R600/SI: Allow commuting with src2_modifiers

llvm-svn: 221911

diff --git a/llvm/lib/Target/R600/SIInstrInfo.cpp b/llvm/lib/Target/R600/SIInstrInfo.cpp
index 4de885c..1c15eb8 100644 (file)
--- a/llvm/lib/Target/R600/SIInstrInfo.cpp
+++ b/llvm/lib/Target/R600/SIInstrInfo.cpp
@@ -718,11 +718,6 @@ MachineInstr *SIInstrInfo::commuteInstruction(MachineInstr *MI,
       return nullptr;
     }
 
-    // TODO: Is there any reason to commute with src2 modifiers?
-    // TODO: Should be able to commute with output modifiers just fine.
-    if (hasModifiersSet(*MI, AMDGPU::OpName::src2_modifiers))
-      return nullptr;
-
     // Be sure to copy the source modifiers to the right place.
     if (MachineOperand *Src0Mods
           = getNamedOperand(*MI, AMDGPU::OpName::src0_modifiers)) {

diff --git a/llvm/test/CodeGen/R600/commute_modifiers.ll b/llvm/test/CodeGen/R600/commute_modifiers.ll
index f2885a7..30c8067 100644 (file)
--- a/llvm/test/CodeGen/R600/commute_modifiers.ll
+++ b/llvm/test/CodeGen/R600/commute_modifiers.ll
@@ -2,6 +2,7 @@
 
 declare i32 @llvm.r600.read.tidig.x() #1
 declare float @llvm.fabs.f32(float) #1
+declare float @llvm.fma.f32(float, float, float) nounwind readnone
 
 ; FUNC-LABEL: @commute_add_imm_fabs_f32
 ; SI: buffer_load_dword [[X:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
@@ -152,5 +153,29 @@ define void @commute_mul_fabs_x_fneg_fabs_y_f32(float addrspace(1)* %out, float
   ret void
 }
 
+; Make sure we commute the multiply part for the constant in src0 even
+; though we have negate modifier on src2.
+
+; SI-LABEL: {{^}}fma_a_2.0_neg_b_f32
+; SI-DAG: buffer_load_dword [[R1:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI-DAG: buffer_load_dword [[R2:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:0x4
+; SI: v_fma_f32 [[RESULT:v[0-9]+]], 2.0, [[R1]], |[[R2]]|
+; SI: buffer_store_dword [[RESULT]]
+define void @fma_a_2.0_neg_b_f32(float addrspace(1)* %out, float addrspace(1)* %in) {
+  %tid = call i32 @llvm.r600.read.tidig.x() nounwind readnone
+  %gep.0 = getelementptr float addrspace(1)* %out, i32 %tid
+  %gep.1 = getelementptr float addrspace(1)* %gep.0, i32 1
+  %gep.out = getelementptr float addrspace(1)* %out, i32 %tid
+
+  %r1 = load float addrspace(1)* %gep.0
+  %r2 = load float addrspace(1)* %gep.1
+
+  %r2.fabs = call float @llvm.fabs.f32(float %r2)
+
+  %r3 = tail call float @llvm.fma.f32(float %r1, float 2.0, float %r2.fabs)
+  store float %r3, float addrspace(1)* %gep.out
+  ret void
+}
+
 attributes #0 = { nounwind }
 attributes #1 = { nounwind readnone }