[SelectionDAG] Rewrite bfloat16 softening to use the "half promotion" path

The main difference is that this preserves intermediate rounding steps,
which the other route doesn't. This aligns bfloat16 more with half
floats, which use this path on most targets.

I didn't understand what the difference was between these softening
approaches when I first added bfloat lowerings, would be nice if we only
had one of them.

Based on @pengfei 's D131502

Differential Revision: https://reviews.llvm.org/D133207

diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
index 56ad9bc..ab60e3d 100644 (file)
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
@@ -2914,6 +2914,9 @@ bool SelectionDAGLegalize::ExpandNode(SDNode *Node) {
         DAG.getConstant(16, dl,
                         TLI.getShiftAmountTy(MVT::i32, DAG.getDataLayout())));
     Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op);
+    // Add fp_extend in case the output is bigger than f32.
+    if (Node->getValueType(0) != MVT::f32)
+      Op = DAG.getNode(ISD::FP_EXTEND, dl, Node->getValueType(0), Op);
     Results.push_back(Op);
     break;
   }

diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp
index a966b69..f1e80ce 100644 (file)
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp
@@ -2748,39 +2748,47 @@ SDValue DAGTypeLegalizer::SoftPromoteHalfRes_FCOPYSIGN(SDNode *N) {
 }
 
 SDValue DAGTypeLegalizer::SoftPromoteHalfRes_FMAD(SDNode *N) {
-  EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+  EVT OVT = N->getValueType(0);
+  EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), OVT);
   SDValue Op0 = GetSoftPromotedHalf(N->getOperand(0));
   SDValue Op1 = GetSoftPromotedHalf(N->getOperand(1));
   SDValue Op2 = GetSoftPromotedHalf(N->getOperand(2));
   SDLoc dl(N);
 
   // Promote to the larger FP type.
-  Op0 = DAG.getNode(ISD::FP16_TO_FP, dl, NVT, Op0);
-  Op1 = DAG.getNode(ISD::FP16_TO_FP, dl, NVT, Op1);
-  Op2 = DAG.getNode(ISD::FP16_TO_FP, dl, NVT, Op2);
+  auto PromotionOpcode = GetPromotionOpcode(OVT, NVT);
+  Op0 = DAG.getNode(PromotionOpcode, dl, NVT, Op0);
+  Op1 = DAG.getNode(PromotionOpcode, dl, NVT, Op1);
+  Op2 = DAG.getNode(PromotionOpcode, dl, NVT, Op2);
 
   SDValue Res = DAG.getNode(N->getOpcode(), dl, NVT, Op0, Op1, Op2);
 
   // Convert back to FP16 as an integer.
-  return DAG.getNode(ISD::FP_TO_FP16, dl, MVT::i16, Res);
+  return DAG.getNode(GetPromotionOpcode(NVT, OVT), dl, MVT::i16, Res);
 }
 
 SDValue DAGTypeLegalizer::SoftPromoteHalfRes_FPOWI(SDNode *N) {
-  EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+  EVT OVT = N->getValueType(0);
+  EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), OVT);
   SDValue Op0 = GetSoftPromotedHalf(N->getOperand(0));
   SDValue Op1 = N->getOperand(1);
   SDLoc dl(N);
 
-  Op0 = DAG.getNode(ISD::FP16_TO_FP, dl, NVT, Op0);
+  // Promote to the larger FP type.
+  Op0 = DAG.getNode(GetPromotionOpcode(OVT, NVT), dl, NVT, Op0);
 
   SDValue Res = DAG.getNode(N->getOpcode(), dl, NVT, Op0, Op1);
 
   // Convert back to FP16 as an integer.
-  return DAG.getNode(ISD::FP_TO_FP16, dl, MVT::i16, Res);
+  return DAG.getNode(GetPromotionOpcode(NVT, OVT), dl, MVT::i16, Res);
 }
 
 SDValue DAGTypeLegalizer::SoftPromoteHalfRes_FP_ROUND(SDNode *N) {
+  EVT RVT = N->getValueType(0);
+  EVT SVT = N->getOperand(0).getValueType();
+
   if (N->isStrictFPOpcode()) {
+    assert(RVT == MVT::f16);
     SDValue Res =
         DAG.getNode(ISD::STRICT_FP_TO_FP16, SDLoc(N), {MVT::i16, MVT::Other},
                     {N->getOperand(0), N->getOperand(1)});
@@ -2788,7 +2796,8 @@ SDValue DAGTypeLegalizer::SoftPromoteHalfRes_FP_ROUND(SDNode *N) {
     return Res;
   }
 
