[PATCH] D68037: [AMDGPU] gfx10 v_fmac_f16 operand folding

[Stanislav Mekhanoshin via Phabricator via llvm-commits]

rampitec created this revision.
rampitec added reviewers: kzhuravl, arsenm.
Herald added subscribers: hiraditya, t-tye, tpr, dstuttard, yaxunl, nhaehnle, wdng, jvesely.
Herald added a project: LLVM.

Fold immediates into v_fmac_f16.


https://reviews.llvm.org/D68037

Files:
  llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
  llvm/test/CodeGen/AMDGPU/fmul-2-combine-multi-use.ll


Index: llvm/test/CodeGen/AMDGPU/fmul-2-combine-multi-use.ll
===================================================================
--- llvm/test/CodeGen/AMDGPU/fmul-2-combine-multi-use.ll
+++ llvm/test/CodeGen/AMDGPU/fmul-2-combine-multi-use.ll
@@ -126,7 +126,7 @@
 ; 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
@@ -152,7 +152,7 @@
 ; 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]]
@@ -174,7 +174,7 @@
 ; 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]]
@@ -201,8 +201,8 @@
 ; 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
Index: llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
+++ llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
@@ -142,16 +142,20 @@
   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);
     }
@@ -314,12 +318,15 @@
     // 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.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D68037.221803.patch
Type: text/x-patch
Size: 4878 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20190925/820b3bae/attachment-0001.bin>