[llvm] Simplify `(a % b) lt/ge (b-1)` into `(a % b) eq/ne (b-1)` (PR #72504)

[Yingwei Zheng via llvm-commits]

================
@@ -10,6 +10,315 @@ declare void @use_i8(i8)
 declare void @use_i32(i32)
 declare void @use_i64(i64)
 
+; tests for (x % c) >=/ < (c - 1), where c >= 0
+define i1 @srem_sge_test1(i64 %x) {
+; CHECK-LABEL: @srem_sge_test1(
+; CHECK-NEXT:    [[Y:%.*]] = srem i64 [[X:%.*]], 34360750831
+; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i64 [[Y]], 34360750830
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+  %Cminus1 = add i64 34360750831, -1
+  %y = srem i64 %x, 34360750831
+  %cmp = icmp sge i64 %y, %Cminus1
----------------
dtcxzyw wrote:

> what is the reason for the check failure?

It is a flang issue. You can push something to rerun the CI.

> I have a question, but it doesn't relate to this patch, If I am developing a new backend instruction that does stack allocation it lowers alloca to the target architecture.
do you have a link for similar patches for any architecture?

For static allocation, `CodeGen` will do the most of things for you. Then you should provide a target-specific implementation of `TargetFrameLowering` to maintain the stack.

For dynamic allocation, SDAG will transform it into `ISD::DYNAMIC_STACKALLOC`. Then you should lower it into target-specific insts.

Please file an issue if you want to ask more about the backend.


https://github.com/llvm/llvm-project/pull/72504