[AMDGPU] add ExpandFeaturePredicates pass to PB

This pass was added to pass builder by 8ea525f but got
reverted by accident by 9e810ad. This causes any
HIP program compiled with --offload-arch=amdgcnspirv
to fail to run.

This patch put ExpandFeaturePredicates pass back
to PB. Also add it to LTO pipeline. Add a lit test
to prevent future regressions.

Fixes: SWDEV-533964

Author: yxsamliu

clang/test/CodeGenHipStdPar/select-accelerator-code-pass-ordering.cpp

@@ -7,6 +7,7 @@
 // HIPSTDPAR-PRE: Running pass: EntryExitInstrumenterPass
 // HIPSTDPAR-PRE-NEXT: Running pass: EntryExitInstrumenterPass
 // HIPSTDPAR-PRE-NOT: Running pass: HipStdParAcceleratorCodeSelectionPass
+// HIPSTDPAR-PRE-NEXT: Running pass: AMDGPUExpandFeaturePredicatesPass
 // HIPSTDPAR-PRE-NEXT: Running pass: AlwaysInlinerPass
 
 // Ensure Pass HipStdParAcceleratorCodeSelectionPass is invoked in PostLink.

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

@@ -812,6 +812,16 @@ void AMDGPUTargetMachine::registerPassBuilderCallbacks(PassBuilder &PB) {
 #define GET_PASS_REGISTRY "AMDGPUPassRegistry.def"
 #include "llvm/Passes/TargetPassRegistry.inc"
 
+  PB.registerPipelineStartEPCallback(
+      [this](ModulePassManager &PM, OptimizationLevel Level) {
+        PM.addPass(AMDGPUExpandFeaturePredicatesPass(*this));
+      });
+
+  PB.registerFullLinkTimeOptimizationEarlyEPCallback(
+      [this](ModulePassManager &PM, OptimizationLevel Level) {
+        PM.addPass(AMDGPUExpandFeaturePredicatesPass(*this));
+      });
+
   PB.registerScalarOptimizerLateEPCallback(
       [](FunctionPassManager &FPM, OptimizationLevel Level) {
         if (Level == OptimizationLevel::O0)

llvm/test/CodeGen/AMDGPU/print-pipeline-passes.ll

@@ -2,16 +2,22 @@
 ; RUN: opt -mtriple=amdgcn--amdhsa -S -passes="lto<O1>" -print-pipeline-passes %s -o - | FileCheck %s
 ; RUN: opt -mtriple=amdgcn--amdhsa -S -passes="lto<O2>" -print-pipeline-passes %s -o - | FileCheck %s
 ; RUN: opt -mtriple=amdgcn--amdhsa -S -passes="lto<O3>" -print-pipeline-passes %s -o - | FileCheck %s
+; RUN: opt -mtriple=amdgcn--amdhsa -S -O0 -print-pipeline-passes %s -o - | FileCheck --check-prefix=O0 %s
+; RUN: opt -mtriple=amdgcn--amdhsa -S -O1 -print-pipeline-passes %s -o - | FileCheck %s
+; RUN: opt -mtriple=amdgcn--amdhsa -S -O2 -print-pipeline-passes %s -o - | FileCheck %s
+; RUN: opt -mtriple=amdgcn--amdhsa -S -O3 -print-pipeline-passes %s -o - | FileCheck %s
 
 ; RUN: opt -mtriple=amdgcn--amdhsa -S -passes="lto-pre-link<O0>" -print-pipeline-passes -amdgpu-internalize-symbols %s -o - | FileCheck --check-prefix=PRE %s
 ; RUN: opt -mtriple=amdgcn--amdhsa -S -passes="lto-pre-link<O1>" -print-pipeline-passes -amdgpu-internalize-symbols %s -o - | FileCheck --check-prefix=PRE %s
 ; RUN: opt -mtriple=amdgcn--amdhsa -S -passes="lto-pre-link<O2>" -print-pipeline-passes -amdgpu-internalize-symbols %s -o - | FileCheck --check-prefix=PRE %s
 ; RUN: opt -mtriple=amdgcn--amdhsa -S -passes="lto-pre-link<O3>" -print-pipeline-passes -amdgpu-internalize-symbols %s -o - | FileCheck --check-prefix=PRE %s
 
-
+; CHECK: amdgpu-expand-feature-predicates
 ; CHECK: amdgpu-attributor
+; O0: amdgpu-expand-feature-predicates
 ; O0-NOT: amdgpu-attributor
 
+; PRE: amdgpu-expand-feature-predicates
 ; PRE-NOT: internalize
 ; PRE-NOT: amdgpu-attributor
 ; PRE-NOT: printfToRuntime