[X86] Improve demanded bits for X86ISD::BEXTR.

[craig.topper]

Summary

If the control is constant we can figure out exactly which bits
of the input are demanded.

[RKSimon]

llvm/lib/Target/X86/X86ISelLowering.cpp

38352

Maybe just call computeKnownBits to avoid code duplication?

[craig.topper]

llvm/lib/Target/X86/X86ISelLowering.cpp

38352

Won’t that be more costly for compile time?

[RKSimon]

llvm/lib/Target/X86/X86ISelLowering.cpp

38341

Return a new zero constant if we manage to get here?

38352

Yeah - probably not worth it

[craig.topper]

llvm/lib/Target/X86/X86ISelLowering.cpp

38341

Doesn't the caller already do that when all demanded bits are known?

[RKSimon]

LGTM

llvm/lib/Target/X86/X86ISelLowering.cpp

if (auto *Cst1 = dyn_cast<ConstantSDNode>(Op1)) {
       const APInt &Val1 = Cst1->getAPIntValue();
       APInt MaskedVal1 = Val1 & 0xFFFF;
       if (MaskedVal1 != Val1) {
         SDLoc DL(Op);
         return TLO.CombineTo(
             Op, TLO.DAG.getNode(X86ISD::BEXTR, DL, VT, Op0,
                                 TLO.DAG.getConstant(MaskedVal1, DL, VT)));
       }
+
+      unsigned Shift = Cst1->getAPIntValue().extractBitsAsZExtValue(8, 0);
+      unsigned Length = Cst1->getAPIntValue().extractBitsAsZExtValue(8, 8);
+
+      // If the length is 0, the result is 0.
+      if (Length == 0) {
+        Known.setAllZero();
+        return false;
+      }
+
+      if ((Shift + Length) <= BitWidth) {
+        APInt DemandedMask = APInt::getBitsSet(BitWidth, Shift, Shift + Length);
+        if (SimplifyDemandedBits(Op0, DemandedMask, Known, TLO, Depth + 1))
+          return true;
+
+        Known = Known.extractBits(Length, Shift);
+        Known = Known.zextOrTrunc(BitWidth);
+        return false;
+      }
     }
 
     KnownBits Known1;
     APInt DemandedMask(APInt::getLowBitsSet(BitWidth, 16));
     if (SimplifyDemandedBits(Op1, DemandedMask, Known1, TLO, Depth + 1))
       return true;
 
     // If the length is 0, replace with 0.

llvm/test/CodeGen/X86/tbm-intrinsics.ll

 ; X64-NEXT:    cmovel %esi, %eax
 ; X64-NEXT:    retq
 entry:
   %0 = tail call i32 @llvm.x86.tbm.bextri.u32(i32 %a, i32 3841)
   %1 = icmp eq i32 %0, 0
   %2 = select i1 %1, i32 %b, i32 %0
   ret i32 %2
 }
+
+define i32 @test_x86_tbm_bextri_demandedbits(i32 %x) nounwind readonly {
+; X86-LABEL: test_x86_tbm_bextri_demandedbits:
+; X86:       # %bb.0:
+; X86-NEXT:    bextrl $3841, {{[0-9]+}}(%esp), %eax # imm = 0xF01
+; X86-NEXT:    retl
+;
+; X64-LABEL: test_x86_tbm_bextri_demandedbits:
+; X64:       # %bb.0:
+; X64-NEXT:    bextrl $3841, %edi, %eax # imm = 0xF01
+; X64-NEXT:    retq
+  %a = or i32 %x, 4294901761
+  %b = tail call i32 @llvm.x86.tbm.bextri.u32(i32 %a, i32 3841)
+  ret i32 %b
+}