Fix mul for negative zero

Author: peterbell10

llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp

@@ -2548,11 +2548,13 @@ SDValue NVPTXTargetLowering::LowerFMUL(SDValue Op, SelectionDAG &DAG) const {
     return PromoteBinOpToF32(Op.getNode(), DAG);
   }
 
-  // FMUL(a, b) -> FMA(a, b, 0.0)
+  // FMUL(a, b) -> FMA(a, b, -0.0)
+  // NOTE: The identity is -0, not 0, because -0 + 0 == 0 for floats
   SDLoc DL(Op);
   auto VT = Op.getValueType();
-  auto Zero = DAG.getConstantFP(0.0, DL, VT);
-  SmallVector<SDValue, 3> Operands{Op->getOperand(0), Op->getOperand(1), Zero};
+  auto NegZero = DAG.getConstantFP(-0.0, DL, VT);
+  SmallVector<SDValue, 3> Operands{Op->getOperand(0), Op->getOperand(1),
+                                   NegZero};
   return DAG.getNode(ISD::FMA, DL, VT, Operands);
 }
 

llvm/test/CodeGen/NVPTX/bf16-instructions.ll

@@ -385,7 +385,7 @@ define <2 x bfloat> @test_fmulx2(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
 ; SM80-NEXT:  // %bb.0:
 ; SM80-NEXT:    ld.param.b32 %r1, [test_fmulx2_param_1];
 ; SM80-NEXT:    ld.param.b32 %r2, [test_fmulx2_param_0];
-; SM80-NEXT:    mov.b32 %r3, 0;
+; SM80-NEXT:    mov.b32 %r3, -2147450880;
 ; SM80-NEXT:    fma.rn.bf16x2 %r4, %r2, %r1, %r3;
 ; SM80-NEXT:    st.param.b32 [func_retval0], %r4;
 ; SM80-NEXT:    ret;

llvm/test/CodeGen/NVPTX/bf16x2-instructions.ll

@@ -107,7 +107,7 @@ define <2 x bfloat> @test_fmulx2(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
 ; SM80-NEXT:  // %bb.0:
 ; SM80-NEXT:    ld.param.b32 %r1, [test_fmulx2_param_1];
 ; SM80-NEXT:    ld.param.b32 %r2, [test_fmulx2_param_0];
-; SM80-NEXT:    mov.b32 %r3, 0;
+; SM80-NEXT:    mov.b32 %r3, -2147450880;
 ; SM80-NEXT:    fma.rn.bf16x2 %r4, %r2, %r1, %r3;
 ; SM80-NEXT:    st.param.b32 [func_retval0], %r4;
 ; SM80-NEXT:    ret;