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

[Guozhi Wei via llvm-commits]

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