feat: add wasm dialect

[mooori]

Sets up the wasm dialect and adds the I32Extend8S op to it. See #971 for context and motivation.

[mooori]

Changes in hir + masm: Below is the diff in tests/integration/expected due to adding the wasm dialect. I've captured it before merging the latest changes which remove the expect tests.

What's changing:

• hir: Wasm's Operator::I32Extend8S gets translated to the new op from the wasm dialect
• masm: trunc and sext ops get added. These were not there before as on the next branch I32Extend8 is still a no-op, see bug: incorrect lowering of wasm's i32.extend8_s, i32.extend16_s #960

Diff in hir and masm

diff --git a/tests/integration/expected/add_i8.hir b/tests/integration/expected/add_i8.hir
index 4d336c54..1d66f702 100644
--- a/tests/integration/expected/add_i8.hir
+++ b/tests/integration/expected/add_i8.hir
@@ -3,7 +3,7 @@ builtin.component root_ns:root@1.0.0 {
         public builtin.function @entrypoint(v0: i32, v1: i32) -> i32 {
         ^block6(v0: i32, v1: i32):
             v5 = arith.add v1, v0 : i32 #[overflow = wrapping];
-            v6 = arith.sext v5 : i32;
+            v6 = wasm.i32_extend_8s v5 : i32;
             builtin.ret v6;
         };

diff --git a/tests/integration/expected/add_i8.masm b/tests/integration/expected/add_i8.masm
index 9bc4bdc6..cfce1329 100644
--- a/tests/integration/expected/add_i8.masm
+++ b/tests/integration/expected/add_i8.masm
@@ -22,5 +22,16 @@ end
 @callconv("C")
 pub proc entrypoint(i32, i32) -> i32
     u32wrapping_add
+    push.255
+    u32and
+    dup.0
+    push.128
+    u32and
+    eq.128
+    push.0
+    push.4294967040
+    movup.2
+    cdrop
+    u32or
 end

diff --git a/tests/integration/expected/neg_i8.hir b/tests/integration/expected/neg_i8.hir
index 6c70b2fc..cc0886c4 100644
--- a/tests/integration/expected/neg_i8.hir
+++ b/tests/integration/expected/neg_i8.hir
@@ -4,7 +4,7 @@ builtin.component root_ns:root@1.0.0 {
         ^block6(v0: i32):
             v6 = arith.constant 0 : i32;
             v4 = arith.sub v6, v0 : i32 #[overflow = wrapping];
-            v5 = arith.sext v4 : i32;
+            v5 = wasm.i32_extend_8s v4 : i32;
             builtin.ret v5;
         };

diff --git a/tests/integration/expected/neg_i8.masm b/tests/integration/expected/neg_i8.masm
index 1bcde10d..7dbd0703 100644
--- a/tests/integration/expected/neg_i8.masm
+++ b/tests/integration/expected/neg_i8.masm
@@ -24,5 +24,16 @@ pub proc entrypoint(i32) -> i32
     push.0
     swap.1
     u32wrapping_sub
+    push.255
+    u32and
+    dup.0
+    push.128
+    u32and
+    eq.128
+    push.0
+    push.4294967040
+    movup.2
+    cdrop
+    u32or
 end

diff --git a/tests/integration/expected/shl_i8.hir b/tests/integration/expected/shl_i8.hir
index bdc582a6..61ccf782 100644
--- a/tests/integration/expected/shl_i8.hir
+++ b/tests/integration/expected/shl_i8.hir
@@ -6,7 +6,7 @@ builtin.component root_ns:root@1.0.0 {
             v6 = arith.band v1, v10 : i32;
             v7 = hir.bitcast v6 : u32;
             v8 = arith.shl v0, v7 : i32;
-            v9 = arith.sext v8 : i32;
+            v9 = wasm.i32_extend_8s v8 : i32;
             builtin.ret v9;
         };

diff --git a/tests/integration/expected/shl_i8.masm b/tests/integration/expected/shl_i8.masm
index 9eb91e97..5ccca7a0 100644
--- a/tests/integration/expected/shl_i8.masm
+++ b/tests/integration/expected/shl_i8.masm
@@ -25,5 +25,16 @@ pub proc entrypoint(i32, i32) -> i32
     movup.2
     u32and
     u32shl
+    push.255
+    u32and
+    dup.0
+    push.128
+    u32and
+    eq.128
+    push.0
+    push.4294967040
+    movup.2
+    cdrop
+    u32or
 end