-  return DAG.getNode(ISD::FP_TO_FP16, SDLoc(N), MVT::i16, N->getOperand(0));
+  return DAG.getNode(GetPromotionOpcode(SVT, RVT), SDLoc(N), MVT::i16,
+                     N->getOperand(0));
 }
 
 SDValue DAGTypeLegalizer::SoftPromoteHalfRes_LOAD(SDNode *N) {
@@ -2823,13 +2832,14 @@ SDValue DAGTypeLegalizer::SoftPromoteHalfRes_SELECT_CC(SDNode *N) {
 }
 
 SDValue DAGTypeLegalizer::SoftPromoteHalfRes_XINT_TO_FP(SDNode *N) {
-  EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+  EVT OVT = N->getValueType(0);
+  EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), OVT);
   SDLoc dl(N);
 
   SDValue Res = DAG.getNode(N->getOpcode(), dl, NVT, N->getOperand(0));
 
   // Round the value to the softened type.
-  return DAG.getNode(ISD::FP_TO_FP16, dl, MVT::i16, Res);
+  return DAG.getNode(GetPromotionOpcode(NVT, OVT), dl, MVT::i16, Res);
 }
 
 SDValue DAGTypeLegalizer::SoftPromoteHalfRes_UNDEF(SDNode *N) {
@@ -2837,33 +2847,36 @@ SDValue DAGTypeLegalizer::SoftPromoteHalfRes_UNDEF(SDNode *N) {
 }
 
 SDValue DAGTypeLegalizer::SoftPromoteHalfRes_UnaryOp(SDNode *N) {
-  EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+  EVT OVT = N->getValueType(0);
+  EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), OVT);
   SDValue Op = GetSoftPromotedHalf(N->getOperand(0));
   SDLoc dl(N);
 
   // Promote to the larger FP type.
-  Op = DAG.getNode(ISD::FP16_TO_FP, dl, NVT, Op);
+  Op = DAG.getNode(GetPromotionOpcode(OVT, NVT), dl, NVT, Op);
 
   SDValue Res = DAG.getNode(N->getOpcode(), dl, NVT, Op);
 
   // Convert back to FP16 as an integer.
-  return DAG.getNode(ISD::FP_TO_FP16, dl, MVT::i16, Res);
+  return DAG.getNode(GetPromotionOpcode(NVT, OVT), dl, MVT::i16, Res);
 }
 
 SDValue DAGTypeLegalizer::SoftPromoteHalfRes_BinOp(SDNode *N) {
-  EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+  EVT OVT = N->getValueType(0);
+  EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), OVT);
   SDValue Op0 = GetSoftPromotedHalf(N->getOperand(0));
   SDValue Op1 = GetSoftPromotedHalf(N->getOperand(1));
   SDLoc dl(N);
 
   // Promote to the larger FP type.
-  Op0 = DAG.getNode(ISD::FP16_TO_FP, dl, NVT, Op0);
-  Op1 = DAG.getNode(ISD::FP16_TO_FP, dl, NVT, Op1);
+  auto PromotionOpcode = GetPromotionOpcode(OVT, NVT);
+  Op0 = DAG.getNode(PromotionOpcode, dl, NVT, Op0);
+  Op1 = DAG.getNode(PromotionOpcode, dl, NVT, Op1);
 
   SDValue Res = DAG.getNode(N->getOpcode(), dl, NVT, Op0, Op1);
 
   // Convert back to FP16 as an integer.
-  return DAG.getNode(ISD::FP_TO_FP16, dl, MVT::i16, Res);
+  return DAG.getNode(GetPromotionOpcode(NVT, OVT), dl, MVT::i16, Res);
 }
 
 SDValue DAGTypeLegalizer::SoftPromoteHalfRes_VECREDUCE(SDNode *N) {
@@ -2947,22 +2960,27 @@ SDValue DAGTypeLegalizer::SoftPromoteHalfOp_FCOPYSIGN(SDNode *N,
                                                       unsigned OpNo) {
   assert(OpNo == 1 && "Only Operand 1 must need promotion here");
   SDValue Op1 = N->getOperand(1);
+  EVT RVT = Op1.getValueType();
   SDLoc dl(N);
 
   EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), Op1.getValueType());
 
