642bde76c3
The ordered reduction support introduced in 94e366ef2060 can cause an infinite loop when processing complex reduction chains. The worklist algorithm re-adds instructions from PossibleOrderedReductionOps when switching to ordered mode, but doesn't track which instructions have already been processed. This allows instructions to be re-added and processed multiple times, creating cycles. Add a Visited set to track processed instructions and skip any that have already been handled, preventing the infinite loop.
127 lines
5.1 KiB
LLVM
127 lines
5.1 KiB
LLVM
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
|
|
; RUN: opt -passes=slp-vectorizer -S < %s | FileCheck %s
|
|
|
|
; Test for infinite loop bug introduced in commit 94e366ef2060 which added
|
|
; support for ordered reductions. The worklist processing could re-add
|
|
; instructions without tracking which were already visited, causing an
|
|
; infinite loop on reduction chains with complex dependencies.
|
|
;
|
|
; This test ensures the pass completes in reasonable time and doesn't hang.
|
|
|
|
define void @pr_slp_ordered_reduction_hang(ptr %inp, ptr %out, i32 %itr) {
|
|
; CHECK-LABEL: @pr_slp_ordered_reduction_hang(
|
|
; CHECK-NEXT: entry:
|
|
; CHECK-NEXT: [[VAL1:%.*]] = load float, ptr [[INP:%.*]], align 4
|
|
; CHECK-NEXT: [[ARRAYIDX1:%.*]] = getelementptr inbounds float, ptr [[INP]], i64 1
|
|
; CHECK-NEXT: [[VAL2:%.*]] = load float, ptr [[ARRAYIDX1]], align 4
|
|
; CHECK-NEXT: [[CMP:%.*]] = icmp sgt i32 [[ITR:%.*]], 0
|
|
; CHECK-NEXT: br i1 [[CMP]], label [[LOOP:%.*]], label [[EXIT:%.*]]
|
|
; CHECK: loop:
|
|
; CHECK-NEXT: [[I:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[INC:%.*]], [[LOOP]] ]
|
|
; CHECK-NEXT: [[X1:%.*]] = phi float [ [[VAL1]], [[ENTRY]] ], [ [[ADD49_7:%.*]], [[LOOP]] ]
|
|
; CHECK-NEXT: [[Y1:%.*]] = phi float [ [[VAL2]], [[ENTRY]] ], [ [[ADD47_7:%.*]], [[LOOP]] ]
|
|
; CHECK: [[INC]] = add nuw nsw i32 [[I]], 1
|
|
; CHECK-NEXT: [[CMP_LOOP:%.*]] = icmp slt i32 [[INC]], [[ITR]]
|
|
; CHECK-NEXT: br i1 [[CMP_LOOP]], label [[LOOP]], label [[EXIT]]
|
|
; CHECK: exit:
|
|
; CHECK-NEXT: [[RESULT:%.*]] = phi float [ [[VAL1]], [[ENTRY]] ], [ [[ADD49_7]], [[LOOP]] ]
|
|
; CHECK-NEXT: store float [[RESULT]], ptr [[OUT:%.*]], align 4
|
|
; CHECK-NEXT: ret void
|
|
;
|
|
entry:
|
|
%val1 = load float, ptr %inp, align 4
|
|
%arrayidx1 = getelementptr inbounds float, ptr %inp, i64 1
|
|
%val2 = load float, ptr %arrayidx1, align 4
|
|
%cmp = icmp sgt i32 %itr, 0
|
|
br i1 %cmp, label %loop, label %exit
|
|
|
|
loop:
|
|
%i = phi i32 [ 0, %entry ], [ %inc, %loop ]
|
|
%x1 = phi float [ %val1, %entry ], [ %add49.7, %loop ]
|
|
%y1 = phi float [ %val2, %entry ], [ %add47.7, %loop ]
|
|
|
|
; Multiple iterations of fibonacci-like fadd chain
|
|
; Each value depends on the previous two, creating a complex
|
|
; dependency pattern that triggers the ordered reduction path
|
|
%add35 = fadd float %y1, %x1
|
|
%add37 = fadd float %x1, %add35
|
|
%add39 = fadd float %add35, %add37
