From fd31a769cef4db0127128386934296da09c5e517 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Sat, 14 Feb 2015 04:24:28 +0000
Subject: [PATCH] R600/SI: Use complex operand folding for div_scale

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229238 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/R600/AMDGPUISelDAGToDAG.cpp     | 19 ++----
 test/CodeGen/R600/llvm.AMDGPU.div_scale.ll | 77 ++++++++++++++++++++++
 2 files changed, 84 insertions(+), 12 deletions(-)

diff --git a/lib/Target/R600/AMDGPUISelDAGToDAG.cpp b/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
index ac05e4ae2ce..b5ab70394f9 100644
--- a/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
+++ b/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
@@ -783,6 +783,8 @@ SDNode *AMDGPUDAGToDAGISel::SelectADD_SUB_I64(SDNode *N) {
   return CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, MVT::i64, Args);
 }
 
+// We need to handle this here because tablegen doesn't support matching
+// instructions with multiple outputs.
 SDNode *AMDGPUDAGToDAGISel::SelectDIV_SCALE(SDNode *N) {
   SDLoc SL(N);
   EVT VT = N->getValueType(0);
@@ -792,19 +794,12 @@ SDNode *AMDGPUDAGToDAGISel::SelectDIV_SCALE(SDNode *N) {
   unsigned Opc
     = (VT == MVT::f64) ? AMDGPU::V_DIV_SCALE_F64 : AMDGPU::V_DIV_SCALE_F32;
 
-  const SDValue Zero = CurDAG->getTargetConstant(0, MVT::i32);
-  const SDValue False = CurDAG->getTargetConstant(0, MVT::i1);
-  SDValue Ops[] = {
-    Zero,             // src0_modifiers
-    N->getOperand(0), // src0
-    Zero,             // src1_modifiers
-    N->getOperand(1), // src1
-    Zero,             // src2_modifiers
-    N->getOperand(2), // src2
-    False,            // clamp
-    Zero              // omod
-  };
+  // src0_modifiers, src0, src1_modifiers, src1, src2_modifiers, src2, clamp, omod
+  SDValue Ops[8];
 
+  SelectVOP3Mods0(N->getOperand(0), Ops[1], Ops[0], Ops[6], Ops[7]);
+  SelectVOP3Mods(N->getOperand(1), Ops[3], Ops[2]);
+  SelectVOP3Mods(N->getOperand(2), Ops[5], Ops[4]);
   return CurDAG->SelectNodeTo(N, Opc, VT, MVT::i1, Ops);
 }
 
diff --git a/test/CodeGen/R600/llvm.AMDGPU.div_scale.ll b/test/CodeGen/R600/llvm.AMDGPU.div_scale.ll
index 32e6eaa1a7a..5773da0bb2e 100644
--- a/test/CodeGen/R600/llvm.AMDGPU.div_scale.ll
+++ b/test/CodeGen/R600/llvm.AMDGPU.div_scale.ll
@@ -3,6 +3,7 @@
 declare i32 @llvm.r600.read.tidig.x() nounwind readnone
 declare { float, i1 } @llvm.AMDGPU.div.scale.f32(float, float, i1) nounwind readnone
 declare { double, i1 } @llvm.AMDGPU.div.scale.f64(double, double, i1) nounwind readnone
+declare float @llvm.fabs.f32(float) nounwind readnone
 
