[AMDGPU] gfx10 v_fmac_f16 operand folding

[rampitec]

Summary

Fold immediates into v_fmac_f16.

[arsenm]

LGTM

[kzhuravl]

LGTM

llvm/trunk/lib/Target/AMDGPU/SIFoldOperands.cpp

static bool isInlineConstantIfFolded(const SIInstrInfo *TII,
                                      const MachineOperand &OpToFold) {
   if (TII->isInlineConstant(UseMI, OpNo, OpToFold))
     return true;
 
   unsigned Opc = UseMI.getOpcode();
   switch (Opc) {
   case AMDGPU::V_MAC_F32_e64:
   case AMDGPU::V_MAC_F16_e64:
-  case AMDGPU::V_FMAC_F32_e64: {
+  case AMDGPU::V_FMAC_F32_e64:
+  case AMDGPU::V_FMAC_F16_e64: {
     // Special case for mac. Since this is replaced with mad when folded into
     // src2, we need to check the legality for the final instruction.
     int Src2Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2);
     if (static_cast<int>(OpNo) == Src2Idx) {
-      bool IsFMA = Opc == AMDGPU::V_FMAC_F32_e64;
-      bool IsF32 = Opc == AMDGPU::V_MAC_F32_e64;
+      bool IsFMA = Opc == AMDGPU::V_FMAC_F32_e64 ||
+                   Opc == AMDGPU::V_FMAC_F16_e64;
+      bool IsF32 = Opc == AMDGPU::V_MAC_F32_e64 ||
+                   Opc == AMDGPU::V_FMAC_F32_e64;
 
       unsigned Opc = IsFMA ?
-        AMDGPU::V_FMA_F32 : (IsF32 ? AMDGPU::V_MAD_F32 : AMDGPU::V_MAD_F16);
+        (IsF32 ? AMDGPU::V_FMA_F32 : AMDGPU::V_FMA_F16_gfx9) :
+        (IsF32 ? AMDGPU::V_MAD_F32 : AMDGPU::V_MAD_F16);
       const MCInstrDesc &MadDesc = TII->get(Opc);
       return TII->isInlineConstant(OpToFold, MadDesc.OpInfo[OpNo].OperandType);
     }
     return false;
   }
   default:
     return false;
   }

 static bool tryAddToFoldList(SmallVectorImpl<FoldCandidate> &FoldList,
                              MachineInstr *MI, unsigned OpNo,
                              MachineOperand *OpToFold,
                              const SIInstrInfo *TII) {
   if (!TII->isOperandLegal(*MI, OpNo, OpToFold)) {
     // Special case for v_mac_{f16, f32}_e64 if we are trying to fold into src2
     unsigned Opc = MI->getOpcode();
     if ((Opc == AMDGPU::V_MAC_F32_e64 || Opc == AMDGPU::V_MAC_F16_e64 ||
-         Opc == AMDGPU::V_FMAC_F32_e64) &&
+         Opc == AMDGPU::V_FMAC_F32_e64 || Opc == AMDGPU::V_FMAC_F16_e64) &&
         (int)OpNo == AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2)) {
-      bool IsFMA = Opc == AMDGPU::V_FMAC_F32_e64;
-      bool IsF32 = Opc == AMDGPU::V_MAC_F32_e64;
+      bool IsFMA = Opc == AMDGPU::V_FMAC_F32_e64 ||
+                   Opc == AMDGPU::V_FMAC_F16_e64;
+      bool IsF32 = Opc == AMDGPU::V_MAC_F32_e64 ||
+                   Opc == AMDGPU::V_FMAC_F32_e64;
       unsigned NewOpc = IsFMA ?
-        AMDGPU::V_FMA_F32 : (IsF32 ? AMDGPU::V_MAD_F32 : AMDGPU::V_MAD_F16);
+        (IsF32 ? AMDGPU::V_FMA_F32 : AMDGPU::V_FMA_F16_gfx9) :
+        (IsF32 ? AMDGPU::V_MAD_F32 : AMDGPU::V_MAD_F16);
 
