[llvm] [DAGCombine] `visitANY_EXTEND` - fold `(aext (freeze (load x)))` -> `(aext (truncate (freeze (extload x))))` (PR #218454)

[Simon Pilgrim via llvm-commits]

================
@@ -2144,58 +2144,58 @@ define <64 x i8> @test_compress_v64i8(<64 x i8> %vec, <64 x i1> %mask, <64 x i8>
 ; AVX2-NEXT:    vpinsrb $3, %ecx, %xmm6, %xmm6
 ; AVX2-NEXT:    vpinsrb $4, %r8d, %xmm6, %xmm6
 ; AVX2-NEXT:    vpinsrb $5, %r9d, %xmm6, %xmm6
-; AVX2-NEXT:    movzbl 16(%rbp), %r10d
-; AVX2-NEXT:    vpinsrb $6, %r10d, %xmm6, %xmm6
-; AVX2-NEXT:    movzbl 24(%rbp), %r11d
-; AVX2-NEXT:    vpinsrb $7, %r11d, %xmm6, %xmm6
-; AVX2-NEXT:    movzbl 32(%rbp), %ebx
-; AVX2-NEXT:    vpinsrb $8, %ebx, %xmm6, %xmm6
-; AVX2-NEXT:    movzbl 40(%rbp), %r14d
-; AVX2-NEXT:    vpinsrb $9, %r14d, %xmm6, %xmm6
-; AVX2-NEXT:    movzbl 48(%rbp), %r15d
+; AVX2-NEXT:    movl 16(%rbp), %edi
+; AVX2-NEXT:    vpinsrb $6, %edi, %xmm6, %xmm6
----------------
RKSimon wrote:

Looks like we're getting closer to finally folding the load :) 

https://github.com/llvm/llvm-project/pull/218454