llvm/test/CodeGen/AMDGPU/GlobalISel/fcanonicalize.bf16.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
; RUN: llc -global-isel=1 -mtriple=amdgpu9.00 -amdgpu-enable-delay-alu=0 -o - %s | FileCheck -check-prefix=GFX9 %s
; RUN: llc -global-isel=1 -mtriple=amdgpu12.00 -amdgpu-enable-delay-alu=0 -o - %s | FileCheck -check-prefix=GFX12 %s
; RUN: llc -global-isel=1 -mtriple=amdgpu12.50 -amdgpu-enable-delay-alu=0 -o - %s | FileCheck -check-prefix=GFX1250 %s

define amdgpu_ps bfloat @fcanonicalize_bf16_v(bfloat %src) {
; GFX9-LABEL: fcanonicalize_bf16_v:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_max_f32_e32 v0, v0, v0
; GFX9-NEXT:    v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT:    v_mov_b32_e32 v3, 0x7fff
; GFX9-NEXT:    v_or_b32_e32 v1, 0x400000, v0
; GFX9-NEXT:    v_add3_u32 v2, v2, v0, v3
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, 0, v0
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: fcanonicalize_bf16_v:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX12-NEXT:    v_max_num_f32_e32 v0, v0, v0
; GFX12-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX12-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, 0, v0
; GFX12-NEXT:    v_add3_u32 v1, v1, v0, 0x7fff
; GFX12-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX12-NEXT:    v_mov_b16_e32 v0.l, v0.h
; GFX12-NEXT:    ; return to shader part epilog
;
; GFX1250-LABEL: fcanonicalize_bf16_v:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-NEXT:    s_mov_b64 s[64:65], 0
; GFX1250-NEXT:    v_nop
; GFX1250-NEXT:    global_prefetch_b8 v0, s[64:65] scope:SCOPE_SE
; GFX1250-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX1250-NEXT:    v_max_num_f32_e32 v0, v0, v0
; GFX1250-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX1250-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX1250-NEXT:    v_cmp_u_f32_e32 vcc_lo, 0, v0
; GFX1250-NEXT:    v_add3_u32 v1, v1, v0, 0x7fff
; GFX1250-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX1250-NEXT:    v_mov_b16_e32 v0.l, v0.h
; GFX1250-NEXT:    ; return to shader part epilog
  %result = call bfloat @llvm.canonicalize.bf16(bfloat %src)
  ret bfloat %result
}

define amdgpu_ps bfloat @fcanonicalize_bf16_s(bfloat inreg %src) {
; GFX9-LABEL: fcanonicalize_bf16_s:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_lshl_b32 s0, s0, 16
; GFX9-NEXT:    v_max_f32_e64 v0, s0, s0
; GFX9-NEXT:    v_readfirstlane_b32 s0, v0
; GFX9-NEXT:    s_bfe_u32 s2, s0, 0x10010
; GFX9-NEXT:    s_or_b32 s1, s0, 0x400000
; GFX9-NEXT:    s_add_i32 s0, s2, s0
; GFX9-NEXT:    s_addk_i32 s0, 0x7fff
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, 0, v0
; GFX9-NEXT:    s_cmp_lg_u64 vcc, 0
; GFX9-NEXT:    s_cselect_b32 s0, s1, s0
; GFX9-NEXT:    s_lshr_b32 s0, s0, 16
; GFX9-NEXT:    v_mov_b32_e32 v0, s0
