diff options
Diffstat (limited to 'llvm/test/CodeGen/AMDGPU/whole-wave-functions.ll')
| -rw-r--r-- | llvm/test/CodeGen/AMDGPU/whole-wave-functions.ll | 2414 |
1 files changed, 2414 insertions, 0 deletions
diff --git a/llvm/test/CodeGen/AMDGPU/whole-wave-functions.ll b/llvm/test/CodeGen/AMDGPU/whole-wave-functions.ll new file mode 100644 index 0000000..53d0292 --- /dev/null +++ b/llvm/test/CodeGen/AMDGPU/whole-wave-functions.ll @@ -0,0 +1,2414 @@ +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5 +; RUN: llc -global-isel=0 -mtriple=amdgcn--amdpal -mcpu=gfx1200 -verify-machineinstrs < %s | FileCheck --check-prefix=DAGISEL %s +; RUN: llc -global-isel=1 -mtriple=amdgcn--amdpal -mcpu=gfx1200 -verify-machineinstrs < %s | FileCheck --check-prefix=GISEL %s +; RUN: llc -global-isel=0 -mtriple=amdgcn--amdpal -mcpu=gfx1200 -mattr=+wavefrontsize64 -verify-machineinstrs < %s | FileCheck --check-prefix=DAGISEL64 %s +; RUN: llc -global-isel=1 -mtriple=amdgcn--amdpal -mcpu=gfx1200 -mattr=+wavefrontsize64 -verify-machineinstrs < %s | FileCheck --check-prefix=GISEL64 %s + +; Make sure the i1 %active is passed through EXEC. +; The EXEC mask should be set to -1 for the duration of the function +; and restored to its original value in the epilogue. +; We will also need to restore the inactive lanes for any allocated VGPRs. +define amdgpu_gfx_whole_wave i32 @basic_test(i1 %active, i32 %a, i32 %b) { +; DAGISEL-LABEL: basic_test: +; DAGISEL: ; %bb.0: +; DAGISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL-NEXT: s_wait_expcnt 0x0 +; DAGISEL-NEXT: s_wait_samplecnt 0x0 +; DAGISEL-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL-NEXT: s_wait_kmcnt 0x0 +; DAGISEL-NEXT: s_xor_saveexec_b32 vcc_lo, -1 +; DAGISEL-NEXT: s_clause 0x1 +; DAGISEL-NEXT: scratch_store_b32 off, v0, s32 +; DAGISEL-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; DAGISEL-NEXT: s_mov_b32 exec_lo, -1 +; DAGISEL-NEXT: s_wait_alu 0xfffe +; DAGISEL-NEXT: v_dual_cndmask_b32 v0, 5, v0 :: v_dual_cndmask_b32 v1, 3, v1 +; DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) +; DAGISEL-NEXT: v_mov_b32_dpp v0, v1 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; DAGISEL-NEXT: s_xor_b32 exec_lo, vcc_lo, -1 +; DAGISEL-NEXT: s_clause 0x1 +; DAGISEL-NEXT: scratch_load_b32 v0, off, s32 +; DAGISEL-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; DAGISEL-NEXT: s_mov_b32 exec_lo, vcc_lo +; DAGISEL-NEXT: s_wait_loadcnt 0x0 +; DAGISEL-NEXT: s_setpc_b64 s[30:31] +; +; GISEL-LABEL: basic_test: +; GISEL: ; %bb.0: +; GISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL-NEXT: s_wait_expcnt 0x0 +; GISEL-NEXT: s_wait_samplecnt 0x0 +; GISEL-NEXT: s_wait_bvhcnt 0x0 +; GISEL-NEXT: s_wait_kmcnt 0x0 +; GISEL-NEXT: s_xor_saveexec_b32 vcc_lo, -1 +; GISEL-NEXT: s_clause 0x1 +; GISEL-NEXT: scratch_store_b32 off, v0, s32 +; GISEL-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; GISEL-NEXT: s_mov_b32 exec_lo, -1 +; GISEL-NEXT: s_wait_alu 0xfffe +; GISEL-NEXT: v_dual_cndmask_b32 v0, 5, v0 :: v_dual_cndmask_b32 v1, 3, v1 +; GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GISEL-NEXT: v_mov_b32_dpp v0, v1 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; GISEL-NEXT: s_xor_b32 exec_lo, vcc_lo, -1 +; GISEL-NEXT: s_clause 0x1 +; GISEL-NEXT: scratch_load_b32 v0, off, s32 +; GISEL-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; GISEL-NEXT: s_mov_b32 exec_lo, vcc_lo +; GISEL-NEXT: s_wait_loadcnt 0x0 +; GISEL-NEXT: s_setpc_b64 s[30:31] +; +; DAGISEL64-LABEL: basic_test: +; DAGISEL64: ; %bb.0: +; DAGISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL64-NEXT: s_wait_expcnt 0x0 +; DAGISEL64-NEXT: s_wait_samplecnt 0x0 +; DAGISEL64-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL64-NEXT: s_wait_kmcnt 0x0 +; DAGISEL64-NEXT: s_xor_saveexec_b64 vcc, -1 +; DAGISEL64-NEXT: s_clause 0x1 +; DAGISEL64-NEXT: scratch_store_b32 off, v0, s32 +; DAGISEL64-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; DAGISEL64-NEXT: s_mov_b64 exec, -1 +; DAGISEL64-NEXT: s_wait_alu 0xfffe +; DAGISEL64-NEXT: v_cndmask_b32_e32 v0, 5, v0, vcc +; DAGISEL64-NEXT: v_cndmask_b32_e32 v1, 3, v1, vcc +; DAGISEL64-NEXT: s_delay_alu instid0(VALU_DEP_1) +; DAGISEL64-NEXT: v_mov_b32_dpp v0, v1 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; DAGISEL64-NEXT: s_xor_b64 exec, vcc, -1 +; DAGISEL64-NEXT: s_clause 0x1 +; DAGISEL64-NEXT: scratch_load_b32 v0, off, s32 +; DAGISEL64-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; DAGISEL64-NEXT: s_mov_b64 exec, vcc +; DAGISEL64-NEXT: s_wait_loadcnt 0x0 +; DAGISEL64-NEXT: s_setpc_b64 s[30:31] +; +; GISEL64-LABEL: basic_test: +; GISEL64: ; %bb.0: +; GISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL64-NEXT: s_wait_expcnt 0x0 +; GISEL64-NEXT: s_wait_samplecnt 0x0 +; GISEL64-NEXT: s_wait_bvhcnt 0x0 +; GISEL64-NEXT: s_wait_kmcnt 0x0 +; GISEL64-NEXT: s_xor_saveexec_b64 vcc, -1 +; GISEL64-NEXT: s_clause 0x1 +; GISEL64-NEXT: scratch_store_b32 off, v0, s32 +; GISEL64-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; GISEL64-NEXT: s_mov_b64 exec, -1 +; GISEL64-NEXT: s_wait_alu 0xfffe +; GISEL64-NEXT: v_cndmask_b32_e32 v0, 5, v0, vcc +; GISEL64-NEXT: v_cndmask_b32_e32 v1, 3, v1, vcc +; GISEL64-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GISEL64-NEXT: v_mov_b32_dpp v0, v1 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; GISEL64-NEXT: s_xor_b64 exec, vcc, -1 +; GISEL64-NEXT: s_clause 0x1 +; GISEL64-NEXT: scratch_load_b32 v0, off, s32 +; GISEL64-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; GISEL64-NEXT: s_mov_b64 exec, vcc +; GISEL64-NEXT: s_wait_loadcnt 0x0 +; GISEL64-NEXT: s_setpc_b64 s[30:31] + %x = select i1 %active, i32 %a, i32 5 + %y = select i1 %active, i32 %b, i32 3 + %ret = call i32 @llvm.amdgcn.update.dpp.i32(i32 %x, i32 %y, i32 1, i32 1, i32 1, i1 false) + ret i32 %ret +} + +; Make sure we don't crash if there's only one use for %active. +define amdgpu_gfx_whole_wave i32 @single_use_of_active(i1 %active, i32 %a, i32 %b) { +; DAGISEL-LABEL: single_use_of_active: +; DAGISEL: ; %bb.0: +; DAGISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL-NEXT: s_wait_expcnt 0x0 +; DAGISEL-NEXT: s_wait_samplecnt 0x0 +; DAGISEL-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL-NEXT: s_wait_kmcnt 0x0 +; DAGISEL-NEXT: s_xor_saveexec_b32 vcc_lo, -1 +; DAGISEL-NEXT: s_clause 0x1 +; DAGISEL-NEXT: scratch_store_b32 off, v0, s32 +; DAGISEL-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; DAGISEL-NEXT: s_mov_b32 exec_lo, -1 +; DAGISEL-NEXT: s_wait_alu 0xfffe +; DAGISEL-NEXT: v_cndmask_b32_e32 v1, 17, v1, vcc_lo +; DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) +; DAGISEL-NEXT: v_mov_b32_dpp v0, v1 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; DAGISEL-NEXT: s_xor_b32 exec_lo, vcc_lo, -1 +; DAGISEL-NEXT: s_clause 0x1 +; DAGISEL-NEXT: scratch_load_b32 v0, off, s32 +; DAGISEL-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; DAGISEL-NEXT: s_mov_b32 exec_lo, vcc_lo +; DAGISEL-NEXT: s_wait_loadcnt 0x0 +; DAGISEL-NEXT: s_setpc_b64 s[30:31] +; +; GISEL-LABEL: single_use_of_active: +; GISEL: ; %bb.0: +; GISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL-NEXT: s_wait_expcnt 0x0 +; GISEL-NEXT: s_wait_samplecnt 0x0 +; GISEL-NEXT: s_wait_bvhcnt 0x0 +; GISEL-NEXT: s_wait_kmcnt 0x0 +; GISEL-NEXT: s_xor_saveexec_b32 vcc_lo, -1 +; GISEL-NEXT: s_clause 0x1 +; GISEL-NEXT: scratch_store_b32 off, v0, s32 +; GISEL-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; GISEL-NEXT: s_mov_b32 exec_lo, -1 +; GISEL-NEXT: s_wait_alu 0xfffe +; GISEL-NEXT: v_cndmask_b32_e32 v1, 17, v1, vcc_lo +; GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GISEL-NEXT: v_mov_b32_dpp v0, v1 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; GISEL-NEXT: s_xor_b32 exec_lo, vcc_lo, -1 +; GISEL-NEXT: s_clause 0x1 +; GISEL-NEXT: scratch_load_b32 v0, off, s32 +; GISEL-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; GISEL-NEXT: s_mov_b32 exec_lo, vcc_lo +; GISEL-NEXT: s_wait_loadcnt 0x0 +; GISEL-NEXT: s_setpc_b64 s[30:31] +; +; DAGISEL64-LABEL: single_use_of_active: +; DAGISEL64: ; %bb.0: +; DAGISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL64-NEXT: s_wait_expcnt 0x0 +; DAGISEL64-NEXT: s_wait_samplecnt 0x0 +; DAGISEL64-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL64-NEXT: s_wait_kmcnt 0x0 +; DAGISEL64-NEXT: s_xor_saveexec_b64 vcc, -1 +; DAGISEL64-NEXT: s_clause 0x1 +; DAGISEL64-NEXT: scratch_store_b32 off, v0, s32 +; DAGISEL64-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; DAGISEL64-NEXT: s_mov_b64 exec, -1 +; DAGISEL64-NEXT: s_wait_alu 0xfffe +; DAGISEL64-NEXT: v_cndmask_b32_e32 v1, 17, v1, vcc +; DAGISEL64-NEXT: s_delay_alu instid0(VALU_DEP_1) +; DAGISEL64-NEXT: v_mov_b32_dpp v0, v1 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; DAGISEL64-NEXT: s_xor_b64 exec, vcc, -1 +; DAGISEL64-NEXT: s_clause 0x1 +; DAGISEL64-NEXT: scratch_load_b32 v0, off, s32 +; DAGISEL64-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; DAGISEL64-NEXT: s_mov_b64 exec, vcc +; DAGISEL64-NEXT: s_wait_loadcnt 0x0 +; DAGISEL64-NEXT: s_setpc_b64 s[30:31] +; +; GISEL64-LABEL: single_use_of_active: +; GISEL64: ; %bb.0: +; GISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL64-NEXT: s_wait_expcnt 0x0 +; GISEL64-NEXT: s_wait_samplecnt 0x0 +; GISEL64-NEXT: s_wait_bvhcnt 0x0 +; GISEL64-NEXT: s_wait_kmcnt 0x0 +; GISEL64-NEXT: s_xor_saveexec_b64 vcc, -1 +; GISEL64-NEXT: s_clause 0x1 +; GISEL64-NEXT: scratch_store_b32 off, v0, s32 +; GISEL64-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; GISEL64-NEXT: s_mov_b64 exec, -1 +; GISEL64-NEXT: s_wait_alu 0xfffe +; GISEL64-NEXT: v_cndmask_b32_e32 v1, 17, v1, vcc +; GISEL64-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GISEL64-NEXT: v_mov_b32_dpp v0, v1 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; GISEL64-NEXT: s_xor_b64 exec, vcc, -1 +; GISEL64-NEXT: s_clause 0x1 +; GISEL64-NEXT: scratch_load_b32 v0, off, s32 +; GISEL64-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; GISEL64-NEXT: s_mov_b64 exec, vcc +; GISEL64-NEXT: s_wait_loadcnt 0x0 +; GISEL64-NEXT: s_setpc_b64 s[30:31] + %y = select i1 %active, i32 %b, i32 17 + %ret = call i32 @llvm.amdgcn.update.dpp.i32(i32 %a, i32 %y, i32 1, i32 1, i32 1, i1 false) + ret i32 %ret +} + +; Make sure we don't crash if %active is not used at all. +define amdgpu_gfx_whole_wave i32 @unused_active(i1 %active, i32 %a, i32 %b) { +; DAGISEL-LABEL: unused_active: +; DAGISEL: ; %bb.0: +; DAGISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL-NEXT: s_wait_expcnt 0x0 +; DAGISEL-NEXT: s_wait_samplecnt 0x0 +; DAGISEL-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL-NEXT: s_wait_kmcnt 0x0 +; DAGISEL-NEXT: s_xor_saveexec_b32 s0, -1 +; DAGISEL-NEXT: scratch_store_b32 off, v0, s32 ; 4-byte Folded Spill +; DAGISEL-NEXT: s_mov_b32 exec_lo, -1 +; DAGISEL-NEXT: v_mov_b32_e32 v0, 14 +; DAGISEL-NEXT: s_wait_alu 0xfffe +; DAGISEL-NEXT: s_xor_b32 exec_lo, s0, -1 +; DAGISEL-NEXT: scratch_load_b32 v0, off, s32 ; 4-byte Folded Reload +; DAGISEL-NEXT: s_mov_b32 exec_lo, s0 +; DAGISEL-NEXT: s_wait_loadcnt 0x0 +; DAGISEL-NEXT: s_setpc_b64 s[30:31] +; +; GISEL-LABEL: unused_active: +; GISEL: ; %bb.0: +; GISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL-NEXT: s_wait_expcnt 0x0 +; GISEL-NEXT: s_wait_samplecnt 0x0 +; GISEL-NEXT: s_wait_bvhcnt 0x0 +; GISEL-NEXT: s_wait_kmcnt 0x0 +; GISEL-NEXT: s_xor_saveexec_b32 s0, -1 +; GISEL-NEXT: scratch_store_b32 off, v0, s32 ; 4-byte Folded Spill +; GISEL-NEXT: s_mov_b32 exec_lo, -1 +; GISEL-NEXT: v_mov_b32_e32 v0, 14 +; GISEL-NEXT: s_wait_alu 0xfffe +; GISEL-NEXT: s_xor_b32 exec_lo, s0, -1 +; GISEL-NEXT: scratch_load_b32 v0, off, s32 ; 4-byte Folded Reload +; GISEL-NEXT: s_mov_b32 exec_lo, s0 +; GISEL-NEXT: s_wait_loadcnt 0x0 +; GISEL-NEXT: s_setpc_b64 s[30:31] +; +; DAGISEL64-LABEL: unused_active: +; DAGISEL64: ; %bb.0: +; DAGISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL64-NEXT: s_wait_expcnt 0x0 +; DAGISEL64-NEXT: s_wait_samplecnt 0x0 +; DAGISEL64-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL64-NEXT: s_wait_kmcnt 0x0 +; DAGISEL64-NEXT: s_xor_saveexec_b64 s[0:1], -1 +; DAGISEL64-NEXT: scratch_store_b32 off, v0, s32 ; 4-byte Folded Spill +; DAGISEL64-NEXT: s_mov_b64 exec, -1 +; DAGISEL64-NEXT: v_mov_b32_e32 v0, 14 +; DAGISEL64-NEXT: s_wait_alu 0xfffe +; DAGISEL64-NEXT: s_xor_b64 exec, s[0:1], -1 +; DAGISEL64-NEXT: scratch_load_b32 v0, off, s32 ; 4-byte Folded Reload +; DAGISEL64-NEXT: s_mov_b64 exec, s[0:1] +; DAGISEL64-NEXT: s_wait_loadcnt 0x0 +; DAGISEL64-NEXT: s_setpc_b64 s[30:31] +; +; GISEL64-LABEL: unused_active: +; GISEL64: ; %bb.0: +; GISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL64-NEXT: