[DAG] Unroll and expand illegal result of LDEXP and POWI instead of widen.

[pengfei]

llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

4090–4095

How about change like this?

fallthrough to WidenVecRes_EXTEND_VECTOR_INREG seems different from the old code that calling WidenVecRes_ExpOp.

[liutianle]

If unroll fails, don't fall through but try widen and break.

[pengfei]

LGTM.

[pengfei]

Maybe add a test case for POWI too?

[liutianle]

Add powi.v2f32.i32 lit test.

llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

case ISD::FCOPYSIGN:
     break;
 
   case ISD::IS_FPCLASS:
     Res = WidenVecRes_IS_FPCLASS(N);
     break;
 
   case ISD::FLDEXP:
   case ISD::FPOWI:
-    Res = WidenVecRes_ExpOp(N);
+    if (unrollExpandedOp())
       break;
+    // If the target has custom/legal support for the scalar FP intrinsic ops
+    // (they are probably not destined to become libcalls), then widen those
+    // like any other unary ops.
+    [[fallthrough]];
 
   case ISD::ANY_EXTEND_VECTOR_INREG:
   case ISD::SIGN_EXTEND_VECTOR_INREG:
   case ISD::ZERO_EXTEND_VECTOR_INREG:
     Res = WidenVecRes_EXTEND_VECTOR_INREG(N);
     break;
 
   case ISD::ANY_EXTEND:

llvm/test/CodeGen/X86/ldexp.ll

; WIN32-NEXT:    retl
   %zext = zext i8 %x to i32
   %ldexp = call double @llvm.ldexp.f64.i32(double 1.000000e+00, i32 %zext)
   ret double %ldexp
 }
 
 define <2 x float> @ldexp_v2f32(<2 x float> %val, <2 x i32> %exp) {
 ; X64-LABEL: ldexp_v2f32:
 ; X64:       # %bb.0:
-; X64-NEXT:    subq $72, %rsp
-; X64-NEXT:    .cfi_def_cfa_offset 80
+; X64-NEXT:    subq $56, %rsp
+; X64-NEXT:    .cfi_def_cfa_offset 64
 ; X64-NEXT:    movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; X64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; X64-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
-; X64-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[3,3,3,3]
-; X64-NEXT:    movd %xmm2, %edi
-; X64-NEXT:    callq ldexpf@PLT
 ; X64-NEXT:    movaps %xmm0, (%rsp) # 16-byte Spill
-; X64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; X64-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1]
-; X64-NEXT:    pshufd $238, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
-; X64-NEXT:    # xmm1 = mem[2,3,2,3]
 ; X64-NEXT:    movd %xmm1, %edi
 ; X64-NEXT:    callq ldexpf@PLT
-; X64-NEXT:    unpcklps (%rsp), %xmm0 # 16-byte Folded Reload
-; X64-NEXT:    # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
-; X64-NEXT:    movaps %xmm0, (%rsp) # 16-byte Spill
-; X64-NEXT:    movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; X64-NEXT:    movd %xmm0, %edi
-; X64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
-; X64-NEXT:    callq ldexpf@PLT
 ; X64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
-; X64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
+; X64-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload
 ; X64-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
 ; X64-NEXT:    pshufd $85, {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
 ; X64-NEXT:    # xmm1 = mem[1,1,1,1]
 ; X64-NEXT:    movd %xmm1, %edi
 ; X64-NEXT:    callq ldexpf@PLT
 ; X64-NEXT:    movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
 ; X64-NEXT:    unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
-; X64-NEXT:    unpcklpd (%rsp), %xmm1 # 16-byte Folded Reload
-; X64-NEXT:    # xmm1 = xmm1[0],mem[0]
 ; X64-NEXT:    movaps %xmm1, %xmm0
-; X64-NEXT:    addq $72, %rsp
+; X64-NEXT:    addq $56, %rsp
 ; X64-NEXT:    .cfi_def_cfa_offset 8
 ; X64-NEXT:    retq
 ;
 ; WIN32-LABEL: ldexp_v2f32:
 ; WIN32:       # %bb.0:
 ; WIN32-NEXT:    pushl %edi
 ; WIN32-NEXT:    pushl %esi
 ; WIN32-NEXT:    subl $8, %esp