[DAGCombiner] Simplifying {si|ui}tofp when only signbit is needed

If we only need the signbit `uitofp` simplified to 0, and `sitofp`
simplifies to `bitcast`.

Author: goldsteinn

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

@@ -816,6 +816,18 @@ SDValue TargetLowering::SimplifyMultipleUseDemandedBits(
     }
     break;
   }
+  case ISD::SINT_TO_FP: {
+    EVT InnerVT = Op.getOperand(0).getValueType();
+    if (DemandedBits.isSignMask() &&
+        VT.getScalarSizeInBits() == InnerVT.getScalarSizeInBits())
+      return DAG.getBitcast(VT, Op.getOperand(0));
+    break;
+  }
+  case ISD::UINT_TO_FP: {
+    if (DemandedBits.isSignMask())
+      return DAG.getConstant(0, SDLoc(Op), VT);
+    break;
+  }
   case ISD::SIGN_EXTEND_INREG: {
     // If none of the extended bits are demanded, eliminate the sextinreg.
     SDValue Op0 = Op.getOperand(0);
@@ -2313,6 +2325,22 @@ bool TargetLowering::SimplifyDemandedBits(
     Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth);
     break;
   }
+  case ISD::SINT_TO_FP: {
+    EVT InnerVT = Op.getOperand(0).getValueType();
+    if (DemandedBits.isSignMask() &&
+        VT.getScalarSizeInBits() == InnerVT.getScalarSizeInBits())
+      return TLO.CombineTo(Op, TLO.DAG.getBitcast(VT, Op.getOperand(0)));
+
+    Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth);
+    break;
+  }
+  case ISD::UINT_TO_FP: {
+    if (DemandedBits.isSignMask())
+      return TLO.CombineTo(Op, TLO.DAG.getConstant(0, dl, VT));
+
+    Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth);
+    break;
+  }
   case ISD::SIGN_EXTEND_INREG: {
     SDValue Op0 = Op.getOperand(0);
     EVT ExVT = cast<VTSDNode>(Op.getOperand(1))->getVT();

llvm/test/CodeGen/X86/combine-sse41-intrinsics.ll

@@ -164,14 +164,13 @@ define <4 x float> @demandedbits_sitofp_blendvps(<4 x float> %a0, <4 x float> %a
 ; SSE-LABEL: demandedbits_sitofp_blendvps:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    movaps %xmm0, %xmm3
-; SSE-NEXT:    cvtdq2ps %xmm2, %xmm0
+; SSE-NEXT:    movaps %xmm2, %xmm0
 ; SSE-NEXT:    blendvps %xmm0, %xmm1, %xmm3
 ; SSE-NEXT:    movaps %xmm3, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: demandedbits_sitofp_blendvps:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vcvtdq2ps %xmm2, %xmm2
 ; AVX-NEXT:    vblendvps %xmm2, %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %cvt = sitofp <4 x i32> %a2 to <4 x float>

llvm/test/CodeGen/X86/int-to-fp-demanded.ll

@@ -7,21 +7,13 @@ declare void @use.i32(i32)
 define i32 @sitofp_signbit_only(i32 %i_in) {
 ; X86-LABEL: sitofp_signbit_only:
 ; X86:       # %bb.0:
-; X86-NEXT:    subl $8, %esp
-; X86-NEXT:    .cfi_def_cfa_offset 12
-; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    movl %eax, (%esp)
-; X86-NEXT:    fildl (%esp)
-; X86-NEXT:    fstps {{[0-9]+}}(%esp)
 ; X86-NEXT:    movl $-2147483648, %eax # imm = 0x80000000
 ; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    addl $8, %esp
-; X86-NEXT:    .cfi_def_cfa_offset 4
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: sitofp_signbit_only:
 ; X64:       # %bb.0:
-; X64-NEXT:    cvtsi2ss %edi, %xmm0
+; X64-NEXT:    movd %edi, %xmm0
 ; X64-NEXT:    movmskps %xmm0, %eax
 ; X64-NEXT:    shll $31, %eax
 ; X64-NEXT:    retq
@@ -48,8 +40,8 @@ define i32 @sitofp_signbit_only_okay_width(i16 %i_in) {
 ;
 ; X64-LABEL: sitofp_signbit_only_okay_width:
 ; X64:       # %bb.0:
-; X64-NEXT:    movswl %di, %eax
-; X64-NEXT:    cvtsi2ss %eax, %xmm0
+; X64-NEXT:    shll $16, %edi
+; X64-NEXT:    movd %edi, %xmm0
 ; X64-NEXT:    movmskps %xmm0, %eax
 ; X64-NEXT:    shll $31, %eax
 ; X64-NEXT:    retq
@@ -89,15 +81,14 @@ define <2 x i16> @sitofp_signbit_only_fail_bad_width2(i32 %i_in) {
 ; X86:       # %bb.0:
 ; X86-NEXT:    subl $8, %esp
 ; X86-NEXT:    .cfi_def_cfa_offset 12
-; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    movl %eax, (%esp)
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X86-NEXT:    movl %edx, (%esp)
 ; X86-NEXT:    fildl (%esp)
 ; X86-NEXT:    fstps {{[0-9]+}}(%esp)
-; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    movl %eax, %edx
 ; X86-NEXT:    shrl $16, %edx
-; X86-NEXT:    andl $32768, %eax # imm = 0x8000
 ; X86-NEXT:    andl $32768, %edx # imm = 0x8000
+; X86-NEXT:    movl $32768, %eax # imm = 0x8000
+; X86-NEXT:    andl {{[0-9]+}}(%esp), %eax
 ; X86-NEXT:    # kill: def $ax killed $ax killed $eax
 ; X86-NEXT:    # kill: def $dx killed $dx killed $edx
 ; X86-NEXT:    addl $8, %esp