[PATCH] D14596: [SROA] Choose more profitable type in findCommonType

[Xinliang David Li via llvm-commits]

you can probably create a smaller test case by simplifying post-inline IR dump.

David

On Tue, Dec 15, 2015 at 5:38 PM, Guozhi Wei <carrot at google.com> wrote:
> Carrot added a comment.
>
> Following is the IR of function foo in David's test case, after the last pass that still has the correct type for %ldd,
>
> define void @_Z3fooi(i32 signext %n) #0 {
> entry:
>
>   %ldd = alloca %"struct.std::complex", align 4
>   %ref.tmp = alloca %"struct.std::complex", align 4
>   %0 = bitcast %"struct.std::complex"* %ldd to i8*
>   call void @llvm.lifetime.start(i64 8, i8* %0) #3
>   call void @_ZNSt7complexIfEC2Eff(%"struct.std::complex"* %ldd, float 0.000000e+00, float 0.000000e+00)
>   br label %for.cond
>
> for.cond:                                         ; preds = %for.body, %entry
>
>   %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
>   %cmp = icmp slt i32 %i.0, %n
>   br i1 %cmp, label %for.body, label %for.end
>
> for.body:                                         ; preds = %for.cond
>
>   %call = call fastcc [2 x float] @_ZL7computeRSt7complexIfES1_()
>   %coerce.dive = getelementptr inbounds %"struct.std::complex", %"struct.std::complex"* %ref.tmp, i32 0, i32 0
>   %1 = bitcast { float, float }* %coerce.dive to [2 x float]*
>   %call.fca.0.gep = getelementptr inbounds [2 x float], [2 x float]* %1, i32 0, i32 0
>   %call.fca.0.extract = extractvalue [2 x float] %call, 0
>   store float %call.fca.0.extract, float* %call.fca.0.gep
>   %call.fca.1.gep = getelementptr inbounds [2 x float], [2 x float]* %1, i32 0, i32 1
>   %call.fca.1.extract = extractvalue [2 x float] %call, 1
>   store float %call.fca.1.extract, float* %call.fca.1.gep
>   %call1 = call dereferenceable(8) %"struct.std::complex"* @_ZNSt7complexIfEpLIfEERS0_RKS_IT_E(%"struct.std::complex"* %ldd, %"struct.std::complex"* dereferenceable(8) %ref.tmp)
>   %inc = add nsw i32 %i.0, 1
>   br label %for.cond
>
> for.end:                                          ; preds = %for.cond
>
>   call void @llvm.memcpy.p0i8.p0i8.i64(i8* bitcast (%"struct.std::complex"* @dd to i8*), i8* %0, i64 8, i32 4, i1 false)
>   call void @llvm.lifetime.end(i64 8, i8* %0) #3
>   ret void
>
> }
>
> In this function there is no explicit floating point operation, all of them are wrapped by function call. So combine pass actually works as intended. The memcpy is lowered to integer ld/st. But later in inlining pass all the complex number function calls are inlined, so in SROA pass it can see both integer ld/st and fp operations.
>
>
> http://reviews.llvm.org/D14596
>
>
>