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: fcanonicalize_bf16_s:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_lshl_b32 s0, s0, 16
; GFX12-NEXT:    v_max_num_f32_e64 v0, s0, s0
; GFX12-NEXT:    v_readfirstlane_b32 s0, v0
; GFX12-NEXT:    s_bfe_u32 s1, s0, 0x10010
; GFX12-NEXT:    s_or_b32 s2, s0, 0x400000
; GFX12-NEXT:    s_wait_alu depctr_sa_sdst(0)
; GFX12-NEXT:    s_add_co_i32 s1, s1, s0
; GFX12-NEXT:    s_wait_alu depctr_sa_sdst(0)
; GFX12-NEXT:    s_addk_co_i32 s1, 0x7fff
; GFX12-NEXT:    s_cmp_u_f32 s0, 0
; GFX12-NEXT:    s_wait_alu depctr_sa_sdst(0)
; GFX12-NEXT:    s_cselect_b32 s0, s2, s1
; GFX12-NEXT:    s_wait_alu depctr_sa_sdst(0)
; GFX12-NEXT:    s_lshr_b32 s0, s0, 16
; GFX12-NEXT:    s_wait_alu depctr_sa_sdst(0)
; GFX12-NEXT:    v_mov_b32_e32 v0, s0
; GFX12-NEXT:    ; return to shader part epilog
;
; GFX1250-LABEL: fcanonicalize_bf16_s:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-NEXT:    s_mov_b64 s[64:65], 0
; GFX1250-NEXT:    v_nop
; GFX1250-NEXT:    global_prefetch_b8 v0, s[64:65] scope:SCOPE_SE
; GFX1250-NEXT:    s_lshl_b32 s0, s0, 16
; GFX1250-NEXT:    v_max_num_f32_e64 v0, s0, s0
; GFX1250-NEXT:    v_readfirstlane_b32 s0, v0
; GFX1250-NEXT:    s_bfe_u32 s1, s0, 0x10010
; GFX1250-NEXT:    s_or_b32 s2, s0, 0x400000
; GFX1250-NEXT:    s_add_co_i32 s1, s1, s0
; GFX1250-NEXT:    s_addk_co_i32 s1, 0x7fff
; GFX1250-NEXT:    s_cmp_u_f32 s0, 0
; GFX1250-NEXT:    s_cselect_b32 s0, s2, s1
; GFX1250-NEXT:    s_lshr_b32 s0, s0, 16
; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
; GFX1250-NEXT:    ; return to shader part epilog
  %result = call bfloat @llvm.canonicalize.bf16(bfloat %src)
  ret bfloat %result
}

define amdgpu_ps <2 x bfloat> @fcanonicalize_v2bf16_v(<2 x bfloat> %src) {
; GFX9-LABEL: fcanonicalize_v2bf16_v:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_max_f32_e32 v0, v0, v0
; GFX9-NEXT:    v_bfe_u32 v3, v0, 16, 1
; GFX9-NEXT:    v_mov_b32_e32 v4, 0x7fff
; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_add3_u32 v3, v3, v0, v4
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, 0, v0
; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v3, v2, vcc
; GFX9-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v1
; GFX9-NEXT:    v_add3_u32 v3, v3, v1, v4
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, 0, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v3, v2, vcc
; GFX9-NEXT:    s_mov_b32 s0, 0x7060302
; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s0
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: fcanonicalize_v2bf16_v:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    v_mov_b16_e32 v1.l, v0.h
; GFX12-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX12-NEXT:    v_dual_max_num_f32 v1, v1, v1 :: v_dual_lshlrev_b32 v0, 16, v0
; GFX12-NEXT:    v_max_num_f32_e32 v0, v0, v0
; GFX12-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX12-NEXT:    v_bfe_u32 v2, v0, 16, 1
