As a result, I've had to make a minor patch to the lineariser code. Given that the lineariser knows the information I need at the point where it generates the instruction, it would seem to make sense to tag the instruction with the type. This is made quite easy by the way that the lineariser has a allocate-typed-instruction function.

Patch enclosed; it's very simple.

Does this seem like a reasonable approach? Is it something that would be useful to have in the base builds? Are there any ramifications I should be aware of?

Signature · 4 lines
┌─── dg@cowlark.com ───── http://www.cowlark.com ─────
│ "I have always wished for my computer to be as easy to use as my
│ telephone; my wish has come true because I can no longer figure out
│ how to use my telephone." --- Bjarne Stroustrup
linearize.c +3 -1
diff -rwu sparse.orig/linearize.c sparse/linearize.c
@@ -55,7 +55,9 @@
 
 static struct instruction *alloc_typed_instruction(int opcode, struct symbol *type)
 {
-	return alloc_instruction(opcode, type_size(type));
+	struct instruction *insn = alloc_instruction(opcode, type_size(type));
+	insn->type = type;
+	return insn;
 }
 
 static struct entrypoint *alloc_entrypoint(void)
linearize.h +1 -0
diff -rwu sparse.orig/linearize.h sparse/linearize.h
@@ -71,6 +71,7 @@
 		 size:24;
 	struct basic_block *bb;
 	struct position pos;
+	struct symbol *type;
 	union {
 		pseudo_t target;
 		pseudo_t cond;		/* for branch and switch */
#1 lore raw
Quoted 13 lines · David Given

I've been having trouble finding the types of pseudos; the approach I've been using, which is to follow the chain of pseudo's definers until I find an instruction with enough type information to use, turns out to fail utterly on OP_LOAD instructions; I can't find any way of getting enough information from the opcode's arguments to determine the type of the target.

As a result, I've had to make a minor patch to the lineariser code. Given that the lineariser knows the information I need at the point where it generates the instruction, it would seem to make sense to tag the instruction with the type. This is made quite easy by the way that the lineariser has a allocate-typed-instruction function.

Quoted 5 lines

Patch enclosed; it's very simple.

Does this seem like a reasonable approach? Is it something that would be useful to have in the base builds? Are there any ramifications I should be aware of?

I can't think of any fundamental reason not to do this, other than the standard reason of data structure size. However, I'd like to hear something from others on the list before accepting this patch.

- Josh Triplett

#2 lore raw
Quoted 2 lines · Josh Triplett

I can't think of any fundamental reason not to do this, other than the standard reason of data structure size.

There is other ways to avoid the size blow up. At the linearized byte code level,
the back end only need to know it is one of the stander C type . So we have an array of stander ctype C type similar to ctype_declaration[].

We just need to replace the size member with the array index in the instruction struct..

Chris