[llvm] [LV][VPlan] Prevent calculate cost for skiped instructions in precomputeCosts(). (PR #127966)
Florian Hahn via llvm-commits
llvm-commits at lists.llvm.org
Wed Feb 19 23:50:49 PST 2025
================
@@ -0,0 +1,86 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -passes=loop-vectorize -mtriple=x86_64-unknown-linux-gnu -S | FileCheck %s
+
+; Check if the vplan-based cost model select same VF to the legacy cost model.
+; Reduced from: https://github.com/llvm/llvm-project/issues/115744#issuecomment-2670479463
+
+; int *e;
+; int f;
+; void g() {
+; for (; e[f]; f++)
+; e[b(f + 9)];
+; }
+
+define void @g(i64 %n) {
+; CHECK-LABEL: @g(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[TMP0:%.*]] = trunc i64 [[N:%.*]] to i32
+; CHECK-NEXT: [[TMP1:%.*]] = add i32 [[TMP0]], 1
+; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[TMP1]], 8
+; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_SCEVCHECK:%.*]]
+; CHECK: vector.scevcheck:
+; CHECK-NEXT: [[TMP2:%.*]] = icmp ugt i64 [[N]], 4294967295
+; CHECK-NEXT: br i1 [[TMP2]], label [[SCALAR_PH]], label [[ENTRY:%.*]]
+; CHECK: vector.ph:
+; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[TMP1]], 8
+; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP1]], [[N_MOD_VF]]
+; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i64> poison, i64 [[N]], i64 0
+; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i64> [[BROADCAST_SPLATINSERT]], <4 x i64> poison, <4 x i32> zeroinitializer
+; CHECK-NEXT: br label [[LOOP:%.*]]
+; CHECK: vector.body:
+; CHECK-NEXT: [[SELECT:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[SELECT_NEXT:%.*]], [[LOOP]] ]
+; CHECK-NEXT: [[VEC_IND:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[ENTRY]] ], [ [[VEC_IND_NEXT:%.*]], [[LOOP]] ]
+; CHECK-NEXT: [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[ENTRY]] ], [ [[TMP9:%.*]], [[LOOP]] ]
+; CHECK-NEXT: [[VEC_PHI1:%.*]] = phi <4 x i32> [ zeroinitializer, [[ENTRY]] ], [ [[TMP10:%.*]], [[LOOP]] ]
+; CHECK-NEXT: [[STEP_ADD:%.*]] = add <4 x i32> [[VEC_IND]], splat (i32 4)
+; CHECK-NEXT: [[TMP3:%.*]] = zext <4 x i32> [[VEC_IND]] to <4 x i64>
+; CHECK-NEXT: [[TMP4:%.*]] = zext <4 x i32> [[STEP_ADD]] to <4 x i64>
+; CHECK-NEXT: [[TMP5:%.*]] = icmp eq <4 x i64> [[BROADCAST_SPLAT]], [[TMP3]]
+; CHECK-NEXT: [[TMP6:%.*]] = icmp eq <4 x i64> [[BROADCAST_SPLAT]], [[TMP4]]
+; CHECK-NEXT: [[TMP7:%.*]] = select <4 x i1> [[TMP5]], <4 x i32> zeroinitializer, <4 x i32> zeroinitializer
+; CHECK-NEXT: [[TMP8:%.*]] = select <4 x i1> [[TMP6]], <4 x i32> zeroinitializer, <4 x i32> zeroinitializer
+; CHECK-NEXT: [[TMP9]] = or <4 x i32> [[TMP7]], [[VEC_PHI]]
+; CHECK-NEXT: [[TMP10]] = or <4 x i32> [[TMP8]], [[VEC_PHI1]]
+; CHECK-NEXT: [[SELECT_NEXT]] = add nuw i32 [[SELECT]], 8
+; CHECK-NEXT: [[VEC_IND_NEXT]] = add <4 x i32> [[STEP_ADD]], splat (i32 4)
+; CHECK-NEXT: [[TMP11:%.*]] = icmp eq i32 [[SELECT_NEXT]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[TMP11]], label [[MIDDLE_BLOCK:%.*]], label [[LOOP]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK: middle.block:
+; CHECK-NEXT: [[BIN_RDX:%.*]] = or <4 x i32> [[TMP10]], [[TMP9]]
+; CHECK-NEXT: [[TMP12:%.*]] = call i32 @llvm.vector.reduce.or.v4i32(<4 x i32> [[BIN_RDX]])
+; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[TMP1]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
+; CHECK: scalar.ph:
+; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY1:%.*]] ], [ 0, [[VECTOR_SCEVCHECK]] ]
+; CHECK-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP12]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY1]] ], [ 0, [[VECTOR_SCEVCHECK]] ]
+; CHECK-NEXT: br label [[LOOP1:%.*]]
+; CHECK: loop:
+; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP1]] ]
+; CHECK-NEXT: [[SELECT1:%.*]] = phi i32 [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[SELECT_NEXT1:%.*]], [[LOOP1]] ]
+; CHECK-NEXT: [[IV_WIDEN:%.*]] = zext i32 [[IV]] to i64
+; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[N]], [[IV_WIDEN]]
+; CHECK-NEXT: [[SELECT_I:%.*]] = select i1 [[EXITCOND]], i32 0, i32 0
+; CHECK-NEXT: [[SELECT_NEXT1]] = or i32 [[SELECT_I]], [[SELECT1]]
+; CHECK-NEXT: [[IV_NEXT]] = add i32 [[IV]], 1
+; CHECK-NEXT: br i1 [[EXITCOND]], label [[EXIT]], label [[LOOP1]], !llvm.loop [[LOOP3:![0-9]+]]
+; CHECK: exit:
+; CHECK-NEXT: [[SPEC_SELECT_I_LCSSA:%.*]] = phi i32 [ [[SELECT_NEXT1]], [[LOOP1]] ], [ [[TMP12]], [[MIDDLE_BLOCK]] ]
+; CHECK-NEXT: ret void
+;
+entry:
+ br label %loop
+
+loop:
+ %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ]
+ %select = phi i32 [ 0, %entry ], [ %select.next, %loop ]
+ %iv.widen = zext i32 %iv to i64
+ %exitcond = icmp eq i64 %n, %iv.widen
+ %select.i = select i1 %exitcond, i32 0, i32 0
+ %select.next = or i32 %select.i, %select
----------------
fhahn wrote:
It would be great if you could update the test to not be a nop; we still want it to work once we add more folds on VPlan. Maybe try `llvm-reduce` with with running `instcombine` before `loop-vectorize?
https://github.com/llvm/llvm-project/pull/127966
More information about the llvm-commits
mailing list