; GFX12-NEXT:    v_or_b32_e32 v4, 0x400000, v0
; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, 0, v0
; GFX12-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; GFX12-NEXT:    v_add3_u32 v3, v3, v1, 0x7fff
; GFX12-NEXT:    v_add3_u32 v2, v2, v0, 0x7fff
; GFX12-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, 0, v1
; GFX12-NEXT:    s_wait_alu depctr_va_vcc(0)
; GFX12-NEXT:    v_cndmask_b32_e32 v0, v3, v5, vcc_lo
; GFX12-NEXT:    v_mov_b16_e32 v0.l, v2.h
; GFX12-NEXT:    ; return to shader part epilog
;
; GFX1250-LABEL: fcanonicalize_v2bf16_v:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-NEXT:    s_mov_b64 s[64:65], 0
; GFX1250-NEXT:    v_nop
; GFX1250-NEXT:    global_prefetch_b8 v0, s[64:65] scope:SCOPE_SE
; GFX1250-NEXT:    v_mov_b16_e32 v1.l, v0.h
; GFX1250-NEXT:    v_dual_lshlrev_b32 v0, 16, v0 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX1250-NEXT:    v_dual_max_num_f32 v0, v0, v0 :: v_dual_max_num_f32 v1, v1, v1
; GFX1250-NEXT:    v_bfe_u32 v2, v0, 16, 1
; GFX1250-NEXT:    v_or_b32_e32 v4, 0x400000, v0
; GFX1250-NEXT:    v_cmp_u_f32_e32 vcc_lo, 0, v0
; GFX1250-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX1250-NEXT:    v_add3_u32 v2, v2, v0, 0x7fff
; GFX1250-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; GFX1250-NEXT:    v_add3_u32 v3, v3, v1, 0x7fff
; GFX1250-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
; GFX1250-NEXT:    v_cmp_u_f32_e32 vcc_lo, 0, v1
; GFX1250-NEXT:    v_cndmask_b32_e32 v0, v3, v5, vcc_lo
; GFX1250-NEXT:    v_mov_b16_e32 v0.l, v2.h
; GFX1250-NEXT:    ; return to shader part epilog
  %result = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> %src)
  ret <2 x bfloat> %result
}

define amdgpu_ps <2 x bfloat> @fcanonicalize_v2bf16_s(<2 x bfloat> inreg %src) {
; GFX9-LABEL: fcanonicalize_v2bf16_s:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_lshr_b32 s1, s0, 16
; GFX9-NEXT:    s_lshl_b32 s0, s0, 16
; GFX9-NEXT:    v_max_f32_e64 v0, s0, s0
; GFX9-NEXT:    v_readfirstlane_b32 s0, v0
; GFX9-NEXT:    s_bfe_u32 s3, s0, 0x10010
; GFX9-NEXT:    s_or_b32 s2, s0, 0x400000
; GFX9-NEXT:    s_add_i32 s0, s3, s0
; GFX9-NEXT:    s_addk_i32 s0, 0x7fff
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, 0, v0
; GFX9-NEXT:    s_cmp_lg_u64 vcc, 0
; GFX9-NEXT:    s_cselect_b32 s0, s2, s0
; GFX9-NEXT:    s_lshl_b32 s1, s1, 16
; GFX9-NEXT:    v_max_f32_e64 v0, s1, s1
; GFX9-NEXT:    v_readfirstlane_b32 s1, v0
; GFX9-NEXT:    s_bfe_u32 s3, s1, 0x10010
; GFX9-NEXT:    s_or_b32 s2, s1, 0x400000
; GFX9-NEXT:    s_add_i32 s1, s3, s1
; GFX9-NEXT:    s_addk_i32 s1, 0x7fff
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, 0, v0
; GFX9-NEXT:    s_cmp_lg_u64 vcc, 0
; GFX9-NEXT:    s_cselect_b32 s1, s2, s1
; GFX9-NEXT:    s_pack_hh_b32_b16 s0, s0, s1
; GFX9-NEXT:    v_mov_b32_e32 v0, s0
