[llvm-bugs] [Bug 38946] New: [IndVarSimplify / LoopVectorizer] IV simplification prohibits vectorization

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=38946

            Bug ID: 38946
           Summary: [IndVarSimplify / LoopVectorizer]  IV simplification
                    prohibits vectorization
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Scalar Optimizations
          Assignee: unassignedbugs at nondot.org
          Reporter: paulsson at linux.vnet.ibm.com
                CC: llvm-bugs at lists.llvm.org

Created attachment 20881
  --> https://bugs.llvm.org/attachment.cgi?id=20881&action=edit
reduced testcase

Similarly to the test case in bug 38935, IndVarSimplify rewrites the IV from
i32 to i64 on SystemZ, and the result is that the LoopVectorizer can no longer
recognize this as a vectorizable loop:

opt tc_indvar_phi.ll -mtriple=i686-unknown-linux-gnu -mcpu=icelake-server -S -o
/dev/null -O3 -debug-only=loop-vectorize 
LV: We can vectorize this loop!

opt tc_indvar_phi.ll -mtriple=systemz-unknown -mcpu=z14 -S -o /dev/null -O3
-debug-only=loop-vectorize
LV: Found an unidentified PHI.  %.125 = phi i32 [ %.01630, %.lr.ph.preheader ],
[ %.2, %14 ]
LV: Not vectorizing: Cannot prove legality.


.lr.ph.preheader:                                 ; preds = %.preheader
  br label %.lr.ph

.lr.ph:                                           ; preds = %.lr.ph.preheader,
%13
  %.125 = phi i32 [ %.2, %13 ], [ %.01630, %.lr.ph.preheader ]
  %.11824 = phi i32 [ %.3, %13 ], [ %.01729, %.lr.ph.preheader ]
  %.12123 = phi i32 [ %14, %13 ], [ %.02028, %.lr.ph.preheader ]
  %6 = sext i32 %.12123 to i64
  %7 = getelementptr inbounds %0, %0* %0, i64 %6, i32 3
  %8 = load i32, i32* %7, align 4, !tbaa !5
  %9 = icmp eq i32 %8, 0
  br i1 %9, label %10, label %13

; <label>:10:                                     ; preds = %.lr.ph
  %11 = icmp slt i32 %.031, %.11824
  %spec.select = select i1 %11, i32 %.031, i32 %.11824
  %12 = icmp slt i32 %.12123, %.125
  %spec.select22 = select i1 %12, i32 %.12123, i32 %.125
  br label %13

; <label>:13:                                     ; preds = %10, %.lr.ph
  %.3 = phi i32 [ %.11824, %.lr.ph ], [ %spec.select, %10 ]
  %.2 = phi i32 [ %.125, %.lr.ph ], [ %spec.select22, %10 ]
  %14 = add nsw i32 %.12123, 1
  %15 = icmp slt i32 %.12123, -1
  br i1 %15, label %.lr.ph, label %._crit_edge.loopexit
....

*** IR Dump After Induction Variable Simplification ***

.lr.ph.preheader:                                 ; preds = %.preheader
  %6 = sext i32 %.02028 to i64
  br label %.lr.ph

; Loop:
.lr.ph:                                           ; preds = %.lr.ph.preheader,
%15
  %indvars.iv = phi i64 [ %6, %.lr.ph.preheader ], [ %indvars.iv.next, %15 ]
  %.125 = phi i32 [ %.2, %15 ], [ %.01630, %.lr.ph.preheader ]
  %.11824 = phi i32 [ %.3, %15 ], [ %.01729, %.lr.ph.preheader ]
  %7 = getelementptr inbounds %0, %0* %0, i64 %indvars.iv, i32 3
  %8 = load i32, i32* %7, align 4, !tbaa !5
  %9 = icmp eq i32 %8, 0
  br i1 %9, label %10, label %15

; <label>:10:                                     ; preds = %.lr.ph
  %11 = icmp slt i32 %.031, %.11824
  %spec.select = select i1 %11, i32 %.031, i32 %.11824
  %12 = sext i32 %.125 to i64
  %13 = icmp slt i64 %indvars.iv, %12
  %14 = trunc i64 %indvars.iv to i32
  %spec.select22 = select i1 %13, i32 %14, i32 %.125
  br label %15

; <label>:15:                                     ; preds = %10, %.lr.ph
  %.3 = phi i32 [ %.11824, %.lr.ph ], [ %spec.select, %10 ]
  %.2 = phi i32 [ %.125, %.lr.ph ], [ %spec.select22, %10 ]
  %indvars.iv.next = add nsw i64 %indvars.iv, 1
  %16 = icmp slt i64 %indvars.iv, -1
  br i1 %16, label %.lr.ph, label %._crit_edge.loopexit


It looks like some analysis can't handle the sext:ed phi.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20180914/99f8413b/attachment.html>