[AMDGPU] Add intrinsics for alignbit and alignbyte instructions

[b-sumner]

This looks fine to me. I was concerned about how successful we would be pattern matching them.

[rampitec]

In D34046#776849, @b-sumner wrote:

This looks fine to me. I was concerned about how successful we would be pattern matching them.

We do not match it after numerous of optimizations, it is better to use intrinsic from the library.

[arsenm]

LGTM. We should try to pattern match these though, I think through ROTR one is matched sometimes already

include/llvm/IR/IntrinsicsAMDGPU.td

def int_amdgcn_readfirstlane :
   Intrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoMem, IntrConvergent]>;
 
 // The lane argument must be uniform across the currently active threads of the
 // current wave. Otherwise, the result is undefined.
 def int_amdgcn_readlane :
   GCCBuiltin<"__builtin_amdgcn_readlane">,
   Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem, IntrConvergent]>;
 
+def int_amdgcn_alignbit : Intrinsic<[llvm_i32_ty],
+  [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty],
+  [IntrNoMem, IntrSpeculatable]
+>;
+
+def int_amdgcn_alignbyte : Intrinsic<[llvm_i32_ty],
+  [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty],
+  [IntrNoMem, IntrSpeculatable]
+>;
+
 //===----------------------------------------------------------------------===//
 // CI+ Intrinsics
 //===----------------------------------------------------------------------===//
 
 def int_amdgcn_s_dcache_inv_vol :
   GCCBuiltin<"__builtin_amdgcn_s_dcache_inv_vol">,
   Intrinsic<[], [], []>;
 

lib/Target/AMDGPU/VOP3Instructions.td

 
 def V_CUBEID_F32 : VOP3Inst <"v_cubeid_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, int_amdgcn_cubeid>;
 def V_CUBESC_F32 : VOP3Inst <"v_cubesc_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, int_amdgcn_cubesc>;
 def V_CUBETC_F32 : VOP3Inst <"v_cubetc_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, int_amdgcn_cubetc>;
 def V_CUBEMA_F32 : VOP3Inst <"v_cubema_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, int_amdgcn_cubema>;
 def V_BFE_U32 : VOP3Inst <"v_bfe_u32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUbfe_u32>;
 def V_BFE_I32 : VOP3Inst <"v_bfe_i32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUbfe_i32>;
 def V_BFI_B32 : VOP3Inst <"v_bfi_b32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUbfi>;
-def V_ALIGNBIT_B32 : VOP3Inst <"v_alignbit_b32", VOP3_Profile<VOP_I32_I32_I32_I32>>;
-def V_ALIGNBYTE_B32 : VOP3Inst <"v_alignbyte_b32", VOP3_Profile<VOP_I32_I32_I32_I32>>;
+def V_ALIGNBIT_B32 : VOP3Inst <"v_alignbit_b32", VOP3_Profile<VOP_I32_I32_I32_I32>, int_amdgcn_alignbit>;
+def V_ALIGNBYTE_B32 : VOP3Inst <"v_alignbyte_b32", VOP3_Profile<VOP_I32_I32_I32_I32>, int_amdgcn_alignbyte>;
 def V_MIN3_F32 : VOP3Inst <"v_min3_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, AMDGPUfmin3>;
 def V_MIN3_I32 : VOP3Inst <"v_min3_i32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUsmin3>;
 def V_MIN3_U32 : VOP3Inst <"v_min3_u32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUumin3>;
 def V_MAX3_F32 : VOP3Inst <"v_max3_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, AMDGPUfmax3>;
 def V_MAX3_I32 : VOP3Inst <"v_max3_i32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUsmax3>;
 def V_MAX3_U32 : VOP3Inst <"v_max3_u32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUumax3>;
 def V_MED3_F32 : VOP3Inst <"v_med3_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, AMDGPUfmed3>;
 def V_MED3_I32 : VOP3Inst <"v_med3_i32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUsmed3>;

test/CodeGen/AMDGPU/llvm.amdgcn.alignb.ll

+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+
+declare i32 @llvm.amdgcn.alignbit(i32, i32, i32) #0
+declare i32 @llvm.amdgcn.alignbyte(i32, i32, i32) #0
+
+; GCN-LABEL: {{^}}v_alignbit_b32:
+; GCN: v_alignbit_b32 {{[vs][0-9]+}}, {{[vs][0-9]+}}, {{[vs][0-9]+}}
+define amdgpu_kernel void @v_alignbit_b32(i32 addrspace(1)* %out, i32 %src1, i32 %src2, i32 %src3) #1 {
+  %val = call i32 @llvm.amdgcn.alignbit(i32 %src1, i32 %src2, i32 %src3) #0
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_alignbyte_b32:
+; GCN: v_alignbyte_b32 {{[vs][0-9]+}}, {{[vs][0-9]+}}, {{[vs][0-9]+}}
+define amdgpu_kernel void @v_alignbyte_b32(i32 addrspace(1)* %out, i32 %src1, i32 %src2, i32 %src3) #1 {
+  %val = call i32 @llvm.amdgcn.alignbyte(i32 %src1, i32 %src2, i32 %src3) #0
+  store i32 %val, i32 addrspace(1)* %out
+  ret void
+}
+
+attributes #0 = { nounwind readnone }
+attributes #1 = { nounwind }