[llvm] r306980 - [InstCombine] look through bswap/bitreverse for equality comparisons

[Philip Reames via llvm-commits]

Nothing wrong with this change, but any chance we could generalize 
here?  Are there other intrinsics whose input/output preserve equality?  
I think any intrinsic which is a 1-to-1 total function preserves this 
property right?  What about sin/cost/tan, sqrt, etc..?

There's also a larger category where equality of inputs implies equality 
of outputs.  That is: f(a) == f(a) is always true for certain functions 
f.  Do we exploit that anywhere?  log, signbit, fabs, etc... all have 
that property.  (I suspect we get this one, but have not checked.)

Philip

On 07/02/2017 07:34 AM, Sanjay Patel via llvm-commits wrote:
> Author: spatel
> Date: Sun Jul  2 07:34:50 2017
> New Revision: 306980
>
> URL: http://llvm.org/viewvc/llvm-project?rev=306980&view=rev
> Log:
> [InstCombine] look through bswap/bitreverse for equality comparisons
>
> I noticed this missed bswap optimization in the CGP memcmp() expansion,
> and then I saw that we don't have the fold in InstCombine.
>
> Differential Revision: https://reviews.llvm.org/D34763
>
> Modified:
>      llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp
>      llvm/trunk/test/Transforms/InstCombine/icmp.ll
>
> Modified: llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp?rev=306980&r1=306979&r2=306980&view=diff
> ==============================================================================
> --- llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp (original)
> +++ llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp Sun Jul  2 07:34:50 2017
> @@ -3438,6 +3438,15 @@ Instruction *InstCombiner::foldICmpEqual
>       }
>     }
>   
> +  // If both operands are byte-swapped or bit-reversed, just compare the
> +  // original values.
> +  // TODO: Move this to a function similar to foldICmpIntrinsicWithConstant()
> +  // and handle more intrinsics.
> +  if ((match(Op0, m_BSwap(m_Value(A))) && match(Op1, m_BSwap(m_Value(B)))) ||
> +      (match(Op0, m_Intrinsic<Intrinsic::bitreverse>(m_Value(A))) &&
> +       match(Op1, m_Intrinsic<Intrinsic::bitreverse>(m_Value(B)))))
> +    return new ICmpInst(Pred, A, B);
> +
>     return nullptr;
>   }
>   
>
> Modified: llvm/trunk/test/Transforms/InstCombine/icmp.ll
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstCombine/icmp.ll?rev=306980&r1=306979&r2=306980&view=diff
> ==============================================================================
> --- llvm/trunk/test/Transforms/InstCombine/icmp.ll (original)
> +++ llvm/trunk/test/Transforms/InstCombine/icmp.ll Sun Jul  2 07:34:50 2017
> @@ -2979,9 +2979,7 @@ declare i32 @llvm.bswap.i32(i32)
>   
>   define i1 @bswap_ne(i32 %x, i32 %y) {
>   ; CHECK-LABEL: @bswap_ne(
> -; CHECK-NEXT:    [[SWAPX:%.*]] = call i32 @llvm.bswap.i32(i32 %x)
> -; CHECK-NEXT:    [[SWAPY:%.*]] = call i32 @llvm.bswap.i32(i32 %y)
> -; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[SWAPX]], [[SWAPY]]
> +; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 %x, %y
>   ; CHECK-NEXT:    ret i1 [[CMP]]
>   ;
>     %swapx = call i32 @llvm.bswap.i32(i32 %x)
> @@ -2994,9 +2992,7 @@ declare <8 x i16> @llvm.bswap.v8i16(<8 x
>   
>   define <8 x i1> @bswap_vec_eq(<8 x i16> %x, <8 x i16> %y) {
>   ; CHECK-LABEL: @bswap_vec_eq(
> -; CHECK-NEXT:    [[SWAPX:%.*]] = call <8 x i16> @llvm.bswap.v8i16(<8 x i16> %x)
> -; CHECK-NEXT:    [[SWAPY:%.*]] = call <8 x i16> @llvm.bswap.v8i16(<8 x i16> %y)
> -; CHECK-NEXT:    [[CMP:%.*]] = icmp eq <8 x i16> [[SWAPX]], [[SWAPY]]
> +; CHECK-NEXT:    [[CMP:%.*]] = icmp eq <8 x i16> %x, %y
>   ; CHECK-NEXT:    ret <8 x i1> [[CMP]]
>   ;
>     %swapx = call <8 x i16> @llvm.bswap.v8i16(<8 x i16> %x)
> @@ -3009,9 +3005,7 @@ declare i64 @llvm.bitreverse.i64(i64)
>   
>   define i1 @bitreverse_eq(i64 %x, i64 %y) {
>   ; CHECK-LABEL: @bitreverse_eq(
> -; CHECK-NEXT:    [[REVX:%.*]] = call i64 @llvm.bitreverse.i64(i64 %x)
> -; CHECK-NEXT:    [[REVY:%.*]] = call i64 @llvm.bitreverse.i64(i64 %y)
> -; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i64 [[REVX]], [[REVY]]
> +; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i64 %x, %y
>   ; CHECK-NEXT:    ret i1 [[CMP]]
>   ;
>     %revx = call i64 @llvm.bitreverse.i64(i64 %x)
> @@ -3024,9 +3018,7 @@ declare <8 x i16> @llvm.bitreverse.v8i16
>   
>   define <8 x i1> @bitreverse_vec_ne(<8 x i16> %x, <8 x i16> %y) {
>   ; CHECK-LABEL: @bitreverse_vec_ne(
> -; CHECK-NEXT:    [[REVX:%.*]] = call <8 x i16> @llvm.bitreverse.v8i16(<8 x i16> %x)
> -; CHECK-NEXT:    [[REVY:%.*]] = call <8 x i16> @llvm.bitreverse.v8i16(<8 x i16> %y)
> -; CHECK-NEXT:    [[CMP:%.*]] = icmp ne <8 x i16> [[REVX]], [[REVY]]
> +; CHECK-NEXT:    [[CMP:%.*]] = icmp ne <8 x i16> %x, %y
>   ; CHECK-NEXT:    ret <8 x i1> [[CMP]]
>   ;
>     %revx = call <8 x i16> @llvm.bitreverse.v8i16(<8 x i16> %x)
>
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits