Fix memory leak on ARM

Summary: We have a memory leak on ARM. When we have an inlined function we're using kBitTest but it sets different flags than x64 does. So we need an architecture specific check to see which flag to check.

Reviewed By: mpage

Differential Revision: D103768786

fbshipit-source-id: 702a5525b6780f4b5ae92ae4c2ed15d82a909016

Author: DinoV

cinderx/Jit/lir/generator.cpp

@@ -4036,14 +4036,24 @@ LIRGenerator::TranslatedBlock LIRGenerator::TranslateOneBasicBlock(
             Ind{callee_frame,
                 (Py_ssize_t)offsetof(FrameHeader, func) -
                     (Py_ssize_t)sizeof(FrameHeader)});
+
         JIT_DCHECK(
             JIT_FRAME_INITIALIZED == 2,
             "JIT_FRAME_INITIALIZED changed"); // this is the bit we're testing
                                               // below
         bbb.appendInstr(Instruction::kBitTest, rtfs_reg, Imm{1});
         auto done_block = bbb.allocateBlock();
         auto not_materialized_block = bbb.allocateBlock();
-        bbb.appendBranch(Instruction::kBranchNC, not_materialized_block);
+        // kBitTest lowers differently per architecture:
+        // - x86: BT sets the Carry flag to the tested bit value, so
+        //   kBranchNC (jnc) branches when the bit is NOT set.
+        // - ARM64: BT is lowered to TST which sets the Zero flag, so
+        //   kBranchE (b.eq) branches when the bit is NOT set.
+        bbb.appendBranch(
+            codegen::arch::kBuildArch == codegen::arch::Arch::kAarch64
+                ? Instruction::kBranchE
+                : Instruction::kBranchNC,
+            not_materialized_block);
         bbb.appendBlock(bbb.allocateBlock());
 
         // The frame was materialized, let's use the unlink helper to clean