       // Check if changing this to a v_mad_{f16, f32} instruction will allow us
       // to fold the operand.
       MI->setDesc(TII->get(NewOpc));
       bool FoldAsMAD = tryAddToFoldList(FoldList, MI, OpNo, OpToFold, TII);
       if (FoldAsMAD) {
         MI->untieRegOperand(OpNo);
         return true;

llvm/trunk/test/CodeGen/AMDGPU/fmul-2-combine-multi-use.ll

 ; GFX8_10:      v_add_f16_e64 v{{[0-9]+}}, s{{[0-9]+}}, -1.0
 ; GFX8_10:      v_add_f16_e64 v{{[0-9]+}}, s{{[0-9]+}}, -1.0
 ; GFX8_10:      v_cmp_gt_f16_e64 {{vcc|vcc_lo}}, |v{{[0-9]+}}|, |v{{[0-9]+}}|
 ; GFX8_10:      v_cndmask_b32_e32
 ; GFX8_10:      v_add_f16_e64 v{{[0-9]+}}, |v{{[0-9]+}}|, |v{{[0-9]+}}|
 ; GFX8_10:      v_mul_f16_e32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
 ; VI-FLUSH:     v_mad_f16 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, 1.0
 ; VI-DENORM:    v_fma_f16 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, 1.0
-; GFX10-DENORM: v_fmac_f16_e64 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}
+; GFX10-DENORM: v_fma_f16 v{{[0-9]+}}, -v{{[0-9]+}}, v{{[0-9]+}}, 1.0
 ; GFX10-FLUSH:  v_sub_f16_e32 v{{[0-9]+}}, 1.0, v{{[0-9]+}}
 define amdgpu_kernel void @multiple_fadd_use_test_f16(half addrspace(1)* %out, i16 zeroext %x.arg, i16 zeroext %y.arg, i16 zeroext %z.arg) #0 {
   %x = bitcast i16 %x.arg to half
   %y = bitcast i16 %y.arg to half
   %z = bitcast i16 %z.arg to half
   %a11 = fadd half %y, -1.0
   %a12 = call half @llvm.fabs.f16(half %a11)
   %a13 = fadd half %x, -1.0

 }
 
 ; GCN-LABEL: {{^}}multiple_use_fadd_fmac_f16:
 ; GCN-DAG: v_add_f16_e64 [[MUL2:v[0-9]+]], [[X:s[0-9]+]], s{{[0-9]+}}
 
 ; VI-FLUSH-DAG:     v_mac_f16_e64 [[MAD:v[0-9]+]], [[X]], 2.0
 ; VI-DENORM-DAG:    v_fma_f16 [[MAD:v[0-9]+]], [[X]], 2.0, v{{[0-9]+}}
 ; GFX10-FLUSH-DAG:  v_add_f16_e32 [[MAD:v[0-9]+]], s{{[0-9]+}}, [[MUL2]]
-; GFX10-DENORM-DAG: v_fmac_f16_e64 [[MAD:v[0-9]+]], [[X]], 2.0
+; GFX10-DENORM-DAG: v_fma_f16 [[MAD:v[0-9]+]], [[X]], 2.0, s{{[0-9]+}}
 
 ; GCN-DAG: buffer_store_short [[MUL2]]
 ; GCN-DAG: buffer_store_short [[MAD]]
 ; GCN: s_endpgm
 define amdgpu_kernel void @multiple_use_fadd_fmac_f16(half addrspace(1)* %out, i16 zeroext %x.arg, i16 zeroext %y.arg) #0 {
   %x = bitcast i16 %x.arg to half
   %y = bitcast i16 %y.arg to half
   %out.gep.1 = getelementptr half, half addrspace(1)* %out, i32 1
   %mul2 = fmul fast half %x, 2.0
   %mad = fadd fast half %mul2, %y
   store volatile half %mul2, half addrspace(1)* %out
   store volatile half %mad, half addrspace(1)* %out.gep.1
   ret void
 }
 
