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

[pengfei]

llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

4092–4093

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:
+    if (!unrollExpandedOp())
       Res = WidenVecRes_ExpOp(N);
     break;
 
   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:
   case ISD::FP_EXTEND:

llvm/test/CodeGen/PowerPC/ldexp.ll

; CHECK-NEXT:    blr
   %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) {
 ; CHECK-LABEL: ldexp_v2f32:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    mflr r0
-; CHECK-NEXT:    stdu r1, -96(r1)
-; CHECK-NEXT:    std r0, 112(r1)
-; CHECK-NEXT:    .cfi_def_cfa_offset 96
+; CHECK-NEXT:    stdu r1, -80(r1)
+; CHECK-NEXT:    std r0, 96(r1)
+; CHECK-NEXT:    .cfi_def_cfa_offset 80
 ; CHECK-NEXT:    .cfi_offset lr, 16
-; CHECK-NEXT:    .cfi_offset v28, -64
 ; CHECK-NEXT:    .cfi_offset v29, -48
 ; CHECK-NEXT:    .cfi_offset v30, -32
 ; CHECK-NEXT:    .cfi_offset v31, -16
-; CHECK-NEXT:    li r3, 12
-; CHECK-NEXT:    xscvspdpn f1, v2
-; CHECK-NEXT:    stxv v28, 32(r1) # 16-byte Folded Spill
-; CHECK-NEXT:    stxv v29, 48(r1) # 16-byte Folded Spill
-; CHECK-NEXT:    stxv v30, 64(r1) # 16-byte Folded Spill
-; CHECK-NEXT:    stxv v31, 80(r1) # 16-byte Folded Spill
+; CHECK-NEXT:    xxsldwi vs0, v2, v2, 3
+; CHECK-NEXT:    li r3, 0
+; CHECK-NEXT:    stxv v29, 32(r1) # 16-byte Folded Spill
+; CHECK-NEXT:    xscvspdpn f1, vs0
+; CHECK-NEXT:    vextuwrx r4, r3, v3
+; CHECK-NEXT:    stxv v30, 48(r1) # 16-byte Folded Spill
+; CHECK-NEXT:    stxv v31, 64(r1) # 16-byte Folded Spill
 ; CHECK-NEXT:    vmr v31, v3
 ; CHECK-NEXT:    vmr v30, v2
-; CHECK-NEXT:    vextuwrx r4, r3, v3
 ; CHECK-NEXT:    bl ldexpf
 ; CHECK-NEXT:    nop
 ; CHECK-NEXT:    xxswapd vs0, v30
 ; CHECK-NEXT:    li r3, 4
-; CHECK-NEXT:    xscpsgndp v29, f1, f1
+; CHECK-NEXT:    xscvdpspn v29, f1
 ; CHECK-NEXT:    xscvspdpn f1, vs0
 ; CHECK-NEXT:    vextuwrx r4, r3, v31
 ; CHECK-NEXT:    bl ldexpf
 ; CHECK-NEXT:    nop
-; CHECK-NEXT:    # kill: def $f1 killed $f1 def $vsl1
-; CHECK-NEXT:    xxmrghd vs0, v29, vs1
-; CHECK-NEXT:    li r3, 0
-; CHECK-NEXT:    vextuwrx r4, r3, v31
-; CHECK-NEXT:    xvcvdpsp v28, vs0
-; CHECK-NEXT:    xxsldwi vs0, v30, v30, 3
-; CHECK-NEXT:    xscvspdpn f1, vs0
-; CHECK-NEXT:    bl ldexpf
-; CHECK-NEXT:    nop
-; CHECK-NEXT:    xxsldwi vs0, v30, v30, 1
-; CHECK-NEXT:    xscpsgndp v29, f1, f1
-; CHECK-NEXT:    mfvsrwz r4, v31
-; CHECK-NEXT:    xscvspdpn f1, vs0
-; CHECK-NEXT:    bl ldexpf
-; CHECK-NEXT:    nop
-; CHECK-NEXT:    # kill: def $f1 killed $f1 def $vsl1
-; CHECK-NEXT:    xxmrghd vs0, vs1, v29
-; CHECK-NEXT:    lxv v31, 80(r1) # 16-byte Folded Reload
-; CHECK-NEXT:    lxv v30, 64(r1) # 16-byte Folded Reload
-; CHECK-NEXT:    lxv v29, 48(r1) # 16-byte Folded Reload
-; CHECK-NEXT:    xvcvdpsp v2, vs0
-; CHECK-NEXT:    vmrgew v2, v28, v2
-; CHECK-NEXT:    lxv v28, 32(r1) # 16-byte Folded Reload
-; CHECK-NEXT:    addi r1, r1, 96
+; CHECK-NEXT:    xscvdpspn vs0, f1
+; CHECK-NEXT:    lxv v31, 64(r1) # 16-byte Folded Reload
+; CHECK-NEXT:    lxv v30, 48(r1) # 16-byte Folded Reload
+; CHECK-NEXT:    xxmrghw v2, vs0, v29
+; CHECK-NEXT:    lxv v29, 32(r1) # 16-byte Folded Reload
+; CHECK-NEXT:    addi r1, r1, 80
 ; CHECK-NEXT:    ld r0, 16(r1)
 ; CHECK-NEXT:    mtlr r0
 ; CHECK-NEXT:    blr
   %1 = call <2 x float> @llvm.ldexp.v2f32.v2i32(<2 x float> %val, <2 x i32> %exp)
   ret <2 x float> %1
 }
 