 ; SI-LABEL @test_div_scale_f32_1:
 ; SI-DAG: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64
@@ -285,3 +286,79 @@ define void @test_div_scale_f64_all_scalar_2(double addrspace(1)* %out, double %
   store double %result0, double addrspace(1)* %out, align 8
   ret void
 }
+
+; SI-LABEL @test_div_scale_f32_inline_imm_num:
+; SI-DAG: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI: v_div_scale_f32 [[RESULT0:v[0-9]+]], [[RESULT1:s\[[0-9]+:[0-9]+\]]], [[A]], [[A]], 1.0
+; SI: buffer_store_dword [[RESULT0]]
+; SI: s_endpgm
+define void @test_div_scale_f32_inline_imm_num(float addrspace(1)* %out, float addrspace(1)* %in) nounwind {
+  %tid = call i32 @llvm.r600.read.tidig.x() nounwind readnone
+  %gep.0 = getelementptr float addrspace(1)* %in, i32 %tid
+  %a = load float addrspace(1)* %gep.0, align 4
+
+  %result = call { float, i1 } @llvm.AMDGPU.div.scale.f32(float 1.0, float %a, i1 false) nounwind readnone
+  %result0 = extractvalue { float, i1 } %result, 0
+  store float %result0, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL @test_div_scale_f32_inline_imm_den:
+; SI-DAG: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
+; SI: v_div_scale_f32 [[RESULT0:v[0-9]+]], [[RESULT1:s\[[0-9]+:[0-9]+\]]], 2.0, 2.0, [[A]]
+; SI: buffer_store_dword [[RESULT0]]
+; SI: s_endpgm
+define void @test_div_scale_f32_inline_imm_den(float addrspace(1)* %out, float addrspace(1)* %in) nounwind {
+  %tid = call i32 @llvm.r600.read.tidig.x() nounwind readnone
+  %gep.0 = getelementptr float addrspace(1)* %in, i32 %tid
+  %a = load float addrspace(1)* %gep.0, align 4
+
+  %result = call { float, i1 } @llvm.AMDGPU.div.scale.f32(float %a, float 2.0, i1 false) nounwind readnone
+  %result0 = extractvalue { float, i1 } %result, 0
+  store float %result0, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL @test_div_scale_f32_fabs_num:
+; SI-DAG: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64
+; SI-DAG: buffer_load_dword [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
+; SI: v_div_scale_f32 [[RESULT0:v[0-9]+]], [[RESULT1:s\[[0-9]+:[0-9]+\]]], [[B]], [[B]], |[[A]]|
+; SI: buffer_store_dword [[RESULT0]]
+; SI: s_endpgm
+define void @test_div_scale_f32_fabs_num(float addrspace(1)* %out, float addrspace(1)* %in) nounwind {
+  %tid = call i32 @llvm.r600.read.tidig.x() nounwind readnone
+  %gep.0 = getelementptr float addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr float addrspace(1)* %gep.0, i32 1
+
+  %a = load float addrspace(1)* %gep.0, align 4
+  %b = load float addrspace(1)* %gep.1, align 4
+
+  %a.fabs = call float @llvm.fabs.f32(float %a) nounwind readnone
+
+  %result = call { float, i1 } @llvm.AMDGPU.div.scale.f32(float %a.fabs, float %b, i1 false) nounwind readnone
+  %result0 = extractvalue { float, i1 } %result, 0
+  store float %result0, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL @test_div_scale_f32_fabs_den:
+; SI-DAG: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64
+; SI-DAG: buffer_load_dword [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
+; SI: v_div_scale_f32 [[RESULT0:v[0-9]+]], [[RESULT1:s\[[0-9]+:[0-9]+\]]], |[[B]]|, |[[B]]|, [[A]]
+; SI: buffer_store_dword [[RESULT0]]
+; SI: s_endpgm
+define void @test_div_scale_f32_fabs_den(float addrspace(1)* %out, float addrspace(1)* %in) nounwind {
+  %tid = call i32 @llvm.r600.read.tidig.x() nounwind readnone
+  %gep.0 = getelementptr float addrspace(1)* %in, i32 %tid
+  %gep.1 = getelementptr float addrspace(1)* %gep.0, i32 1
+
+  %a = load float addrspace(1)* %gep.0, align 4
+  %b = load float addrspace(1)* %gep.1, align 4
+
+  %b.fabs = call float @llvm.fabs.f32(float %b) nounwind readnone
+
+  %result = call { float, i1 } @llvm.AMDGPU.div.scale.f32(float %a, float %b.fabs, i1 false) nounwind readnone
+  %result0 = extractvalue { float, i1 } %result, 0
+  store float %result0, float addrspace(1)* %out, align 4
+  ret void
+}
-- 
2.34.1