[x86] propagate FMF from x86-specific intrinsic nodes to others during lowering

This is another fast-math-flags failure exposed by D90901.

Author: rotateright

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -25049,6 +25049,9 @@ SDValue X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
   MVT VT = Op.getSimpleValueType();
   const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
 
+  // Propagate flags from original node to transformed node(s).
+  SelectionDAG::FlagInserter FlagsInserter(DAG, Op->getFlags());
+
   if (IntrData) {
     switch(IntrData->Type) {
     case INTR_TYPE_1OP: {

llvm/test/CodeGen/X86/fmf-propagation.ll

@@ -47,13 +47,15 @@ define float @fmf_setcc_canon(float %x, float %y) {
 
 declare <16 x float> @llvm.x86.avx512.vfmadd.ps.512(<16 x float>, <16 x float>, <16 x float>, i32)
 
+; Check that FMF are propagated twice: from IR to x86-specific node and from x86-specific node to generic node.
+
 ; CHECK-LABEL: Initial selection DAG: %bb.0 'fmf_target_intrinsic:'
 ; CHECK:       v16f32 = llvm.x86.avx512.vfmadd.ps.512 ninf nsz TargetConstant:i64<{{.*}}>
 ; CHECK:       v16f32 = llvm.x86.avx512.vfmadd.ps.512 nsz TargetConstant:i64<{{.*}}>
 
 ; CHECK-LABEL: Legalized selection DAG: %bb.0 'fmf_target_intrinsic:'
-; CHECK:       v16f32 = fma t{{.*}}
-; CHECK:       v16f32 = fma t{{.*}}
+; CHECK:       v16f32 = fma ninf nsz t{{.*}}
+; CHECK:       v16f32 = fma nsz t{{.*}}
 
 define <16 x float> @fmf_target_intrinsic(<16 x float> %a, <16 x float> %b, <16 x float> %c) nounwind {
   %t0 = tail call ninf nsz <16 x float> @llvm.x86.avx512.vfmadd.ps.512(<16 x float> %a, <16 x float> %b, <16 x float> %c, i32 4)