[llvm] [DAG] Generalize setcc(setcc) fold to use known bits. (PR #66503)

[David Green via llvm-commits]

================
@@ -403,14 +403,19 @@ define i32 @PR17487(i1 %tobool) {
 ;
 ; X64-LIN-LABEL: PR17487:
 ; X64-LIN:       # %bb.0:
-; X64-LIN-NEXT:    movl %edi, %eax
-; X64-LIN-NEXT:    andl $1, %eax
+; X64-LIN-NEXT:    movd %edi, %xmm0
+; X64-LIN-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
+; X64-LIN-NEXT:    pand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
+; X64-LIN-NEXT:    pextrw $4, %xmm0, %eax
 ; X64-LIN-NEXT:    retq
 ;
 ; X64-WIN-LABEL: PR17487:
 ; X64-WIN:       # %bb.0:
-; X64-WIN-NEXT:    andb $1, %cl
 ; X64-WIN-NEXT:    movzbl %cl, %eax
+; X64-WIN-NEXT:    movd %eax, %xmm0
+; X64-WIN-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
+; X64-WIN-NEXT:    pand __xmm at 00000000000000010000000000000001(%rip), %xmm0
+; X64-WIN-NEXT:    pextrw $4, %xmm0, %eax
----------------
davemgreen wrote:

Yeah these I wasn't sure about - how important they were. They get folded away under normal optimizations: https://godbolt.org/z/P4vMYd5oc, and it looked like the test was only added for correctness.

I think if the combines happened in topo order it might do better. The DAG looks like this:
```
          t56: v4i32 = BUILD_VECTOR undef:i32, undef:i32, t2, undef:i32
        t57: v2i64 = bitcast t56                                       
        t36: v2i64 = BUILD_VECTOR Constant:i64<1>, Constant:i64<1>     
      t39: v2i64 = and t57, t36                                        
    t50: v16i8 = bitcast t39                                           
  t51: i8 = extract_vector_elt t50, Constant:i64<8>                    
t45: i32 = zero_extend t51                                             
```

https://github.com/llvm/llvm-project/pull/66503