AMDGPU: Move libcall simplify into PeepholeEP

We were running this immediately on the incoming IR, which
is still littered with temporary allocas obscuring trivial values.
This needs to run after initial SROA to handle sincos insertion.

Fixes: SWDEV-456865

Author: arsenm

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

@@ -655,9 +655,6 @@ void AMDGPUTargetMachine::registerPassBuilderCallbacks(
   PB.registerPipelineStartEPCallback(
       [](ModulePassManager &PM, OptimizationLevel Level) {
         FunctionPassManager FPM;
-        FPM.addPass(AMDGPUUseNativeCallsPass());
-        if (EnableLibCallSimplify && Level != OptimizationLevel::O0)
-          FPM.addPass(AMDGPUSimplifyLibCallsPass());
         PM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
         if (EnableHipStdPar)
           PM.addPass(HipStdParAcceleratorCodeSelectionPass());
@@ -681,6 +678,16 @@ void AMDGPUTargetMachine::registerPassBuilderCallbacks(
           PM.addPass(AMDGPUAlwaysInlinePass());
       });
 
+  PB.registerPeepholeEPCallback(
+      [](FunctionPassManager &FPM, OptimizationLevel Level) {
+        if (Level == OptimizationLevel::O0)
+          return;
+
+        FPM.addPass(AMDGPUUseNativeCallsPass());
+        if (EnableLibCallSimplify)
+          FPM.addPass(AMDGPUSimplifyLibCallsPass());
+      });
+
   PB.registerCGSCCOptimizerLateEPCallback(
       [this](CGSCCPassManager &PM, OptimizationLevel Level) {
         if (Level == OptimizationLevel::O0)

llvm/test/CodeGen/AMDGPU/amdgpu-libcall-sincos-pass-ordering.ll

@@ -12,9 +12,10 @@ define protected amdgpu_kernel void @swdev456865(ptr addrspace(1) %out0, ptr add
 ; CHECK-LABEL: define protected amdgpu_kernel void @swdev456865(
 ; CHECK-SAME: ptr addrspace(1) nocapture writeonly [[OUT0:%.*]], ptr addrspace(1) nocapture writeonly [[OUT1:%.*]], ptr addrspace(1) nocapture writeonly [[OUT2:%.*]], float noundef [[X:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[I_I:%.*]] = tail call float asm "pseudo-libcall-sin [[TMP0:%.*]], %1", "=v,v"(float noundef [[X]]) #[[ATTR1:[0-9]+]]
-; CHECK-NEXT:    [[I_I1:%.*]] = tail call float asm "pseudo-libcall-cos [[TMP0]], %1", "=v,v"(float noundef [[X]]) #[[ATTR1]]
-; CHECK-NEXT:    [[ADD:%.*]] = fadd float [[I_I]], [[I_I1]]
+; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
+; CHECK-NEXT:    [[I_I:%.*]] = call float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]]) #[[ATTR1:[0-9]+]]
+; CHECK-NEXT:    [[I_I2:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
+; CHECK-NEXT:    [[ADD:%.*]] = fadd float [[I_I]], [[I_I2]]
 ; CHECK-NEXT:    [[CONV:%.*]] = fpext float [[X]] to double
 ; CHECK-NEXT:    [[CONV5:%.*]] = fpext float [[ADD]] to double
 ; CHECK-NEXT:    store double [[CONV]], ptr addrspace(1) [[OUT0]], align 8