 define <4 x float> @ldexp_v4f32(<4 x float> %val, <4 x i32> %exp) {

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

llvm/test/CodeGen/X86/powi.ll

 ; X64-NEXT:    .cfi_adjust_cfa_offset 8
 ; X64-NEXT:    popq %rdi
 ; X64-NEXT:    .cfi_adjust_cfa_offset -8
 ; X64-NEXT:    jmp __powidf2@PLT # TAILCALL
   %ret = tail call double @llvm.powi.f64.i32(double %a, i32 15) nounwind ; <double> [#uses=1]
   ret double %ret
 }
 
+define <2 x float> @powi_v2f32(<2 x float> %a) minsize {
+; X86-X87-LABEL: powi_v2f32:
+; X86-X87:       # %bb.0:
+; X86-X87-NEXT:    pushl %esi
+; X86-X87-NEXT:    .cfi_def_cfa_offset 8
+; X86-X87-NEXT:    subl $16, %esp
+; X86-X87-NEXT:    .cfi_def_cfa_offset 24
+; X86-X87-NEXT:    .cfi_offset %esi, -8
+; X86-X87-NEXT:    flds {{[0-9]+}}(%esp)
+; X86-X87-NEXT:    fstps {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill
+; X86-X87-NEXT:    flds {{[0-9]+}}(%esp)
+; X86-X87-NEXT:    pushl $15
+; X86-X87-NEXT:    .cfi_adjust_cfa_offset 4
+; X86-X87-NEXT:    popl %esi
+; X86-X87-NEXT:    .cfi_adjust_cfa_offset -4
+; X86-X87-NEXT:    movl %esi, {{[0-9]+}}(%esp)
+; X86-X87-NEXT:    fstps (%esp)
+; X86-X87-NEXT:    calll __powisf2
+; X86-X87-NEXT:    fstps {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Spill
+; X86-X87-NEXT:    movl %esi, {{[0-9]+}}(%esp)
+; X86-X87-NEXT:    flds {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Reload
+; X86-X87-NEXT:    fstps (%esp)
+; X86-X87-NEXT:    calll __powisf2
+; X86-X87-NEXT:    flds {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Reload
+; X86-X87-NEXT:    fxch %st(1)
+; X86-X87-NEXT:    addl $16, %esp
+; X86-X87-NEXT:    .cfi_def_cfa_offset 8
+; X86-X87-NEXT:    popl %esi
+; X86-X87-NEXT:    .cfi_def_cfa_offset 4
+; X86-X87-NEXT:    retl
+;
+; X86-SSE-LABEL: powi_v2f32:
+; X86-SSE:       # %bb.0:
+; X86-SSE-NEXT:    pushl %esi
+; X86-SSE-NEXT:    .cfi_def_cfa_offset 8
+; X86-SSE-NEXT:    subl $32, %esp
+; X86-SSE-NEXT:    .cfi_def_cfa_offset 40
+; X86-SSE-NEXT:    .cfi_offset %esi, -8
+; X86-SSE-NEXT:    movups %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
+; X86-SSE-NEXT:    pushl $15
+; X86-SSE-NEXT:    .cfi_adjust_cfa_offset 4
+; X86-SSE-NEXT:    popl %esi
+; X86-SSE-NEXT:    .cfi_adjust_cfa_offset -4
+; X86-SSE-NEXT:    movl %esi, {{[0-9]+}}(%esp)