   Op1 = GetSoftPromotedHalf(Op1);
-  Op1 = DAG.getNode(ISD::FP16_TO_FP, dl, NVT, Op1);
+  Op1 = DAG.getNode(GetPromotionOpcode(RVT, NVT), dl, NVT, Op1);
 
   return DAG.getNode(N->getOpcode(), dl, N->getValueType(0), N->getOperand(0),
                      Op1);
 }
 
 SDValue DAGTypeLegalizer::SoftPromoteHalfOp_FP_EXTEND(SDNode *N) {
+  EVT RVT = N->getValueType(0);
   bool IsStrict = N->isStrictFPOpcode();
-  SDValue Op = GetSoftPromotedHalf(N->getOperand(IsStrict ? 1 : 0));
+  SDValue Op = N->getOperand(IsStrict ? 1 : 0);
+  EVT SVT = Op.getValueType();
+  Op = GetSoftPromotedHalf(N->getOperand(IsStrict ? 1 : 0));
 
   if (IsStrict) {
+    assert(SVT == MVT::f16);
     SDValue Res =
         DAG.getNode(ISD::STRICT_FP16_TO_FP, SDLoc(N),
                     {N->getValueType(0), MVT::Other}, {N->getOperand(0), Op});
@@ -2971,31 +2989,35 @@ SDValue DAGTypeLegalizer::SoftPromoteHalfOp_FP_EXTEND(SDNode *N) {
     return SDValue();
   }
 
-  return DAG.getNode(ISD::FP16_TO_FP, SDLoc(N), N->getValueType(0), Op);
+  return DAG.getNode(GetPromotionOpcode(SVT, RVT), SDLoc(N), RVT, Op);
 }
 
 SDValue DAGTypeLegalizer::SoftPromoteHalfOp_FP_TO_XINT(SDNode *N) {
+  EVT RVT = N->getValueType(0);
   SDValue Op = N->getOperand(0);
+  EVT SVT = Op.getValueType();
   SDLoc dl(N);
 
   EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), Op.getValueType());
 
   Op = GetSoftPromotedHalf(Op);
 
-  SDValue Res = DAG.getNode(ISD::FP16_TO_FP, dl, NVT, Op);
+  SDValue Res = DAG.getNode(GetPromotionOpcode(SVT, RVT), dl, NVT, Op);
 
   return DAG.getNode(N->getOpcode(), dl, N->getValueType(0), Res);
 }
 
 SDValue DAGTypeLegalizer::SoftPromoteHalfOp_FP_TO_XINT_SAT(SDNode *N) {
+  EVT RVT = N->getValueType(0);
   SDValue Op = N->getOperand(0);
+  EVT SVT = Op.getValueType();
   SDLoc dl(N);
 
   EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), Op.getValueType());
 
   Op = GetSoftPromotedHalf(Op);
 
-  SDValue Res = DAG.getNode(ISD::FP16_TO_FP, dl, NVT, Op);
+  SDValue Res = DAG.getNode(GetPromotionOpcode(SVT, RVT), dl, NVT, Op);
 
   return DAG.getNode(N->getOpcode(), dl, N->getValueType(0), Res,
                      N->getOperand(1));
@@ -3008,14 +3030,16 @@ SDValue DAGTypeLegalizer::SoftPromoteHalfOp_SELECT_CC(SDNode *N,
   SDValue Op1 = N->getOperand(1);
   SDLoc dl(N);
 
-  EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), Op0.getValueType());
+  EVT SVT = Op0.getValueType();
+  EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), SVT);
 
   Op0 = GetSoftPromotedHalf(Op0);
   Op1 = GetSoftPromotedHalf(Op1);
 
   // Promote to the larger FP type.
-  Op0 = DAG.getNode(ISD::FP16_TO_FP, dl, NVT, Op0);
-  Op1 = DAG.getNode(ISD::FP16_TO_FP, dl, NVT, Op1);
+  auto PromotionOpcode = GetPromotionOpcode(SVT, NVT);
+  Op0 = DAG.getNode(PromotionOpcode, dl, NVT, Op0);
+  Op1 = DAG.getNode(PromotionOpcode, dl, NVT, Op1);
 
   return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N->getValueType(0), Op0, Op1,
                      N->getOperand(2), N->getOperand(3), N->getOperand(4));
@@ -3027,14 +3051,16 @@ SDValue DAGTypeLegalizer::SoftPromoteHalfOp_SETCC(SDNode *N) {
   ISD::CondCode CCCode = cast<CondCodeSDNode>(N->getOperand(2))->get();
   SDLoc dl(N);
 
