[clang] [CIR] Upstream pointer subtraction handling (PR #163306)

[Andy Kaylor via cfe-commits]

================
@@ -0,0 +1,35 @@
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fclangir -emit-cir %s -o %t.cir
+// RUN: FileCheck --check-prefix=CIR --input-file=%t.cir %s
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fclangir -emit-llvm %s -o %t.ll
+// RUN: FileCheck --check-prefix=LLVM --input-file=%t.ll %s
+
+int addrcmp(const void* a, const void* b) {
+  // CIR-LABEL: addrcmp
+  // CIR: %[[R:.*]] = cir.ptr_diff
+  // CIR: cir.cast integral %[[R]] : !s64i -> !s32i
+
+  // LLVM-LABEL: define dso_local i32 @addrcmp(
+  // LLVM: %[[PTR_A:.*]] = ptrtoint ptr {{.*}} to i64
+  // LLVM: %[[PTR_B:.*]] = ptrtoint ptr {{.*}} to i64
+  // LLVM: %[[SUB:.*]] = sub i64 %[[PTR_A]], %[[PTR_B]]
+  // LLVM-NOT: sdiv
+  // LLVM: trunc i64 %[[SUB]] to i32
+  return *(const void**)a - *(const void**)b;
+}
+
+unsigned long long test_ptr_diff(int *a, int* b) {
+  // CIR-LABEL: test_ptr_diff
+  // CIR: %[[D:.*]] = cir.ptr_diff {{.*}} : !cir.ptr<!s32i> -> !s64i
+  // CIR: %[[U:.*]] = cir.cast integral %[[D]] : !s64i -> !u64i
+  // CIR: cir.return {{.*}} : !u64i
+
+  // LLVM-LABEL: define dso_local i64 @test_ptr_diff(
+  // LLVM: %[[IA:.*]] = ptrtoint ptr %{{.*}} to i64
+  // LLVM: %[[IB:.*]] = ptrtoint ptr %{{.*}} to i64
+  // LLVM: %[[SUB:.*]] = sub i64 %[[IA]], %[[IB]]
+  // LLVM: %[[Q:.*]] = sdiv{{( exact)?}} i64 %[[SUB]], 4
----------------
andykaylor wrote:

This check will pass either with or without the `exact` keyword being present. I don't think we want that in this case. In cases where keywords may be present that genuinely are not relevant to the test (such as `align` on loads and stores in tests that aren't testing load and store alignment), we generally use the more general form, `{{.*}}` to tolerate any keywords, but in this case I think the presence or absence of `exact` is relevant to the test.

Classic codegen (OGCG) is producing `exact` here, and so the CIR path should also, but we currently don't. I think you can fix this by calling `setExact` on the SDiv/UDiv operation in the lowering code. You'll probably need to introduce a new temporary variable to hold the result where you do this though.

https://github.com/llvm/llvm-project/pull/163306