[llvm] [RISCV] Improve loop by extract reduction instruction (PR #179215)

[via llvm-commits]

================
@@ -0,0 +1,68 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; loop-reduce-motion-test.ll
+; RUN: opt -passes=loop-reduce-motion -S < %s | FileCheck %s
+
+define i32 @pixel_asd8(ptr %pix1, i64 %stride1, ptr %pix2, i64 %stride2, i32 %height) {
+; CHECK-LABEL: @pixel_asd8(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[CMP21:%.*]] = icmp sgt i32 [[HEIGHT:%.*]], 0
+; CHECK-NEXT:    br i1 [[CMP21]], label [[FOR_COND1_PREHEADER_PREHEADER:%.*]], label [[FOR_COND_CLEANUP:%.*]]
+; CHECK:       for.cond1.preheader.preheader:
+; CHECK-NEXT:    br label [[FOR_COND1_PREHEADER:%.*]]
+; CHECK:       for.cond1.preheader:
+; CHECK-NEXT:    [[Y_025:%.*]] = phi i32 [ [[INC9:%.*]], [[FOR_COND1_PREHEADER]] ], [ 0, [[FOR_COND1_PREHEADER_PREHEADER]] ]
+; CHECK-NEXT:    [[PIX1_ADDR_023:%.*]] = phi ptr [ [[ADD_PTR:%.*]], [[FOR_COND1_PREHEADER]] ], [ [[PIX1:%.*]], [[FOR_COND1_PREHEADER_PREHEADER]] ]
+; CHECK-NEXT:    [[PIX2_ADDR_022:%.*]] = phi ptr [ [[ADD_PTR10:%.*]], [[FOR_COND1_PREHEADER]] ], [ [[PIX2:%.*]], [[FOR_COND1_PREHEADER_PREHEADER]] ]
+; CHECK-NEXT:    [[VEC_SUM_PHI:%.*]] = phi <8 x i32> [ zeroinitializer, [[FOR_COND1_PREHEADER_PREHEADER]] ], [ [[VEC_SUM_NEXT:%.*]], [[FOR_COND1_PREHEADER]] ]
+; CHECK-NEXT:    [[TMP0:%.*]] = load <8 x i8>, ptr [[PIX1_ADDR_023]], align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = load <8 x i8>, ptr [[PIX2_ADDR_022]], align 8
+; CHECK-NEXT:    [[TMP2:%.*]] = zext <8 x i8> [[TMP0]] to <8 x i32>
+; CHECK-NEXT:    [[TMP3:%.*]] = zext <8 x i8> [[TMP1]] to <8 x i32>
+; CHECK-NEXT:    [[TMP4:%.*]] = sub nsw <8 x i32> [[TMP2]], [[TMP3]]
+; CHECK-NEXT:    [[VEC_SUM_NEXT]] = add <8 x i32> [[VEC_SUM_PHI]], [[TMP4]]
+; CHECK-NEXT:    [[INC9]] = add nuw nsw i32 [[Y_025]], 1
+; CHECK-NEXT:    [[ADD_PTR]] = getelementptr inbounds i8, ptr [[PIX1_ADDR_023]], i64 [[STRIDE1:%.*]]
+; CHECK-NEXT:    [[ADD_PTR10]] = getelementptr inbounds i8, ptr [[PIX2_ADDR_022]], i64 [[STRIDE2:%.*]]
+; CHECK-NEXT:    [[EXITCOND_NOT:%.*]] = icmp eq i32 [[INC9]], [[HEIGHT]]
+; CHECK-NEXT:    br i1 [[EXITCOND_NOT]], label [[LOOP_EXIT_LANDING:%.*]], label [[FOR_COND1_PREHEADER]]
+; CHECK:       loop.exit.landing:
+; CHECK-NEXT:    [[SCALAR_TOTAL_SUM:%.*]] = call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> [[VEC_SUM_NEXT]])
+; CHECK-NEXT:    [[TMP5:%.*]] = add i32 0, [[SCALAR_TOTAL_SUM]]
+; CHECK-NEXT:    br label [[FOR_COND_CLEANUP_LOOPEXIT:%.*]]
----------------
Anjian-Wen wrote:

It seems depend on whether the only exit block has phi nodes which use the reduce_sum we try to sink. If the exit block has, I think it is hard to insert the reduce_sum in front of it. maybe we need to make additional changes ?
if we create a new block we can ignore this changes

https://github.com/llvm/llvm-project/pull/179215