+  EVT SVT = Op0.getValueType();
   EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), Op0.getValueType());
 
   Op0 = GetSoftPromotedHalf(Op0);
   Op1 = GetSoftPromotedHalf(Op1);
 
   // Promote to the larger FP type.
-  Op0 = DAG.getNode(ISD::FP16_TO_FP, dl, NVT, Op0);
-  Op1 = DAG.getNode(ISD::FP16_TO_FP, dl, NVT, Op1);
+  auto PromotionOpcode = GetPromotionOpcode(SVT, NVT);
+  Op0 = DAG.getNode(PromotionOpcode, dl, NVT, Op0);
+  Op1 = DAG.getNode(PromotionOpcode, dl, NVT, Op1);
 
   return DAG.getSetCC(SDLoc(N), N->getValueType(0), Op0, Op1, CCCode);
 }

diff --git a/llvm/lib/CodeGen/TargetLoweringBase.cpp b/llvm/lib/CodeGen/TargetLoweringBase.cpp
index 4cf81cd..9bc2a1e 100644 (file)
--- a/llvm/lib/CodeGen/TargetLoweringBase.cpp
+++ b/llvm/lib/CodeGen/TargetLoweringBase.cpp
@@ -1385,7 +1385,7 @@ void TargetLoweringBase::computeRegisterProperties(
     NumRegistersForVT[MVT::bf16] = NumRegistersForVT[MVT::f32];
     RegisterTypeForVT[MVT::bf16] = RegisterTypeForVT[MVT::f32];
     TransformToType[MVT::bf16] = MVT::f32;
-    ValueTypeActions.setTypeAction(MVT::bf16, TypePromoteFloat);
+    ValueTypeActions.setTypeAction(MVT::bf16, TypeSoftPromoteHalf);
   }
 
   // Loop over all of the vector value types to see which need transformations.

