[AMDGPU] Add regbankselect rules for G_FSHR

Author: gandhi56

llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp

@@ -526,6 +526,10 @@ RegBankLegalizeRules::RegBankLegalizeRules(const GCNSubtarget &_ST,
       .Div(S32, {{Vgpr32}, {Vgpr32, Vgpr32}})
       .Div(S64, {{Vgpr64}, {Vgpr64, Vgpr32}});
 
+  addRulesForGOpcs({G_FSHR}, Standard)
+      .Uni(S32, {{UniInVgprS32}, {Vgpr32, Vgpr32, Vgpr32}})
+      .Div(S32, {{Vgpr32}, {Vgpr32, Vgpr32, Vgpr32}});
+
   addRulesForGOpcs({G_FRAME_INDEX}).Any({{UniP5, _}, {{SgprP5}, {None}}});
 
   addRulesForGOpcs({G_UBFX, G_SBFX}, Standard)

llvm/test/CodeGen/AMDGPU/GlobalISel/fshr-new-regbank-select.ll

@@ -0,0 +1,35 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
+; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 < %s | FileCheck %s
+
+define amdgpu_ps void @uniform_fshr_i32(i32 inreg %lhs, i32 inreg %rhs, i32 inreg %amt, ptr %resptr) {
+; CHECK-LABEL: uniform_fshr_i32:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    v_mov_b32_e32 v2, s2
+; CHECK-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; CHECK-NEXT:    v_alignbit_b32 v2, s0, s1, v2
+; CHECK-NEXT:    v_readfirstlane_b32 s0, v2
+; CHECK-NEXT:    s_add_co_i32 s0, s0, s0
+; CHECK-NEXT:    s_wait_alu 0xfffe
+; CHECK-NEXT:    v_mov_b32_e32 v2, s0
+; CHECK-NEXT:    flat_store_b32 v[0:1], v2
+; CHECK-NEXT:    s_endpgm
+  %vres = call i32 @llvm.fshr.i32(i32 %lhs, i32 %rhs, i32 %amt)
+  %add = add i32 %vres, %vres
+  store i32 %add, ptr %resptr
+  ret void
+}
+
+declare i32 @llvm.amdgcn.readfirstlane.i32(i32)
+
+define amdgpu_ps void @divergent_fshr_i32(i32 %lhs, i32 %rhs, i32 %amt, ptr %resptr) {
+; CHECK-LABEL: divergent_fshr_i32:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    v_alignbit_b32 v0, v0, v1, v2
+; CHECK-NEXT:    flat_store_b32 v[3:4], v0
+; CHECK-NEXT:    s_endpgm
+  %result = call i32 @llvm.fshr.i32(i32 %lhs, i32 %rhs, i32 %amt)
+  store i32 %result, ptr %resptr
+  ret void
+}
+
+declare i32 @llvm.fshr.i32(i32, i32, i32)