AMDGPU: Add codegen support for gfx950 v_ashr_pk_i8/u8_i32 (#118304)

Co-authored-by: Sirish Pande <[email protected]>

Author: arsenm

llvm/lib/Target/AMDGPU/VOP3Instructions.td

@@ -1457,6 +1457,23 @@ let SubtargetPredicate = HasAshrPkInsts, isReMaterializable = 1 in {
   defm V_ASHR_PK_U8_I32 : VOP3Inst<"v_ashr_pk_u8_i32", VOP3_Profile<VOP_I16_I32_I32_I32, VOP3_OPSEL_ONLY>, int_amdgcn_ashr_pk_u8_i32>;
 } // End SubtargetPredicate = HasAshrPkInsts, isReMaterializable = 1
 
+class AshrPkI8Pat<VOP3_Pseudo inst, int lo, int hi>: GCNPat<
+    (i16 (or (i16 (shl (i16 (trunc (i32 (AMDGPUsmed3 (i32 (sra i32:$src1, i32:$src2)), (i32 lo), (i32 hi))))), (i16 8))),
+             (i16 (and (i16 (trunc (i32 (AMDGPUsmed3 (i32 (sra i32:$src0, i32:$src2)), (i32 lo), (i32 hi))))), (i16 255))))),
+    (inst 0, VSrc_b32:$src0, 0,  VSrc_b32:$src1, 0, VSrc_b32:$src2, 0 )
+>;
+
+class AshrPkU8Pat<VOP3_Pseudo inst, int lo, int hi>: GCNPat<
+    (i16 (or (i16 (shl (i16 (trunc (i32 (AMDGPUsmed3 (i32 (sra i32:$src1, i32:$src2)), (i32 lo), (i32 hi))))), (i16 8))),
+             (i16 (trunc (i32 (AMDGPUsmed3 (i32 (sra i32:$src0, i32:$src2)), (i32 lo), (i32 hi))))))),
+    (inst 0, VSrc_b32:$src0, 0,  VSrc_b32:$src1, 0, VSrc_b32:$src2, 0 )
+>;
+
+let SubtargetPredicate = HasAshrPkInsts in {
+  def : AshrPkI8Pat<V_ASHR_PK_I8_I32_e64, -128, 127>;
+  def : AshrPkU8Pat<V_ASHR_PK_U8_I32_e64, 0, 255>;
+}
+
 //===----------------------------------------------------------------------===//
 // Integer Clamp Patterns
 //===----------------------------------------------------------------------===//

llvm/test/CodeGen/AMDGPU/v_ashr_pk.ll

@@ -0,0 +1,62 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
+; RUN: llc -mtriple=amdgcn -mcpu=gfx950 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX950 %s
+define amdgpu_kernel void @v_ashr_pk_i8_i32(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) #0 {
+; GFX950-LABEL: v_ashr_pk_i8_i32:
+; GFX950:       ; %bb.0:
+; GFX950-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
+; GFX950-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x24
+; GFX950-NEXT:    v_mov_b32_e32 v1, 0xffffff80
+; GFX950-NEXT:    v_mov_b32_e32 v2, 0x7f
+; GFX950-NEXT:    v_mov_b32_e32 v0, 0
+; GFX950-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX950-NEXT:    s_ashr_i32 s1, s1, s2
+; GFX950-NEXT:    s_ashr_i32 s0, s0, s2
+; GFX950-NEXT:    v_med3_i32 v3, s0, v1, v2
+; GFX950-NEXT:    v_med3_i32 v1, s1, v1, v2
+; GFX950-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
+; GFX950-NEXT:    v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX950-NEXT:    global_store_short v0, v1, s[6:7]
+; GFX950-NEXT:    s_endpgm
+  %insert.0 = insertelement <2 x i32> poison, i32 %src0, i64 0
+  %build_vector = insertelement <2 x i32> %insert.0, i32 %src1, i64 1
+  %src2.clamp = and i32 %src2, 31
+  %insert.1 = insertelement <2 x i32> poison, i32 %src2.clamp, i64 0
+  %src2.broadcast = shufflevector <2 x i32> %insert.1, <2 x i32> poison, <2 x i32> zeroinitializer
+  %ashr = ashr <2 x i32> %build_vector, %src2.broadcast
+  %sat.low = tail call <2 x i32> @llvm.smax.v2i32(<2 x i32> %ashr, <2 x i32> <i32 -128, i32 -128>)
+  %sat.hi = tail call <2 x i32> @llvm.smin.v2i32(<2 x i32> %sat.low, <2 x i32> <i32 127, i32 127>)
+  %trunc = trunc nsw <2 x i32> %sat.hi to <2 x i8>
+  %ret = bitcast <2 x i8> %trunc to i16
+  store i16 %ret, ptr addrspace(1) %out
+  ret void
+}
+
+define amdgpu_kernel void @v_ashr_pk_u8_i32(ptr addrspace(1) %out, i32 %src0, i32 %src1, i32 %src2) #0 {
+; GFX950-LABEL: v_ashr_pk_u8_i32:
+; GFX950:       ; %bb.0:
+; GFX950-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
+; GFX950-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x24
+; GFX950-NEXT:    v_mov_b32_e32 v1, 0xff
+; GFX950-NEXT:    v_mov_b32_e32 v0, 0
+; GFX950-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX950-NEXT:    s_ashr_i32 s1, s1, s2
+; GFX950-NEXT:    s_ashr_i32 s0, s0, s2
+; GFX950-NEXT:    v_med3_i32 v2, s0, 0, v1
+; GFX950-NEXT:    v_med3_i32 v1, s1, 0, v1
+; GFX950-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
+; GFX950-NEXT:    v_or_b32_e32 v1, v2, v1
+; GFX950-NEXT:    global_store_short v0, v1, s[6:7]
+; GFX950-NEXT:    s_endpgm
+  %insert.0 = insertelement <2 x i32> poison, i32 %src0, i64 0
+  %build_vector = insertelement <2 x i32> %insert.0, i32 %src1, i64 1
+  %src2.clamp = and i32 %src2, 31
+  %insert.1 = insertelement <2 x i32> poison, i32 %src2.clamp, i64 0
+  %src2.broadcast = shufflevector <2 x i32> %insert.1, <2 x i32> poison, <2 x i32> zeroinitializer
+  %ashr = ashr <2 x i32> %build_vector, %src2.broadcast
+  %sat.low = tail call <2 x i32> @llvm.smax.v2i32(<2 x i32> %ashr, <2 x i32> <i32 0, i32 0>)
+  %sat.hi = tail call <2 x i32> @llvm.smin.v2i32(<2 x i32> %sat.low, <2 x i32> <i32 255, i32 255>)
+  %trunc = trunc nsw <2 x i32> %sat.hi to <2 x i8>
+  %ret = bitcast <2 x i8> %trunc to i16
+  store i16 %ret, ptr addrspace(1) %out
+  ret void
+}