diff --git a/llvm/test/CodeGen/X86/bfloat.ll b/llvm/test/CodeGen/X86/bfloat.ll
index e3f88f0..391833d 100644 (file)
--- a/llvm/test/CodeGen/X86/bfloat.ll
+++ b/llvm/test/CodeGen/X86/bfloat.ll
@@ -6,10 +6,10 @@ define void @add(ptr %pa, ptr %pb, ptr %pc) nounwind {
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    pushq %rbx
 ; CHECK-NEXT:    movq %rdx, %rbx
-; CHECK-NEXT:    movzwl (%rdi), %eax
+; CHECK-NEXT:    movzwl (%rsi), %eax
 ; CHECK-NEXT:    shll $16, %eax
 ; CHECK-NEXT:    movd %eax, %xmm1
-; CHECK-NEXT:    movzwl (%rsi), %eax
+; CHECK-NEXT:    movzwl (%rdi), %eax
 ; CHECK-NEXT:    shll $16, %eax
 ; CHECK-NEXT:    movd %eax, %xmm0
 ; CHECK-NEXT:    addss %xmm1, %xmm0
@@ -29,10 +29,10 @@ define bfloat @add2(bfloat %a, bfloat %b) nounwind {
 ; CHECK-LABEL: add2:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    pushq %rax
-; CHECK-NEXT:    movd %xmm1, %eax
-; CHECK-NEXT:    shll $16, %eax
-; CHECK-NEXT:    movd %eax, %xmm1
 ; CHECK-NEXT:    movd %xmm0, %eax
+; CHECK-NEXT:    movd %xmm1, %ecx
+; CHECK-NEXT:    shll $16, %ecx
+; CHECK-NEXT:    movd %ecx, %xmm1
 ; CHECK-NEXT:    shll $16, %eax
 ; CHECK-NEXT:    movd %eax, %xmm0
 ; CHECK-NEXT:    addss %xmm1, %xmm0
@@ -46,27 +46,31 @@ define bfloat @add2(bfloat %a, bfloat %b) nounwind {
 define void @add_double(ptr %pa, ptr %pb, ptr %pc) nounwind {
 ; CHECK-LABEL: add_double:
 ; CHECK:       # %bb.0:
+; CHECK-NEXT:    pushq %rbp
 ; CHECK-NEXT:    pushq %r14
 ; CHECK-NEXT:    pushq %rbx
-; CHECK-NEXT:    pushq %rax
 ; CHECK-NEXT:    movq %rdx, %r14
 ; CHECK-NEXT:    movq %rsi, %rbx
 ; CHECK-NEXT:    movq {{.*#+}} xmm0 = mem[0],zero
 ; CHECK-NEXT:    callq __truncdfbf2@PLT
-; CHECK-NEXT:    movd %xmm0, %eax
-; CHECK-NEXT:    shll $16, %eax
-; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
+; CHECK-NEXT:    movd %xmm0, %ebp
 ; CHECK-NEXT:    movq {{.*#+}} xmm0 = mem[0],zero
 ; CHECK-NEXT:    callq __truncdfbf2@PLT
 ; CHECK-NEXT:    movd %xmm0, %eax
 ; CHECK-NEXT:    shll $16, %eax
+; CHECK-NEXT:    movd %eax, %xmm1
+; CHECK-NEXT:    shll $16, %ebp
+; CHECK-NEXT:    movd %ebp, %xmm0
+; CHECK-NEXT:    addss %xmm1, %xmm0
+; CHECK-NEXT:    callq __truncsfbf2@PLT
+; CHECK-NEXT:    movd %xmm0, %eax
+; CHECK-NEXT:    shll $16, %eax
 ; CHECK-NEXT:    movd %eax, %xmm0
-; CHECK-NEXT:    addss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Folded Reload
 ; CHECK-NEXT:    cvtss2sd %xmm0, %xmm0
 ; CHECK-NEXT:    movsd %xmm0, (%r14)
-; CHECK-NEXT:    addq $8, %rsp
 ; CHECK-NEXT:    popq %rbx
 ; CHECK-NEXT:    popq %r14
+; CHECK-NEXT:    popq %rbp
 ; CHECK-NEXT:    retq
   %la = load double, ptr %pa
   %a = fptrunc double %la to bfloat
@@ -81,21 +85,27 @@ define void @add_double(ptr %pa, ptr %pb, ptr %pc) nounwind {
 define double @add_double2(double %da, double %db) nounwind {
 ; CHECK-LABEL: add_double2:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    subq $24, %rsp
+; CHECK-NEXT:    pushq %rbx
+; CHECK-NEXT:    subq $16, %rsp
 ; CHECK-NEXT:    movsd %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
 ; CHECK-NEXT:    callq __truncdfbf2@PLT
-; CHECK-NEXT:    movd %xmm0, %eax
-; CHECK-NEXT:    shll $16, %eax
-; CHECK-NEXT:    movl %eax, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
+; CHECK-NEXT:    movd %xmm0, %ebx
 ; CHECK-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 8-byte Folded Reload
 ; CHECK-NEXT:    # xmm0 = mem[0],zero
 ; CHECK-NEXT:    callq __truncdfbf2@PLT