s_wait_expcnt 0x0 +; GISEL64-NEXT: s_wait_samplecnt 0x0 +; GISEL64-NEXT: s_wait_bvhcnt 0x0 +; GISEL64-NEXT: s_wait_kmcnt 0x0 +; GISEL64-NEXT: s_xor_saveexec_b64 s[0:1], -1 +; GISEL64-NEXT: scratch_store_b32 off, v0, s32 ; 4-byte Folded Spill +; GISEL64-NEXT: s_mov_b64 exec, -1 +; GISEL64-NEXT: v_mov_b32_e32 v0, 14 +; GISEL64-NEXT: s_wait_alu 0xfffe +; GISEL64-NEXT: s_xor_b64 exec, s[0:1], -1 +; GISEL64-NEXT: scratch_load_b32 v0, off, s32 ; 4-byte Folded Reload +; GISEL64-NEXT: s_mov_b64 exec, s[0:1] +; GISEL64-NEXT: s_wait_loadcnt 0x0 +; GISEL64-NEXT: s_setpc_b64 s[30:31] + ret i32 14 +} + +; For any used VGPRs (including those used for SGPR spills), we need to restore the inactive lanes. +; For CSR VGPRs, we need to restore all lanes. +define amdgpu_gfx_whole_wave i32 @csr(i1 %active, i32 %a, i32 %b) { +; DAGISEL-LABEL: csr: +; DAGISEL: ; %bb.0: +; DAGISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL-NEXT: s_wait_expcnt 0x0 +; DAGISEL-NEXT: s_wait_samplecnt 0x0 +; DAGISEL-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL-NEXT: s_wait_kmcnt 0x0 +; DAGISEL-NEXT: s_xor_saveexec_b32 vcc_lo, -1 +; DAGISEL-NEXT: s_clause 0x3 +; DAGISEL-NEXT: scratch_store_b32 off, v2, s32 +; DAGISEL-NEXT: scratch_store_b32 off, v0, s32 offset:4 +; DAGISEL-NEXT: scratch_store_b32 off, v1, s32 offset:8 +; DAGISEL-NEXT: scratch_store_b32 off, v49, s32 offset:16 +; DAGISEL-NEXT: s_mov_b32 exec_lo, -1 +; DAGISEL-NEXT: scratch_store_b32 off, v40, s32 offset:12 ; 4-byte Folded Spill +; DAGISEL-NEXT: ;;#ASMSTART +; DAGISEL-NEXT: ; clobber CSR +; DAGISEL-NEXT: ;;#ASMEND +; DAGISEL-NEXT: v_writelane_b32 v2, s20, 0 +; DAGISEL-NEXT: ;;#ASMSTART +; DAGISEL-NEXT: ; clobber non-CSR +; DAGISEL-NEXT: ;;#ASMEND +; DAGISEL-NEXT: scratch_load_b32 v40, off, s32 offset:12 ; 4-byte Folded Reload +; DAGISEL-NEXT: s_wait_alu 0xfffe +; DAGISEL-NEXT: v_dual_cndmask_b32 v0, 5, v0 :: v_dual_cndmask_b32 v1, 3, v1 +; DAGISEL-NEXT: v_readlane_b32 s20, v2, 0 +; DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_2) +; DAGISEL-NEXT: v_mov_b32_dpp v0, v1 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; DAGISEL-NEXT: s_xor_b32 exec_lo, vcc_lo, -1 +; DAGISEL-NEXT: s_clause 0x3 +; DAGISEL-NEXT: scratch_load_b32 v2, off, s32 +; DAGISEL-NEXT: scratch_load_b32 v0, off, s32 offset:4 +; DAGISEL-NEXT: scratch_load_b32 v1, off, s32 offset:8 +; DAGISEL-NEXT: scratch_load_b32 v49, off, s32 offset:16 +; DAGISEL-NEXT: s_mov_b32 exec_lo, vcc_lo +; DAGISEL-NEXT: s_wait_loadcnt 0x0 +; DAGISEL-NEXT: s_wait_alu 0xf1ff +; DAGISEL-NEXT: s_setpc_b64 s[30:31] +; +; GISEL-LABEL: csr: +; GISEL: ; %bb.0: +; GISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL-NEXT: s_wait_expcnt 0x0 +; GISEL-NEXT: s_wait_samplecnt 0x0 +; GISEL-NEXT: s_wait_bvhcnt 0x0 +; GISEL-NEXT: s_wait_kmcnt 0x0 +; GISEL-NEXT: s_xor_saveexec_b32 vcc_lo, -1 +; GISEL-NEXT: s_clause 0x3 +; GISEL-NEXT: scratch_store_b32 off, v2, s32 +; GISEL-NEXT: scratch_store_b32 off, v0, s32 offset:4 +; GISEL-NEXT: scratch_store_b32 off, v1, s32 offset:8 +; GISEL-NEXT: scratch_store_b32 off, v49, s32 offset:16 +; GISEL-NEXT: s_mov_b32 exec_lo, -1 +; GISEL-NEXT: scratch_store_b32 off, v40, s32 offset:12 ; 4-byte Folded Spill +; GISEL-NEXT: ;;#ASMSTART +; GISEL-NEXT: ; clobber CSR +; GISEL-NEXT: ;;#ASMEND +; GISEL-NEXT: v_writelane_b32 v2, s20, 0 +; GISEL-NEXT: ;;#ASMSTART +; GISEL-NEXT: ; clobber non-CSR +; GISEL-NEXT: ;;#ASMEND +; GISEL-NEXT: scratch_load_b32 v40, off, s32 offset:12 ; 4-byte Folded Reload +; GISEL-NEXT: s_wait_alu 0xfffe +; GISEL-NEXT: v_dual_cndmask_b32 v0, 5, v0 :: v_dual_cndmask_b32 v1, 3, v1 +; GISEL-NEXT: v_readlane_b32 s20, v2, 0 +; GISEL-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GISEL-NEXT: v_mov_b32_dpp v0, v1 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; GISEL-NEXT: s_xor_b32 exec_lo, vcc_lo, -1 +; GISEL-NEXT: s_clause 0x3 +; GISEL-NEXT: scratch_load_b32 v2, off, s32 +; GISEL-NEXT: scratch_load_b32 v0, off, s32 offset:4 +; GISEL-NEXT: scratch_load_b32 v1, off, s32 offset:8 +; GISEL-NEXT: scratch_load_b32 v49, off, s32 offset:16 +; GISEL-NEXT: s_mov_b32 exec_lo, vcc_lo +; GISEL-NEXT: s_wait_loadcnt 0x0 +; GISEL-NEXT: s_wait_alu 0xf1ff +; GISEL-NEXT: s_setpc_b64 s[30:31] +; +; DAGISEL64-LABEL: csr: +; DAGISEL64: ; %bb.0: +; DAGISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL64-NEXT: s_wait_expcnt 0x0 +; DAGISEL64-NEXT: s_wait_samplecnt 0x0 +; DAGISEL64-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL64-NEXT: s_wait_kmcnt 0x0 +; DAGISEL64-NEXT: s_xor_saveexec_b64 vcc, -1 +; DAGISEL64-NEXT: s_clause 0x3 +; DAGISEL64-NEXT: scratch_store_b32 off, v2, s32 +; DAGISEL64-NEXT: scratch_store_b32 off, v0, s32 offset:4 +; DAGISEL64-NEXT: scratch_store_b32 off, v1, s32 offset:8 +; DAGISEL64-NEXT: scratch_store_b32 off, v49, s32 offset:16 +; DAGISEL64-NEXT: s_mov_b64 exec, -1 +; DAGISEL64-NEXT: scratch_store_b32 off, v40, s32 offset:12 ; 4-byte Folded Spill +; DAGISEL64-NEXT: ;;#ASMSTART +; DAGISEL64-NEXT: ; clobber CSR +; DAGISEL64-NEXT: ;;#ASMEND +; DAGISEL64-NEXT: v_writelane_b32 v2, s20, 0 +; DAGISEL64-NEXT: ;;#ASMSTART +; DAGISEL64-NEXT: ; clobber non-CSR +; DAGISEL64-NEXT: ;;#ASMEND +; DAGISEL64-NEXT: scratch_load_b32 v40, off, s32 offset:12 ; 4-byte Folded Reload +; DAGISEL64-NEXT: s_wait_alu 0xfffe +; DAGISEL64-NEXT: v_cndmask_b32_e32 v0, 5, v0, vcc +; DAGISEL64-NEXT: v_cndmask_b32_e32 v1, 3, v1, vcc +; DAGISEL64-NEXT: v_readlane_b32 s20, v2, 0 +; DAGISEL64-NEXT: s_delay_alu instid0(VALU_DEP_2) +; DAGISEL64-NEXT: v_mov_b32_dpp v0, v1 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; DAGISEL64-NEXT: s_xor_b64 exec, vcc, -1 +; DAGISEL64-NEXT: s_clause 0x3 +; DAGISEL64-NEXT: scratch_load_b32 v2, off, s32 +; DAGISEL64-NEXT: scratch_load_b32 v0, off, s32 offset:4 +; DAGISEL64-NEXT: scratch_load_b32 v1, off, s32 offset:8 +; DAGISEL64-NEXT: scratch_load_b32 v49, off, s32 offset:16 +; DAGISEL64-NEXT: s_mov_b64 exec, vcc +; DAGISEL64-NEXT: s_wait_loadcnt 0x0 +; DAGISEL64-NEXT: s_wait_alu 0xf1ff +; DAGISEL64-NEXT: s_setpc_b64 s[30:31] +; +; GISEL64-LABEL: csr: +; GISEL64: ; %bb.0: +; GISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL64-NEXT: s_wait_expcnt 0x0 +; GISEL64-NEXT: s_wait_samplecnt 0x0 +; GISEL64-NEXT: s_wait_bvhcnt 0x0 +; GISEL64-NEXT: s_wait_kmcnt 0x0 +; GISEL64-NEXT: s_xor_saveexec_b64 vcc, -1 +; GISEL64-NEXT: s_clause 0x3 +; GISEL64-NEXT: scratch_store_b32 off, v2, s32 +; GISEL64-NEXT: scratch_store_b32 off, v0, s32 offset:4 +; GISEL64-NEXT: scratch_store_b32 off, v1, s32 offset:8 +; GISEL64-NEXT: scratch_store_b32 off, v49, s32 offset:16 +; GISEL64-NEXT: s_mov_b64 exec, -1 +; GISEL64-NEXT: scratch_store_b32 off, v40, s32 offset:12 ; 4-byte Folded Spill +; GISEL64-NEXT: ;;#ASMSTART +; GISEL64-NEXT: ; clobber CSR +; GISEL64-NEXT: ;;#ASMEND +; GISEL64-NEXT: v_writelane_b32 v2, s20, 0 +; GISEL64-NEXT: ;;#ASMSTART +; GISEL64-NEXT: ; clobber non-CSR +; GISEL64-NEXT: ;;#ASMEND +; GISEL64-NEXT: scratch_load_b32 v40, off, s32 offset:12 ; 4-byte Folded Reload +; GISEL64-NEXT: s_wait_alu 0xfffe +; GISEL64-NEXT: v_cndmask_b32_e32 v0, 5, v0, vcc +; GISEL64-NEXT: v_cndmask_b32_e32 v1, 3, v1, vcc +; GISEL64-NEXT: v_readlane_b32 s20, v2, 0 +; GISEL64-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GISEL64-NEXT: v_mov_b32_dpp v0, v1 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; GISEL64-NEXT: s_xor_b64 exec, vcc, -1 +; GISEL64-NEXT: s_clause 0x3 +; GISEL64-NEXT: scratch_load_b32 v2, off, s32 +; GISEL64-NEXT: scratch_load_b32 v0, off, s32 offset:4 +; GISEL64-NEXT: scratch_load_b32 v1, off, s32 offset:8 +; GISEL64-NEXT: scratch_load_b32 v49, off, s32 offset:16 +; GISEL64-NEXT: s_mov_b64 exec, vcc +; GISEL64-NEXT: s_wait_loadcnt 0x0 +; GISEL64-NEXT: s_wait_alu 0xf1ff +; GISEL64-NEXT: s_setpc_b64 s[30:31] + %x = select i1 %active, i32 %a, i32 5 + %y = select i1 %active, i32 %b, i32 3 + call void asm sideeffect "; clobber CSR", "~{v40},~{s48}"() + call void asm sideeffect "; clobber non-CSR", "~{v49},~{s20}"() + %ret = call i32 @llvm.amdgcn.update.dpp.i32(i32 %x, i32 %y, i32 1, i32 1, i32 1, i1 false) + ret i32 %ret +} + +; Save and restore all lanes of v40. +define amdgpu_gfx_whole_wave void @csr_vgpr_only(i1 %active, i32 %a, i32 %b) { +; DAGISEL-LABEL: csr_vgpr_only: +; DAGISEL: ; %bb.0: +; DAGISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL-NEXT: s_wait_expcnt 0x0 +; DAGISEL-NEXT: s_wait_samplecnt 0x0 +; DAGISEL-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL-NEXT: s_wait_kmcnt 0x0 +; DAGISEL-NEXT: s_or_saveexec_b32 s0, -1 +; DAGISEL-NEXT: scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill +; DAGISEL-NEXT: ;;#ASMSTART +; DAGISEL-NEXT: ; clobber CSR VGPR +; DAGISEL-NEXT: ;;#ASMEND +; DAGISEL-NEXT: scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload +; DAGISEL-NEXT: s_wait_alu 0xfffe +; DAGISEL-NEXT: s_mov_b32 exec_lo, s0 +; DAGISEL-NEXT: s_wait_loadcnt 0x0 +; DAGISEL-NEXT: s_setpc_b64 s[30:31] +; +; GISEL-LABEL: csr_vgpr_only: +; GISEL: ; %bb.0: +; GISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL-NEXT: s_wait_expcnt 0x0 +; GISEL-NEXT: s_wait_samplecnt 0x0 +; GISEL-NEXT: s_wait_bvhcnt 0x0 +; GISEL-NEXT: s_wait_kmcnt 0x0 +; GISEL-NEXT: s_or_saveexec_b32 s0, -1 +; GISEL-NEXT: scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill +; GISEL-NEXT: ;;#ASMSTART +; GISEL-NEXT: ; clobber CSR VGPR +; GISEL-NEXT: ;;#ASMEND +; GISEL-NEXT: scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload +; GISEL-NEXT: s_wait_alu 0xfffe +; GISEL-NEXT: s_mov_b32 exec_lo, s0 +; GISEL-NEXT: s_wait_loadcnt 0x0 +; GISEL-NEXT: s_setpc_b64 s[30:31] +; +; DAGISEL64-LABEL: csr_vgpr_only: +; DAGISEL64: ; %bb.0: +; DAGISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL64-NEXT: s_wait_expcnt 0x0 +; DAGISEL64-NEXT: s_wait_samplecnt 0x0 +; DAGISEL64-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL64-NEXT: s_wait_kmcnt 0x0 +; DAGISEL64-NEXT: s_or_saveexec_b64 s[0:1], -1 +; DAGISEL64-NEXT: scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill +; DAGISEL64-NEXT: ;;#ASMSTART +; DAGISEL64-NEXT: ; clobber CSR VGPR +; DAGISEL64-NEXT: ;;#ASMEND +; DAGISEL64-NEXT: scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload +; DAGISEL64-NEXT: s_wait_alu 0xfffe +; DAGISEL64-NEXT: s_mov_b64 exec, s[0:1] +; DAGISEL64-NEXT: s_wait_loadcnt 0x0 +; DAGISEL64-NEXT: s_setpc_b64 s[30:31] +; +; GISEL64-LABEL: csr_vgpr_only: +; GISEL64: ; %bb.0: +; GISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL64-NEXT: s_wait_expcnt 0x0 +; GISEL64-NEXT: s_wait_samplecnt 0x0 +; GISEL64-NEXT: s_wait_bvhcnt 0x0 +; GISEL64-NEXT: s_wait_kmcnt 0x0 +; GISEL64-NEXT: s_or_saveexec_b64 s[0:1], -1 +; GISEL64-NEXT: scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill +; GISEL64-NEXT: ;;#ASMSTART +; GISEL64-NEXT: ; clobber CSR VGPR +; GISEL64-NEXT: ;;#ASMEND +; GISEL64-NEXT: scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload +; GISEL64-NEXT: s_wait_alu 0xfffe +; GISEL64-NEXT: s_mov_b64 exec, s[0:1] +; GISEL64-NEXT: s_wait_loadcnt 0x0 +; GISEL64-NEXT: s_setpc_b64 s[30:31] + call void asm sideeffect "; clobber CSR VGPR", "~{v40}"() + ret void +} + +define amdgpu_gfx_whole_wave void @sgpr_spill_only(i1 %active, i32 %a, i32 %b) { +; DAGISEL-LABEL: sgpr_spill_only: +; DAGISEL: ; %bb.0: +; DAGISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL-NEXT: s_wait_expcnt 0x0 +; DAGISEL-NEXT: s_wait_samplecnt 0x0 +; DAGISEL-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL-NEXT: s_wait_kmcnt 0x0 +; DAGISEL-NEXT: s_xor_saveexec_b32 s0, -1 +; DAGISEL-NEXT: scratch_store_b32 off, v0, s32 ; 4-byte Folded Spill +; DAGISEL-NEXT: s_mov_b32 exec_lo, -1 +; DAGISEL-NEXT: v_writelane_b32 v0, s68, 0 +; DAGISEL-NEXT: ;;#ASMSTART +; DAGISEL-NEXT: ; clobber CSR SGPR +; DAGISEL-NEXT: ;;#ASMEND +; DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) +; DAGISEL-NEXT: v_readlane_b32 s68, v0, 0 +; DAGISEL-NEXT: s_wait_alu 0xfffe +; DAGISEL-NEXT: s_xor_b32 exec_lo, s0, -1 +; DAGISEL-NEXT: scratch_load_b32 v0, off, s32 ; 4-byte Folded Reload +; DAGISEL-NEXT: s_mov_b32 exec_lo, s0 +; DAGISEL-NEXT: s_wait_loadcnt 0x0 +; DAGISEL-NEXT: s_setpc_b64 s[30:31] +; +; GISEL-LABEL: sgpr_spill_only: +; GISEL: ; %bb.0: +; GISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL-NEXT: s_wait_expcnt 0x0 +; GISEL-NEXT: s_wait_samplecnt 0x0 +; GISEL-NEXT: s_wait_bvhcnt 0x0 +; GISEL-NEXT: s_wait_kmcnt 0x0 +; GISEL-NEXT: s_xor_saveexec_b32 s0, -1 +; GISEL-NEXT: scratch_store_b32 off, v0, s32 ; 4-byte Folded Spill +; GISEL-NEXT: s_mov_b32 exec_lo, -1 +; GISEL-NEXT: v_writelane_b32 v0, s68, 0 +; GISEL-NEXT: ;;#ASMSTART +; GISEL-NEXT: ; clobber CSR SGPR +; GISEL-NEXT: ;;#ASMEND +; GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GISEL-NEXT: