From 8a70e28114f50b121c122a20a1cbafaa7b7e790d Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Fri, 26 Sep 2014 17:54:52 +0000
Subject: [PATCH] R600/SI: Don't move operands that are required to be SGPRs

e.g. v_cndmask_b32 requires the condition operand be an SGPR.
If one of the source operands were an SGPR, that would be considered
the one SGPR use and the condition operand would be illegally moved.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@218529 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/R600/SIInstrInfo.cpp | 21 ++++++++++++++++-
 test/CodeGen/R600/v_cndmask.ll  | 42 ++++++++++++++++++++++++++-------
 2 files changed, 53 insertions(+), 10 deletions(-)

diff --git a/lib/Target/R600/SIInstrInfo.cpp b/lib/Target/R600/SIInstrInfo.cpp
index bc3f9ddf1df..b2d4a1c2f00 100644
--- a/lib/Target/R600/SIInstrInfo.cpp
+++ b/lib/Target/R600/SIInstrInfo.cpp
@@ -1357,8 +1357,27 @@ void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
   // XXX - Do any VOP3 instructions read VCC?
   // Legalize VOP3
   if (isVOP3(MI->getOpcode())) {
-    int VOP3Idx[3] = {Src0Idx, Src1Idx, Src2Idx};
+    const MCInstrDesc &Desc = get(MI->getOpcode());
     unsigned SGPRReg = AMDGPU::NoRegister;
+
+    int VOP3Idx[3] = { Src0Idx, Src1Idx, Src2Idx };
+
+    // First we need to consider the instruction's operand requirements before
+    // legalizing. Some operands are required to be SGPRs, but we are still
+    // bound by the constant bus requirement to only use one.
+    //
+    // If the operand's class is an SGPR, we can never move it.
+    for (unsigned i = 0; i < 3; ++i) {
+      int Idx = VOP3Idx[i];
+      if (Idx == -1)
+        break;
+
+      if (RI.isSGPRClassID(Desc.OpInfo[Idx].RegClass)) {
+        SGPRReg = MI->getOperand(Idx).getReg();
+        break;
+      }
+    }
+
     for (unsigned i = 0; i < 3; ++i) {
       int Idx = VOP3Idx[i];
       if (Idx == -1)
diff --git a/test/CodeGen/R600/v_cndmask.ll b/test/CodeGen/R600/v_cndmask.ll
index 84087ee78d5..51e3d8b080f 100644
--- a/test/CodeGen/R600/v_cndmask.ll
+++ b/test/CodeGen/R600/v_cndmask.ll
@@ -1,14 +1,38 @@
-; RUN: llc < %s -march=r600 -mcpu=SI -verify-machineinstrs | FileCheck --check-prefix=SI %s
+; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
 
-; SI: @v_cnd_nan
-; SI: V_CNDMASK_B32_e64 v{{[0-9]}},
+declare i32 @llvm.r600.read.tidig.x() #1
+
+; SI-LABEL: @v_cnd_nan_nosgpr
+; SI: V_CNDMASK_B32_e64 v{{[0-9]}}, v{{[0-9]}}, -1, s{{\[[0-9]+:[0-9]+\]}}
 ; SI-DAG: v{{[0-9]}}
 ; All nan values are converted to 0xffffffff
-; SI-DAG: -1
-define void @v_cnd_nan(float addrspace(1)* %out, i32 %c, float %f) {
-entry:
-  %0 = icmp ne i32 %c, 0
-  %1 = select i1 %0, float 0xFFFFFFFFE0000000, float %f
-  store float %1, float addrspace(1)* %out
+; SI: S_ENDPGM
+define void @v_cnd_nan_nosgpr(float addrspace(1)* %out, i32 %c, float addrspace(1)* %fptr) #0 {
+  %idx = call i32 @llvm.r600.read.tidig.x() #1
+  %f.gep = getelementptr float addrspace(1)* %fptr, i32 %idx
+  %f = load float addrspace(1)* %fptr
+  %setcc = icmp ne i32 %c, 0
+  %select = select i1 %setcc, float 0xFFFFFFFFE0000000, float %f
+  store float %select, float addrspace(1)* %out
   ret void
 }
+
+
+; This requires slightly trickier SGPR operand legalization since the
+; single constant bus SGPR usage is the last operand, and it should
+; never be moved.
+
+; SI-LABEL: @v_cnd_nan
+; SI: V_CNDMASK_B32_e64 v{{[0-9]}}, v{{[0-9]}}, -1, s{{\[[0-9]+:[0-9]+\]}}
+; SI-DAG: v{{[0-9]}}
+; All nan values are converted to 0xffffffff
+; SI: S_ENDPGM
+define void @v_cnd_nan(float addrspace(1)* %out, i32 %c, float %f) #0 {
+  %setcc = icmp ne i32 %c, 0
+  %select = select i1 %setcc, float 0xFFFFFFFFE0000000, float %f
+  store float %select, float addrspace(1)* %out
+  ret void
+}
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind readnone }
-- 
2.34.1