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: fcanonicalize_v2bf16_s:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_lshl_b32 s1, s0, 16
; GFX12-NEXT:    s_lshr_b32 s0, s0, 16
; GFX12-NEXT:    v_max_num_f32_e64 v0, s1, s1
; GFX12-NEXT:    v_readfirstlane_b32 s1, v0
; GFX12-NEXT:    s_bfe_u32 s2, s1, 0x10010
; GFX12-NEXT:    s_or_b32 s3, s1, 0x400000
; GFX12-NEXT:    s_add_co_i32 s2, s2, s1
; GFX12-NEXT:    s_addk_co_i32 s2, 0x7fff
; GFX12-NEXT:    s_cmp_u_f32 s1, 0
; GFX12-NEXT:    s_cselect_b32 s1, s3, s2
; GFX12-NEXT:    s_lshl_b32 s0, s0, 16
; GFX12-NEXT:    s_wait_alu depctr_sa_sdst(0)
; GFX12-NEXT:    s_lshr_b32 s1, s1, 16
; GFX12-NEXT:    v_max_num_f32_e64 v0, s0, s0
; GFX12-NEXT:    v_readfirstlane_b32 s0, v0
; GFX12-NEXT:    s_bfe_u32 s2, s0, 0x10010
; GFX12-NEXT:    s_or_b32 s3, s0, 0x400000
; GFX12-NEXT:    s_add_co_i32 s2, s2, s0
; GFX12-NEXT:    s_addk_co_i32 s2, 0x7fff
; GFX12-NEXT:    s_cmp_u_f32 s0, 0
; GFX12-NEXT:    s_cselect_b32 s0, s3, s2
; GFX12-NEXT:    s_wait_alu depctr_sa_sdst(0)
; GFX12-NEXT:    s_lshr_b32 s0, s0, 16
; GFX12-NEXT:    s_wait_alu depctr_sa_sdst(0)
; GFX12-NEXT:    s_pack_ll_b32_b16 s0, s1, s0
; GFX12-NEXT:    s_wait_alu depctr_sa_sdst(0)
; GFX12-NEXT:    v_mov_b32_e32 v0, s0
; GFX12-NEXT:    ; return to shader part epilog
;
; GFX1250-LABEL: fcanonicalize_v2bf16_s:
; GFX1250:       ; %bb.0:
; GFX1250-NEXT:    s_setreg_imm32_b32 hwreg(HW_REG_WAVE_MODE, 25, 1), 1 ; msbs: dst=0 src0=0 src1=0 src2=0
; GFX1250-NEXT:    s_mov_b64 s[64:65], 0
; GFX1250-NEXT:    v_nop
; GFX1250-NEXT:    global_prefetch_b8 v0, s[64:65] scope:SCOPE_SE
; GFX1250-NEXT:    s_lshl_b32 s1, s0, 16
; GFX1250-NEXT:    s_lshr_b32 s0, s0, 16
; GFX1250-NEXT:    v_max_num_f32_e64 v0, s1, s1
; GFX1250-NEXT:    v_readfirstlane_b32 s1, v0
; GFX1250-NEXT:    s_bfe_u32 s2, s1, 0x10010
; GFX1250-NEXT:    s_or_b32 s3, s1, 0x400000
; GFX1250-NEXT:    s_add_co_i32 s2, s2, s1
; GFX1250-NEXT:    s_addk_co_i32 s2, 0x7fff
; GFX1250-NEXT:    s_cmp_u_f32 s1, 0
; GFX1250-NEXT:    s_cselect_b32 s1, s3, s2
; GFX1250-NEXT:    s_lshl_b32 s0, s0, 16
; GFX1250-NEXT:    s_lshr_b32 s1, s1, 16
; GFX1250-NEXT:    v_max_num_f32_e64 v0, s0, s0
; GFX1250-NEXT:    v_readfirstlane_b32 s0, v0
; GFX1250-NEXT:    s_bfe_u32 s2, s0, 0x10010
; GFX1250-NEXT:    s_or_b32 s3, s0, 0x400000
; GFX1250-NEXT:    s_add_co_i32 s2, s2, s0
; GFX1250-NEXT:    s_addk_co_i32 s2, 0x7fff
; GFX1250-NEXT:    s_cmp_u_f32 s0, 0
; GFX1250-NEXT:    s_cselect_b32 s0, s3, s2
; GFX1250-NEXT:    s_lshr_b32 s0, s0, 16
; GFX1250-NEXT:    s_pack_ll_b32_b16 s0, s1, s0
; GFX1250-NEXT:    v_mov_b32_e32 v0, s0
; GFX1250-NEXT:    ; return to shader part epilog
  %result = call <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat> %src)
  ret <2 x bfloat> %result
}

declare bfloat @llvm.canonicalize.bf16(bfloat)
declare <2 x bfloat> @llvm.canonicalize.v2bf16(<2 x bfloat>)