[llvm] [VPlan] Constant-fold ActiveLaneMask intrinsics (PR #208852)

[Ramkumar Ramachandra via llvm-commits]

================
@@ -0,0 +1,171 @@
+; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py UTC_ARGS: --filter-out-after "^scalar.ph" --version 6
+; RUN: opt -passes=loop-vectorize -tail-folding-policy=must-fold-tail -force-tail-folding-style=data -force-vector-width=4 -disable-output -vplan-print-after=printOptimizedVPlan %s 2>&1 | FileCheck %s --check-prefix=FIXED
+; RUN: opt -passes=loop-vectorize -tail-folding-policy=must-fold-tail -force-tail-folding-style=data -force-vector-width="vscale x 4" -disable-output -vplan-print-after=printOptimizedVPlan %s 2>&1 | FileCheck %s --check-prefix=SCALABLE
+
+; We never emit fixed-vector active.lane.masks foldable to false.
+define i16 @foldable_fixed_alm(i32 %x) {
+; FIXED-LABEL: VPlan for loop in 'foldable_fixed_alm'
+; FIXED:  VPlan 'Initial VPlan for VF={4},UF>=1' {
+; FIXED-NEXT:  Live-in vp<[[VP0:%[0-9]+]]> = VF * UF
+; FIXED-NEXT:  Live-in vp<[[VP1:%[0-9]+]]> = vector-trip-count
+; FIXED-NEXT:  Live-in ir<4> = original trip-count
+; FIXED-EMPTY:
+; FIXED-NEXT:  ir-bb<entry>:
+; FIXED-NEXT:  Successor(s): scalar.ph, vector.ph
+; FIXED-EMPTY:
+; FIXED-NEXT:  vector.ph:
+; FIXED-NEXT:    vp<[[VP2:%[0-9]+]]> = DERIVED-IV ir<0> + vp<[[VP1]]> * ir<%x>
+; FIXED-NEXT:  Successor(s): vector loop
+; FIXED-EMPTY:
+; FIXED-NEXT:  <x1> vector loop: {
+; FIXED-NEXT:  vp<[[VP3:%[0-9]+]]> = CANONICAL-IV
+; FIXED-EMPTY:
+; FIXED-NEXT:    vector.body:
+; FIXED-NEXT:      EMIT vp<%index.next> = add nuw vp<[[VP3]]>, vp<[[VP0]]>
+; FIXED-NEXT:      EMIT branch-on-count vp<%index.next>, vp<[[VP1]]>
+; FIXED-NEXT:    No successors
+; FIXED-NEXT:  }
+; FIXED-NEXT:  Successor(s): middle.block
+; FIXED-EMPTY:
+; FIXED-NEXT:  middle.block:
+; FIXED-NEXT:    EMIT-SCALAR vp<[[VP5:%[0-9]+]]> = trunc ir<%x> to i16
+; FIXED-NEXT:    EMIT vp<[[VP6:%[0-9]+]]> = sub nuw vp<[[VP1]]>, ir<1>
+; FIXED-NEXT:    vp<[[VP7:%[0-9]+]]> = DERIVED-IV ir<0> + vp<[[VP6]]> * vp<[[VP5]]>
+; FIXED-NEXT:    EMIT vp<%cmp.n> = icmp eq ir<4>, vp<[[VP1]]>
+; FIXED-NEXT:    EMIT branch-on-cond vp<%cmp.n>
+; FIXED-NEXT:  Successor(s): ir-bb<exit>, scalar.ph
+; FIXED-EMPTY:
+; FIXED-NEXT:  ir-bb<exit>:
+; FIXED-NEXT:    IR   %t.lcssa = phi i16 [ %t, %loop ] (extra operand: vp<[[VP7]]> from middle.block)
+; FIXED-NEXT:  No successors
+; FIXED-EMPTY:
+; FIXED-NEXT:  scalar.ph:
+;