+; X86-SSE-NEXT:    movss %xmm0, (%esp)
+; X86-SSE-NEXT:    calll __powisf2
+; X86-SSE-NEXT:    movl %esi, {{[0-9]+}}(%esp)
+; X86-SSE-NEXT:    movups {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 # 16-byte Reload
+; X86-SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; X86-SSE-NEXT:    movss %xmm0, (%esp)
+; X86-SSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; X86-SSE-NEXT:    calll __powisf2
+; X86-SSE-NEXT:    fstps {{[0-9]+}}(%esp)
+; X86-SSE-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; X86-SSE-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; X86-SSE-NEXT:    unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
+; X86-SSE-NEXT:    addl $32, %esp
+; X86-SSE-NEXT:    .cfi_def_cfa_offset 8
+; X86-SSE-NEXT:    popl %esi
+; X86-SSE-NEXT:    .cfi_def_cfa_offset 4
+; X86-SSE-NEXT:    retl
+;
+; X64-LABEL: powi_v2f32:
+; X64:       # %bb.0:
+; X64-NEXT:    pushq %rbx
+; X64-NEXT:    .cfi_def_cfa_offset 16
+; X64-NEXT:    subq $32, %rsp
+; X64-NEXT:    .cfi_def_cfa_offset 48
+; X64-NEXT:    .cfi_offset %rbx, -16
+; X64-NEXT:    movaps %xmm0, (%rsp) # 16-byte Spill
+; X64-NEXT:    pushq $15
+; X64-NEXT:    .cfi_adjust_cfa_offset 8
+; X64-NEXT:    popq %rbx
+; X64-NEXT:    .cfi_adjust_cfa_offset -8
+; X64-NEXT:    movl %ebx, %edi
+; X64-NEXT:    callq __powisf2@PLT
+; X64-NEXT:    movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
+; X64-NEXT:    movaps (%rsp), %xmm0 # 16-byte Reload
+; X64-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
+; X64-NEXT:    movl %ebx, %edi
+; X64-NEXT:    callq __powisf2@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:    movaps %xmm1, %xmm0
+; X64-NEXT:    addq $32, %rsp
+; X64-NEXT:    .cfi_def_cfa_offset 16
+; X64-NEXT:    popq %rbx
+; X64-NEXT:    .cfi_def_cfa_offset 8
+; X64-NEXT:    retq
+  %ret = tail call < 2 x float> @llvm.powi.v2f32.i32(<2 x float> %a, i32 15) nounwind ;
+  ret <2 x float> %ret
+}
+
 declare double @llvm.powi.f64.i32(double, i32) nounwind readonly
+declare < 2 x float> @llvm.powi.v2f32.i32(<2 x float>, i32) nounwind readonly
 
 !llvm.module.flags = !{!0}
 !0 = !{i32 1, !"ProfileSummary", !1}
 !1 = !{!2, !3, !4, !5, !6, !7, !8, !9}
 !2 = !{!"ProfileFormat", !"InstrProf"}
 !3 = !{!"TotalCount", i64 10000}
 !4 = !{!"MaxCount", i64 10}
 !5 = !{!"MaxInternalCount", i64 1}