|
|
%add41 = fadd float %add37, %add39
|
|
%add43 = fadd float %add39, %add41
|
|
%add45 = fadd float %add41, %add43
|
|
%add47 = fadd float %add43, %add45
|
|
%add49 = fadd float %add45, %add47
|
|
|
|
%add35.1 = fadd float %add47, %add49
|
|
%add37.1 = fadd float %add49, %add35.1
|
|
%add39.1 = fadd float %add35.1, %add37.1
|
|
%add41.1 = fadd float %add37.1, %add39.1
|
|
%add43.1 = fadd float %add39.1, %add41.1
|
|
%add45.1 = fadd float %add41.1, %add43.1
|
|
%add47.1 = fadd float %add43.1, %add45.1
|
|
%add49.1 = fadd float %add45.1, %add47.1
|
|
|
|
%add35.2 = fadd float %add47.1, %add49.1
|
|
%add37.2 = fadd float %add49.1, %add35.2
|
|
%add39.2 = fadd float %add35.2, %add37.2
|
|
%add41.2 = fadd float %add37.2, %add39.2
|
|
%add43.2 = fadd float %add39.2, %add41.2
|
|
%add45.2 = fadd float %add41.2, %add43.2
|
|
%add47.2 = fadd float %add43.2, %add45.2
|
|
%add49.2 = fadd float %add45.2, %add47.2
|
|
|
|
%add35.3 = fadd float %add47.2, %add49.2
|
|
%add37.3 = fadd float %add49.2, %add35.3
|
|
%add39.3 = fadd float %add35.3, %add37.3
|
|
%add41.3 = fadd float %add37.3, %add39.3
|
|
%add43.3 = fadd float %add39.3, %add41.3
|
|
%add45.3 = fadd float %add41.3, %add43.3
|
|
%add47.3 = fadd float %add43.3, %add45.3
|
|
%add49.3 = fadd float %add45.3, %add47.3
|
|
|
|
%add35.4 = fadd float %add47.3, %add49.3
|
|
%add37.4 = fadd float %add49.3, %add35.4
|
|
%add39.4 = fadd float %add35.4, %add37.4
|
|
%add41.4 = fadd float %add37.4, %add39.4
|
|
%add43.4 = fadd float %add39.4, %add41.4
|
|
%add45.4 = fadd float %add41.4, %add43.4
|
|
%add47.4 = fadd float %add43.4, %add45.4
|
|
%add49.4 = fadd float %add45.4, %add47.4
|
|
|
|
%add35.5 = fadd float %add47.4, %add49.4
|
|
%add37.5 = fadd float %add49.4, %add35.5
|
|
%add39.5 = fadd float %add35.5, %add37.5
|
|
%add41.5 = fadd float %add37.5, %add39.5
|
|
%add43.5 = fadd float %add39.5, %add41.5
|
|
%add45.5 = fadd float %add41.5, %add43.5
|
|
%add47.5 = fadd float %add43.5, %add45.5
|
|
%add49.5 = fadd float %add45.5, %add47.5
|
|
|
|
%add35.6 = fadd float %add47.5, %add49.5
|
|
%add37.6 = fadd float %add49.5, %add35.6
|
|
%add39.6 = fadd float %add35.6, %add37.6
|
|
%add41.6 = fadd float %add37.6, %add39.6
|
|
%add43.6 = fadd float %add39.6, %add41.6
|
|
%add45.6 = fadd float %add41.6, %add43.6
|
|
%add47.6 = fadd float %add43.6, %add45.6
|
|
%add49.6 = fadd float %add45.6, %add47.6
|
|
|
|
%add35.7 = fadd float %add47.6, %add49.6
|
|
%add37.7 = fadd float %add49.6, %add35.7
|
|
%add39.7 = fadd float %add35.7, %add37.7
|
|
%add41.7 = fadd float %add37.7, %add39.7
|
|
%add43.7 = fadd float %add39.7, %add41.7
|
|
%add45.7 = fadd float %add41.7, %add43.7
|
|
%add47.7 = fadd float %add43.7, %add45.7
|
|
%add49.7 = fadd float %add45.7, %add47.7
|
|
|
|
%inc = add nuw nsw i32 %i, 1
|
|
%cmp.loop = icmp slt i32 %inc, %itr
|
|
br i1 %cmp.loop, label %loop, label %exit
|
|
|
|
exit:
|
|
%result = phi float [ %val1, %entry ], [ %add49.7, %loop ]
|
|
store float %result, ptr %out, align 4
|
|
ret void
|
|
}
|