+; SCALABLE-LABEL: VPlan for loop in 'foldable_fixed_alm'
+; SCALABLE:  VPlan 'Initial VPlan for VF={vscale x 4},UF>=1' {
+; SCALABLE-NEXT:  Live-in vp<[[VP0:%[0-9]+]]> = VF * UF
+; SCALABLE-NEXT:  Live-in vp<[[VP1:%[0-9]+]]> = vector-trip-count
+; SCALABLE-EMPTY:
+; SCALABLE-NEXT:  ir-bb<entry>:
+; SCALABLE-NEXT:  Successor(s): scalar.ph, vector.ph
+; SCALABLE-EMPTY:
+; SCALABLE-NEXT:  vector.ph:
+; SCALABLE-NEXT:  Successor(s): vector loop
+; SCALABLE-EMPTY:
+; SCALABLE-NEXT:  <x1> vector loop: {
+; SCALABLE-NEXT:  vp<[[VP2:%[0-9]+]]> = CANONICAL-IV
+; SCALABLE-EMPTY:
+; SCALABLE-NEXT:    vector.body:
+; SCALABLE-NEXT:      EMIT vp<%index.next> = add vp<[[VP2]]>, vp<[[VP0]]>
+; SCALABLE-NEXT:      EMIT branch-on-count vp<%index.next>, vp<[[VP1]]>
+; SCALABLE-NEXT:    No successors
+; SCALABLE-NEXT:  }
+; SCALABLE-NEXT:  Successor(s): middle.block
+; SCALABLE-EMPTY:
+; SCALABLE-NEXT:  middle.block:
+; SCALABLE-NEXT:    EMIT-SCALAR vp<[[VP5:%[0-9]+]]> = trunc ir<%x> to i16
+; SCALABLE-NEXT:    vp<[[VP6:%[0-9]+]]> = DERIVED-IV ir<0> + ir<3> * vp<[[VP5]]>
+; SCALABLE-NEXT:  Successor(s): ir-bb<exit>
+; SCALABLE-EMPTY:
+; SCALABLE-NEXT:  ir-bb<exit>:
+; SCALABLE-NEXT:    IR   %t.lcssa = phi i16 [ %t, %loop ] (extra operand: vp<[[VP6]]> from middle.block)
+; SCALABLE-NEXT:  No successors
+; SCALABLE-EMPTY:
+; SCALABLE-NEXT:  scalar.ph:
+;
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+  %red = phi i32 [ 0, %entry ], [ %add, %loop ]
+  %t = trunc i32 %red to i16
+  %add = add i32 %red, %x
+  %iv.next = add i64 %iv, 1
+  %ec = icmp eq i64 %iv, 3
+  br i1 %ec, label %exit, label %loop
+
+exit:
+  ret i16 %t
+}
+
+define i16 @tc_overflow(i32 %x) {
+; FIXED-LABEL: VPlan for loop in 'tc_overflow'
+; FIXED:  VPlan 'Initial VPlan for VF={4},UF>=1' {
+; FIXED-NEXT:  Live-in vp<[[VP0:%[0-9]+]]> = VF * UF
+; FIXED-NEXT:  Live-in vp<[[VP1:%[0-9]+]]> = vector-trip-count
+; FIXED-EMPTY:
+; FIXED-NEXT:  ir-bb<entry>:
+; FIXED-NEXT:  Successor(s): scalar.ph, vector.ph
+; FIXED-EMPTY:
+; FIXED-NEXT:  vector.ph:
+; FIXED-NEXT:  Successor(s): vector loop
+; FIXED-EMPTY:
+; FIXED-NEXT:  <x1> vector loop: {
+; FIXED-NEXT:  vp<[[VP2:%[0-9]+]]> = CANONICAL-IV
+; FIXED-EMPTY:
+; FIXED-NEXT:    vector.body:
+; FIXED-NEXT:      EMIT vp<%index.next> = add vp<[[VP2]]>, vp<[[VP0]]>
----------------
artagnon wrote:

Ah, it's just testing that we don't crash on very large TCs: will add a comment?

https://github.com/llvm/llvm-project/pull/208852