From: Tom Stellard <thomas.stellard@amd.com>
Date: Fri, 1 Aug 2014 00:32:28 +0000 (+0000)
Subject: R600/SI: Fix incorrect commute operation in shrink instructions pass
X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=commitdiff_plain;h=e5fc4208eba3a3ca3489d6bdbdcfff9c45b1e5ab

R600/SI: Fix incorrect commute operation in shrink instructions pass

We were commuting the instruction by still shrinking it using the
original opcode.

NOTE: This is a candidate for the 3.5 branch.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@214463 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Target/R600/SIInstrInfo.cpp b/lib/Target/R600/SIInstrInfo.cpp
index 0249ad07208..fc395ee4919 100644
--- a/lib/Target/R600/SIInstrInfo.cpp
+++ b/lib/Target/R600/SIInstrInfo.cpp
@@ -668,6 +668,10 @@ bool SIInstrInfo::isImmOperandLegal(const MachineInstr *MI, unsigned OpNo,
   return RI.regClassCanUseImmediate(OpInfo.RegClass);
 }
 
+bool SIInstrInfo::hasVALU32BitEncoding(unsigned Opcode) const {
+  return AMDGPU::getVOPe32(Opcode) != -1;
+}
+
 bool SIInstrInfo::verifyInstruction(const MachineInstr *MI,
                                     StringRef &ErrInfo) const {
   uint16_t Opcode = MI->getOpcode();
diff --git a/lib/Target/R600/SIInstrInfo.h b/lib/Target/R600/SIInstrInfo.h
index c53298add02..48d6ca5fcf3 100644
--- a/lib/Target/R600/SIInstrInfo.h
+++ b/lib/Target/R600/SIInstrInfo.h
@@ -115,6 +115,10 @@ public:
   bool isImmOperandLegal(const MachineInstr *MI, unsigned OpNo,
                          const MachineOperand &MO) const;
 
+  /// \brief Return true if this 64-bit VALU instruction has a 32-bit encoding.
+  /// This function will return false if you pass it a 32-bit instruction.
+  bool hasVALU32BitEncoding(unsigned Opcode) const;
+
   bool verifyInstruction(const MachineInstr *MI,
                          StringRef &ErrInfo) const override;
 
diff --git a/lib/Target/R600/SIShrinkInstructions.cpp b/lib/Target/R600/SIShrinkInstructions.cpp
index 362a5c1e4e0..745c4b65644 100644
--- a/lib/Target/R600/SIShrinkInstructions.cpp
+++ b/lib/Target/R600/SIShrinkInstructions.cpp
@@ -125,9 +125,7 @@ bool SIShrinkInstructions::runOnMachineFunction(MachineFunction &MF) {
       Next = std::next(I);
       MachineInstr &MI = *I;
 
-      int Op32 = AMDGPU::getVOPe32(MI.getOpcode());
-
-      if (Op32 == -1)
+      if (!TII->hasVALU32BitEncoding(MI.getOpcode()))
         continue;
 
       if (!canShrink(MI, TII, TRI, MRI)) {
@@ -138,6 +136,13 @@ bool SIShrinkInstructions::runOnMachineFunction(MachineFunction &MF) {
           continue;
       }
 
+      int Op32 = AMDGPU::getVOPe32(MI.getOpcode());
+
+      // Op32 could be -1 here if we started with an instruction that had a
+      // a 32-bit encoding and then commuted it to an instruction that did not.
+      if (Op32 == -1)
+        continue;
+
       if (TII->isVOPC(Op32)) {
         unsigned DstReg = MI.getOperand(0).getReg();
         if (TargetRegisterInfo::isVirtualRegister(DstReg)) {
diff --git a/test/CodeGen/R600/vop-shrink.ll b/test/CodeGen/R600/vop-shrink.ll
new file mode 100644
index 00000000000..54e588d8084
--- /dev/null
+++ b/test/CodeGen/R600/vop-shrink.ll
@@ -0,0 +1,41 @@
+; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs< %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
+; XXX: This testis for a bug in the SIShrinkInstruction pass and it will be
+;       relevant once we are selecting 64-bit instructions.  We are
+;       currently selecting mostly 32-bit instruction, so the
+;       SIShrinkInstructions pass isn't doing much.
+; XFAIL: *
+
+; Test that we correctly commute a sub instruction
+; FUNC-LABEL: @sub_rev
+; SI-NOT: V_SUB_I32_e32 v{{[0-9]+}}, s
+; SI: V_SUBREV_I32_e32 v{{[0-9]+}}, s
+
+; ModuleID = 'vop-shrink.ll'
+
+define void @sub_rev(i32 addrspace(1)* %out, <4 x i32> %sgpr, i32 %cond) {
+entry:
+  %vgpr = call i32 @llvm.r600.read.tidig.x() #1
+  %tmp = icmp eq i32 %cond, 0
+  br i1 %tmp, label %if, label %else
+
+if:                                               ; preds = %entry
+  %tmp1 = getelementptr i32 addrspace(1)* %out, i32 1
+  %tmp2 = extractelement <4 x i32> %sgpr, i32 1
+  store i32 %tmp2, i32 addrspace(1)* %out
+  br label %endif
+
+else:                                             ; preds = %entry
+  %tmp3 = extractelement <4 x i32> %sgpr, i32 2
+  %tmp4 = sub i32 %vgpr, %tmp3
+  store i32 %tmp4, i32 addrspace(1)* %out
+  br label %endif
+
+endif:                                            ; preds = %else, %if
+  ret void
+}
+
+; Function Attrs: nounwind readnone
+declare i32 @llvm.r600.read.tidig.x() #0
+
+attributes #0 = { nounwind readnone }
+attributes #1 = { readnone }