Merge pull request #957 from vext01/first-block-bug

First block bug

tests/c/simple.newcg.c

@@ -1,7 +1,6 @@
 // # Currently this test breaks CI entirely, so we temporarily ignore it
 // # completely.
-// ignore-if: true
-// # ignore-if: test $YK_JIT_COMPILER != "yk"
+// ignore-if: test $YK_JIT_COMPILER != "yk"
 // Run-time:
 //   env-var: YKD_PRINT_IR=aot
 //   env-var: YKD_SERIALISE_COMPILATION=1

ykrt/src/compile/jitc_yk/trace_builder.rs

@@ -171,7 +171,22 @@ impl<'a> TraceBuilder<'a> {
             Some(b) => Ok(b),
             None => Err(CompilationError::Unrecoverable("empty trace".into())),
         }?;
-        let firstblk = self.lookup_aot_block(first_blk);
+
+        // Find the block containing the control point call. This is the (sole) predecessor of the
+        // first (guaranteed mappable) block in the trace.
+        let prev = match first_blk {
+            TracedAOTBlock::Mapped { func_name, bb } => {
+                debug_assert!(*bb > 0);
+                // It's `- 1` due to the way the ykllvm block splitting pass works.
+                TracedAOTBlock::Mapped {
+                    func_name: func_name.clone(),
+                    bb: bb - 1,
+                }
+            }
+            TracedAOTBlock::Unmappable => panic!(),
+        };
+
+        let firstblk = self.lookup_aot_block(&prev);
         // FIXME: This unwrap assumes the first block is mappable, but Laurie just merged a change
         // that strips the initial block (the block we return to from the control point), so I
         // don't think this assumption necessarily holds any more. Investigate.