Messy code for insertelement on AVX2 vectors

[efriedma-quic]

Bugzilla Link	50971
Version	trunk
OS	Windows NT
CC	@topperc,@LebedevRI,@RKSimon,@phoebewang,@rotateright

Extended Description

Examples:

#include <immintrin.h>
#define IDX 3
__m256d float1(__m256d a, __m256d b) { a[IDX] = b[0]; return a; }
__m256d float2(__m256d a, double b) { a[IDX] = b; return a; }
__m256d float3(__m256d a, double b) { return _mm256_blend_epi32(a, _mm256_set1_pd(b), 3 << IDX * 2); }
__m256i int1(__m256i a, __m256i b) { a[IDX] = b[0]; return a; }
__m256i int2(__m256i a, long b) { a[IDX] = b; return a; }
__m256i int3(__m256i a, long b) { return _mm256_blend_epi32(a, _mm256_set1_epi64x(b), 3 << IDX * 2); }
__m256i int4(__m256i a, long *b) { a[IDX] = *b; return a; }
__m256i int5(__m256i a, long *b) { return _mm256_blend_epi32(a, _mm256_set1_epi64x(*b), 3 << IDX * 2); }

It looks like the lowering extracts an 128-bit vector, inserts into that, then reconstructs the 256-bit vector. A broadcast+blend is almost always going to be more efficient, I think.

[LebedevRI]

Candidate patch: https://reviews.llvm.org/D105390

[RKSimon]

Current Codegen: https://godbolt.org/z/6aKGbdbTe

[LebedevRI]

Current Codegen: https://godbolt.org/z/6aKGbdbTe

godbolt hasn't updated yet, but we only still have the problem
with the first snippet there, float1().

[LebedevRI]

Something along the lines of https://reviews.llvm.org/D108253 should deal with that last motivational pattern.