 ; CHECK-NEXT:    movd %xmm0, %eax
 ; CHECK-NEXT:    shll $16, %eax
+; CHECK-NEXT:    movd %eax, %xmm1
+; CHECK-NEXT:    shll $16, %ebx
+; CHECK-NEXT:    movd %ebx, %xmm0
+; CHECK-NEXT:    addss %xmm1, %xmm0
+; CHECK-NEXT:    callq __truncsfbf2@PLT
+; CHECK-NEXT:    movd %xmm0, %eax
+; CHECK-NEXT:    shll $16, %eax
 ; CHECK-NEXT:    movd %eax, %xmm0
-; CHECK-NEXT:    addss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Folded Reload
 ; CHECK-NEXT:    cvtss2sd %xmm0, %xmm0
-; CHECK-NEXT:    addq $24, %rsp
+; CHECK-NEXT:    addq $16, %rsp
+; CHECK-NEXT:    popq %rbx
 ; CHECK-NEXT:    retq
   %a = fptrunc double %da to bfloat
   %b = fptrunc double %db to bfloat
@@ -174,134 +184,135 @@ define <8 x bfloat> @addv(<8 x bfloat> %a, <8 x bfloat> %b) nounwind {
 ; CHECK-LABEL: addv:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    pushq %rbp
+; CHECK-NEXT:    pushq %r15
 ; CHECK-NEXT:    pushq %r14
+; CHECK-NEXT:    pushq %r13
+; CHECK-NEXT:    pushq %r12
 ; CHECK-NEXT:    pushq %rbx
-; CHECK-NEXT:    subq $32, %rsp
-; CHECK-NEXT:    movq %xmm1, %rax
-; CHECK-NEXT:    movq %rax, %rcx
-; CHECK-NEXT:    shrq $32, %rcx
-; CHECK-NEXT:    shll $16, %ecx
-; CHECK-NEXT:    movd %ecx, %xmm2
+; CHECK-NEXT:    subq $56, %rsp
 ; CHECK-NEXT:    movq %xmm0, %rcx
-; CHECK-NEXT:    movq %rcx, %rdx
-; CHECK-NEXT:    shrq $32, %rdx
-; CHECK-NEXT:    shll $16, %edx
-; CHECK-NEXT:    movd %edx, %xmm3
-; CHECK-NEXT:    addss %xmm2, %xmm3
-; CHECK-NEXT:    movss %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; CHECK-NEXT:    movq %rax, %rdx
-; CHECK-NEXT:    shrq $48, %rdx
-; CHECK-NEXT:    shll $16, %edx
-; CHECK-NEXT:    movd %edx, %xmm2
-; CHECK-NEXT:    movq %rcx, %rdx
-; CHECK-NEXT:    shrq $48, %rdx
-; CHECK-NEXT:    shll $16, %edx
-; CHECK-NEXT:    movd %edx, %xmm3
-; CHECK-NEXT:    addss %xmm2, %xmm3
-; CHECK-NEXT:    movss %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; CHECK-NEXT:    movl %eax, %edx
-; CHECK-NEXT:    shll $16, %edx
-; CHECK-NEXT:    movd %edx, %xmm2
-; CHECK-NEXT:    movl %ecx, %edx
-; CHECK-NEXT:    shll $16, %edx
-; CHECK-NEXT:    movd %edx, %xmm3
-; CHECK-NEXT:    addss %xmm2, %xmm3
-; CHECK-NEXT:    movss %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; CHECK-NEXT:    andl $-65536, %eax # imm = 0xFFFF0000
-; CHECK-NEXT:    movd %eax, %xmm2
-; CHECK-NEXT:    andl $-65536, %ecx # imm = 0xFFFF0000
-; CHECK-NEXT:    movd %ecx, %xmm3
-; CHECK-NEXT:    addss %xmm2, %xmm3
-; CHECK-NEXT:    movss %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; CHECK-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
-; CHECK-NEXT:    movq %xmm1, %rax
-; CHECK-NEXT:    movq %rax, %rcx
-; CHECK-NEXT:    shrq $32, %rcx
-; CHECK-NEXT:    shll $16, %ecx
-; CHECK-NEXT:    movd %ecx, %xmm1
+; CHECK-NEXT:    movq %rcx, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
+; CHECK-NEXT:    movq %rcx, %rax
+; CHECK-NEXT:    shrq $32, %rax
+; CHECK-NEXT:    movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
+; CHECK-NEXT:    movq %xmm1, %rdx
+; CHECK-NEXT:    movq %rdx, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
+; CHECK-NEXT:    movq %rdx, %rax
+; CHECK-NEXT:    shrq $32, %rax
+; CHECK-NEXT:    movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
+; CHECK-NEXT:    movq %rcx, %rax
+; CHECK-NEXT:    shrq $48, %rax
+; CHECK-NEXT:    movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