 ; GCN-LABEL: {{^}}multiple_use_fadd_fmad_f16:
 ; GCN-DAG: v_add_f16_e64 [[MUL2:v[0-9]+]], |[[X:s[0-9]+]]|, |s{{[0-9]+}}|
 
 ; VI-FLUSH-DAG:     v_mad_f16 [[MAD:v[0-9]+]], |[[X]]|, 2.0, v{{[0-9]+}}
 ; VI-DENORM-DAG:    v_fma_f16 [[MAD:v[0-9]+]], |[[X]]|, 2.0, v{{[0-9]+}}
 ; GFX10-FLUSH-DAG:  v_add_f16_e32 [[MAD:v[0-9]+]], s{{[0-9]+}}, [[MUL2]]
-; GFX10-DENORM-DAG: v_fmac_f16_e64 [[MAD:v[0-9]+]], |[[X]]|, 2.0
+; GFX10-DENORM-DAG: v_fma_f16 [[MAD:v[0-9]+]], |[[X]]|, 2.0, s{{[0-9]+}}
 
 ; GCN-DAG: buffer_store_short [[MUL2]]
 ; GCN-DAG: buffer_store_short [[MAD]]
 ; GCN: s_endpgm
 define amdgpu_kernel void @multiple_use_fadd_fmad_f16(half addrspace(1)* %out, i16 zeroext %x.arg, i16 zeroext %y.arg) #0 {
   %x = bitcast i16 %x.arg to half
   %y = bitcast i16 %y.arg to half
   %out.gep.1 = getelementptr half, half addrspace(1)* %out, i32 1

 ; VI-FLUSH: v_mad_f16 {{v[0-9]+}}, |[[X]]|, 2.0, v{{[0-9]+}}
 
 ; VI-DENORM: v_fma_f16 {{v[0-9]+}}, |[[X:s[0-9]+]]|, 2.0, v{{[0-9]+}}
 ; VI-DENORM: v_fma_f16 {{v[0-9]+}}, |[[X]]|, 2.0, v{{[0-9]+}}
 
 ; GFX10-FLUSH:  v_add_f16_e64 [[MUL2:v[0-9]+]], |[[X:s[0-9]+]]|, |{{s[0-9]+}}|
 ; GFX10-FLUSH:  v_add_f16_e32 {{v[0-9]+}}, {{s[0-9]+}}, [[MUL2]]
 ; GFX10-FLUSH:  v_add_f16_e32 {{v[0-9]+}}, {{s[0-9]+}}, [[MUL2]]
-; GFX10-DENORM: v_fmac_f16_e64 {{v[0-9]+}}, |[[X:s[0-9]+]]|, 2.0
-; GFX10-DENORM: v_fmac_f16_e64 {{v[0-9]+}}, |[[X]]|, 2.0
+; GFX10-DENORM: v_fma_f16 {{v[0-9]+}}, |[[X:s[0-9]+]]|, 2.0, s{{[0-9]+}}
+; GFX10-DENORM: v_fma_f16 {{v[0-9]+}}, |[[X]]|, 2.0, s{{[0-9]+}}
 
 define amdgpu_kernel void @multiple_use_fadd_multi_fmad_f16(half addrspace(1)* %out, i16 zeroext %x.arg, i16 zeroext %y.arg, i16 zeroext %z.arg) #0 {
   %x = bitcast i16 %x.arg to half
   %y = bitcast i16 %y.arg to half
   %z = bitcast i16 %z.arg to half
   %out.gep.1 = getelementptr half, half addrspace(1)* %out, i32 1
   %x.abs = call half @llvm.fabs.f16(half %x)
   %mul2 = fmul fast half %x.abs, 2.0