[llvm-bugs] [Bug 32778] New: [Polly] [ScopDetect/Info] Incorrect Invariant Load Hoisting

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=32778

            Bug ID: 32778
           Summary: [Polly] [ScopDetect/Info] Incorrect Invariant Load
                    Hoisting
           Product: Polly
           Version: unspecified
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Optimizer
          Assignee: polly-dev at googlegroups.com
          Reporter: huihuiz at codeaurora.org
                CC: llvm-bugs at lists.llvm.org

Created attachment 18346
  --> https://bugs.llvm.org/attachment.cgi?id=18346&action=edit
generetated output file

For Polly test case "test/Isl/CodeGen/reduction_2.ll"

The patch "[ScopDetect/Info] Allow unconditional hoisting of loads from
dereferenceable ptrs"

git-svn-id: https://llvm.org/svn/llvm-project/polly/trunk@297375

Actually generate incorrect output code. 

Run command: opt  -polly-process-unprofitable  -polly-remarks-minimal  -basicaa
-polly-ast -analyze  -polly-invariant-load-hoisting=true <
./test/Isl/CodeGen/reduction_2.ll -polly-scops

Shows that RED[0] is invariant access
Function: main
    Region: %for.cond---%if.end
    Max Loop Depth:  1
    Invariant Accesses: {
            ReadAccess :=       [Reduction Type: NONE] [Scalar: 0]
                [tmp34] -> { Stmt_for_end[] -> MemRef_RED[0] };
            Execution Context: [tmp34] -> {  :  }
    }
...

Run command: opt  -polly-process-unprofitable  -polly-remarks-minimal  -basicaa
 -polly-invariant-load-hoisting=true < ./test/Isl/CodeGen/reduction_2.ll
-polly-codegen -S -o test.ll
to get the generated ll output.

For BB polly.stmt.for.body, "%p_add30 = add nsw i32 %polly.access.RED.load,
%tmp26_p_scalar_" 
is using the old value %polly.access.RED.load, which only get preloaded once,
and never reloaded.
While the "store i32 %p_add30, i32* %polly.access.cast.RED, !alias.scope !0,
!noalias !2",
stores to %polly.access.cast.RED


polly.stmt.for.body:                              ; preds = %polly.loop_header
  %scevgep1 = getelementptr i32, i32* %scevgep, i64 %polly.indvar
  %tmp10_p_scalar_ = load i32, i32* %scevgep1, !alias.scope !3, !noalias !5
  %scevgep2 = getelementptr [1021 x i32], [1021 x i32]* %A, i64 0, i64
%polly.indvar
  %tmp16_p_scalar_ = load i32, i32* %scevgep2, !alias.scope !3, !noalias !5
  %p_add = add nsw i32 %tmp10_p_scalar_, %tmp16_p_scalar_
  %scevgep4 = getelementptr i32, i32* %scevgep3, i64 %polly.indvar
  store i32 %p_add, i32* %scevgep4, !alias.scope !3, !noalias !5
  %tmp26_p_scalar_ = load i32, i32* %scevgep2, !alias.scope !3, !noalias !5
  %p_add30 = add nsw i32 %polly.access.RED.load, %tmp26_p_scalar_
  store i32 %p_add30, i32* %polly.access.cast.RED, !alias.scope !0, !noalias !2
  %polly.indvar_next = add nsw i64 %polly.indvar, 1
  %polly.loop_cond = icmp sle i64 %polly.indvar, 1017
  br i1 %polly.loop_cond, label %polly.loop_header, label %polly.loop_exit


polly.preload.begin:                              ; preds =
%polly.split_new_and_old
  %0 = bitcast [1 x i32]* undef to i32*
  %polly.access.RED = getelementptr i32, i32* %polly.access.cast.RED, i64 0
  %polly.access.RED.load = load i32, i32* %polly.access.RED, !alias.scope !0,
!noalias !2
  store i32 %polly.access.RED.load, i32* %tmp34.preload.s2a
  br i1 true, label %polly.start, label %for.cond.pre_entry_bb



I attached the output test.ll for easier verification.

The correct generated ll should look something like:
polly.stmt.for.body:
...
  %tmp1 = load i32, i32* %RED
  %p_add = add nsw i32 %tmp2, %tmp1
  store i32 %p_add, i32* %RED
...

-- 
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/20170424/1b93e7b6/attachment.html>