[x86] fix miscompile from wrongly identified fneg

[spatel]

Summary

We may need to peek through a bitcast when identifying an fneg idiom via its pool constant, but we can't allow a different-sized constant in that match.

This is noted in issue #55758 with an example that needs fast-math, but as the test here shows, this has potential to miscompile more generally (no fast-math required).

[RKSimon]

LGTM - its annoying that we don't see the XOR constants, but the 2 XORs (mostly) demonstrate what's happening - if it folded to a single XOR it'd be trickier.

llvm/lib/Target/X86/X86ISelLowering.cpp

case X86ISD::FXOR: {
     // sign bit masks. Ignore the undef elements.
     if (getTargetConstantBitsFromNode(Op1, ScalarSize, UndefElts, EltBits,
                                       /* AllowWholeUndefs */ true,
                                       /* AllowPartialUndefs */ false)) {
       for (unsigned I = 0, E = EltBits.size(); I < E; I++)
         if (!UndefElts[I] && !EltBits[I].isSignMask())
           return SDValue();
 
-      return peekThroughBitcasts(Op0);
-    }
-  }
+      // Only allow bitcast from correctly-sized constant.
+      Op0 = peekThroughBitcasts(Op0);
+      if (Op0.getScalarValueSizeInBits() == ScalarSize)
+        return Op0;
     }
+    break;
+  } // case
+  } // switch
 
   return SDValue();
 }
 
 static unsigned negateFMAOpcode(unsigned Opcode, bool NegMul, bool NegAcc,
                                 bool NegRes) {
   if (NegMul) {
     switch (Opcode) {

llvm/test/CodeGen/X86/neg_fp.ll

; CHECK-NEXT:    retl
   %mul2 = fmul fast float %sub1, %a2
   %neg = fneg fast float %a0
   %add3 = fadd fast float %a1, %neg
   %sub4 = fadd fast float %add3, %a2
   %div5 = fdiv fast float %sub4, %mul2
   ret float %div5
 }
 
-; FIXME: PR55758 - this is not -(-X)
+; PR55758 - this is not -(-X)
 
 define <2 x i64> @fneg_mismatched_sizes(<4 x float> %x) {
 ; CHECK-LABEL: fneg_mismatched_sizes:
 ; CHECK:       # %bb.0:
+; CHECK-NEXT:    xorps {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
+; CHECK-NEXT:    xorps {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0
 ; CHECK-NEXT:    retl
   %n = fneg <4 x float> %x
   %b = bitcast <4 x float> %n to <2 x i64>
   %r = xor <2 x i64> %b, <i64 -9223372036854775808, i64 -9223372036854775808>
   ret <2 x i64> %r
 }