+; CHECK-NEXT:    movq %rdx, %rax
+; CHECK-NEXT:    shrq $48, %rax
+; CHECK-NEXT:    movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
 ; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
-; CHECK-NEXT:    movq %xmm0, %rcx
-; CHECK-NEXT:    movq %rcx, %rdx
-; CHECK-NEXT:    shrq $32, %rdx
-; CHECK-NEXT:    shll $16, %edx
-; CHECK-NEXT:    movd %edx, %xmm0
-; CHECK-NEXT:    addss %xmm1, %xmm0
-; CHECK-NEXT:    movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; CHECK-NEXT:    movq %rax, %rdx
-; CHECK-NEXT:    shrq $48, %rdx
-; CHECK-NEXT:    shll $16, %edx
-; CHECK-NEXT:    movd %edx, %xmm0
-; CHECK-NEXT:    movq %rcx, %rdx
-; CHECK-NEXT:    shrq $48, %rdx
-; CHECK-NEXT:    shll $16, %edx
-; CHECK-NEXT:    movd %edx, %xmm1
-; CHECK-NEXT:    addss %xmm0, %xmm1
-; CHECK-NEXT:    movss %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
-; CHECK-NEXT:    movl %eax, %edx
-; CHECK-NEXT:    shll $16, %edx
-; CHECK-NEXT:    movd %edx, %xmm0
-; CHECK-NEXT:    movl %ecx, %edx
-; CHECK-NEXT:    shll $16, %edx
-; CHECK-NEXT:    movd %edx, %xmm1
-; CHECK-NEXT:    addss %xmm0, %xmm1
-; CHECK-NEXT:    movss %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
+; CHECK-NEXT:    movq %xmm0, %rbx
+; CHECK-NEXT:    movq %rbx, %rax
+; CHECK-NEXT:    shrq $32, %rax
+; CHECK-NEXT:    movq %rax, (%rsp) # 8-byte Spill
+; CHECK-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,2,3]
+; CHECK-NEXT:    movq %xmm0, %rbp
+; CHECK-NEXT:    movq %rbp, %r15
+; CHECK-NEXT:    shrq $32, %r15
+; CHECK-NEXT:    movq %rbx, %r13
+; CHECK-NEXT:    shrq $48, %r13
+; CHECK-NEXT:    movq %rbp, %r12
+; CHECK-NEXT:    shrq $48, %r12
+; CHECK-NEXT:    movl %ebp, %eax
 ; CHECK-NEXT:    andl $-65536, %eax # imm = 0xFFFF0000
 ; CHECK-NEXT:    movd %eax, %xmm1
-; CHECK-NEXT:    andl $-65536, %ecx # imm = 0xFFFF0000
-; CHECK-NEXT:    movd %ecx, %xmm0
+; CHECK-NEXT:    movl %ebx, %eax
+; CHECK-NEXT:    andl $-65536, %eax # imm = 0xFFFF0000
+; CHECK-NEXT:    movd %eax, %xmm0
 ; CHECK-NEXT:    addss %xmm1, %xmm0
 ; CHECK-NEXT:    callq __truncsfbf2@PLT
-; CHECK-NEXT:    movd %xmm0, %ebx
+; CHECK-NEXT:    movd %xmm0, %r14d
+; CHECK-NEXT:    shll $16, %r14d
+; CHECK-NEXT:    shll $16, %ebp
+; CHECK-NEXT:    movd %ebp, %xmm1
 ; CHECK-NEXT:    shll $16, %ebx
-; CHECK-NEXT:    movd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Folded Reload
-; CHECK-NEXT:    # xmm0 = mem[0],zero,zero,zero
+; CHECK-NEXT:    movd %ebx, %xmm0
+; CHECK-NEXT:    addss %xmm1, %xmm0
 ; CHECK-NEXT:    callq __truncsfbf2@PLT
 ; CHECK-NEXT:    movd %xmm0, %eax
-; CHECK-NEXT:    movzwl %ax, %r14d
-; CHECK-NEXT:    orl %ebx, %r14d
-; CHECK-NEXT:    movd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Folded Reload
-; CHECK-NEXT:    # xmm0 = mem[0],zero,zero,zero
+; CHECK-NEXT:    movzwl %ax, %ebx
+; CHECK-NEXT:    orl %r14d, %ebx
+; CHECK-NEXT:    shll $16, %r12d
+; CHECK-NEXT:    movd %r12d, %xmm1
+; CHECK-NEXT:    shll $16, %r13d
+; CHECK-NEXT:    movd %r13d, %xmm0
+; CHECK-NEXT:    addss %xmm1, %xmm0
 ; CHECK-NEXT:    callq __truncsfbf2@PLT
 ; CHECK-NEXT:    movd %xmm0, %ebp
 ; CHECK-NEXT:    shll $16, %ebp
-; CHECK-NEXT:    movd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Folded Reload
-; CHECK-NEXT:    # xmm0 = mem[0],zero,zero,zero
+; CHECK-NEXT:    shll $16, %r15d
+; CHECK-NEXT:    movd %r15d, %xmm1
+; CHECK-NEXT:    movq (%rsp), %rax # 8-byte Reload
+; CHECK-NEXT:    shll $16, %eax