v_readlane_b32 s68, v0, 0 +; GISEL-NEXT: s_wait_alu 0xfffe +; GISEL-NEXT: s_xor_b32 exec_lo, s0, -1 +; GISEL-NEXT: scratch_load_b32 v0, off, s32 ; 4-byte Folded Reload +; GISEL-NEXT: s_mov_b32 exec_lo, s0 +; GISEL-NEXT: s_wait_loadcnt 0x0 +; GISEL-NEXT: s_setpc_b64 s[30:31] +; +; DAGISEL64-LABEL: sgpr_spill_only: +; DAGISEL64: ; %bb.0: +; DAGISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL64-NEXT: s_wait_expcnt 0x0 +; DAGISEL64-NEXT: s_wait_samplecnt 0x0 +; DAGISEL64-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL64-NEXT: s_wait_kmcnt 0x0 +; DAGISEL64-NEXT: s_xor_saveexec_b64 s[0:1], -1 +; DAGISEL64-NEXT: scratch_store_b32 off, v0, s32 ; 4-byte Folded Spill +; DAGISEL64-NEXT: s_mov_b64 exec, -1 +; DAGISEL64-NEXT: v_writelane_b32 v0, s68, 0 +; DAGISEL64-NEXT: ;;#ASMSTART +; DAGISEL64-NEXT: ; clobber CSR SGPR +; DAGISEL64-NEXT: ;;#ASMEND +; DAGISEL64-NEXT: s_delay_alu instid0(VALU_DEP_1) +; DAGISEL64-NEXT: v_readlane_b32 s68, v0, 0 +; DAGISEL64-NEXT: s_wait_alu 0xfffe +; DAGISEL64-NEXT: s_xor_b64 exec, s[0:1], -1 +; DAGISEL64-NEXT: scratch_load_b32 v0, off, s32 ; 4-byte Folded Reload +; DAGISEL64-NEXT: s_mov_b64 exec, s[0:1] +; DAGISEL64-NEXT: s_wait_loadcnt 0x0 +; DAGISEL64-NEXT: s_setpc_b64 s[30:31] +; +; GISEL64-LABEL: sgpr_spill_only: +; GISEL64: ; %bb.0: +; GISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL64-NEXT: s_wait_expcnt 0x0 +; GISEL64-NEXT: s_wait_samplecnt 0x0 +; GISEL64-NEXT: s_wait_bvhcnt 0x0 +; GISEL64-NEXT: s_wait_kmcnt 0x0 +; GISEL64-NEXT: s_xor_saveexec_b64 s[0:1], -1 +; GISEL64-NEXT: scratch_store_b32 off, v0, s32 ; 4-byte Folded Spill +; GISEL64-NEXT: s_mov_b64 exec, -1 +; GISEL64-NEXT: v_writelane_b32 v0, s68, 0 +; GISEL64-NEXT: ;;#ASMSTART +; GISEL64-NEXT: ; clobber CSR SGPR +; GISEL64-NEXT: ;;#ASMEND +; GISEL64-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GISEL64-NEXT: v_readlane_b32 s68, v0, 0 +; GISEL64-NEXT: s_wait_alu 0xfffe +; GISEL64-NEXT: s_xor_b64 exec, s[0:1], -1 +; GISEL64-NEXT: scratch_load_b32 v0, off, s32 ; 4-byte Folded Reload +; GISEL64-NEXT: s_mov_b64 exec, s[0:1] +; GISEL64-NEXT: s_wait_loadcnt 0x0 +; GISEL64-NEXT: s_setpc_b64 s[30:31] + call void asm sideeffect "; clobber CSR SGPR", "~{s68}"() + ret void +} + +define amdgpu_gfx_whole_wave i32 @multiple_blocks(i1 %active, i32 %a, i32 %b) { +; DAGISEL-LABEL: multiple_blocks: +; DAGISEL: ; %bb.0: +; DAGISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL-NEXT: s_wait_expcnt 0x0 +; DAGISEL-NEXT: s_wait_samplecnt 0x0 +; DAGISEL-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL-NEXT: s_wait_kmcnt 0x0 +; DAGISEL-NEXT: s_xor_saveexec_b32 vcc_lo, -1 +; DAGISEL-NEXT: s_clause 0x1 +; DAGISEL-NEXT: scratch_store_b32 off, v0, s32 +; DAGISEL-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; DAGISEL-NEXT: s_mov_b32 exec_lo, -1 +; DAGISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) +; DAGISEL-NEXT: s_mov_b32 s1, exec_lo +; DAGISEL-NEXT: v_cmpx_eq_u32_e64 v0, v1 +; DAGISEL-NEXT: ; %bb.1: ; %if.then +; DAGISEL-NEXT: v_add_nc_u32_e32 v1, v0, v1 +; DAGISEL-NEXT: ; %bb.2: ; %if.end +; DAGISEL-NEXT: s_wait_alu 0xfffe +; DAGISEL-NEXT: s_or_b32 exec_lo, exec_lo, s1 +; DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) +; DAGISEL-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc_lo +; DAGISEL-NEXT: s_xor_b32 exec_lo, vcc_lo, -1 +; DAGISEL-NEXT: s_clause 0x1 +; DAGISEL-NEXT: scratch_load_b32 v0, off, s32 +; DAGISEL-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; DAGISEL-NEXT: s_mov_b32 exec_lo, vcc_lo +; DAGISEL-NEXT: s_wait_loadcnt 0x0 +; DAGISEL-NEXT: s_setpc_b64 s[30:31] +; +; GISEL-LABEL: multiple_blocks: +; GISEL: ; %bb.0: +; GISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL-NEXT: s_wait_expcnt 0x0 +; GISEL-NEXT: s_wait_samplecnt 0x0 +; GISEL-NEXT: s_wait_bvhcnt 0x0 +; GISEL-NEXT: s_wait_kmcnt 0x0 +; GISEL-NEXT: s_xor_saveexec_b32 vcc_lo, -1 +; GISEL-NEXT: s_clause 0x1 +; GISEL-NEXT: scratch_store_b32 off, v0, s32 +; GISEL-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; GISEL-NEXT: s_mov_b32 exec_lo, -1 +; GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) +; GISEL-NEXT: s_mov_b32 s1, exec_lo +; GISEL-NEXT: v_cmpx_eq_u32_e64 v0, v1 +; GISEL-NEXT: ; %bb.1: ; %if.then +; GISEL-NEXT: v_add_nc_u32_e32 v1, v0, v1 +; GISEL-NEXT: ; %bb.2: ; %if.end +; GISEL-NEXT: s_wait_alu 0xfffe +; GISEL-NEXT: s_or_b32 exec_lo, exec_lo, s1 +; GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GISEL-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc_lo +; GISEL-NEXT: s_xor_b32 exec_lo, vcc_lo, -1 +; GISEL-NEXT: s_clause 0x1 +; GISEL-NEXT: scratch_load_b32 v0, off, s32 +; GISEL-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; GISEL-NEXT: s_mov_b32 exec_lo, vcc_lo +; GISEL-NEXT: s_wait_loadcnt 0x0 +; GISEL-NEXT: s_setpc_b64 s[30:31] +; +; DAGISEL64-LABEL: multiple_blocks: +; DAGISEL64: ; %bb.0: +; DAGISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL64-NEXT: s_wait_expcnt 0x0 +; DAGISEL64-NEXT: s_wait_samplecnt 0x0 +; DAGISEL64-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL64-NEXT: s_wait_kmcnt 0x0 +; DAGISEL64-NEXT: s_xor_saveexec_b64 vcc, -1 +; DAGISEL64-NEXT: s_clause 0x1 +; DAGISEL64-NEXT: scratch_store_b32 off, v0, s32 +; DAGISEL64-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; DAGISEL64-NEXT: s_mov_b64 exec, -1 +; DAGISEL64-NEXT: s_delay_alu instid0(SALU_CYCLE_1) +; DAGISEL64-NEXT: s_mov_b64 s[2:3], exec +; DAGISEL64-NEXT: v_cmpx_eq_u32_e64 v0, v1 +; DAGISEL64-NEXT: ; %bb.1: ; %if.then +; DAGISEL64-NEXT: v_add_nc_u32_e32 v1, v0, v1 +; DAGISEL64-NEXT: ; %bb.2: ; %if.end +; DAGISEL64-NEXT: s_wait_alu 0xfffe +; DAGISEL64-NEXT: s_or_b64 exec, exec, s[2:3] +; DAGISEL64-NEXT: s_delay_alu instid0(VALU_DEP_1) +; DAGISEL64-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc +; DAGISEL64-NEXT: s_xor_b64 exec, vcc, -1 +; DAGISEL64-NEXT: s_clause 0x1 +; DAGISEL64-NEXT: scratch_load_b32 v0, off, s32 +; DAGISEL64-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; DAGISEL64-NEXT: s_mov_b64 exec, vcc +; DAGISEL64-NEXT: s_wait_loadcnt 0x0 +; DAGISEL64-NEXT: s_setpc_b64 s[30:31] +; +; GISEL64-LABEL: multiple_blocks: +; GISEL64: ; %bb.0: +; GISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL64-NEXT: s_wait_expcnt 0x0 +; GISEL64-NEXT: s_wait_samplecnt 0x0 +; GISEL64-NEXT: s_wait_bvhcnt 0x0 +; GISEL64-NEXT: s_wait_kmcnt 0x0 +; GISEL64-NEXT: s_xor_saveexec_b64 vcc, -1 +; GISEL64-NEXT: s_clause 0x1 +; GISEL64-NEXT: scratch_store_b32 off, v0, s32 +; GISEL64-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; GISEL64-NEXT: s_mov_b64 exec, -1 +; GISEL64-NEXT: s_delay_alu instid0(SALU_CYCLE_1) +; GISEL64-NEXT: s_mov_b64 s[2:3], exec +; GISEL64-NEXT: v_cmpx_eq_u32_e64 v0, v1 +; GISEL64-NEXT: ; %bb.1: ; %if.then +; GISEL64-NEXT: v_add_nc_u32_e32 v1, v0, v1 +; GISEL64-NEXT: ; %bb.2: ; %if.end +; GISEL64-NEXT: s_wait_alu 0xfffe +; GISEL64-NEXT: s_or_b64 exec, exec, s[2:3] +; GISEL64-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GISEL64-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc +; GISEL64-NEXT: s_xor_b64 exec, vcc, -1 +; GISEL64-NEXT: s_clause 0x1 +; GISEL64-NEXT: scratch_load_b32 v0, off, s32 +; GISEL64-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; GISEL64-NEXT: s_mov_b64 exec, vcc +; GISEL64-NEXT: s_wait_loadcnt 0x0 +; GISEL64-NEXT: s_setpc_b64 s[30:31] + %c = icmp eq i32 %a, %b + br i1 %c, label %if.then, label %if.end + +if.then: ; preds = %0 + %d = add i32 %a, %b + br label %if.end + +if.end: + %f = phi i32 [ %d, %if.then ], [ %b, %0 ] + %e = select i1 %active, i32 %a, i32 %f + ret i32 %e +} + +define amdgpu_gfx_whole_wave i64 @ret_64(i1 %active, i64 %a, i64 %b) { +; DAGISEL-LABEL: ret_64: +; DAGISEL: ; %bb.0: +; DAGISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL-NEXT: s_wait_expcnt 0x0 +; DAGISEL-NEXT: s_wait_samplecnt 0x0 +; DAGISEL-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL-NEXT: s_wait_kmcnt 0x0 +; DAGISEL-NEXT: s_xor_saveexec_b32 vcc_lo, -1 +; DAGISEL-NEXT: s_clause 0x3 +; DAGISEL-NEXT: scratch_store_b32 off, v0, s32 +; DAGISEL-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; DAGISEL-NEXT: scratch_store_b32 off, v2, s32 offset:8 +; DAGISEL-NEXT: scratch_store_b32 off, v3, s32 offset:12 +; DAGISEL-NEXT: s_mov_b32 exec_lo, -1 +; DAGISEL-NEXT: s_wait_alu 0xfffe +; DAGISEL-NEXT: v_dual_cndmask_b32 v1, 0, v1 :: v_dual_cndmask_b32 v0, 5, v0 +; DAGISEL-NEXT: v_dual_cndmask_b32 v2, 3, v2 :: v_dual_cndmask_b32 v3, 0, v3 +; DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2) +; DAGISEL-NEXT: v_mov_b32_dpp v0, v2 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; DAGISEL-NEXT: v_mov_b32_dpp v1, v3 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; DAGISEL-NEXT: s_xor_b32 exec_lo, vcc_lo, -1 +; DAGISEL-NEXT: s_clause 0x3 +; DAGISEL-NEXT: scratch_load_b32 v0, off, s32 +; DAGISEL-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; DAGISEL-NEXT: scratch_load_b32 v2, off, s32 offset:8 +; DAGISEL-NEXT: scratch_load_b32 v3, off, s32 offset:12 +; DAGISEL-NEXT: s_mov_b32 exec_lo, vcc_lo +; DAGISEL-NEXT: s_wait_loadcnt 0x0 +; DAGISEL-NEXT: s_setpc_b64 s[30:31] +; +; GISEL-LABEL: ret_64: +; GISEL: ; %bb.0: +; GISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL-NEXT: s_wait_expcnt 0x0 +; GISEL-NEXT: s_wait_samplecnt 0x0 +; GISEL-NEXT: s_wait_bvhcnt 0x0 +; GISEL-NEXT: s_wait_kmcnt 0x0 +; GISEL-NEXT: s_xor_saveexec_b32 vcc_lo, -1 +; GISEL-NEXT: s_clause 0x3 +; GISEL-NEXT: scratch_store_b32 off, v0, s32 +; GISEL-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; GISEL-NEXT: scratch_store_b32 off, v2, s32 offset:8 +; GISEL-NEXT: scratch_store_b32 off, v3, s32 offset:12 +; GISEL-NEXT: s_mov_b32 exec_lo, -1 +; GISEL-NEXT: s_wait_alu 0xfffe +; GISEL-NEXT: v_dual_cndmask_b32 v0, 5, v0 :: v_dual_cndmask_b32 v1, 0, v1 +; GISEL-NEXT: v_dual_cndmask_b32 v2, 3, v2 :: v_dual_cndmask_b32 v3, 0, v3 +; GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2) +; GISEL-NEXT: v_mov_b32_dpp v0, v2 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; GISEL-NEXT: v_mov_b32_dpp v1, v3 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; GISEL-NEXT: s_xor_b32 exec_lo, vcc_lo, -1 +; GISEL-NEXT: s_clause 0x3 +; GISEL-NEXT: scratch_load_b32 v0, off, s32 +; GISEL-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; GISEL-NEXT: scratch_load_b32 v2, off, s32 offset:8 +; GISEL-NEXT: scratch_load_b32 v3, off, s32 offset:12 +; GISEL-NEXT: s_mov_b32 exec_lo, vcc_lo +; GISEL-NEXT: s_wait_loadcnt 0x0 +; GISEL-NEXT: s_setpc_b64 s[30:31] +; +; DAGISEL64-LABEL: ret_64: +; DAGISEL64: ; %bb.0: +; DAGISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL64-NEXT: s_wait_expcnt 0x0 +; DAGISEL64-NEXT: s_wait_samplecnt 0x0 +; DAGISEL64-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL64-NEXT: s_wait_kmcnt 0x0 +; DAGISEL64-NEXT: s_xor_saveexec_b64 vcc, -1 +; DAGISEL64-NEXT: s_clause 0x3 +; DAGISEL64-NEXT: scratch_store_b32 off, v0, s32 +; DAGISEL64-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; DAGISEL64-NEXT: scratch_store_b32 off, v2, s32 offset:8 +; DAGISEL64-NEXT: scratch_store_b32 off, v3, s32 offset:12 +; DAGISEL64-NEXT: s_mov_b64 exec, -1 +; DAGISEL64-NEXT: s_wait_alu 0xfffe +; DAGISEL64-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc +; DAGISEL64-NEXT: v_cndmask_b32_e32 v0, 5, v0, vcc +; DAGISEL64-NEXT: v_cndmask_b32_e32 v2, 3, v2, vcc +; DAGISEL64-NEXT: v_cndmask_b32_e32 v3, 0, v3, vcc +; DAGISEL64-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; DAGISEL64-NEXT: v_mov_b32_dpp v0, v2 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; DAGISEL64-NEXT: v_mov_b32_dpp v1, v3 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; DAGISEL64-NEXT: s_xor_b64 exec, vcc, -1 +; DAGISEL64-NEXT: s_clause 0x3 +; DAGISEL64-NEXT: scratch_load_b32 v0, off, s32 +; DAGISEL64-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; DAGISEL64-NEXT: scratch_load_b32 v2, off, s32 offset:8 +; DAGISEL64-NEXT: scratch_load_b32 v3, off, s32 offset:12 +; DAGISEL64-NEXT: s_mov_b64 exec, vcc +; DAGISEL64-NEXT: s_wait_loadcnt 0x0 +; DAGISEL64-NEXT: s_setpc_b64 s[30:31] +; +; GISEL64-LABEL: ret_64: +; GISEL64: ; %bb.0: +; GISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL64-NEXT: s_wait_expcnt 0x0 +; GISEL64-NEXT: s_wait_samplecnt 0x0 +; GISEL64-NEXT: s_wait_bvhcnt 0x0 +; GISEL64-NEXT: s_wait_kmcnt 0x0 +; GISEL64-NEXT: s_xor_saveexec_b64 vcc, -1 +; GISEL64-NEXT: s_clause 0x3 +; GISEL64-NEXT: scratch_store_b32 off, v0, s32 +; GISEL64-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; GISEL64-NEXT: scratch_store_b32 off, v2, s32 offset:8 +; GISEL64-NEXT: scratch_store_b32 off, v3, s32 offset:12 +; GISEL64-NEXT: s_mov_b64 exec, -1 +; GISEL64-NEXT: s_wait_alu 0xfffe +; GISEL64-NEXT: v_cndmask_b32_e32 v0, 5, v0, vcc +; GISEL64-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc +; GISEL64-NEXT: v_cndmask_b32_e32 v2, 3, v2, vcc +; GISEL64-NEXT: v_cndmask_b32_e32 v3, 0, v3, vcc +; GISEL64-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GISEL64-NEXT: v_mov_b32_dpp v0, v2 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; GISEL64-NEXT: v_mov_b32_dpp v1, v3 quad_perm:[1,0,0,0] row_mask:0x1 bank_mask:0x1 +; GISEL64-NEXT: s_xor_b64 exec, vcc, -1 +; GISEL64-NEXT: s_clause 0x3 +; GISEL64-NEXT: scratch_load_b32 v0, off, s32 +; GISEL64-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; GISEL64-NEXT: scratch_load_b32 v2, off, s32 offset:8 +; GISEL64-NEXT: scratch_load_b32 v3, off, s32 offset:12 +; GISEL64-NEXT: s_mov_b64 exec, vcc +; GISEL64-NEXT: s_wait_loadcnt 0x0 +; GISEL64-NEXT: s_setpc_b64 s[30:31] + %x = select i1 %active, i64 %a, i64 5 + %y = select i1 %active, i64 %b, i64 3 + %ret = call i64 @llvm.amdgcn.update.dpp.i64(i64 %x, i64 %y, i32 1, i32 1, i32 1, i1 false) + ret i64 %ret +} + +define amdgpu_gfx_whole_wave void @inreg_args(i1 %active, i32 inreg %i32, <4 x i32> inreg %v4i32, float inreg %float, ptr addrspace(5) inreg %ptr, ptr addrspace(5) inreg %ptr2) { +; DAGISEL-LABEL: inreg_args: +; DAGISEL: ; %bb.0: +; DAGISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL-NEXT: s_wait_expcnt 0x0 +; DAGISEL-NEXT: s_wait_samplecnt 0x0 +; DAGISEL-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL-NEXT: s_wait_kmcnt 0x0 +; DAGISEL-NEXT: s_xor_saveexec_b32 s0, -1 +; DAGISEL-NEXT: s_clause 0x5 +; DAGISEL-NEXT: scratch_store_b32 off, v0, s32 +; DAGISEL-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; DAGISEL-NEXT: scratch_store_b32 off, v2, s32 offset:8 +; DAGISEL-NEXT: scratch_store_b32 off, v3, s32 offset:12 +; DAGISEL-NEXT: scratch_store_b32 off, v4, s32 offset:16 +; DAGISEL-NEXT: scratch_store_b32 off, v5, s32 offset:20 +; DAGISEL-NEXT: s_mov_b32 exec_lo, -1 +; DAGISEL-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s9 +; DAGISEL-NEXT: v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s6 +; DAGISEL-NEXT: v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v3, s8 +; DAGISEL-NEXT: scratch_store_b32 off, v4, s10 +; DAGISEL-NEXT: s_clause 0x1 +; DAGISEL-NEXT: scratch_store_b128 off, v[0:3], s11 +; DAGISEL-NEXT: scratch_store_b32 off, v5, s11 +; DAGISEL-NEXT: s_wait_alu 0xfffe +; DAGISEL-NEXT: s_xor_b32 exec_lo, s0, -1 +; DAGISEL-NEXT: s_clause 0x5 +; DAGISEL-NEXT: scratch_load_b32 v0, off, s32 +; DAGISEL-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; DAGISEL-NEXT: scratch_load_b32 v2, off, s32 offset:8 +; DAGISEL-NEXT: scratch_load_b32 v3, off, s32 offset:12 +; DAGISEL-NEXT: scratch_load_b32 v4, off, s32 offset:16 +; DAGISEL-NEXT: scratch_load_b32 v5, off, s32 offset:20 +; DAGISEL-NEXT: s_mov_b32 exec_lo, s0 +; DAGISEL-NEXT: s_wait_loadcnt 0x0 +; DAGISEL-NEXT: s_setpc_b64 s[30:31] +; +; GISEL-LABEL: inreg_args: +; GISEL: ; %bb.0: +; GISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL-NEXT: s_wait_expcnt 0x0 +; GISEL-NEXT: s_wait_samplecnt 0x0 +; GISEL-NEXT: s_wait_bvhcnt 0x0 +; GISEL-NEXT: s_wait_kmcnt 0x0 +; GISEL-NEXT: s_xor_saveexec_b32 s34, -1 +; GISEL-NEXT: s_clause 0x5 +; GISEL-NEXT: scratch_store_b32 off, v0, s32 +; GISEL-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; GISEL-NEXT: scratch_store_b32 off, v2, s32 offset:8 +; GISEL-NEXT: scratch_store_b32 off, v3, s32 offset:12 +; GISEL-NEXT: scratch_store_b32 off, v4, s32 offset:16 +; GISEL-NEXT: scratch_store_b32 off, v5, s32 offset:20 +; GISEL-NEXT: s_mov_b32 exec_lo, -1 +; GISEL-NEXT: s_mov_b32 s0, s5 +; GISEL-NEXT: s_mov_b32 s1, s6 +; GISEL-NEXT: s_mov_b32 s2, s7 +; GISEL-NEXT: s_mov_b32 s3, s8 +; GISEL-NEXT: v_mov_b32_e32 v4, s4 +; GISEL-NEXT: s_wait_alu 0xfffe +; GISEL-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v3, s3 +; GISEL-NEXT: v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v2, s2 +; GISEL-NEXT: v_mov_b32_e32 v5, s9 +; GISEL-NEXT: scratch_store_b32 off, v4, s10 +; GISEL-NEXT: s_clause 0x1 +; GISEL-NEXT: scratch_store_b128 off, v[0:3], s11 +; GISEL-NEXT: scratch_store_b32 off, v5, s11 +; GISEL-NEXT: s_xor_b32 exec_lo, s34, -1 +; GISEL-NEXT: s_clause 0x5 +; GISEL-NEXT: scratch_load_b32 v0, off, s32 +; GISEL-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; GISEL-NEXT: scratch_load_b32 v2, off, s32 offset:8 +; GISEL-NEXT: scratch_load_b32 v3, off, s32 offset:12 +; GISEL-NEXT: scratch_load_b32 v4, off, s32 offset:16 +; GISEL-NEXT: scratch_load_b32 v5, off, s32 offset:20 +; GISEL-NEXT: s_mov_b32 exec_lo, s34 +; GISEL-NEXT: s_wait_loadcnt 0x0 +; GISEL-NEXT: s_setpc_b64 s[30:31] +; +; DAGISEL64-LABEL: inreg_args: +; DAGISEL64: ; %bb.0: +; DAGISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL64-NEXT: s_wait_expcnt 0x0 +; DAGISEL64-NEXT: s_wait_samplecnt 0x0 +; DAGISEL64-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL64-NEXT: s_wait_kmcnt 0x0 +; DAGISEL64-NEXT: s_xor_saveexec_b64 s[0:1], -1 +; DAGISEL64-NEXT: s_clause 0x5 +; DAGISEL64-NEXT: scratch_store_b32 off, v0, s32 +; DAGISEL64-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; DAGISEL64-NEXT: scratch_store_b32 off, v2, s32 offset:8 +; DAGISEL64-NEXT: scratch_store_b32 off, v3, s32 offset:12 +; DAGISEL64-NEXT: scratch_store_b32 off, v4, s32 offset:16 +; DAGISEL64-NEXT: scratch_store_b32 off, v5, s32 offset:20 +; DAGISEL64-NEXT: s_mov_b64 exec, -1 +; DAGISEL64-NEXT: v_mov_b32_e32 v4, s4 +; DAGISEL64-NEXT: v_mov_b32_e32 v0, s5 +; DAGISEL64-NEXT: v_mov_b32_e32 v1, s6 +; DAGISEL64-NEXT: v_mov_b32_e32 v2, s7 +; DAGISEL64-NEXT: v_mov_b32_e32 v3, s8 +; DAGISEL64-NEXT: v_mov_b32_e32 v5, s9 +; DAGISEL64-NEXT: scratch_store_b32 off, v4, s10 +; DAGISEL64-NEXT: s_clause 0x1 +; DAGISEL64-NEXT: scratch_store_b128 off, v[0:3], s11 +; DAGISEL64-NEXT: scratch_store_b32 off, v5, s11 +; DAGISEL64-NEXT: s_wait_alu 0xfffe +; DAGISEL64-NEXT: s_xor_b64 exec, s[0:1], -1 +; DAGISEL64-NEXT: s_clause 0x5 +; DAGISEL64-NEXT: scratch_load_b32 v0, off, s32 +; DAGISEL64-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; DAGISEL64-NEXT: scratch_load_b32 v2, off, s32 offset:8 +; DAGISEL64-NEXT: scratch_load_b32 v3, off, s32 offset:12 +; DAGISEL64-NEXT: scratch_load_b32 v4, off, s32 offset:16 +; DAGISEL64-NEXT: scratch_load_b32 v5, off, s32 offset:20 +; DAGISEL64-NEXT: s_mov_b64 exec, s[0:1] +; DAGISEL64-NEXT: s_wait_loadcnt 0x0 +; DAGISEL64-NEXT: s_setpc_b64 s[30:31] +; +; GISEL64-LABEL: inreg_args: +; GISEL64: ; %bb.0: +; GISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL64-NEXT: s_wait_expcnt 0x0 +; GISEL64-NEXT: s_wait_samplecnt 0x0 +; GISEL64-NEXT: s_wait_bvhcnt 0x0 +; GISEL64-NEXT: s_wait_kmcnt 0x0 +; GISEL64-NEXT: s_xor_saveexec_b64 s[34:35], -1 +; GISEL64-NEXT: s_clause 0x5 +; GISEL64-NEXT: scratch_store_b32 off, v0, s32 +; GISEL64-NEXT: scratch_store_b32 off, v1, s32 offset:4 +; GISEL64-NEXT: scratch_store_b32 off, v2, s32 offset:8 +; GISEL64-NEXT: scratch_store_b32 off, v3, s32 offset:12 +; GISEL64-NEXT: scratch_store_b32 off, v4, s32 offset:16 +; GISEL64-NEXT: scratch_store_b32 off, v5, s32 offset:20 +; GISEL64-NEXT: s_mov_b64 exec, -1 +; GISEL64-NEXT: s_mov_b32 s0, s5 +; GISEL64-NEXT: s_mov_b32 s1, s6 +; GISEL64-NEXT: s_mov_b32 s2, s7 +; GISEL64-NEXT: s_mov_b32 s3, s8 +; GISEL64-NEXT: v_mov_b32_e32 v4, s4 +; GISEL64-NEXT: s_wait_alu 0xfffe +; GISEL64-NEXT: v_mov_b32_e32 v0, s0 +; GISEL64-NEXT: v_mov_b32_e32 v1, s1 +; GISEL64-NEXT: v_mov_b32_e32 v2, s2 +; GISEL64-NEXT: v_mov_b32_e32 v3, s3 +; GISEL64-NEXT: v_mov_b32_e32 v5, s9 +; GISEL64-NEXT: scratch_store_b32 off, v4, s10 +; GISEL64-NEXT: s_clause 0x1 +; GISEL64-NEXT: scratch_store_b128 off, v[0:3], s11 +; GISEL64-NEXT: scratch_store_b32 off, v5, s11 +; GISEL64-NEXT: s_xor_b64 exec, s[34:35], -1 +; GISEL64-NEXT: s_clause 0x5 +; GISEL64-NEXT: scratch_load_b32 v0, off, s32 +; GISEL64-NEXT: scratch_load_b32 v1, off, s32 offset:4 +; GISEL64-NEXT: scratch_load_b32 v2, off, s32 offset:8 +; GISEL64-NEXT: scratch_load_b32 v3, off, s32 offset:12 +; GISEL64-NEXT: scratch_load_b32 v4, off, s32 offset:16 +; GISEL64-NEXT: scratch_load_b32 v5, off, s32 offset:20 +; GISEL64-NEXT: s_mov_b64 exec, s[34:35] +; GISEL64-NEXT: s_wait_loadcnt 0x0 +; GISEL64-NEXT: s_setpc_b64 s[30:31] + store i32 %i32, ptr addrspace(5) %ptr + store <4 x i32> %v4i32, ptr addrspace(5) %ptr2 + store float %float, ptr addrspace(5) %ptr2 + ret void +} + +declare amdgpu_gfx <2 x half> @gfx_callee(<2 x half> %x, <2 x half> %y) + +define amdgpu_gfx_whole_wave <2 x half> @call_gfx_from_whole_wave(i1 %active, <2 x half> %x, <2 x half> %y) { +; DAGISEL-LABEL: call_gfx_from_whole_wave: +; DAGISEL: ; %bb.0: +; DAGISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL-NEXT: s_wait_expcnt 0x0 +; DAGISEL-NEXT: s_wait_samplecnt 0x0 +; DAGISEL-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL-NEXT: s_wait_kmcnt 0x0 +; DAGISEL-NEXT: s_mov_b32 s0, s33 +; DAGISEL-NEXT: s_mov_b32 s33, s32 +; DAGISEL-NEXT: s_xor_saveexec_b32 s4, -1 +; DAGISEL-NEXT: s_clause 0x1f +; DAGISEL-NEXT: scratch_store_b32 off, v0, s33 offset:4 +; DAGISEL-NEXT: scratch_store_b32 off, v1, s33 offset:8 +; DAGISEL-NEXT: scratch_store_b32 off, v2, s33 offset:12 +; DAGISEL-NEXT: scratch_store_b32 off, v3, s33 offset:16 +; DAGISEL-NEXT: scratch_store_b32 off, v4, s33 offset:20 +; DAGISEL-NEXT: scratch_store_b32 off, v5, s33 offset:24 +; DAGISEL-NEXT: scratch_store_b32 off, v6, s33 offset:28 +; DAGISEL-NEXT: scratch_store_b32 off, v7, s33 offset:32 +; DAGISEL-NEXT: scratch_store_b32 off, v8, s33 offset:36 +; DAGISEL-NEXT: scratch_store_b32 off, v9, s33 offset:40 +; DAGISEL-NEXT: scratch_store_b32 off, v10, s33 offset:44 +; DAGISEL-NEXT: scratch_store_b32 off, v11, s33 offset:48 +; DAGISEL-NEXT: scratch_store_b32 off, v12, s33 offset:52 +; DAGISEL-NEXT: scratch_store_b32 off, v13, s33 offset:56 +; DAGISEL-NEXT: scratch_store_b32 off, v14, s33 offset:60 +; DAGISEL-NEXT: scratch_store_b32 off, v15, s33 offset:64 +; DAGISEL-NEXT: scratch_store_b32 off, v16, s33 offset:68 +; DAGISEL-NEXT: scratch_store_b32 off, v17, s33 offset:72 +; DAGISEL-NEXT: scratch_store_b32 off, v18, s33 offset:76 +; DAGISEL-NEXT: scratch_store_b32 off, v19, s33 offset:80 +; DAGISEL-NEXT: scratch_store_b32 off, v20, s33 offset:84 +; DAGISEL-NEXT: scratch_store_b32 off, v21, s33 offset:88 +; DAGISEL-NEXT: scratch_store_b32 off, v22, s33 offset:92 +; DAGISEL-NEXT: scratch_store_b32 off, v23, s33 offset:96 +; DAGISEL-NEXT: scratch_store_b32 off, v24, s33 offset:100 +; DAGISEL-NEXT: scratch_store_b32 off, v25, s33 offset:104 +; DAGISEL-NEXT: scratch_store_b32 off, v26, s33 offset:108 +; DAGISEL-NEXT: scratch_store_b32 off, v27, s33 offset:112 +; DAGISEL-NEXT: scratch_store_b32 off, v28, s33 offset:116 +; DAGISEL-NEXT: scratch_store_b32 off, v29, s33 offset:120 +; DAGISEL-NEXT: scratch_store_b32 off, v30, s33 offset:124 +; DAGISEL-NEXT: scratch_store_b32 off, v31, s33 offset:128 +; DAGISEL-NEXT: s_clause 0x1f +; DAGISEL-NEXT: scratch_store_b32 off, v32, s33 offset:132 +; DAGISEL-NEXT: scratch_store_b32 off, v33, s33 offset:136 +; DAGISEL-NEXT: scratch_store_b32 off, v34, s33 offset:140 +; DAGISEL-NEXT: scratch_store_b32 off, v35, s33 offset:144 +; DAGISEL-NEXT: scratch_store_b32 off, v36, s33 offset:148 +; DAGISEL-NEXT: scratch_store_b32 off, v37, s33 offset:152 +; DAGISEL-NEXT: scratch_store_b32 off, v38, s33 offset:156 +; DAGISEL-NEXT: scratch_store_b32 off, v39, s33 offset:160 +; DAGISEL-NEXT: scratch_store_b32 off, v48, s33 offset:164 +; DAGISEL-NEXT: scratch_store_b32 off, v49, s33 offset:168 +; DAGISEL-NEXT: scratch_store_b32 off, v50, s33 offset:172 +; DAGISEL-NEXT: scratch_store_b32 off, v51, s33 offset:176 +; DAGISEL-NEXT: scratch_store_b32 off, v52, s33 offset:180 +; DAGISEL-NEXT: scratch_store_b32 off, v53, s33 offset:184 +; DAGISEL-NEXT: scratch_store_b32 off, v54, s33 offset:188 +; DAGISEL-NEXT: scratch_store_b32 off, v55, s33 offset:192 +; DAGISEL-NEXT: scratch_store_b32 off, v64, s33 offset:196 +; DAGISEL-NEXT: scratch_store_b32 off, v65, s33 offset:200 +; DAGISEL-NEXT: scratch_store_b32 off, v66, s33 offset:204 +; DAGISEL-NEXT: scratch_store_b32 off, v67, s33 offset:208 +; DAGISEL-NEXT: scratch_store_b32 off, v68, s33 offset:212 +; DAGISEL-NEXT: scratch_store_b32 off, v69, s33 offset:216 +; DAGISEL-NEXT: scratch_store_b32 off, v70, s33 offset:220 +; DAGISEL-NEXT: scratch_store_b32 off, v71, s33 offset:224 +; DAGISEL-NEXT: scratch_store_b32 off, v80, s33 offset:228 +; DAGISEL-NEXT: scratch_store_b32 off, v81, s33 offset:232 +; DAGISEL-NEXT: scratch_store_b32 off, v82, s33 offset:236 +; DAGISEL-NEXT: scratch_store_b32 off, v83, s33 offset:240 +; DAGISEL-NEXT: scratch_store_b32 off, v84, s33 offset:244 +; DAGISEL-NEXT: scratch_store_b32 off, v85, s33 offset:248 +; DAGISEL-NEXT: scratch_store_b32 off, v86, s33 offset:252 +; DAGISEL-NEXT: scratch_store_b32 off, v87, s33 offset:256 +; DAGISEL-NEXT: s_clause 0x1f +; DAGISEL-NEXT: scratch_store_b32 off, v96, s33 offset:260 +; DAGISEL-NEXT: scratch_store_b32 off, v97, s33 offset:264 +; DAGISEL-NEXT: scratch_store_b32 off, v98, s33 offset:268 +; DAGISEL-NEXT: scratch_store_b32 off, v99, s33 offset:272 +; DAGISEL-NEXT: scratch_store_b32 off, v100, s33 offset:276 +; DAGISEL-NEXT: scratch_store_b32 off, v101, s33 offset:280 +; DAGISEL-NEXT: scratch_store_b32 off, v102, s33 offset:284 +; DAGISEL-NEXT: scratch_store_b32 off, v103, s33 offset:288 +; DAGISEL-NEXT: scratch_store_b32 off, v112, s33 offset:292 +; DAGISEL-NEXT: scratch_store_b32 off, v113, s33 offset:296 +; DAGISEL-NEXT: scratch_store_b32 off, v114, s33 offset:300 +; DAGISEL-NEXT: scratch_store_b32 off, v115, s33 offset:304 +; DAGISEL-NEXT: scratch_store_b32 off, v116, s33 offset:308 +; DAGISEL-NEXT: scratch_store_b32 off, v117, s33 offset:312 +; DAGISEL-NEXT: scratch_store_b32 off, v118, s33 offset:316 +; DAGISEL-NEXT: scratch_store_b32 off, v119, s33 offset:320 +; DAGISEL-NEXT: scratch_store_b32 off, v128, s33 offset:324 +; DAGISEL-NEXT: scratch_store_b32 off, v129, s33 offset:328 +; DAGISEL-NEXT: scratch_store_b32 off, v130, s33 offset:332 +; DAGISEL-NEXT: scratch_store_b32 off, v131, s33 offset:336 +; DAGISEL-NEXT: scratch_store_b32 off, v132, s33 offset:340 +; DAGISEL-NEXT: scratch_store_b32 off, v133, s33 offset:344 +; DAGISEL-NEXT: scratch_store_b32 off, v134, s33 offset:348 +; DAGISEL-NEXT: scratch_store_b32 off, v135, s33 offset:352 +; DAGISEL-NEXT: scratch_store_b32 off, v144, s33 offset:356 +; DAGISEL-NEXT: scratch_store_b32 off, v145, s33 offset:360 +; DAGISEL-NEXT: scratch_store_b32 off, v146, s33 offset:364 +; DAGISEL-NEXT: scratch_store_b32 off, v147, s33 offset:368 +; DAGISEL-NEXT: scratch_store_b32 off, v148, s33 offset:372 +; DAGISEL-NEXT: scratch_store_b32 off, v149, s33 offset:376 +; DAGISEL-NEXT: scratch_store_b32 off, v150, s33 offset:380 +; DAGISEL-NEXT: scratch_store_b32 off, v151, s33 offset:384 +; DAGISEL-NEXT: s_clause 0x1f +; DAGISEL-NEXT: scratch_store_b32 off, v160, s33 offset:388 +; DAGISEL-NEXT: scratch_store_b32 off, v161, s33 offset:392 +; DAGISEL-NEXT: scratch_store_b32 off, v162, s33 offset:396 +; DAGISEL-NEXT: scratch_store_b32 off, v163, s33 offset:400 +; DAGISEL-NEXT: scratch_store_b32 off, v164, s33 offset:404 +; DAGISEL-NEXT: scratch_store_b32 off, v165, s33 offset:408 +; DAGISEL-NEXT: scratch_store_b32 off, v166, s33 offset:412 +; DAGISEL-NEXT: scratch_store_b32 off, v167, s33 offset:416 +; DAGISEL-NEXT: scratch_store_b32 off, v176, s33 offset:420 +; DAGISEL-NEXT: scratch_store_b32 off, v177, s33 offset:424 +; DAGISEL-NEXT: scratch_store_b32 off, v178, s33 offset:428 +; DAGISEL-NEXT: scratch_store_b32 off, v179, s33 offset:432 +; DAGISEL-NEXT: scratch_store_b32 off, v180, s33 offset:436 +; DAGISEL-NEXT: scratch_store_b32 off, v181, s33 offset:440 +; DAGISEL-NEXT: scratch_store_b32 off, v182, s33 offset:444 +; DAGISEL-NEXT: scratch_store_b32 off, v183, s33 offset:448 +; DAGISEL-NEXT: scratch_store_b32 off, v192, s33 offset:452 +; DAGISEL-NEXT: scratch_store_b32 off, v193, s33 offset:456 +; DAGISEL-NEXT: scratch_store_b32 off, v194, s33 offset:460 +; DAGISEL-NEXT: scratch_store_b32 off, v195, s33 offset:464 +; DAGISEL-NEXT: scratch_store_b32 off, v196, s33 offset:468 +; DAGISEL-NEXT: scratch_store_b32 off, v197, s33 offset:472 +; DAGISEL-NEXT: scratch_store_b32 off, v198, s33 offset:476 +; DAGISEL-NEXT: scratch_store_b32 off, v199, s33 offset:480 +; DAGISEL-NEXT: scratch_store_b32 off, v208, s33 offset:484 +; DAGISEL-NEXT: scratch_store_b32 off, v209, s33 offset:488 +; DAGISEL-NEXT: scratch_store_b32 off, v210, s33 offset:492 +; DAGISEL-NEXT: scratch_store_b32 off, v211, s33 offset:496 +; DAGISEL-NEXT: scratch_store_b32 off, v212, s33 offset:500 +; DAGISEL-NEXT: scratch_store_b32 off, v213, s33 offset:504 +; DAGISEL-NEXT: scratch_store_b32 off, v214, s33 offset:508 +; DAGISEL-NEXT: scratch_store_b32 off, v215, s33 offset:512 +; DAGISEL-NEXT: s_clause 0xf +; DAGISEL-NEXT: scratch_store_b32 off, v224, s33 offset:516 +; DAGISEL-NEXT: scratch_store_b32 off, v225, s33 offset:520 +; DAGISEL-NEXT: scratch_store_b32 off, v226, s33 offset:524 +; DAGISEL-NEXT: scratch_store_b32 off, v227, s33 offset:528 +; DAGISEL-NEXT: scratch_store_b32 off, v228, s33 offset:532 +; DAGISEL-NEXT: scratch_store_b32 off, v229, s33 offset:536 +; DAGISEL-NEXT: scratch_store_b32 off, v230, s33 offset:540 +; DAGISEL-NEXT: scratch_store_b32 off, v231, s33 offset:544 +; DAGISEL-NEXT: scratch_store_b32 off, v240, s33 offset:548 +; DAGISEL-NEXT: scratch_store_b32 off, v241, s33 offset:552 +; DAGISEL-NEXT: scratch_store_b32 off, v242, s33 offset:556 +; DAGISEL-NEXT: scratch_store_b32 off, v243, s33 offset:560 +; DAGISEL-NEXT: scratch_store_b32 off, v244, s33 offset:564 +; DAGISEL-NEXT: scratch_store_b32 off, v245, s33 offset:568 +; DAGISEL-NEXT: scratch_store_b32 off, v246, s33 offset:572 +; DAGISEL-NEXT: scratch_store_b32 off, v247, s33 offset:576 +; DAGISEL-NEXT: s_mov_b32 exec_lo, -1 +; DAGISEL-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill +; DAGISEL-NEXT: s_wait_alu 0xfffe +; DAGISEL-NEXT: v_writelane_b32 v40, s0, 3 +; DAGISEL-NEXT: v_mov_b32_e32 v2, v0 +; DAGISEL-NEXT: v_swap_b32 v0, v1 +; DAGISEL-NEXT: s_mov_b32 s1, gfx_callee@abs32@hi +; DAGISEL-NEXT: v_writelane_b32 v40, s4, 0 +; DAGISEL-NEXT: s_mov_b32 s0, gfx_callee@abs32@lo +; DAGISEL-NEXT: s_addk_co_i32 s32, 0x250 +; DAGISEL-NEXT: v_writelane_b32 v40, s30, 1 +; DAGISEL-NEXT: v_writelane_b32 v40, s31, 2 +; DAGISEL-NEXT: s_wait_alu 0xfffe +; DAGISEL-NEXT: s_swappc_b64 s[30:31], s[0:1] +; DAGISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) +; DAGISEL-NEXT: v_readlane_b32 s31, v40, 2 +; DAGISEL-NEXT: v_readlane_b32 s30, v40, 1 +; DAGISEL-NEXT: v_readlane_b32 s4, v40, 0 +; DAGISEL-NEXT: v_readlane_b32 s0, v40, 3 +; DAGISEL-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload +; DAGISEL-NEXT: s_mov_b32 s32, s33 +; DAGISEL-NEXT: s_xor_b32 exec_lo, s4, -1 +; DAGISEL-NEXT: s_clause 0x1f +; DAGISEL-NEXT: scratch_load_b32 v0, off, s33 offset:4 +; DAGISEL-NEXT: scratch_load_b32 v1, off, s33 offset:8 +; DAGISEL-NEXT: scratch_load_b32 v2, off, s33 offset:12 +; DAGISEL-NEXT: scratch_load_b32 v3, off, s33 offset:16 +; DAGISEL-NEXT: scratch_load_b32 v4, off, s33 offset:20 +; DAGISEL-NEXT: scratch_load_b32 v5, off, s33 offset:24 +; DAGISEL-NEXT: scratch_load_b32 v6, off, s33 offset:28 +; DAGISEL-NEXT: scratch_load_b32 v7, off, s33 offset:32 +; DAGISEL-NEXT: scratch_load_b32 v8, off, s33 offset:36 +; DAGISEL-NEXT: scratch_load_b32 v9, off, s33 offset:40 +; DAGISEL-NEXT: scratch_load_b32 v10, off, s33 offset:44 +; DAGISEL-NEXT: scratch_load_b32 v11, off, s33 offset:48 +; DAGISEL-NEXT: scratch_load_b32 v12, off, s33 offset:52 +; DAGISEL-NEXT: scratch_load_b32 v13, off, s33 offset:56 +; DAGISEL-NEXT: scratch_load_b32 v14, off, s33 offset:60 +; DAGISEL-NEXT: scratch_load_b32 v15, off, s33 offset:64 +; DAGISEL-NEXT: scratch_load_b32 v16, off, s33 offset:68 +; DAGISEL-NEXT: scratch_load_b32 v17, off, s33 offset:72 +; DAGISEL-NEXT: scratch_load_b32 v18, off, s33 offset:76 +; DAGISEL-NEXT: scratch_load_b32 v19, off, s33 offset:80 +; DAGISEL-NEXT: scratch_load_b32 v20, off, s33 offset:84 +; DAGISEL-NEXT: scratch_load_b32 v21, off, s33 offset:88 +; DAGISEL-NEXT: scratch_load_b32 v22, off, s33 offset:92 +; DAGISEL-NEXT: scratch_load_b32 v23, off, s33 offset:96 +; DAGISEL-NEXT: scratch_load_b32 v24, off, s33 offset:100 +; DAGISEL-NEXT: scratch_load_b32 v25, off, s33 offset:104 +; DAGISEL-NEXT: scratch_load_b32 v26, off, s33 offset:108 +; DAGISEL-NEXT: scratch_load_b32 v27, off, s33 offset:112 +; DAGISEL-NEXT: scratch_load_b32 v28, off, s33 offset:116 +; DAGISEL-NEXT: scratch_load_b32 v29, off, s33 offset:120 +; DAGISEL-NEXT: scratch_load_b32 v30, off, s33 offset:124 +; DAGISEL-NEXT: scratch_load_b32 v31, off, s33 offset:128 +; DAGISEL-NEXT: s_clause 0x1f +; DAGISEL-NEXT: scratch_load_b32 v32, off, s33 offset:132 +; DAGISEL-NEXT: scratch_load_b32 v33, off, s33 offset:136 +; DAGISEL-NEXT: scratch_load_b32 v34, off, s33 offset:140 +; DAGISEL-NEXT: scratch_load_b32 v35, off, s33 offset:144 +; DAGISEL-NEXT: scratch_load_b32 v36, off, s33 offset:148 +; DAGISEL-NEXT: scratch_load_b32 v37, off, s33 offset:152 +; DAGISEL-NEXT: scratch_load_b32 v38, off, s33 offset:156 +; DAGISEL-NEXT: scratch_load_b32 v39, off, s33 offset:160 +; DAGISEL-NEXT: scratch_load_b32 v48, off, s33 offset:164 +; DAGISEL-NEXT: scratch_load_b32 v49, off, s33 offset:168 +; DAGISEL-NEXT: scratch_load_b32 v50, off, s33 offset:172 +; DAGISEL-NEXT: scratch_load_b32 v51, off, s33 offset:176 +; DAGISEL-NEXT: scratch_load_b32 v52, off, s33 offset:180 +; DAGISEL-NEXT: scratch_load_b32 v53, off, s33 offset:184 +; DAGISEL-NEXT: scratch_load_b32 v54, off, s33 offset:188 +; DAGISEL-NEXT: scratch_load_b32 v55, off, s33 offset:192 +; DAGISEL-NEXT: scratch_load_b32 v64, off, s33 offset:196 +; DAGISEL-NEXT: scratch_load_b32 v65, off, s33 offset:200 +; DAGISEL-NEXT: scratch_load_b32 v66, off, s33 offset:204 +; DAGISEL-NEXT: scratch_load_b32 v67, off, s33 offset:208 +; DAGISEL-NEXT: scratch_load_b32 v68, off, s33 offset:212 +; DAGISEL-NEXT: scratch_load_b32 v69, off, s33 offset:216 +; DAGISEL-NEXT: scratch_load_b32 v70, off, s33 offset:220 +; DAGISEL-NEXT: scratch_load_b32 v71, off, s33 offset:224 +; DAGISEL-NEXT: scratch_load_b32 v80, off, s33 offset:228 +; DAGISEL-NEXT: scratch_load_b32 v81, off, s33 offset:232 +; DAGISEL-NEXT: scratch_load_b32 v82, off, s33 offset:236 +; DAGISEL-NEXT: scratch_load_b32 v83, off, s33 offset:240 +; DAGISEL-NEXT: scratch_load_b32 v84, off, s33 offset:244 +; DAGISEL-NEXT: scratch_load_b32 v85, off, s33 offset:248 +; DAGISEL-NEXT: scratch_load_b32 v86, off, s33 offset:252 +; DAGISEL-NEXT: scratch_load_b32 v87, off, s33 offset:256 +; DAGISEL-NEXT: s_clause 0x1f +; DAGISEL-NEXT: scratch_load_b32 v96, off, s33 offset:260 +; DAGISEL-NEXT: scratch_load_b32 v97, off, s33 offset:264 +; DAGISEL-NEXT: scratch_load_b32 v98, off, s33 offset:268 +; DAGISEL-NEXT: scratch_load_b32 v99, off, s33 offset:272 +; DAGISEL-NEXT: scratch_load_b32 v100, off, s33 offset:276 +; DAGISEL-NEXT: scratch_load_b32 v101, off, s33 offset:280 +; DAGISEL-NEXT: scratch_load_b32 v102, off, s33 offset:284 +; DAGISEL-NEXT: scratch_load_b32 v103, off, s33 offset:288 +; DAGISEL-NEXT: scratch_load_b32 v112, off, s33 offset:292 +; DAGISEL-NEXT: scratch_load_b32 v113, off, s33 offset:296 +; DAGISEL-NEXT: scratch_load_b32 v114, off, s33 offset:300 +; DAGISEL-NEXT: scratch_load_b32 v115, off, s33 offset:304 +; DAGISEL-NEXT: scratch_load_b32 v116, off, s33 offset:308 +; DAGISEL-NEXT: scratch_load_b32 v117, off, s33 offset:312 +; DAGISEL-NEXT: scratch_load_b32 v118, off, s33 offset:316 +; DAGISEL-NEXT: scratch_load_b32 v119, off, s33 offset:320 +; DAGISEL-NEXT: scratch_load_b32 v128, off, s33 offset:324 +; DAGISEL-NEXT: scratch_load_b32 v129, off, s33 offset:328 +; DAGISEL-NEXT: scratch_load_b32 v130, off, s33 offset:332 +; DAGISEL-NEXT: scratch_load_b32 v131, off, s33 offset:336 +; DAGISEL-NEXT: scratch_load_b32 v132, off, s33 offset:340 +; DAGISEL-NEXT: scratch_load_b32 v133, off, s33 offset:344 +; DAGISEL-NEXT: scratch_load_b32 v134, off, s33 offset:348 +; DAGISEL-NEXT: scratch_load_b32 v135, off, s33 offset:352 +; DAGISEL-NEXT: scratch_load_b32 v144, off, s33 offset:356 +; DAGISEL-NEXT: scratch_load_b32 v145, off, s33 offset:360 +; DAGISEL-NEXT: scratch_load_b32 v146, off, s33 offset:364 +; DAGISEL-NEXT: scratch_load_b32 v147, off, s33 offset:368 +; DAGISEL-NEXT: scratch_load_b32 v148, off, s33 offset:372 +; DAGISEL-NEXT: scratch_load_b32 v149, off, s33 offset:376 +; DAGISEL-NEXT: scratch_load_b32 v150, off, s33 offset:380 +; DAGISEL-NEXT: scratch_load_b32 v151, off, s33 offset:384 +; DAGISEL-NEXT: s_clause 0x1f +; DAGISEL-NEXT: scratch_load_b32 v160, off, s33 offset:388 +; DAGISEL-NEXT: scratch_load_b32 v161, off, s33 offset:392 +; DAGISEL-NEXT: scratch_load_b32 v162, off, s33 offset:396 +; DAGISEL-NEXT: scratch_load_b32 v163, off, s33 offset:400 +; DAGISEL-NEXT: scratch_load_b32 v164, off, s33 offset:404 +; DAGISEL-NEXT: scratch_load_b32 v165, off, s33 offset:408 +; DAGISEL-NEXT: scratch_load_b32 v166, off, s33 offset:412 +; DAGISEL-NEXT: scratch_load_b32 v167, off, s33 offset:416 +; DAGISEL-NEXT: scratch_load_b32 v176, off, s33 offset:420 +; DAGISEL-NEXT: scratch_load_b32 v177, off, s33 offset:424 +; DAGISEL-NEXT: scratch_load_b32 v178, off, s33 offset:428 +; DAGISEL-NEXT: scratch_load_b32 v179, off, s33 offset:432 +; DAGISEL-NEXT: scratch_load_b32 v180, off, s33 offset:436 +; DAGISEL-NEXT: scratch_load_b32 v181, off, s33 offset:440 +; DAGISEL-NEXT: scratch_load_b32 v182, off, s33 offset:444 +; DAGISEL-NEXT: scratch_load_b32 v183, off, s33 offset:448 +; DAGISEL-NEXT: scratch_load_b32 v192, off, s33 offset:452 +; DAGISEL-NEXT: scratch_load_b32 v193, off, s33 offset:456 +; DAGISEL-NEXT: scratch_load_b32 v194, off, s33 offset:460 +; DAGISEL-NEXT: scratch_load_b32 v195, off, s33 offset:464 +; DAGISEL-NEXT: scratch_load_b32 v196, off, s33 offset:468 +; DAGISEL-NEXT: scratch_load_b32 v197, off, s33 offset:472 +; DAGISEL-NEXT: scratch_load_b32 v198, off, s33 offset:476 +; DAGISEL-NEXT: scratch_load_b32 v199, off, s33 offset:480 +; DAGISEL-NEXT: scratch_load_b32 v208, off, s33 offset:484 +; DAGISEL-NEXT: scratch_load_b32 v209, off, s33 offset:488 +; DAGISEL-NEXT: scratch_load_b32 v210, off, s33 offset:492 +; DAGISEL-NEXT: scratch_load_b32 v211, off, s33 offset:496 +; DAGISEL-NEXT: scratch_load_b32 v212, off, s33 offset:500 +; DAGISEL-NEXT: scratch_load_b32 v213, off, s33 offset:504 +; DAGISEL-NEXT: scratch_load_b32 v214, off, s33 offset:508 +; DAGISEL-NEXT: scratch_load_b32 v215, off, s33 offset:512 +; DAGISEL-NEXT: s_clause 0xf +; DAGISEL-NEXT: scratch_load_b32 v224, off, s33 offset:516 +; DAGISEL-NEXT: scratch_load_b32 v225, off, s33 offset:520 +; DAGISEL-NEXT: scratch_load_b32 v226, off, s33 offset:524 +; DAGISEL-NEXT: scratch_load_b32 v227, off, s33 offset:528 +; DAGISEL-NEXT: scratch_load_b32 v228, off, s33 offset:532 +; DAGISEL-NEXT: scratch_load_b32 v229, off, s33 offset:536 +; DAGISEL-NEXT: scratch_load_b32 v230, off, s33 offset:540 +; DAGISEL-NEXT: scratch_load_b32 v231, off, s33 offset:544 +; DAGISEL-NEXT: scratch_load_b32 v240, off, s33 offset:548 +; DAGISEL-NEXT: scratch_load_b32 v241, off, s33 offset:552 +; DAGISEL-NEXT: scratch_load_b32 v242, off, s33 offset:556 +; DAGISEL-NEXT: scratch_load_b32 v243, off, s33 offset:560 +; DAGISEL-NEXT: scratch_load_b32 v244, off, s33 offset:564 +; DAGISEL-NEXT: scratch_load_b32 v245, off, s33 offset:568 +; DAGISEL-NEXT: scratch_load_b32 v246, off, s33 offset:572 +; DAGISEL-NEXT: scratch_load_b32 v247, off, s33 offset:576 +; DAGISEL-NEXT: s_mov_b32 exec_lo, s4 +; DAGISEL-NEXT: s_mov_b32 s33, s0 +; DAGISEL-NEXT: s_wait_loadcnt 0x0 +; DAGISEL-NEXT: s_wait_alu 0xfffe +; DAGISEL-NEXT: s_setpc_b64 s[30:31] +; +; GISEL-LABEL: call_gfx_from_whole_wave: +; GISEL: ; %bb.0: +; GISEL-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL-NEXT: s_wait_expcnt 0x0 +; GISEL-NEXT: s_wait_samplecnt 0x0 +; GISEL-NEXT: s_wait_bvhcnt 0x0 +; GISEL-NEXT: s_wait_kmcnt 0x0 +; GISEL-NEXT: s_mov_b32 s0, s33 +; GISEL-NEXT: s_mov_b32 s33, s32 +; GISEL-NEXT: s_xor_saveexec_b32 s4, -1 +; GISEL-NEXT: s_clause 0x1f +; GISEL-NEXT: scratch_store_b32 off, v0, s33 offset:4 +; GISEL-NEXT: scratch_store_b32 off, v1, s33 offset:8 +; GISEL-NEXT: scratch_store_b32 off, v2, s33 offset:12 +; GISEL-NEXT: scratch_store_b32 off, v3, s33 offset:16 +; GISEL-NEXT: scratch_store_b32 off, v4, s33 offset:20 +; GISEL-NEXT: scratch_store_b32 off, v5, s33 offset:24 +; GISEL-NEXT: scratch_store_b32 off, v6, s33 offset:28 +; GISEL-NEXT: scratch_store_b32 off, v7, s33 offset:32 +; GISEL-NEXT: scratch_store_b32 off, v8, s33 offset:36 +; GISEL-NEXT: scratch_store_b32 off, v9, s33 offset:40 +; GISEL-NEXT: scratch_store_b32 off, v10, s33 offset:44 +; GISEL-NEXT: scratch_store_b32 off, v11, s33 offset:48 +; GISEL-NEXT: scratch_store_b32 off, v12, s33 offset:52 +; GISEL-NEXT: scratch_store_b32 off, v13, s33 offset:56 +; GISEL-NEXT: scratch_store_b32 off, v14, s33 offset:60 +; GISEL-NEXT: scratch_store_b32 off, v15, s33 offset:64 +; GISEL-NEXT: scratch_store_b32 off, v16, s33 offset:68 +; GISEL-NEXT: scratch_store_b32 off, v17, s33 offset:72 +; GISEL-NEXT: scratch_store_b32 off, v18, s33 offset:76 +; GISEL-NEXT: scratch_store_b32 off, v19, s33 offset:80 +; GISEL-NEXT: scratch_store_b32 off, v20, s33 offset:84 +; GISEL-NEXT: scratch_store_b32 off, v21, s33 offset:88 +; GISEL-NEXT: scratch_store_b32 off, v22, s33 offset:92 +; GISEL-NEXT: scratch_store_b32 off, v23, s33 offset:96 +; GISEL-NEXT: scratch_store_b32 off, v24, s33 offset:100 +; GISEL-NEXT: scratch_store_b32 off, v25, s33 offset:104 +; GISEL-NEXT: scratch_store_b32 off, v26, s33 offset:108 +; GISEL-NEXT: scratch_store_b32 off, v27, s33 offset:112 +; GISEL-NEXT: scratch_store_b32 off, v28, s33 offset:116 +; GISEL-NEXT: scratch_store_b32 off, v29, s33 offset:120 +; GISEL-NEXT: scratch_store_b32 off, v30, s33 offset:124 +; GISEL-NEXT: scratch_store_b32 off, v31, s33 offset:128 +; GISEL-NEXT: s_clause 0x1f +; GISEL-NEXT: scratch_store_b32 off, v32, s33 offset:132 +; GISEL-NEXT: scratch_store_b32 off, v33, s33 offset:136 +; GISEL-NEXT: scratch_store_b32 off, v34, s33 offset:140 +; GISEL-NEXT: scratch_store_b32 off, v35, s33 offset:144 +; GISEL-NEXT: scratch_store_b32 off, v36, s33 offset:148 +; GISEL-NEXT: scratch_store_b32 off, v37, s33 offset:152 +; GISEL-NEXT: scratch_store_b32 off, v38, s33 offset:156 +; GISEL-NEXT: scratch_store_b32 off, v39, s33 offset:160 +; GISEL-NEXT: scratch_store_b32 off, v48, s33 offset:164 +; GISEL-NEXT: scratch_store_b32 off, v49, s33 offset:168 +; GISEL-NEXT: scratch_store_b32 off, v50, s33 offset:172 +; GISEL-NEXT: scratch_store_b32 off, v51, s33 offset:176 +; GISEL-NEXT: scratch_store_b32 off, v52, s33 offset:180 +; GISEL-NEXT: scratch_store_b32 off, v53, s33 offset:184 +; GISEL-NEXT: scratch_store_b32 off, v54, s33 offset:188 +; GISEL-NEXT: scratch_store_b32 off, v55, s33 offset:192 +; GISEL-NEXT: scratch_store_b32 off, v64, s33 offset:196 +; GISEL-NEXT: scratch_store_b32 off, v65, s33 offset:200 +; GISEL-NEXT: scratch_store_b32 off, v66, s33 offset:204 +; GISEL-NEXT: scratch_store_b32 off, v67, s33 offset:208 +; GISEL-NEXT: scratch_store_b32 off, v68, s33 offset:212 +; GISEL-NEXT: scratch_store_b32 off, v69, s33 offset:216 +; GISEL-NEXT: scratch_store_b32 off, v70, s33 offset:220 +; GISEL-NEXT: scratch_store_b32 off, v71, s33 offset:224 +; GISEL-NEXT: scratch_store_b32 off, v80, s33 offset:228 +; GISEL-NEXT: scratch_store_b32 off, v81, s33 offset:232 +; GISEL-NEXT: scratch_store_b32 off, v82, s33 offset:236 +; GISEL-NEXT: scratch_store_b32 off, v83, s33 offset:240 +; GISEL-NEXT: scratch_store_b32 off, v84, s33 offset:244 +; GISEL-NEXT: scratch_store_b32 off, v85, s33 offset:248 +; GISEL-NEXT: scratch_store_b32 off, v86, s33 offset:252 +; GISEL-NEXT: scratch_store_b32 off, v87, s33 offset:256 +; GISEL-NEXT: s_clause 0x1f +; GISEL-NEXT: scratch_store_b32 off, v96, s33 offset:260 +; GISEL-NEXT: scratch_store_b32 off, v97, s33 offset:264 +; GISEL-NEXT: scratch_store_b32 off, v98, s33 offset:268 +; GISEL-NEXT: scratch_store_b32 off, v99, s33 offset:272 +; GISEL-NEXT: scratch_store_b32 off, v100, s33 offset:276 +; GISEL-NEXT: scratch_store_b32 off, v101, s33 offset:280 +; GISEL-NEXT: scratch_store_b32 off, v102, s33 offset:284 +; GISEL-NEXT: scratch_store_b32 off, v103, s33 offset:288 +; GISEL-NEXT: scratch_store_b32 off, v112, s33 offset:292 +; GISEL-NEXT: scratch_store_b32 off, v113, s33 offset:296 +; GISEL-NEXT: scratch_store_b32 off, v114, s33 offset:300 +; GISEL-NEXT: scratch_store_b32 off, v115, s33 offset:304 +; GISEL-NEXT: scratch_store_b32 off, v116, s33 offset:308 +; GISEL-NEXT: scratch_store_b32 off, v117, s33 offset:312 +; GISEL-NEXT: scratch_store_b32 off, v118, s33 offset:316 +; GISEL-NEXT: scratch_store_b32 off, v119, s33 offset:320 +; GISEL-NEXT: scratch_store_b32 off, v128, s33 offset:324 +; GISEL-NEXT: scratch_store_b32 off, v129, s33 offset:328 +; GISEL-NEXT: scratch_store_b32 off, v130, s33 offset:332 +; GISEL-NEXT: scratch_store_b32 off, v131, s33 offset:336 +; GISEL-NEXT: scratch_store_b32 off, v132, s33 offset:340 +; GISEL-NEXT: scratch_store_b32 off, v133, s33 offset:344 +; GISEL-NEXT: scratch_store_b32 off, v134, s33 offset:348 +; GISEL-NEXT: scratch_store_b32 off, v135, s33 offset:352 +; GISEL-NEXT: scratch_store_b32 off, v144, s33 offset:356 +; GISEL-NEXT: scratch_store_b32 off, v145, s33 offset:360 +; GISEL-NEXT: scratch_store_b32 off, v146, s33 offset:364 +; GISEL-NEXT: scratch_store_b32 off, v147, s33 offset:368 +; GISEL-NEXT: scratch_store_b32 off, v148, s33 offset:372 +; GISEL-NEXT: scratch_store_b32 off, v149, s33 offset:376 +; GISEL-NEXT: scratch_store_b32 off, v150, s33 offset:380 +; GISEL-NEXT: scratch_store_b32 off, v151, s33 offset:384 +; GISEL-NEXT: s_clause 0x1f +; GISEL-NEXT: scratch_store_b32 off, v160, s33 offset:388 +; GISEL-NEXT: scratch_store_b32 off, v161, s33 offset:392 +; GISEL-NEXT: scratch_store_b32 off, v162, s33 offset:396 +; GISEL-NEXT: scratch_store_b32 off, v163, s33 offset:400 +; GISEL-NEXT: scratch_store_b32 off, v164, s33 offset:404 +; GISEL-NEXT: scratch_store_b32 off, v165, s33 offset:408 +; GISEL-NEXT: scratch_store_b32 off, v166, s33 offset:412 +; GISEL-NEXT: scratch_store_b32 off, v167, s33 offset:416 +; GISEL-NEXT: scratch_store_b32 off, v176, s33 offset:420 +; GISEL-NEXT: scratch_store_b32 off, v177, s33 offset:424 +; GISEL-NEXT: scratch_store_b32 off, v178, s33 offset:428 +; GISEL-NEXT: scratch_store_b32 off, v179, s33 offset:432 +; GISEL-NEXT: scratch_store_b32 off, v180, s33 offset:436 +; GISEL-NEXT: scratch_store_b32 off, v181, s33 offset:440 +; GISEL-NEXT: scratch_store_b32 off, v182, s33 offset:444 +; GISEL-NEXT: scratch_store_b32 off, v183, s33 offset:448 +; GISEL-NEXT: scratch_store_b32 off, v192, s33 offset:452 +; GISEL-NEXT: scratch_store_b32 off, v193, s33 offset:456 +; GISEL-NEXT: scratch_store_b32 off, v194, s33 offset:460 +; GISEL-NEXT: scratch_store_b32 off, v195, s33 offset:464 +; GISEL-NEXT: scratch_store_b32 off, v196, s33 offset:468 +; GISEL-NEXT: scratch_store_b32 off, v197, s33 offset:472 +; GISEL-NEXT: scratch_store_b32 off, v198, s33 offset:476 +; GISEL-NEXT: scratch_store_b32 off, v199, s33 offset:480 +; GISEL-NEXT: scratch_store_b32 off, v208, s33 offset:484 +; GISEL-NEXT: scratch_store_b32 off, v209, s33 offset:488 +; GISEL-NEXT: scratch_store_b32 off, v210, s33 offset:492 +; GISEL-NEXT: scratch_store_b32 off, v211, s33 offset:496 +; GISEL-NEXT: scratch_store_b32 off, v212, s33 offset:500 +; GISEL-NEXT: scratch_store_b32 off, v213, s33 offset:504 +; GISEL-NEXT: scratch_store_b32 off, v214, s33 offset:508 +; GISEL-NEXT: scratch_store_b32 off, v215, s33 offset:512 +; GISEL-NEXT: s_clause 0xf +; GISEL-NEXT: scratch_store_b32 off, v224, s33 offset:516 +; GISEL-NEXT: scratch_store_b32 off, v225, s33 offset:520 +; GISEL-NEXT: scratch_store_b32 off, v226, s33 offset:524 +; GISEL-NEXT: scratch_store_b32 off, v227, s33 offset:528 +; GISEL-NEXT: scratch_store_b32 off, v228, s33 offset:532 +; GISEL-NEXT: scratch_store_b32 off, v229, s33 offset:536 +; GISEL-NEXT: scratch_store_b32 off, v230, s33 offset:540 +; GISEL-NEXT: scratch_store_b32 off, v231, s33 offset:544 +; GISEL-NEXT: scratch_store_b32 off, v240, s33 offset:548 +; GISEL-NEXT: scratch_store_b32 off, v241, s33 offset:552 +; GISEL-NEXT: scratch_store_b32 off, v242, s33 offset:556 +; GISEL-NEXT: scratch_store_b32 off, v243, s33 offset:560 +; GISEL-NEXT: scratch_store_b32 off, v244, s33 offset:564 +; GISEL-NEXT: scratch_store_b32 off, v245, s33 offset:568 +; GISEL-NEXT: scratch_store_b32 off, v246, s33 offset:572 +; GISEL-NEXT: scratch_store_b32 off, v247, s33 offset:576 +; GISEL-NEXT: s_mov_b32 exec_lo, -1 +; GISEL-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill +; GISEL-NEXT: s_wait_alu 0xfffe +; GISEL-NEXT: v_writelane_b32 v40, s0, 3 +; GISEL-NEXT: v_mov_b32_e32 v2, v0 +; GISEL-NEXT: v_swap_b32 v0, v1 +; GISEL-NEXT: s_mov_b32 s0, gfx_callee@abs32@lo +; GISEL-NEXT: v_writelane_b32 v40, s4, 0 +; GISEL-NEXT: s_mov_b32 s1, gfx_callee@abs32@hi +; GISEL-NEXT: s_addk_co_i32 s32, 0x250 +; GISEL-NEXT: v_writelane_b32 v40, s30, 1 +; GISEL-NEXT: v_writelane_b32 v40, s31, 2 +; GISEL-NEXT: s_wait_alu 0xfffe +; GISEL-NEXT: s_swappc_b64 s[30:31], s[0:1] +; GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GISEL-NEXT: v_readlane_b32 s31, v40, 2 +; GISEL-NEXT: v_readlane_b32 s30, v40, 1 +; GISEL-NEXT: v_readlane_b32 s4, v40, 0 +; GISEL-NEXT: v_readlane_b32 s0, v40, 3 +; GISEL-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload +; GISEL-NEXT: s_mov_b32 s32, s33 +; GISEL-NEXT: s_xor_b32 exec_lo, s4, -1 +; GISEL-NEXT: s_clause 0x1f +; GISEL-NEXT: scratch_load_b32 v0, off, s33 offset:4 +; GISEL-NEXT: scratch_load_b32 v1, off, s33 offset:8 +; GISEL-NEXT: scratch_load_b32 v2, off, s33 offset:12 +; GISEL-NEXT: scratch_load_b32 v3, off, s33 offset:16 +; GISEL-NEXT: scratch_load_b32 v4, off, s33 offset:20 +; GISEL-NEXT: scratch_load_b32 