[NewGVN][InstCombine] missed load elimination

[nurmukhametov]

Consider the following code

%v8_uniform_FVector4 = type { [4 x double] }

define void @foo(ptr %ret_mem, <4 x double> %a, <4 x double> %b) {
  %ptr = alloca %v8_uniform_FVector4, align 8
  store <4 x double> %a, ptr %ptr, align 8
  %e0 = load double, ptr %ptr, align 8
  %of1 = getelementptr inbounds %v8_uniform_FVector4, ptr %ptr, i64 0, i32 0, i64 1
  %e1 = load double, ptr %of1, align 8
  %o2 = getelementptr inbounds %v8_uniform_FVector4, ptr %ptr, i64 0, i32 0, i64 2
  %e2 = load double, ptr %o2, align 8
  %o3 = getelementptr inbounds %v8_uniform_FVector4, ptr %ptr, i64 0, i32 0, i64 3
  %e3 = load double, ptr %o3, align 8
  %vec0 = insertelement <4 x double> poison, double %e1, i64 0
  %vec1 = insertelement <4 x double> %vec0, double %e2, i64 1
  %vec2 = insertelement <4 x double> %vec1, double %e0, i64 2
  %vec3 = insertelement <4 x double> %vec2, double %e3, i64 3
  store <4 x double> %vec3, ptr %ret_mem, align 8
  ret void
}

NewGVN and InstCombine miss to optimize loads, whereas GVN and InstCombine managed to do that as follows

define void @foo(ptr %ret_mem, <4 x double> %a, <4 x double> %b) {
  %vec3 = shufflevector <4 x double> %a, <4 x double> poison, <4 x i32> <i32 1, i32 2, i32 0, i32 3>
  store <4 x double> %vec3, ptr %ret_mem, align 8
  ret void
}

Compiler explorer link: https://godbolt.org/z/sbKv81a6s

[dtcxzyw]

https://godbolt.org/z/a4ef1c47j

SLP-vectorizer should convert insertelement + load chain into a vector load, then InstCombine will eliminate this store-load pair.