llvm dev - Dec 2015 - LICM doesn't work for IntrReadMem intrinsic function

Thanks for your reply, escha,

Yes, -loop-rorate makes it work.

Regards,

Xiangyang

2015-12-01 12:48 GMT-05:00 <escha at apple.com>:
>
> > On Dec 1, 2015, at 9:30 AM, Xiangyang Guo via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
> >
> > Hi, All,
> >
> > Suppose I define one memory read only intrinsic function
"foo" in
> Intrinsics.td like this
> >
> > def int_foo : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty],
> [IntrReadMem]>;
> >
> > Suppose I have the following IR, which means the intrinsic function
> "foo" is called 10 times in a loop. Since the parameters of
function "foo"
> are invariant and the function "foo" is defined as
"IntrReadMem". I assume
> the '-licm' can move the function call 'foo' out of the
loop body. However,
> after I use "opt test.ll -basicaa -licm -S", the function call
'foo' is
> still inside the loop body. Do I misunderstand something here? Any
> suggestions are appreciated. Thanks a lot.
>
> Does it work if you call -loop-rotate first? Loop unrolling requires loops
> to be in canonicalized rotated form first, so I wouldn’t be surprised if
> LICM does as well.
>
> —escha
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20151201/cb989800/attachment.html>

Hi,

if I have the following IR, LICM doesn't work again, even if I use
'-loop-rotate' firstly. In this IR, the return value of intrinsic
function
"foo" is used by another function "func2" as parameter. 
However, for the
intrinsic function "foo", the parameters are still loop invariant. Do
you
know why this happens? Thanks a lot.

Regards,

Xiangyang

; Function Attrs: uwtable
define void @_Z5func1ii(i32 %a, i32 %b) #0 {
  br label %1

; <label>:1                                       ; preds = %0, %1
  %i.01 = phi i32 [ 0, %0 ], [ %6, %1 ]
  %2 = call i32 @llvm.foo(i32 %a, i32 1)
  %3 = call i32 @_Z5func2ii(i32 %b, i32 %2)
  %4 = call i32 @llvm.foo(i32 %a, i32 2)
  %5 = call i32 @_Z5func2ii(i32 %3, i32 %4)
  %6 = add nsw i32 %i.01, 1
  %7 = icmp slt i32 %6, 10
  br i1 %7, label %1, label %8

; <label>:8                                       ; preds = %1
  ret void
}

; Function Attrs: nounwind readonly
declare i32 @llvm.foo(i32, i32) #1

declare i32 @_Z5func2ii(i32, i32) #2

2015-12-01 13:16 GMT-05:00 Xiangyang Guo <eceguo at gmail.com>:
> Thanks for your reply, escha,
>
> Yes, -loop-rorate makes it work.
>
> Regards,
>
> Xiangyang
>
> 2015-12-01 12:48 GMT-05:00 <escha at apple.com>:
>
>>
>> > On Dec 1, 2015, at 9:30 AM, Xiangyang Guo via llvm-dev <
>> llvm-dev at lists.llvm.org> wrote:
>> >
>> > Hi, All,
>> >
>> > Suppose I define one memory read only intrinsic function
"foo" in
>> Intrinsics.td like this
>> >
>> > def int_foo : Intrinsic<[llvm_i32_ty], [llvm_i32_ty,
llvm_i32_ty],
>> [IntrReadMem]>;
>> >
>> > Suppose I have the following IR, which means the intrinsic
function
>> "foo" is called 10 times in a loop. Since the parameters of
function "foo"
>> are invariant and the function "foo" is defined as
"IntrReadMem". I assume
>> the '-licm' can move the function call 'foo' out of the
loop body. However,
>> after I use "opt test.ll -basicaa -licm -S", the function
call 'foo' is
>> still inside the loop body. Do I misunderstand something here? Any
>> suggestions are appreciated. Thanks a lot.
>>
>> Does it work if you call -loop-rotate first? Loop unrolling requires
>> loops to be in canonicalized rotated form first, so I wouldn’t be
surprised
>> if LICM does as well.
>>
>> —escha
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20151201/0908be32/attachment.html>

On 1 December 2015 at 12:33, Xiangyang Guo via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
> if I have the following IR, LICM doesn't work again, even if I use
> '-loop-rotate' firstly. In this IR, the return value of intrinsic
function
> "foo" is used by another function "func2" as parameter.
However, for the
> intrinsic function "foo", the parameters are still loop
invariant. Do you
> know why this happens? Thanks a lot.
Is there anything telling LLVM that func2 won't write to memory and
change the value returned by llvm.foo?

Cheers.

Tim.