From 7d8f1710a314e5bbc580a9e210d7dc6fa1b50510 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Fri, 17 Oct 2014 18:00:48 +0000
Subject: [PATCH] R600/SI: Allow commuting with source modifiers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@220066 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/R600/SIInstrInfo.cpp        | 33 +++++++++++------
 test/CodeGen/R600/commute_modifiers.ll | 50 ++++++++++++++++++++++++++
 2 files changed, 72 insertions(+), 11 deletions(-)
 create mode 100644 test/CodeGen/R600/commute_modifiers.ll

diff --git a/lib/Target/R600/SIInstrInfo.cpp b/lib/Target/R600/SIInstrInfo.cpp
index 691e108442b..ae8a23e6fdb 100644
--- a/lib/Target/R600/SIInstrInfo.cpp
+++ b/lib/Target/R600/SIInstrInfo.cpp
@@ -718,19 +718,30 @@ MachineInstr *SIInstrInfo::commuteInstruction(MachineInstr *MI,
       return nullptr;
     }
 
-    // XXX: Commute VOP3 instructions with abs and neg set .
-    const MachineOperand *Src0Mods = getNamedOperand(*MI,
-                                          AMDGPU::OpName::src0_modifiers);
-    const MachineOperand *Src1Mods = getNamedOperand(*MI,
-                                          AMDGPU::OpName::src1_modifiers);
-    const MachineOperand *Src2Mods = getNamedOperand(*MI,
-                                          AMDGPU::OpName::src2_modifiers);
-
-    if ((Src0Mods && Src0Mods->getImm()) ||
-        (Src1Mods && Src1Mods->getImm()) ||
-        (Src2Mods && Src2Mods->getImm()))
+    // 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)) {
+      MachineOperand *Src1Mods
+        = getNamedOperand(*MI, AMDGPU::OpName::src1_modifiers);
+
+      int Src0ModsVal = Src0Mods->getImm();
+      if (!Src1Mods && Src0ModsVal != 0)
+        return nullptr;
+
+      // XXX - This assert might be a lie. It might be useful to have a neg
+      // modifier with 0.0.
+      int Src1ModsVal = Src1Mods->getImm();
+      assert((Src1ModsVal == 0) && "Not expecting modifiers with immediates");
+
+      Src1Mods->setImm(Src0ModsVal);
+      Src0Mods->setImm(Src1ModsVal);
+    }
+
     unsigned Reg = Src0.getReg();
     unsigned SubReg = Src0.getSubReg();
     if (Src1.isImm())
diff --git a/test/CodeGen/R600/commute_modifiers.ll b/test/CodeGen/R600/commute_modifiers.ll
new file mode 100644
index 00000000000..97038f19003
--- /dev/null
+++ b/test/CodeGen/R600/commute_modifiers.ll
@@ -0,0 +1,50 @@
+; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
+
+declare i32 @llvm.r600.read.tidig.x() #1
+declare float @llvm.fabs.f32(float) #1
+
+; 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{{$}}
+; SI: V_ADD_F32_e64 [[REG:v[0-9]+]], 2.0, |[[X]]|
+; SI-NEXT: BUFFER_STORE_DWORD [[REG]]
+define void @commute_add_imm_fabs_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
+  %tid = call i32 @llvm.r600.read.tidig.x() #1
+  %gep.0 = getelementptr float addrspace(1)* %in, i32 %tid
+  %x = load float addrspace(1)* %gep.0
+  %x.fabs = call float @llvm.fabs.f32(float %x) #1
+  %z = fadd float 2.0, %x.fabs
+  store float %z, float addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: @commute_mul_imm_fneg_fabs_f32
+; SI: BUFFER_LOAD_DWORD [[X:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI: V_MUL_F32_e64 [[REG:v[0-9]+]], -4.0, |[[X]]|
+; SI-NEXT: BUFFER_STORE_DWORD [[REG]]
+define void @commute_mul_imm_fneg_fabs_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
+  %tid = call i32 @llvm.r600.read.tidig.x() #1
+  %gep.0 = getelementptr float addrspace(1)* %in, i32 %tid
+  %x = load float addrspace(1)* %gep.0
+  %x.fabs = call float @llvm.fabs.f32(float %x) #1
+  %x.fneg.fabs = fsub float -0.000000e+00, %x.fabs
+  %z = fmul float 4.0, %x.fneg.fabs
+  store float %z, float addrspace(1)* %out
+  ret void
+}
+
+; FUNC-LABEL: @commute_mul_imm_fneg_f32
+; SI: BUFFER_LOAD_DWORD [[X:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI: V_MUL_F32_e32 [[REG:v[0-9]+]], -4.0, [[X]]
+; SI-NEXT: BUFFER_STORE_DWORD [[REG]]
+define void @commute_mul_imm_fneg_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
+  %tid = call i32 @llvm.r600.read.tidig.x() #1
+  %gep.0 = getelementptr float addrspace(1)* %in, i32 %tid
+  %x = load float addrspace(1)* %gep.0
+  %x.fneg = fsub float -0.000000e+00, %x
+  %z = fmul float 4.0, %x.fneg
+  store float %z, float addrspace(1)* %out
+  ret void
+}
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind readnone }
-- 
2.34.1