[llvm] [LegalizeDAG][X86][AArch64][LoongArch] Freeze index when converting extract_elt/extract_subvector to load/store on stack. (PR #88985)

[Simon Pilgrim via llvm-commits]

================
@@ -490,69 +519,92 @@ define <4 x float> @var_shuffle_v4f32(<4 x float> %v, <4 x i32> %indices) nounwi
 define <16 x i8> @var_shuffle_v16i8_from_v16i8_v32i8(<16 x i8> %v, <32 x i8> %indices) nounwind {
 ; SSE3-LABEL: var_shuffle_v16i8_from_v16i8_v32i8:
 ; SSE3:       # %bb.0:
+; SSE3-NEXT:    pushq %rbp
+; SSE3-NEXT:    pushq %r15
+; SSE3-NEXT:    pushq %r14
+; SSE3-NEXT:    pushq %r13
+; SSE3-NEXT:    pushq %r12
+; SSE3-NEXT:    pushq %rbx
 ; SSE3-NEXT:    movaps %xmm1, -{{[0-9]+}}(%rsp)
-; SSE3-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
 ; SSE3-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; SSE3-NEXT:    movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
+; SSE3-NEXT:    movzbl -{{[0-9]+}}(%rsp), %ecx
+; SSE3-NEXT:    movzbl -{{[0-9]+}}(%rsp), %edx
+; SSE3-NEXT:    movzbl -{{[0-9]+}}(%rsp), %esi
+; SSE3-NEXT:    movzbl -{{[0-9]+}}(%rsp), %edi
+; SSE3-NEXT:    movzbl -{{[0-9]+}}(%rsp), %r8d
+; SSE3-NEXT:    movzbl -{{[0-9]+}}(%rsp), %r9d
+; SSE3-NEXT:    movzbl -{{[0-9]+}}(%rsp), %r10d
+; SSE3-NEXT:    movzbl -{{[0-9]+}}(%rsp), %r11d
+; SSE3-NEXT:    movzbl -{{[0-9]+}}(%rsp), %ebx
+; SSE3-NEXT:    movzbl -{{[0-9]+}}(%rsp), %ebp
+; SSE3-NEXT:    movzbl -{{[0-9]+}}(%rsp), %r14d
+; SSE3-NEXT:    movzbl -{{[0-9]+}}(%rsp), %r15d
+; SSE3-NEXT:    movzbl -{{[0-9]+}}(%rsp), %r12d
+; SSE3-NEXT:    movzbl -{{[0-9]+}}(%rsp), %r13d
+; SSE3-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; SSE3-NEXT:    movaps %xmm0, -{{[0-9]+}}(%rsp)
+; SSE3-NEXT:    movzbl %al, %eax
 ; SSE3-NEXT:    andl $15, %eax
 ; SSE3-NEXT:    movzbl -24(%rsp,%rax), %eax
 ; SSE3-NEXT:    movd %eax, %xmm1
-; SSE3-NEXT:    movzbl -{{[0-9]+}}(%rsp), %eax
+; SSE3-NEXT:    movzbl %r13b, %eax
----------------
RKSimon wrote:

The (ext (freeze (load))) regressions are starting to appear more and more now that DAG is embracing freeze, it'd be useful to get this sorted even if we have to always fold aext to zext

https://github.com/llvm/llvm-project/pull/88985