+; CHECK-NEXT:    movd %eax, %xmm0
+; CHECK-NEXT:    addss %xmm1, %xmm0
 ; CHECK-NEXT:    callq __truncsfbf2@PLT
 ; CHECK-NEXT:    movd %xmm0, %eax
-; CHECK-NEXT:    movzwl %ax, %ebx
-; CHECK-NEXT:    orl %ebp, %ebx
-; CHECK-NEXT:    shlq $32, %rbx
-; CHECK-NEXT:    orq %r14, %rbx
-; CHECK-NEXT:    movd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Folded Reload
-; CHECK-NEXT:    # xmm0 = mem[0],zero,zero,zero
+; CHECK-NEXT:    movzwl %ax, %r14d
+; CHECK-NEXT:    orl %ebp, %r14d
+; CHECK-NEXT:    shlq $32, %r14
+; CHECK-NEXT:    orq %rbx, %r14
+; CHECK-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %r15 # 8-byte Reload
+; CHECK-NEXT:    movl %r15d, %eax
+; CHECK-NEXT:    andl $-65536, %eax # imm = 0xFFFF0000
+; CHECK-NEXT:    movd %eax, %xmm1
+; CHECK-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rbx # 8-byte Reload
+; CHECK-NEXT:    movl %ebx, %eax
+; CHECK-NEXT:    andl $-65536, %eax # imm = 0xFFFF0000
+; CHECK-NEXT:    movd %eax, %xmm0
+; CHECK-NEXT:    addss %xmm1, %xmm0
 ; CHECK-NEXT:    callq __truncsfbf2@PLT
 ; CHECK-NEXT:    movd %xmm0, %ebp
 ; CHECK-NEXT:    shll $16, %ebp
-; CHECK-NEXT:    movd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Folded Reload
-; CHECK-NEXT:    # xmm0 = mem[0],zero,zero,zero
+; CHECK-NEXT:    movq %r15, %rax
+; CHECK-NEXT:    shll $16, %eax
+; CHECK-NEXT:    movd %eax, %xmm1
+; CHECK-NEXT:    movq %rbx, %rax
+; CHECK-NEXT:    shll $16, %eax
+; CHECK-NEXT:    movd %eax, %xmm0
+; CHECK-NEXT:    addss %xmm1, %xmm0
 ; CHECK-NEXT:    callq __truncsfbf2@PLT
 ; CHECK-NEXT:    movd %xmm0, %eax
-; CHECK-NEXT:    movzwl %ax, %r14d
-; CHECK-NEXT:    orl %ebp, %r14d
-; CHECK-NEXT:    movd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Folded Reload
-; CHECK-NEXT:    # xmm0 = mem[0],zero,zero,zero
+; CHECK-NEXT:    movzwl %ax, %ebx
+; CHECK-NEXT:    orl %ebp, %ebx
+; CHECK-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
+; CHECK-NEXT:    shll $16, %eax
+; CHECK-NEXT:    movd %eax, %xmm1
+; CHECK-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
+; CHECK-NEXT:    shll $16, %eax
+; CHECK-NEXT:    movd %eax, %xmm0
+; CHECK-NEXT:    addss %xmm1, %xmm0
 ; CHECK-NEXT:    callq __truncsfbf2@PLT
 ; CHECK-NEXT:    movd %xmm0, %ebp
 ; CHECK-NEXT:    shll $16, %ebp
-; CHECK-NEXT:    movd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Folded Reload
-; CHECK-NEXT:    # xmm0 = mem[0],zero,zero,zero
+; CHECK-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
+; CHECK-NEXT:    shll $16, %eax
+; CHECK-NEXT:    movd %eax, %xmm1
+; CHECK-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
+; CHECK-NEXT:    shll $16, %eax
+; CHECK-NEXT:    movd %eax, %xmm0
+; CHECK-NEXT:    addss %xmm1, %xmm0
 ; CHECK-NEXT:    callq __truncsfbf2@PLT
 ; CHECK-NEXT:    movd %xmm0, %eax
 ; CHECK-NEXT:    movzwl %ax, %eax
 ; CHECK-NEXT:    orl %ebp, %eax
 ; CHECK-NEXT:    shlq $32, %rax
-; CHECK-NEXT:    orq %r14, %rax
+; CHECK-NEXT:    orq %rbx, %rax
 ; CHECK-NEXT:    movq %rax, %xmm0
-; CHECK-NEXT:    movq %rbx, %xmm1
+; CHECK-NEXT:    movq %r14, %xmm1
 ; CHECK-NEXT:    punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; CHECK-NEXT:    addq $32, %rsp
+; CHECK-NEXT:    addq $56, %rsp
 ; CHECK-NEXT:    popq %rbx
+; CHECK-NEXT:    popq %r12
+; CHECK-NEXT:    popq %r13
 ; CHECK-NEXT:    popq %r14
+; CHECK-NEXT:    popq %r15
 ; CHECK-NEXT:    popq %rbp
 ; CHECK-NEXT:    retq
   %add = fadd <8 x bfloat> %a, %b