[PATCH] D20789: Consecutive memory access in Loop Vectorizer - fixed and simplified
Elena Demikhovsky via llvm-commits
llvm-commits at lists.llvm.org
Tue Aug 2 10:34:46 PDT 2016
delena added a comment.
In https://reviews.llvm.org/D20789#503518, @anemet wrote:
> In https://reviews.llvm.org/D20789#503264, @delena wrote:
>
> > This is the output of LoopAccessAnalysis:
> >
> > W:\ver2>build\Debug\bin\opt.exe -O2 test.ll -S -o test_opt.ll --debug-only=loop-accesses
> > LAA: Found a loop in _Z23IncrementalCopyFastPathPKcPci: while.body
> > LAA: Processing memory accesses...
> >
> > AST: Alias Set Tracker: 1 alias sets for 2 pointer values.
> > AliasSet[0x24f30a0, 2] may alias, No access Pointers: (i64* %2, 18446744073709551615), (i64* %0, 18446744073709551615)
> >
> >
> > LAA: Accesses(2):
> >
> > %2 = bitcast i8* %op.addr.06 to i64* (write)
> > %0 = bitcast i8* %src.addr.05 to i64* (read-only)
> >
> > Underlying objects for pointer %2 = bitcast i8* %op.addr.06 to i64*
> >
> > i8* %op
> >
> > Underlying objects for pointer %0 = bitcast i8* %src.addr.05 to i64*
> >
> > i8* %src
> >
> > LAA: Found a runtime check ptr: %2 = bitcast i8* %op.addr.06 to i64*
> > LAA: Found a runtime check ptr: %0 = bitcast i8* %src.addr.05 to i64*
> > LAA: We need to do 1 pointer comparisons.
>
>
> LAA says we need a run-time alias check. Didn't we generate one?
>
> > LAA: We can perform a memory runtime check if needed.
>
> > LAA: Checking memory dependencies
>
> > Total Dependences: 0
>
No. It says " Total Dependences: 0" because it checks only pointers with the same leader, like A[i] and A[i+5].
IMO, this is a bug that should be fixed. In a separate patch, of course.
> > LAA: No unsafe dependent memory operations in loop. We need runtime memory checks.
>
> > LAA: Adding RT check for range:
>
> > Start: %op End: ((8 * (zext i32 ((8 + (-9 smax (-1 + (-1 * %len))) + %len) /u 8) to i64)) + %op)
>
> > LAA: Adding RT check for range:
>
> > Start: %src End: ((8 * (zext i32 ((8 + (-9 smax (-1 + (-1 * %len))) + %len) /u 8) to i64)) + %src)
>
> > LAA: Found a loop in _Z23IncrementalCopyFastPathPKcPci: while.body
>
> > LAA: Processing memory accesses...
>
> >
>
> > AST: Alias Set Tracker: 1 alias sets for 2 pointer values.
>
> > AliasSet[0x24f3eb0, 2] may alias, No access Pointers: (i64* %231, 18446744073709551615), (i64* %229, 18446744073709551615)
>
> >
>
> >
>
> > LAA: Accesses(2):
>
> >
>
> > %231 = bitcast i8* %op.addr.06 to i64* (write)
>
> > %229 = bitcast i8* %src.addr.05 to i64* (read-only)
>
> >
>
> > Underlying objects for pointer %231 = bitcast i8* %op.addr.06 to i64*
>
> >
>
> > i8* %op
>
> >
>
> > Underlying objects for pointer %229 = bitcast i8* %src.addr.05 to i64*
>
> >
>
> > i8* %src
>
> >
>
> > LAA: Found a runtime check ptr: %231 = bitcast i8* %op.addr.06 to i64*
>
> > LAA: Found a runtime check ptr: %229 = bitcast i8* %src.addr.05 to i64*
>
> > LAA: We need to do 1 pointer comparisons.
>
> > LAA: We can perform a memory runtime check if needed.
>
> > LAA: Checking memory dependencies
>
> > Total Dependences: 0
>
> > LAA: No unsafe dependent memory operations in loop. We need runtime memory checks.
>
> > LAA: Found a loop in _Z23IncrementalCopyFastPathPKcPci: vector.body
>
> > LAA: SCEV could not compute the loop exit count.
>
Repository:
rL LLVM
https://reviews.llvm.org/D20789
More information about the llvm-commits
mailing list