[llvm] [AMDGPU] Implement hasBitTest to Optimize Bit Testing Operations (PR #112652)

Harrison Hao via llvm-commits llvm-commits at lists.llvm.org
Wed Oct 16 20:31:19 PDT 2024


https://github.com/harrisonGPU created https://github.com/llvm/llvm-project/pull/112652

>From https://github.com/llvm/llvm-project/issues/112550

This PR implements the `hasBitTest` function for AMDGPU in the `TargetLowering` class, enabling LLVM to effectively recognize and optimize bit test operations for scalar 32-bit and 64-bit integer types (`i32` and `i64`). 

For example:
```bash
%mask = i32 16       ; Binary: 0b00010000, tests the 5th bit
%test = and i32 %X, %mask
%cmp = icmp eq i32 %test, 0
br i1 %cmp, label %if_zero, label %if_nonzero

if_zero:
  ; Execute action if the 5th bit in %X is 0
  ...
if_nonzero:
  ; Execute alternative action if the 5th bit in %X is not 0
  ...
```
This use of `hasBitTest` ensures direct application of `S_BITCMP0_B32`, streamlining the generation of machine code for bit testing in AMDGPU targets.

>From 0354045aeacb1ddbeffe6f57ec2ec2937033b143 Mon Sep 17 00:00:00 2001
From: Harrison Hao <tsworld1314 at gmail.com>
Date: Thu, 17 Oct 2024 11:25:11 +0800
Subject: [PATCH] [AMDGPU] Implement hasBitTest to Optimize Bit Testing
 Operations

---
 llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp | 21 +++++++++++++++++++
 llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h   |  2 ++
 2 files changed, 23 insertions(+)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index 0f65df0763cc83..0f90fc1dfa750a 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -6043,3 +6043,24 @@ bool AMDGPUTargetLowering::isReassocProfitable(MachineRegisterInfo &MRI,
                                                Register N0, Register N1) const {
   return MRI.hasOneNonDBGUse(N0); // FIXME: handle regbanks
 }
+
+bool AMDGPUTargetLowering::hasBitTest(SDValue X, SDValue Y) const {
+  if (X->isDivergent() || Y->isDivergent())
+    return false;
+
+  EVT VT = X.getValueType();
+
+  if (VT != MVT::i32 && VT != MVT::i64)
+    return false;
+
+  auto *ConstantMaskNode = dyn_cast<ConstantSDNode>(Y);
+  if (!ConstantMaskNode)
+    return false;
+
+  APInt MaskValue = ConstantMaskNode->getAPIntValue();
+
+  if (!MaskValue.isPowerOf2())
+    return false;
+
+  return true;
+}
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
index b2fd31cb2346eb..73240bea4175a9 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
@@ -387,6 +387,8 @@ class AMDGPUTargetLowering : public TargetLowering {
   MVT getFenceOperandTy(const DataLayout &DL) const override {
     return MVT::i32;
   }
+
+  bool hasBitTest(SDValue X, SDValue Y) const override;
 };
 
 namespace AMDGPUISD {



More information about the llvm-commits mailing list