[llvm] [X86, SimplifyCFG] Support hoisting load/store with conditional faulting (Part I) (PR #96878)

[Antonio Frighetto via llvm-commits]

================
@@ -3214,6 +3238,112 @@ bool SimplifyCFGOpt::speculativelyExecuteBB(BranchInst *BI,
   BB->splice(BI->getIterator(), ThenBB, ThenBB->begin(),
              std::prev(ThenBB->end()));
 
+  // If the target supports conditional faulting,
+  // we are looking for code like the following:
+  // \code
+  //   BB:
+  //     ...
+  //     %cond = icmp ult %x, %y
+  //     br i1 %cond, label %TrueBB, label %FalseBB
+  //   FalseBB:
+  //     store i32 1, ptr %q, align 4
+  //     ...
+  //   TrueBB:
+  //     %0 = load i32, ptr %b, align 4
+  //     store i32 %0, ptr %p, align 4
+  //     ...
+  // \endcode
+  //
+  // and transform it into:
+  //
+  // \code
+  //   BB:
+  //     ...
+  //     %cond = icmp ult %x, %y
+  //     %0 = cload i32, ptr %b, %cond
+  //     cstore i32 %0, ptr %p, %cond
+  //     cstore i32 1, ptr %q, ~%cond
+  //     br i1 %cond, label %TrueBB, label %FalseBB
+  //   FalseBB:
+  //     ...
+  //   TrueBB:
+  //     ...
+  // \endcode
+  //
+  // where cload/cstore is represented by intrinsic like llvm.masked.load/store,
----------------
antoniofrighetto wrote:

```suggestion
  // where cload/cstore are represented by llvm.masked.load/store intrinsics,
```

https://github.com/llvm/llvm-project/pull/96878