diff --git a/tests/integration/expected/sub_i8.hir b/tests/integration/expected/sub_i8.hir
index a7d64c84..13c29d49 100644
--- a/tests/integration/expected/sub_i8.hir
+++ b/tests/integration/expected/sub_i8.hir
@@ -3,7 +3,7 @@ builtin.component root_ns:root@1.0.0 {
         public builtin.function @entrypoint(v0: i32, v1: i32) -> i32 {
         ^block6(v0: i32, v1: i32):
             v5 = arith.sub v0, v1 : i32 #[overflow = wrapping];
-            v6 = arith.sext v5 : i32;
+            v6 = wasm.i32_extend_8s v5 : i32;
             builtin.ret v6;
         };

diff --git a/tests/integration/expected/sub_i8.masm b/tests/integration/expected/sub_i8.masm
index 11c83459..b0c5a0c3 100644
--- a/tests/integration/expected/sub_i8.masm
+++ b/tests/integration/expected/sub_i8.masm
@@ -23,5 +23,16 @@ end
 pub proc entrypoint(i32, i32) -> i32
     swap.1
     u32wrapping_sub
+    push.255
+    u32and
+    dup.0
+    push.128
+    u32and
+    eq.128
+    push.0
+    push.4294967040
+    movup.2
+    cdrop
+    u32or
 end

[mooori]

All comments addressed and I've implemented Eval and materialize_constant.

After changing the lowering to emit a pop instead of trunc, as expected the trunc related instructions are removed from the generated masm.

Example: Masm generated for addition of `i8`

@locals("2")
@callconv("C")
pub proc entrypoint(i32, i32) -> i32
    u32wrapping_add
    dup.0
    push.128
    u32and
    eq.128
    push.0
    push.4294967040
    movup.2
    cdrop
    u32or
end

[bitwalker]

You need to manipulate the logical operand stack if the emitter directly here (i.e. via stack_mut), or this will emit actual stack manipulation instructions, which we don't want in this case.

[mooori]

Changed it to modify the OperandStack in 5362735.

[mooori]

Actually sign extension might not work as expected when changing the type to I8 only on the logical stack as in 5362735. Since there are multiple things in flight, it's getting difficult to keep tabs on everything. So I would like to suggest the following:

1. For now, land the initial wasm dialect with inst_emitter.trunc(I8) followed by inst_emitter.sext(I32) in the I32Extend8S lowering. As I understand it, that's inefficient but correct.
2. Open a follow-up issue to optimize the wasm::I32Extend8S lowering by getting rid of the trunc.
3. Fix the encoding of smallints: miden-debug-i32

I've assembled these changes locally and then the overflowing arith tests pass for i8 (along with all other tests). Proceeding from there would be easier, I think.

[bitwalker]

I'd like to try and pin down the cause of the issue before we merge - there shouldn't be a difference in observable behavior between emitting trunc and just "casting" the type of the value on top of the operand stack to i8 and sign-extending, if we assume that the lowering from the frontend is correct. The fact that something isn't working means there is a bug somewhere being papered over by emitting the trunc - I'd like us to to dig a bit deeper before we merge to figure that out.

[mooori]

@bitwalker could you take a look at the test added in e954808? I think it shows that:

• The lowering for wasm::I32Extend8S is correct, because test_i32_extend8_s passes when the lowering is implemented as follows with trunc instruction

impl HirLowering for wasm::I32Extend8S {
    fn emit(&self, emitter: &mut BlockEmitter<'_>) -> Result<(), Report> {
        let mut inst_emitter = emitter.inst_emitter(self.as_operation());
        inst_emitter.trunc(&Type::I8, self.span());
        inst_emitter.sext(&Type::I32, self.span());
        Ok(())
    }
}

• However test_i32_extend8_s fails when trunc is not emitted and instead only the type of the operand on the logical stack is modified, as in 5362735.

This should narrow down the problem to "something is wrong when only modifying the operand stack". I'll hold of opening an issue for it, as it's my first time using the builder pattern to construct a package.

[greenhat]

Besides Paul's notes LGTM!