[llvm] [Delinearization] Add function for fixed size array without relying on GEP (PR #145050)

[Ryotaro Kasuga via llvm-commits]

================
@@ -0,0 +1,446 @@
+; RUN: opt < %s -passes='print<delinearization>' -disable-output -delinearize-use-fixed-size-array-heuristic 2>&1 | FileCheck %s
+
+; void f(int A[][8][32]) {
+;   for (i = 0; i < 42; i++)
+;    for (j = 0; j < 8; j++)
+;     for (k = 0; k < 32; k++)
+;       A[i][j][k] = 1;
+; }
+
+; CHECK:      Delinearization on function a_i_j_k:
+; CHECK:      Base offset: %a
+; CHECK-NEXT: ArrayDecl[UnknownSize][8][32] with elements of 4 bytes.
+; CHECK-NEXT: ArrayRef[{0,+,1}<nuw><nsw><%for.i.header>][{0,+,1}<nuw><nsw><%for.j.header>][{0,+,1}<nuw><nsw><%for.k>]
+define void @a_i_j_k(ptr %a) {
+entry:
+  br label %for.i.header
+
+for.i.header:
+  %i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ]
+  br label %for.j.header
+
+for.j.header:
+  %j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j.latch ]
+  br label %for.k
+
+for.k:
+  %k = phi i32 [ 0, %for.j.header ], [ %k.inc, %for.k ]
+  %idx = getelementptr [8 x [32 x i32]], ptr %a, i32 %i, i32 %j, i32 %k
+  store i32 1, ptr %idx
+  %k.inc = add i32 %k, 1
+  %cmp.k = icmp slt i32 %k.inc, 32
+  br i1 %cmp.k, label %for.k, label %for.j.latch
+
+for.j.latch:
+  %j.inc = add i32 %j, 1
+  %cmp.j = icmp slt i32 %j.inc, 8
+  br i1 %cmp.j, label %for.j.header, label %for.i.latch
+
+for.i.latch:
+  %i.inc = add i32 %i, 1
+  %cmp.i = icmp slt i32 %i.inc, 42
+  br i1 %cmp.i, label %for.i.header, label %exit
+
+exit:
+  ret void
+}
+
+; void f(int A[][8][32]) {
+;   for (i = 0; i < 42; i++)
+;    for (j = 0; j < 8; j++)
+;     for (k = 0; k < 32; k++)
+;       A[i][7-j][k] = 1;
+; }
+
+; CHECK:      Delinearization on function a_i_nj_k:
+; CHECK:      Base offset: %a
+; CHECK-NEXT: ArrayDecl[UnknownSize][8][32] with elements of 4 bytes.
+; CHECK-NEXT: ArrayRef[{0,+,1}<nuw><nsw><%for.i.header>][{7,+,-1}<nsw><%for.j.header>][{0,+,1}<nuw><nsw><%for.k>]
+define void @a_i_nj_k(ptr %a) {
+entry:
+  br label %for.i.header
+
+for.i.header:
+  %i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ]
+  br label %for.j.header
+
+for.j.header:
+  %j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j.latch ]
+  %j.subscript = sub i32 7, %j
+  br label %for.k
+
+for.k:
+  %k = phi i32 [ 0, %for.j.header ], [ %k.inc, %for.k ]
+  %idx = getelementptr [8 x [32 x i32]], ptr %a, i32 %i, i32 %j.subscript, i32 %k
+  store i32 1, ptr %idx
+  %k.inc = add i32 %k, 1
+  %cmp.k = icmp slt i32 %k.inc, 32
+  br i1 %cmp.k, label %for.k, label %for.j.latch
+
+for.j.latch:
+  %j.inc = add i32 %j, 1
+  %cmp.j = icmp slt i32 %j.inc, 8
+  br i1 %cmp.j, label %for.j.header, label %for.i.latch
+
+for.i.latch:
+  %i.inc = add i32 %i, 1
+  %cmp.i = icmp slt i32 %i.inc, 42
+  br i1 %cmp.i, label %for.i.header, label %exit
+
+exit:
+  ret void
+}
+
+; In the following code, the access functions for both stores are represented
+; in the same way in SCEV, so the delinearization results are also the same. We
+; don't have any type information of the underlying objects.
+;
+; void f(int A[][4][64], int B[][8][32]) {
+;   for (i = 0; i < 42; i++)
+;    for (j = 0; j < 4; j++)
+;     for (k = 0; k < 32; k++) {
+;       A[i][j][k] = 1;
+;       B[i][2*j][k] = 1;
+;     }
+; }
+
+; CHECK:      Delinearization on function a_ijk_b_i2jk:
+; CHECK:      Base offset: %a
+; CHECK-NEXT: ArrayDecl[UnknownSize][4][64] with elements of 4 bytes.
+; CHECK-NEXT: ArrayRef[{0,+,1}<nuw><nsw><%for.i.header>][{0,+,1}<nuw><nsw><%for.j.header>][{0,+,1}<nuw><nsw><%for.k>]
+; CHECK:      Base offset: %b
+; CHECK-NEXT: ArrayDecl[UnknownSize][4][64] with elements of 4 bytes.
+; CHECK-NEXT: ArrayRef[{0,+,1}<nuw><nsw><%for.i.header>][{0,+,1}<nuw><nsw><%for.j.header>][{0,+,1}<nuw><nsw><%for.k>]
+define void @a_ijk_b_i2jk(ptr %a, ptr %b) {
+entry:
+  br label %for.i.header
+
+for.i.header:
+  %i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ]
+  br label %for.j.header
+
+for.j.header:
+  %j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j.latch ]
+  %j2 = shl i32 %j, 1
+  br label %for.k
+
+for.k:
+  %k = phi i32 [ 0, %for.j.header ], [ %k.inc, %for.k ]
+  %a.idx = getelementptr [4 x [64 x i32]], ptr %a, i32 %i, i32 %j, i32 %k
+  %b.idx = getelementptr [8 x [32 x i32]], ptr %b, i32 %i, i32 %j2, i32 %k
+  store i32 1, ptr %a.idx
+  store i32 1, ptr %b.idx
+  %k.inc = add i32 %k, 1
+  %cmp.k = icmp slt i32 %k.inc, 32
+  br i1 %cmp.k, label %for.k, label %for.j.latch
+
+for.j.latch:
+  %j.inc = add i32 %j, 1
+  %cmp.j = icmp slt i32 %j.inc, 4
+  br i1 %cmp.j, label %for.j.header, label %for.i.latch
+
+for.i.latch:
+  %i.inc = add i32 %i, 1
+  %cmp.i = icmp slt i32 %i.inc, 42
+  br i1 %cmp.i, label %for.i.header, label %exit
+
+exit:
+  ret void
+}
+
+; The type information of the underlying object is not available, so the
+; delinearization result is different from it.
----------------
kasuga-fj wrote:

Added

https://github.com/llvm/llvm-project/pull/145050