[Mlir-commits] [mlir] [WIP][mlir][EmitC] Model lvalues as a type in EmitC (PR #91475)

[Gil Rapaport]

================
@@ -2,15 +2,15 @@
 // RUN: mlir-translate -mlir-to-cpp -declare-variables-at-top %s | FileCheck %s -check-prefix=CPP-DECLTOP
 
 func.func @emitc_variable() {
-  %c0 = "emitc.variable"(){value = #emitc.opaque<"">} : () -> i32
-  %c1 = "emitc.variable"(){value = 42 : i32} : () -> i32
-  %c2 = "emitc.variable"(){value = -1 : i32} : () -> i32
-  %c3 = "emitc.variable"(){value = -1 : si8} : () -> si8
-  %c4 = "emitc.variable"(){value = 255 : ui8} : () -> ui8
-  %c5 = "emitc.variable"(){value = #emitc.opaque<"">} : () -> !emitc.ptr<i32>
-  %c6 = "emitc.variable"(){value = #emitc.opaque<"NULL">} : () -> !emitc.ptr<i32>
-  %c7 = "emitc.variable"(){value = #emitc.opaque<"">} : () -> !emitc.array<3x7xi32>
-  %c8 = "emitc.variable"(){value = #emitc.opaque<"">} : () -> !emitc.array<5x!emitc.ptr<i8>>
+  %c0 = "emitc.variable"(){value = #emitc.opaque<"">} : () -> !emitc.lvalue<i32>
+  %c1 = "emitc.variable"(){value = 42 : i32} : () -> !emitc.lvalue<i32>
+  %c2 = "emitc.variable"(){value = -1 : i32} : () -> !emitc.lvalue<i32>
+  %c3 = "emitc.variable"(){value = -1 : si8} : () -> !emitc.lvalue<si8>
+  %c4 = "emitc.variable"(){value = 255 : ui8} : () -> !emitc.lvalue<ui8>
+  %c5 = "emitc.variable"(){value = #emitc.opaque<"">} : () -> !emitc.lvalue<!emitc.ptr<i32>>
+  %c6 = "emitc.variable"(){value = #emitc.opaque<"NULL">} : () -> !emitc.lvalue<!emitc.ptr<i32>>
+  %c7 = "emitc.variable"(){value = #emitc.opaque<"">} : () -> !emitc.lvalue<!emitc.array<3x7xi32>>
----------------
aniragil wrote:

IIUC the proposed `lvalue<>` type models primarily modifiable lvalues (which array names are not part of), as we use it to mark locations that can be assigned to and loaded from. Perhaps we should have a `mlvalue<>` type and a `lvalue` type trait/interface? The former would be used for variables, array subscriptions etc, and the latter for `!emitc.array`, `!emitc.mlvalue` etc.

https://github.com/llvm/llvm-project/pull/91475