[PATCH] D156100: [AMDGPU] Have a command-line flag to enable true True16 codegen.

[Ivan Kosarev via Phabricator via llvm-commits]

kosarev updated this revision to Diff 543651.
kosarev added a comment.

Update the commit message.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D156100/new/

https://reviews.llvm.org/D156100

Files:
  llvm/lib/Target/AMDGPU/AMDGPU.td
  llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
  llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
  llvm/lib/Target/AMDGPU/SIISelLowering.cpp


Index: llvm/lib/Target/AMDGPU/SIISelLowering.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -146,8 +146,13 @@
   addRegisterClass(MVT::v16f64, TRI->getVGPRClassForBitWidth(1024));
 
   if (Subtarget->has16BitInsts()) {
-    addRegisterClass(MVT::i16, &AMDGPU::SReg_32RegClass);
-    addRegisterClass(MVT::f16, &AMDGPU::SReg_32RegClass);
+    if (Subtarget->useTrue16BitInsts()) {
+      addRegisterClass(MVT::i16, &AMDGPU::VGPR_16RegClass);
+      addRegisterClass(MVT::f16, &AMDGPU::VGPR_16RegClass);
+    } else {
+      addRegisterClass(MVT::i16, &AMDGPU::SReg_32RegClass);
+      addRegisterClass(MVT::f16, &AMDGPU::SReg_32RegClass);
+    }
 
     // Unless there are also VOP3P operations, not operations are really legal.
     addRegisterClass(MVT::v2i16, &AMDGPU::SReg_32RegClass);
Index: llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
===================================================================
--- llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
+++ llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
@@ -49,6 +49,7 @@
   bool GCN3Encoding = false;
   bool Has16BitInsts = false;
   bool HasTrue16BitInsts = false;
+  bool EnableTrue16Codegen = false;
   bool HasMadMixInsts = false;
   bool HasMadMacF32Insts = false;
   bool HasDsSrc2Insts = false;
@@ -155,6 +156,10 @@
 
   bool hasTrue16BitInsts() const { return HasTrue16BitInsts; }
 
+  // TODO: Remove and use hasTrue16BitInsts() instead once True16 is
+  // fully supported.
+  bool useTrue16BitInsts() const;
+
   bool hasMadMixInsts() const {
     return HasMadMixInsts;
   }
Index: llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
+++ llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
@@ -166,6 +166,10 @@
 
 AMDGPUSubtarget::AMDGPUSubtarget(const Triple &TT) : TargetTriple(TT) {}
 
+bool AMDGPUSubtarget::useTrue16BitInsts() const {
+  return hasTrue16BitInsts() && EnableTrue16Codegen;
+}
+
 GCNSubtarget::GCNSubtarget(const Triple &TT, StringRef GPU, StringRef FS,
                            const GCNTargetMachine &TM)
     : // clang-format off
Index: llvm/lib/Target/AMDGPU/AMDGPU.td
===================================================================
--- llvm/lib/Target/AMDGPU/AMDGPU.td
+++ llvm/lib/Target/AMDGPU/AMDGPU.td
@@ -375,6 +375,12 @@
   "True 16-bit operand instructions"
 >;
 
+def FeatureTrue16Codegen : SubtargetFeature<"true16-codegen",
+  "EnableTrue16Codegen",
+  "true",
+  "Use true 16-bit registers"
+>;
+
 def FeatureVOP3P : SubtargetFeature<"vop3p",
   "HasVOP3PInsts",
   "true",


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D156100.543651.patch
Type: text/x-patch
Size: 2690 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230724/f380e51c/attachment.bin>