[llvm] [InstCombine] Factorise Add and Min/Max using Distributivity (PR #101717)

[Yingwei Zheng via llvm-commits]

================
@@ -0,0 +1,228 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -S -passes=instcombine < %s 2>&1 | FileCheck %s
+
+
+define i8 @umax_of_add_nuw(i8 %a, i8 %b, i8 %c) {
+; CHECK-LABEL: define i8 @umax_of_add_nuw(
+; CHECK-SAME: i8 [[A:%.*]], i8 [[B:%.*]], i8 [[C:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = call i8 @llvm.umax.i8(i8 [[B]], i8 [[C]])
+; CHECK-NEXT:    [[MAX:%.*]] = add nuw i8 [[TMP1]], [[A]]
+; CHECK-NEXT:    ret i8 [[MAX]]
+;
+  %add1 = add nuw i8 %b, %a
+  %add2 = add nuw i8 %c, %a
+  %max = call i8 @llvm.umax.i8(i8 %add1, i8 %add2)
+  ret i8 %max
+}
+
+define i8 @umax_of_add_nuw_comm(i8 %a, i8 %b, i8 %c) {
+; CHECK-LABEL: define i8 @umax_of_add_nuw_comm(
+; CHECK-SAME: i8 [[A:%.*]], i8 [[B:%.*]], i8 [[C:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = call i8 @llvm.umax.i8(i8 [[B]], i8 [[C]])
+; CHECK-NEXT:    [[MAX:%.*]] = add nuw i8 [[TMP1]], [[A]]
+; CHECK-NEXT:    ret i8 [[MAX]]
+;
+  %add1 = add nuw i8 %a, %b
----------------
dtcxzyw wrote:

You need a trick to avoid complexity-based canonicalization :)
https://llvm.org/docs/InstCombineContributorGuide.html#add-commuted-tests


https://github.com/llvm/llvm-project/pull/101717