[llvm] [LoopFusion] Extending SIV to handle separate loops (PR #146383)

[via llvm-commits]

================
@@ -0,0 +1,185 @@
+; RUN: opt -S -passes=loop-fusion -da-disable-delinearization-checks < %s | FileCheck %s
+
+; The two inner loops have no dependency and are allowed to be fused as in the
+; outer loops, different levels are accessed to.
+
+; C Code
+;
+;;  for (long int i = 0; i < n; i++) {
+;;    for (long int j = 0; j < n; j++) {
+;;      for (long int k = 0; k < n; k++) {
+;;        A[i][j][k] = i;
+;;      }
+;;      for (long int k = 0; k < n; k++) {
+;;        temp = A[i + 3][j + 2][k + 1];
+
+define void @backward_dep0(i64 %n, ptr %A) nounwind uwtable ssp {
+entry:
+  %cmp10 = icmp sgt i64 %n, 0
+  br i1 %cmp10, label %for.cond1.preheader.preheader, label %for.end26
+
+; CHECK-LABEL: backward_dep
+; CHECK-COUNT-1: for.body{{[0-9]+}}:
+; CHECK-NOT:     for.body{{[0-9]+}}:
+
+for.cond1.preheader.preheader:                    ; preds = %entry
+  br label %for.cond1.preheader
+
+for.cond1.preheader:                              ; preds = %for.cond1.preheader.preheader, %for.inc24
+  %i.011 = phi i64 [ %inc25, %for.inc24 ], [ 0, %for.cond1.preheader.preheader ]
+  %cmp26 = icmp sgt i64 %n, 0
+  br i1 %cmp26, label %for.cond4.preheader.preheader, label %for.inc24
+
+for.cond4.preheader.preheader:                    ; preds = %for.cond1.preheader
+  br label %for.cond4.preheader
+
+for.cond4.preheader:                              ; preds = %for.cond4.preheader.preheader, %for.inc21
+  %j.07 = phi i64 [ %inc22, %for.inc21 ], [ 0, %for.cond4.preheader.preheader ]
+  %cmp51 = icmp sgt i64 %n, 0
+  br i1 %cmp51, label %for.body6.preheader, label %for.cond10.loopexit
+
+for.body6.preheader:                              ; preds = %for.cond4.preheader
+  br label %for.body6
+
+for.body6:                                        ; preds = %for.body6.preheader, %for.body6
+  %k.02 = phi i64 [ %inc, %for.body6 ], [ 0, %for.body6.preheader ]
+  %arrayidx8 = getelementptr inbounds [100 x [100 x i64]], ptr %A, i64 %i.011, i64 %j.07, i64 %k.02
+  store i64 %i.011, ptr %arrayidx8, align 8
+  %inc = add nsw i64 %k.02, 1
+  %exitcond13 = icmp ne i64 %inc, %n
+  br i1 %exitcond13, label %for.body6, label %for.cond10.loopexit.loopexit
+
+for.cond10.loopexit.loopexit:                     ; preds = %for.body6
+  br label %for.cond10.loopexit
+
+for.cond10.loopexit:                              ; preds = %for.cond10.loopexit.loopexit, %for.cond4.preheader
+  %cmp113 = icmp sgt i64 %n, 0
+  br i1 %cmp113, label %for.body12.preheader, label %for.inc21
+
+for.body12.preheader:                             ; preds = %for.cond10.loopexit
+  br label %for.body12
+
+for.body12:                                       ; preds = %for.body12.preheader, %for.body12
+  %k9.05 = phi i64 [ %inc19, %for.body12 ], [ 0, %for.body12.preheader ]
+  %add = add nsw i64 %k9.05, 1
+  %add13 = add nsw i64 %j.07, 2
+  %add14 = add nsw i64 %i.011, 3
+  %arrayidx17 = getelementptr inbounds [100 x [100 x i64]], ptr %A, i64 %add14, i64 %add13, i64 %add
+  %0 = load i64, ptr %arrayidx17, align 8
+  %inc19 = add nsw i64 %k9.05, 1
+  %exitcond = icmp ne i64 %inc19, %n
+  br i1 %exitcond, label %for.body12, label %for.inc21.loopexit
+
+for.inc21.loopexit:                               ; preds = %for.body12
+  br label %for.inc21
+
+for.inc21:                                        ; preds = %for.inc21.loopexit, %for.cond10.loopexit
+  %inc22 = add nsw i64 %j.07, 1
+  %exitcond14 = icmp ne i64 %inc22, %n
+  br i1 %exitcond14, label %for.cond4.preheader, label %for.inc24.loopexit
+
+for.inc24.loopexit:                               ; preds = %for.inc21
+  br label %for.inc24
+
+for.inc24:                                        ; preds = %for.inc24.loopexit, %for.cond1.preheader
+  %inc25 = add nsw i64 %i.011, 1
+  %exitcond15 = icmp ne i64 %inc25, %n
+  br i1 %exitcond15, label %for.cond1.preheader, label %for.end26.loopexit
+
+for.end26.loopexit:                               ; preds = %for.inc24
+  br label %for.end26
+
+for.end26:                                        ; preds = %for.end26.loopexit, %entry
+  ret void
+}
+
+; The two inner loops have a backward loop-carried dependency, allowing them
+; to be fused.
+
+; C Code
+;
+;;  for (long int i = 0; i < n; i++) {
+;;    for (long int j = 0; j < n; j++) {
+;;      for (long int k = 0; k < n; k++) {
+;;        A[i][j][k] = i;
+;;      }
+;;      for (long int k = 0; k < n; k++) {
+;;        temp = A[i][j][k - 1];
+
+define void @backward_dep1(i64 %n, ptr %A) nounwind uwtable ssp {
----------------
amehsan wrote:

Similarly incomplete C code

https://github.com/llvm/llvm-project/pull/146383