[PATCH] D80652: AMDGPU: Make S_DENORM_MODE not be a scheduling boundary

[Matt Arsenault via Phabricator via llvm-commits]

arsenm created this revision.
arsenm added reviewers: rampitec, kerbowa.
Herald added subscribers: hiraditya, t-tye, tpr, dstuttard, yaxunl, nhaehnle, wdng, jvesely, kzhuravl.
Herald added a project: LLVM.

Now that the mode register uses/defs should be properly modeled, we
don't need to treat the FP mode switch as an arbitrary side effect.


https://reviews.llvm.org/D80652

Files:
  llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
  llvm/lib/Target/AMDGPU/SIInstrInfo.td
  llvm/lib/Target/AMDGPU/SOPInstructions.td


Index: llvm/lib/Target/AMDGPU/SOPInstructions.td
===================================================================
--- llvm/lib/Target/AMDGPU/SOPInstructions.td
+++ llvm/lib/Target/AMDGPU/SOPInstructions.td
@@ -1218,8 +1218,7 @@
   def S_WAITCNT_DEPCTR :
     SOPP <0x023, (ins s16imm:$simm16), "s_waitcnt_depctr $simm16">;
 
-  let hasSideEffects = 1, Uses = [MODE], Defs = [MODE] in {
-    // FIXME: Should remove hasSideEffects
+  let hasSideEffects = 0, Uses = [MODE], Defs = [MODE] in {
     def S_ROUND_MODE :
       SOPP<0x024, (ins s16imm:$simm16), "s_round_mode $simm16">;
     def S_DENORM_MODE :
Index: llvm/lib/Target/AMDGPU/SIInstrInfo.td
===================================================================
--- llvm/lib/Target/AMDGPU/SIInstrInfo.td
+++ llvm/lib/Target/AMDGPU/SIInstrInfo.td
@@ -264,7 +264,7 @@
 
 def SIdenorm_mode : SDNode<"AMDGPUISD::DENORM_MODE",
   SDTypeProfile<0 ,1, [SDTCisInt<0>]>,
-  [SDNPHasChain, SDNPSideEffect, SDNPOptInGlue, SDNPOutGlue]
+  [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]
 >;
 
 //===----------------------------------------------------------------------===//
Index: llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -2950,10 +2950,12 @@
   // Target-independent instructions do not have an implicit-use of EXEC, even
   // when they operate on VGPRs. Treating EXEC modifications as scheduling
   // boundaries prevents incorrect movements of such instructions.
+
+  // TODO: Don't treat setreg with known constant that only changes MODE as
+  // barrier.
   return MI.modifiesRegister(AMDGPU::EXEC, &RI) ||
          MI.getOpcode() == AMDGPU::S_SETREG_IMM32_B32 ||
          MI.getOpcode() == AMDGPU::S_SETREG_B32 ||
-         MI.getOpcode() == AMDGPU::S_DENORM_MODE ||
          changesVGPRIndexingMode(MI);
 }
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D80652.266619.patch
Type: text/x-patch
Size: 1919 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200527/04247e47/attachment-0001.bin>