v5, off, s33 offset:24 +; GISEL-NEXT: scratch_load_b32 v6, off, s33 offset:28 +; GISEL-NEXT: scratch_load_b32 v7, off, s33 offset:32 +; GISEL-NEXT: scratch_load_b32 v8, off, s33 offset:36 +; GISEL-NEXT: scratch_load_b32 v9, off, s33 offset:40 +; GISEL-NEXT: scratch_load_b32 v10, off, s33 offset:44 +; GISEL-NEXT: scratch_load_b32 v11, off, s33 offset:48 +; GISEL-NEXT: scratch_load_b32 v12, off, s33 offset:52 +; GISEL-NEXT: scratch_load_b32 v13, off, s33 offset:56 +; GISEL-NEXT: scratch_load_b32 v14, off, s33 offset:60 +; GISEL-NEXT: scratch_load_b32 v15, off, s33 offset:64 +; GISEL-NEXT: scratch_load_b32 v16, off, s33 offset:68 +; GISEL-NEXT: scratch_load_b32 v17, off, s33 offset:72 +; GISEL-NEXT: scratch_load_b32 v18, off, s33 offset:76 +; GISEL-NEXT: scratch_load_b32 v19, off, s33 offset:80 +; GISEL-NEXT: scratch_load_b32 v20, off, s33 offset:84 +; GISEL-NEXT: scratch_load_b32 v21, off, s33 offset:88 +; GISEL-NEXT: scratch_load_b32 v22, off, s33 offset:92 +; GISEL-NEXT: scratch_load_b32 v23, off, s33 offset:96 +; GISEL-NEXT: scratch_load_b32 v24, off, s33 offset:100 +; GISEL-NEXT: scratch_load_b32 v25, off, s33 offset:104 +; GISEL-NEXT: scratch_load_b32 v26, off, s33 offset:108 +; GISEL-NEXT: scratch_load_b32 v27, off, s33 offset:112 +; GISEL-NEXT: scratch_load_b32 v28, off, s33 offset:116 +; GISEL-NEXT: scratch_load_b32 v29, off, s33 offset:120 +; GISEL-NEXT: scratch_load_b32 v30, off, s33 offset:124 +; GISEL-NEXT: scratch_load_b32 v31, off, s33 offset:128 +; GISEL-NEXT: s_clause 0x1f +; GISEL-NEXT: scratch_load_b32 v32, off, s33 offset:132 +; GISEL-NEXT: scratch_load_b32 v33, off, s33 offset:136 +; GISEL-NEXT: scratch_load_b32 v34, off, s33 offset:140 +; GISEL-NEXT: scratch_load_b32 v35, off, s33 offset:144 +; GISEL-NEXT: scratch_load_b32 v36, off, s33 offset:148 +; GISEL-NEXT: scratch_load_b32 v37, off, s33 offset:152 +; GISEL-NEXT: scratch_load_b32 v38, off, s33 offset:156 +; GISEL-NEXT: scratch_load_b32 v39, off, s33 offset:160 +; GISEL-NEXT: scratch_load_b32 v48, off, s33 offset:164 +; GISEL-NEXT: scratch_load_b32 v49, off, s33 offset:168 +; GISEL-NEXT: scratch_load_b32 v50, off, s33 offset:172 +; GISEL-NEXT: scratch_load_b32 v51, off, s33 offset:176 +; GISEL-NEXT: scratch_load_b32 v52, off, s33 offset:180 +; GISEL-NEXT: scratch_load_b32 v53, off, s33 offset:184 +; GISEL-NEXT: scratch_load_b32 v54, off, s33 offset:188 +; GISEL-NEXT: scratch_load_b32 v55, off, s33 offset:192 +; GISEL-NEXT: scratch_load_b32 v64, off, s33 offset:196 +; GISEL-NEXT: scratch_load_b32 v65, off, s33 offset:200 +; GISEL-NEXT: scratch_load_b32 v66, off, s33 offset:204 +; GISEL-NEXT: scratch_load_b32 v67, off, s33 offset:208 +; GISEL-NEXT: scratch_load_b32 v68, off, s33 offset:212 +; GISEL-NEXT: scratch_load_b32 v69, off, s33 offset:216 +; GISEL-NEXT: scratch_load_b32 v70, off, s33 offset:220 +; GISEL-NEXT: scratch_load_b32 v71, off, s33 offset:224 +; GISEL-NEXT: scratch_load_b32 v80, off, s33 offset:228 +; GISEL-NEXT: scratch_load_b32 v81, off, s33 offset:232 +; GISEL-NEXT: scratch_load_b32 v82, off, s33 offset:236 +; GISEL-NEXT: scratch_load_b32 v83, off, s33 offset:240 +; GISEL-NEXT: scratch_load_b32 v84, off, s33 offset:244 +; GISEL-NEXT: scratch_load_b32 v85, off, s33 offset:248 +; GISEL-NEXT: scratch_load_b32 v86, off, s33 offset:252 +; GISEL-NEXT: scratch_load_b32 v87, off, s33 offset:256 +; GISEL-NEXT: s_clause 0x1f +; GISEL-NEXT: scratch_load_b32 v96, off, s33 offset:260 +; GISEL-NEXT: scratch_load_b32 v97, off, s33 offset:264 +; GISEL-NEXT: scratch_load_b32 v98, off, s33 offset:268 +; GISEL-NEXT: scratch_load_b32 v99, off, s33 offset:272 +; GISEL-NEXT: scratch_load_b32 v100, off, s33 offset:276 +; GISEL-NEXT: scratch_load_b32 v101, off, s33 offset:280 +; GISEL-NEXT: scratch_load_b32 v102, off, s33 offset:284 +; GISEL-NEXT: scratch_load_b32 v103, off, s33 offset:288 +; GISEL-NEXT: scratch_load_b32 v112, off, s33 offset:292 +; GISEL-NEXT: scratch_load_b32 v113, off, s33 offset:296 +; GISEL-NEXT: scratch_load_b32 v114, off, s33 offset:300 +; GISEL-NEXT: scratch_load_b32 v115, off, s33 offset:304 +; GISEL-NEXT: scratch_load_b32 v116, off, s33 offset:308 +; GISEL-NEXT: scratch_load_b32 v117, off, s33 offset:312 +; GISEL-NEXT: scratch_load_b32 v118, off, s33 offset:316 +; GISEL-NEXT: scratch_load_b32 v119, off, s33 offset:320 +; GISEL-NEXT: scratch_load_b32 v128, off, s33 offset:324 +; GISEL-NEXT: scratch_load_b32 v129, off, s33 offset:328 +; GISEL-NEXT: scratch_load_b32 v130, off, s33 offset:332 +; GISEL-NEXT: scratch_load_b32 v131, off, s33 offset:336 +; GISEL-NEXT: scratch_load_b32 v132, off, s33 offset:340 +; GISEL-NEXT: scratch_load_b32 v133, off, s33 offset:344 +; GISEL-NEXT: scratch_load_b32 v134, off, s33 offset:348 +; GISEL-NEXT: scratch_load_b32 v135, off, s33 offset:352 +; GISEL-NEXT: scratch_load_b32 v144, off, s33 offset:356 +; GISEL-NEXT: scratch_load_b32 v145, off, s33 offset:360 +; GISEL-NEXT: scratch_load_b32 v146, off, s33 offset:364 +; GISEL-NEXT: scratch_load_b32 v147, off, s33 offset:368 +; GISEL-NEXT: scratch_load_b32 v148, off, s33 offset:372 +; GISEL-NEXT: scratch_load_b32 v149, off, s33 offset:376 +; GISEL-NEXT: scratch_load_b32 v150, off, s33 offset:380 +; GISEL-NEXT: scratch_load_b32 v151, off, s33 offset:384 +; GISEL-NEXT: s_clause 0x1f +; GISEL-NEXT: scratch_load_b32 v160, off, s33 offset:388 +; GISEL-NEXT: scratch_load_b32 v161, off, s33 offset:392 +; GISEL-NEXT: scratch_load_b32 v162, off, s33 offset:396 +; GISEL-NEXT: scratch_load_b32 v163, off, s33 offset:400 +; GISEL-NEXT: scratch_load_b32 v164, off, s33 offset:404 +; GISEL-NEXT: scratch_load_b32 v165, off, s33 offset:408 +; GISEL-NEXT: scratch_load_b32 v166, off, s33 offset:412 +; GISEL-NEXT: scratch_load_b32 v167, off, s33 offset:416 +; GISEL-NEXT: scratch_load_b32 v176, off, s33 offset:420 +; GISEL-NEXT: scratch_load_b32 v177, off, s33 offset:424 +; GISEL-NEXT: scratch_load_b32 v178, off, s33 offset:428 +; GISEL-NEXT: scratch_load_b32 v179, off, s33 offset:432 +; GISEL-NEXT: scratch_load_b32 v180, off, s33 offset:436 +; GISEL-NEXT: scratch_load_b32 v181, off, s33 offset:440 +; GISEL-NEXT: scratch_load_b32 v182, off, s33 offset:444 +; GISEL-NEXT: scratch_load_b32 v183, off, s33 offset:448 +; GISEL-NEXT: scratch_load_b32 v192, off, s33 offset:452 +; GISEL-NEXT: scratch_load_b32 v193, off, s33 offset:456 +; GISEL-NEXT: scratch_load_b32 v194, off, s33 offset:460 +; GISEL-NEXT: scratch_load_b32 v195, off, s33 offset:464 +; GISEL-NEXT: scratch_load_b32 v196, off, s33 offset:468 +; GISEL-NEXT: scratch_load_b32 v197, off, s33 offset:472 +; GISEL-NEXT: scratch_load_b32 v198, off, s33 offset:476 +; GISEL-NEXT: scratch_load_b32 v199, off, s33 offset:480 +; GISEL-NEXT: scratch_load_b32 v208, off, s33 offset:484 +; GISEL-NEXT: scratch_load_b32 v209, off, s33 offset:488 +; GISEL-NEXT: scratch_load_b32 v210, off, s33 offset:492 +; GISEL-NEXT: scratch_load_b32 v211, off, s33 offset:496 +; GISEL-NEXT: scratch_load_b32 v212, off, s33 offset:500 +; GISEL-NEXT: scratch_load_b32 v213, off, s33 offset:504 +; GISEL-NEXT: scratch_load_b32 v214, off, s33 offset:508 +; GISEL-NEXT: scratch_load_b32 v215, off, s33 offset:512 +; GISEL-NEXT: s_clause 0xf +; GISEL-NEXT: scratch_load_b32 v224, off, s33 offset:516 +; GISEL-NEXT: scratch_load_b32 v225, off, s33 offset:520 +; GISEL-NEXT: scratch_load_b32 v226, off, s33 offset:524 +; GISEL-NEXT: scratch_load_b32 v227, off, s33 offset:528 +; GISEL-NEXT: scratch_load_b32 v228, off, s33 offset:532 +; GISEL-NEXT: scratch_load_b32 v229, off, s33 offset:536 +; GISEL-NEXT: scratch_load_b32 v230, off, s33 offset:540 +; GISEL-NEXT: scratch_load_b32 v231, off, s33 offset:544 +; GISEL-NEXT: scratch_load_b32 v240, off, s33 offset:548 +; GISEL-NEXT: scratch_load_b32 v241, off, s33 offset:552 +; GISEL-NEXT: scratch_load_b32 v242, off, s33 offset:556 +; GISEL-NEXT: scratch_load_b32 v243, off, s33 offset:560 +; GISEL-NEXT: scratch_load_b32 v244, off, s33 offset:564 +; GISEL-NEXT: scratch_load_b32 v245, off, s33 offset:568 +; GISEL-NEXT: scratch_load_b32 v246, off, s33 offset:572 +; GISEL-NEXT: scratch_load_b32 v247, off, s33 offset:576 +; GISEL-NEXT: s_mov_b32 exec_lo, s4 +; GISEL-NEXT: s_mov_b32 s33, s0 +; GISEL-NEXT: s_wait_loadcnt 0x0 +; GISEL-NEXT: s_wait_alu 0xfffe +; GISEL-NEXT: s_setpc_b64 s[30:31] +; +; DAGISEL64-LABEL: call_gfx_from_whole_wave: +; DAGISEL64: ; %bb.0: +; DAGISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; DAGISEL64-NEXT: s_wait_expcnt 0x0 +; DAGISEL64-NEXT: s_wait_samplecnt 0x0 +; DAGISEL64-NEXT: s_wait_bvhcnt 0x0 +; DAGISEL64-NEXT: s_wait_kmcnt 0x0 +; DAGISEL64-NEXT: s_mov_b32 s0, s33 +; DAGISEL64-NEXT: s_mov_b32 s33, s32 +; DAGISEL64-NEXT: s_xor_saveexec_b64 s[4:5], -1 +; DAGISEL64-NEXT: s_clause 0x1f +; DAGISEL64-NEXT: scratch_store_b32 off, v0, s33 offset:4 +; DAGISEL64-NEXT: scratch_store_b32 off, v1, s33 offset:8 +; DAGISEL64-NEXT: scratch_store_b32 off, v2, s33 offset:12 +; DAGISEL64-NEXT: scratch_store_b32 off, v3, s33 offset:16 +; DAGISEL64-NEXT: scratch_store_b32 off, v4, s33 offset:20 +; DAGISEL64-NEXT: scratch_store_b32 off, v5, s33 offset:24 +; DAGISEL64-NEXT: scratch_store_b32 off, v6, s33 offset:28 +; DAGISEL64-NEXT: scratch_store_b32 off, v7, s33 offset:32 +; DAGISEL64-NEXT: scratch_store_b32 off, v8, s33 offset:36 +; DAGISEL64-NEXT: scratch_store_b32 off, v9, s33 offset:40 +; DAGISEL64-NEXT: scratch_store_b32 off, v10, s33 offset:44 +; DAGISEL64-NEXT: scratch_store_b32 off, v11, s33 offset:48 +; DAGISEL64-NEXT: scratch_store_b32 off, v12, s33 offset:52 +; DAGISEL64-NEXT: scratch_store_b32 off, v13, s33 offset:56 +; DAGISEL64-NEXT: scratch_store_b32 off, v14, s33 offset:60 +; DAGISEL64-NEXT: scratch_store_b32 off, v15, s33 offset:64 +; DAGISEL64-NEXT: scratch_store_b32 off, v16, s33 offset:68 +; DAGISEL64-NEXT: scratch_store_b32 off, v17, s33 offset:72 +; DAGISEL64-NEXT: scratch_store_b32 off, v18, s33 offset:76 +; DAGISEL64-NEXT: scratch_store_b32 off, v19, s33 offset:80 +; DAGISEL64-NEXT: scratch_store_b32 off, v20, s33 offset:84 +; DAGISEL64-NEXT: scratch_store_b32 off, v21, s33 offset:88 +; DAGISEL64-NEXT: scratch_store_b32 off, v22, s33 offset:92 +; DAGISEL64-NEXT: scratch_store_b32 off, v23, s33 offset:96 +; DAGISEL64-NEXT: scratch_store_b32 off, v24, s33 offset:100 +; DAGISEL64-NEXT: scratch_store_b32 off, v25, s33 offset:104 +; DAGISEL64-NEXT: scratch_store_b32 off, v26, s33 offset:108 +; DAGISEL64-NEXT: scratch_store_b32 off, v27, s33 offset:112 +; DAGISEL64-NEXT: scratch_store_b32 off, v28, s33 offset:116 +; DAGISEL64-NEXT: scratch_store_b32 off, v29, s33 offset:120 +; DAGISEL64-NEXT: scratch_store_b32 off, v30, s33 offset:124 +; DAGISEL64-NEXT: scratch_store_b32 off, v31, s33 offset:128 +; DAGISEL64-NEXT: s_clause 0x1f +; DAGISEL64-NEXT: scratch_store_b32 off, v32, s33 offset:132 +; DAGISEL64-NEXT: scratch_store_b32 off, v33, s33 offset:136 +; DAGISEL64-NEXT: scratch_store_b32 off, v34, s33 offset:140 +; DAGISEL64-NEXT: scratch_store_b32 off, v35, s33 offset:144 +; DAGISEL64-NEXT: scratch_store_b32 off, v36, s33 offset:148 +; DAGISEL64-NEXT: scratch_store_b32 off, v37, s33 offset:152 +; DAGISEL64-NEXT: scratch_store_b32 off, v38, s33 offset:156 +; DAGISEL64-NEXT: scratch_store_b32 off, v39, s33 offset:160 +; DAGISEL64-NEXT: scratch_store_b32 off, v48, s33 offset:164 +; DAGISEL64-NEXT: scratch_store_b32 off, v49, s33 offset:168 +; DAGISEL64-NEXT: scratch_store_b32 off, v50, s33 offset:172 +; DAGISEL64-NEXT: scratch_store_b32 off, v51, s33 offset:176 +; DAGISEL64-NEXT: scratch_store_b32 off, v52, s33 offset:180 +; DAGISEL64-NEXT: scratch_store_b32 off, v53, s33 offset:184 +; DAGISEL64-NEXT: scratch_store_b32 off, v54, s33 offset:188 +; DAGISEL64-NEXT: scratch_store_b32 off, v55, s33 offset:192 +; DAGISEL64-NEXT: scratch_store_b32 off, v64, s33 offset:196 +; DAGISEL64-NEXT: scratch_store_b32 off, v65, s33 offset:200 +; DAGISEL64-NEXT: scratch_store_b32 off, v66, s33 offset:204 +; DAGISEL64-NEXT: scratch_store_b32 off, v67, s33 offset:208 +; DAGISEL64-NEXT: scratch_store_b32 off, v68, s33 offset:212 +; DAGISEL64-NEXT: scratch_store_b32 off, v69, s33 offset:216 +; DAGISEL64-NEXT: scratch_store_b32 off, v70, s33 offset:220 +; DAGISEL64-NEXT: scratch_store_b32 off, v71, s33 offset:224 +; DAGISEL64-NEXT: scratch_store_b32 off, v80, s33 offset:228 +; DAGISEL64-NEXT: scratch_store_b32 off, v81, s33 offset:232 +; DAGISEL64-NEXT: scratch_store_b32 off, v82, s33 offset:236 +; DAGISEL64-NEXT: scratch_store_b32 off, v83, s33 offset:240 +; DAGISEL64-NEXT: scratch_store_b32 off, v84, s33 offset:244 +; DAGISEL64-NEXT: scratch_store_b32 off, v85, s33 offset:248 +; DAGISEL64-NEXT: scratch_store_b32 off, v86, s33 offset:252 +; DAGISEL64-NEXT: scratch_store_b32 off, v87, s33 offset:256 +; DAGISEL64-NEXT: s_clause 0x1f +; DAGISEL64-NEXT: scratch_store_b32 off, v96, s33 offset:260 +; DAGISEL64-NEXT: scratch_store_b32 off, v97, s33 offset:264 +; DAGISEL64-NEXT: scratch_store_b32 off, v98, s33 offset:268 +; DAGISEL64-NEXT: scratch_store_b32 off, v99, s33 offset:272 +; DAGISEL64-NEXT: scratch_store_b32 off, v100, s33 offset:276 +; DAGISEL64-NEXT: scratch_store_b32 off, v101, s33 offset:280 +; DAGISEL64-NEXT: scratch_store_b32 off, v102, s33 offset:284 +; DAGISEL64-NEXT: scratch_store_b32 off, v103, s33 offset:288 +; DAGISEL64-NEXT: scratch_store_b32 off, v112, s33 offset:292 +; DAGISEL64-NEXT: scratch_store_b32 off, v113, s33 offset:296 +; DAGISEL64-NEXT: scratch_store_b32 off, v114, s33 offset:300 +; DAGISEL64-NEXT: scratch_store_b32 off, v115, s33 offset:304 +; DAGISEL64-NEXT: scratch_store_b32 off, v116, s33 offset:308 +; DAGISEL64-NEXT: scratch_store_b32 off, v117, s33 offset:312 +; DAGISEL64-NEXT: scratch_store_b32 off, v118, s33 offset:316 +; DAGISEL64-NEXT: scratch_store_b32 off, v119, s33 offset:320 +; DAGISEL64-NEXT: scratch_store_b32 off, v128, s33 offset:324 +; DAGISEL64-NEXT: scratch_store_b32 off, v129, s33 offset:328 +; DAGISEL64-NEXT: scratch_store_b32 off, v130, s33 offset:332 +; DAGISEL64-NEXT: scratch_store_b32 off, v131, s33 offset:336 +; DAGISEL64-NEXT: scratch_store_b32 off, v132, s33 offset:340 +; DAGISEL64-NEXT: scratch_store_b32 off, v133, s33 offset:344 +; DAGISEL64-NEXT: scratch_store_b32 off, v134, s33 offset:348 +; DAGISEL64-NEXT: scratch_store_b32 off, v135, s33 offset:352 +; DAGISEL64-NEXT: scratch_store_b32 off, v144, s33 offset:356 +; DAGISEL64-NEXT: scratch_store_b32 off, v145, s33 offset:360 +; DAGISEL64-NEXT: scratch_store_b32 off, v146, s33 offset:364 +; DAGISEL64-NEXT: scratch_store_b32 off, v147, s33 offset:368 +; DAGISEL64-NEXT: scratch_store_b32 off, v148, s33 offset:372 +; DAGISEL64-NEXT: scratch_store_b32 off, v149, s33 offset:376 +; DAGISEL64-NEXT: scratch_store_b32 off, v150, s33 offset:380 +; DAGISEL64-NEXT: scratch_store_b32 off, v151, s33 offset:384 +; DAGISEL64-NEXT: s_clause 0x1f +; DAGISEL64-NEXT: scratch_store_b32 off, v160, s33 offset:388 +; DAGISEL64-NEXT: scratch_store_b32 off, v161, s33 offset:392 +; DAGISEL64-NEXT: scratch_store_b32 off, v162, s33 offset:396 +; DAGISEL64-NEXT: scratch_store_b32 off, v163, s33 offset:400 +; DAGISEL64-NEXT: scratch_store_b32 off, v164, s33 offset:404 +; DAGISEL64-NEXT: scratch_store_b32 off, v165, s33 offset:408 +; DAGISEL64-NEXT: scratch_store_b32 off, v166, s33 offset:412 +; DAGISEL64-NEXT: scratch_store_b32 off, v167, s33 offset:416 +; DAGISEL64-NEXT: scratch_store_b32 off, v176, s33 offset:420 +; DAGISEL64-NEXT: scratch_store_b32 off, v177, s33 offset:424 +; DAGISEL64-NEXT: scratch_store_b32 off, v178, s33 offset:428 +; DAGISEL64-NEXT: scratch_store_b32 off, v179, s33 offset:432 +; DAGISEL64-NEXT: scratch_store_b32 off, v180, s33 offset:436 +; DAGISEL64-NEXT: scratch_store_b32 off, v181, s33 offset:440 +; DAGISEL64-NEXT: scratch_store_b32 off, v182, s33 offset:444 +; DAGISEL64-NEXT: scratch_store_b32 off, v183, s33 offset:448 +; DAGISEL64-NEXT: scratch_store_b32 off, v192, s33 offset:452 +; DAGISEL64-NEXT: scratch_store_b32 off, v193, s33 offset:456 +; DAGISEL64-NEXT: scratch_store_b32 off, v194, s33 offset:460 +; DAGISEL64-NEXT: scratch_store_b32 off, v195, s33 offset:464 +; DAGISEL64-NEXT: scratch_store_b32 off, v196, s33 offset:468 +; DAGISEL64-NEXT: scratch_store_b32 off, v197, s33 offset:472 +; DAGISEL64-NEXT: scratch_store_b32 off, v198, s33 offset:476 +; DAGISEL64-NEXT: scratch_store_b32 off, v199, s33 offset:480 +; DAGISEL64-NEXT: scratch_store_b32 off, v208, s33 offset:484 +; DAGISEL64-NEXT: scratch_store_b32 off, v209, s33 offset:488 +; DAGISEL64-NEXT: scratch_store_b32 off, v210, s33 offset:492 +; DAGISEL64-NEXT: scratch_store_b32 off, v211, s33 offset:496 +; DAGISEL64-NEXT: scratch_store_b32 off, v212, s33 offset:500 +; DAGISEL64-NEXT: scratch_store_b32 off, v213, s33 offset:504 +; DAGISEL64-NEXT: scratch_store_b32 off, v214, s33 offset:508 +; DAGISEL64-NEXT: scratch_store_b32 off, v215, s33 offset:512 +; DAGISEL64-NEXT: s_clause 0xf +; DAGISEL64-NEXT: scratch_store_b32 off, v224, s33 offset:516 +; DAGISEL64-NEXT: scratch_store_b32 off, v225, s33 offset:520 +; DAGISEL64-NEXT: scratch_store_b32 off, v226, s33 offset:524 +; DAGISEL64-NEXT: scratch_store_b32 off, v227, s33 offset:528 +; DAGISEL64-NEXT: scratch_store_b32 off, v228, s33 offset:532 +; DAGISEL64-NEXT: scratch_store_b32 off, v229, s33 offset:536 +; DAGISEL64-NEXT: scratch_store_b32 off, v230, s33 offset:540 +; DAGISEL64-NEXT: scratch_store_b32 off, v231, s33 offset:544 +; DAGISEL64-NEXT: scratch_store_b32 off, v240, s33 offset:548 +; DAGISEL64-NEXT: scratch_store_b32 off, v241, s33 offset:552 +; DAGISEL64-NEXT: scratch_store_b32 off, v242, s33 offset:556 +; DAGISEL64-NEXT: scratch_store_b32 off, v243, s33 offset:560 +; DAGISEL64-NEXT: scratch_store_b32 off, v244, s33 offset:564 +; DAGISEL64-NEXT: scratch_store_b32 off, v245, s33 offset:568 +; DAGISEL64-NEXT: scratch_store_b32 off, v246, s33 offset:572 +; DAGISEL64-NEXT: scratch_store_b32 off, v247, s33 offset:576 +; DAGISEL64-NEXT: s_mov_b64 exec, -1 +; DAGISEL64-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill +; DAGISEL64-NEXT: s_wait_alu 0xfffe +; DAGISEL64-NEXT: v_writelane_b32 v40, s0, 4 +; DAGISEL64-NEXT: v_mov_b32_e32 v2, v0 +; DAGISEL64-NEXT: v_swap_b32 v0, v1 +; DAGISEL64-NEXT: s_mov_b32 s1, gfx_callee@abs32@hi +; DAGISEL64-NEXT: v_writelane_b32 v40, s4, 0 +; DAGISEL64-NEXT: s_mov_b32 s0, gfx_callee@abs32@lo +; DAGISEL64-NEXT: s_addk_co_i32 s32, 0x250 +; DAGISEL64-NEXT: v_writelane_b32 v40, s5, 1 +; DAGISEL64-NEXT: v_writelane_b32 v40, s30, 2 +; DAGISEL64-NEXT: v_writelane_b32 v40, s31, 3 +; DAGISEL64-NEXT: s_wait_alu 0xfffe +; DAGISEL64-NEXT: s_swappc_b64 s[30:31], s[0:1] +; DAGISEL64-NEXT: s_delay_alu instid0(VALU_DEP_1) +; DAGISEL64-NEXT: v_readlane_b32 s31, v40, 3 +; DAGISEL64-NEXT: v_readlane_b32 s30, v40, 2 +; DAGISEL64-NEXT: v_readlane_b32 s5, v40, 1 +; DAGISEL64-NEXT: v_readlane_b32 s4, v40, 0 +; DAGISEL64-NEXT: v_readlane_b32 s0, v40, 4 +; DAGISEL64-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload +; DAGISEL64-NEXT: s_mov_b32 s32, s33 +; DAGISEL64-NEXT: s_xor_b64 exec, s[4:5], -1 +; DAGISEL64-NEXT: s_clause 0x1f +; DAGISEL64-NEXT: scratch_load_b32 v0, off, s33 offset:4 +; DAGISEL64-NEXT: scratch_load_b32 v1, off, s33 offset:8 +; DAGISEL64-NEXT: scratch_load_b32 v2, off, s33 offset:12 +; DAGISEL64-NEXT: scratch_load_b32 v3, off, s33 offset:16 +; DAGISEL64-NEXT: scratch_load_b32 v4, off, s33 offset:20 +; DAGISEL64-NEXT: scratch_load_b32 v5, off, s33 offset:24 +; DAGISEL64-NEXT: scratch_load_b32 v6, off, s33 offset:28 +; DAGISEL64-NEXT: scratch_load_b32 v7, off, s33 offset:32 +; DAGISEL64-NEXT: scratch_load_b32 v8, off, s33 offset:36 +; DAGISEL64-NEXT: scratch_load_b32 v9, off, s33 offset:40 +; DAGISEL64-NEXT: scratch_load_b32 v10, off, s33 offset:44 +; DAGISEL64-NEXT: scratch_load_b32 v11, off, s33 offset:48 +; DAGISEL64-NEXT: scratch_load_b32 v12, off, s33 offset:52 +; DAGISEL64-NEXT: scratch_load_b32 v13, off, s33 offset:56 +; DAGISEL64-NEXT: scratch_load_b32 v14, off, s33 offset:60 +; DAGISEL64-NEXT: scratch_load_b32 v15, off, s33 offset:64 +; DAGISEL64-NEXT: scratch_load_b32 v16, off, s33 offset:68 +; DAGISEL64-NEXT: scratch_load_b32 v17, off, s33 offset:72 +; DAGISEL64-NEXT: scratch_load_b32 v18, off, s33 offset:76 +; DAGISEL64-NEXT: scratch_load_b32 v19, off, s33 offset:80 +; DAGISEL64-NEXT: scratch_load_b32 v20, off, s33 offset:84 +; DAGISEL64-NEXT: scratch_load_b32 v21, off, s33 offset:88 +; DAGISEL64-NEXT: scratch_load_b32 v22, off, s33 offset:92 +; DAGISEL64-NEXT: scratch_load_b32 v23, off, s33 offset:96 +; DAGISEL64-NEXT: scratch_load_b32 v24, off, s33 offset:100 +; DAGISEL64-NEXT: scratch_load_b32 v25, off, s33 offset:104 +; DAGISEL64-NEXT: scratch_load_b32 v26, off, s33 offset:108 +; DAGISEL64-NEXT: scratch_load_b32 v27, off, s33 offset:112 +; DAGISEL64-NEXT: scratch_load_b32 v28, off, s33 offset:116 +; DAGISEL64-NEXT: scratch_load_b32 v29, off, s33 offset:120 +; DAGISEL64-NEXT: scratch_load_b32 v30, off, s33 offset:124 +; DAGISEL64-NEXT: scratch_load_b32 v31, off, s33 offset:128 +; DAGISEL64-NEXT: s_clause 0x1f +; DAGISEL64-NEXT: scratch_load_b32 v32, off, s33 offset:132 +; DAGISEL64-NEXT: scratch_load_b32 v33, off, s33 offset:136 +; DAGISEL64-NEXT: scratch_load_b32 v34, off, s33 offset:140 +; DAGISEL64-NEXT: scratch_load_b32 v35, off, s33 offset:144 +; DAGISEL64-NEXT: scratch_load_b32 v36, off, s33 offset:148 +; DAGISEL64-NEXT: scratch_load_b32 v37, off, s33 offset:152 +; DAGISEL64-NEXT: scratch_load_b32 v38, off, s33 offset:156 +; DAGISEL64-NEXT: scratch_load_b32 v39, off, s33 offset:160 +; DAGISEL64-NEXT: scratch_load_b32 v48, off, s33 offset:164 +; DAGISEL64-NEXT: scratch_load_b32 v49, off, s33 offset:168 +; DAGISEL64-NEXT: scratch_load_b32 v50, off, s33 offset:172 +; DAGISEL64-NEXT: scratch_load_b32 v51, off, s33 offset:176 +; DAGISEL64-NEXT: scratch_load_b32 v52, off, s33 offset:180 +; DAGISEL64-NEXT: scratch_load_b32 v53, off, s33 offset:184 +; DAGISEL64-NEXT: scratch_load_b32 v54, off, s33 offset:188 +; DAGISEL64-NEXT: scratch_load_b32 v55, off, s33 offset:192 +; DAGISEL64-NEXT: scratch_load_b32 v64, off, s33 offset:196 +; DAGISEL64-NEXT: scratch_load_b32 v65, off, s33 offset:200 +; DAGISEL64-NEXT: scratch_load_b32 v66, off, s33 offset:204 +; DAGISEL64-NEXT: scratch_load_b32 v67, off, s33 offset:208 +; DAGISEL64-NEXT: scratch_load_b32 v68, off, s33 offset:212 +; DAGISEL64-NEXT: scratch_load_b32 v69, off, s33 offset:216 +; DAGISEL64-NEXT: scratch_load_b32 v70, off, s33 offset:220 +; DAGISEL64-NEXT: scratch_load_b32 v71, off, s33 offset:224 +; DAGISEL64-NEXT: scratch_load_b32 v80, off, s33 offset:228 +; DAGISEL64-NEXT: scratch_load_b32 v81, off, s33 offset:232 +; DAGISEL64-NEXT: scratch_load_b32 v82, off, s33 offset:236 +; DAGISEL64-NEXT: scratch_load_b32 v83, off, s33 offset:240 +; DAGISEL64-NEXT: scratch_load_b32 v84, off, s33 offset:244 +; DAGISEL64-NEXT: scratch_load_b32 v85, off, s33 offset:248 +; DAGISEL64-NEXT: scratch_load_b32 v86, off, s33 offset:252 +; DAGISEL64-NEXT: scratch_load_b32 v87, off, s33 offset:256 +; DAGISEL64-NEXT: s_clause 0x1f +; DAGISEL64-NEXT: scratch_load_b32 v96, off, s33 offset:260 +; DAGISEL64-NEXT: scratch_load_b32 v97, off, s33 offset:264 +; DAGISEL64-NEXT: scratch_load_b32 v98, off, s33 offset:268 +; DAGISEL64-NEXT: scratch_load_b32 v99, off, s33 offset:272 +; DAGISEL64-NEXT: scratch_load_b32 v100, off, s33 offset:276 +; DAGISEL64-NEXT: scratch_load_b32 v101, off, s33 offset:280 +; DAGISEL64-NEXT: scratch_load_b32 v102, off, s33 offset:284 +; DAGISEL64-NEXT: scratch_load_b32 v103, off, s33 offset:288 +; DAGISEL64-NEXT: scratch_load_b32 v112, off, s33 offset:292 +; DAGISEL64-NEXT: scratch_load_b32 v113, off, s33 offset:296 +; DAGISEL64-NEXT: scratch_load_b32 v114, off, s33 offset:300 +; DAGISEL64-NEXT: scratch_load_b32 v115, off, s33 offset:304 +; DAGISEL64-NEXT: scratch_load_b32 v116, off, s33 offset:308 +; DAGISEL64-NEXT: scratch_load_b32 v117, off, s33 offset:312 +; DAGISEL64-NEXT: scratch_load_b32 v118, off, s33 offset:316 +; DAGISEL64-NEXT: scratch_load_b32 v119, off, s33 offset:320 +; DAGISEL64-NEXT: scratch_load_b32 v128, off, s33 offset:324 +; DAGISEL64-NEXT: scratch_load_b32 v129, off, s33 offset:328 +; DAGISEL64-NEXT: scratch_load_b32 v130, off, s33 offset:332 +; DAGISEL64-NEXT: scratch_load_b32 v131, off, s33 offset:336 +; DAGISEL64-NEXT: scratch_load_b32 v132, off, s33 offset:340 +; DAGISEL64-NEXT: scratch_load_b32 v133, off, s33 offset:344 +; DAGISEL64-NEXT: scratch_load_b32 v134, off, s33 offset:348 +; DAGISEL64-NEXT: scratch_load_b32 v135, off, s33 offset:352 +; DAGISEL64-NEXT: scratch_load_b32 v144, off, s33 offset:356 +; DAGISEL64-NEXT: scratch_load_b32 v145, off, s33 offset:360 +; DAGISEL64-NEXT: scratch_load_b32 v146, off, s33 offset:364 +; DAGISEL64-NEXT: scratch_load_b32 v147, off, s33 offset:368 +; DAGISEL64-NEXT: scratch_load_b32 v148, off, s33 offset:372 +; DAGISEL64-NEXT: scratch_load_b32 v149, off, s33 offset:376 +; DAGISEL64-NEXT: scratch_load_b32 v150, off, s33 offset:380 +; DAGISEL64-NEXT: scratch_load_b32 v151, off, s33 offset:384 +; DAGISEL64-NEXT: s_clause 0x1f +; DAGISEL64-NEXT: scratch_load_b32 v160, off, s33 offset:388 +; DAGISEL64-NEXT: scratch_load_b32 v161, off, s33 offset:392 +; DAGISEL64-NEXT: scratch_load_b32 v162, off, s33 offset:396 +; DAGISEL64-NEXT: scratch_load_b32 v163, off, s33 offset:400 +; DAGISEL64-NEXT: scratch_load_b32 v164, off, s33 offset:404 +; DAGISEL64-NEXT: scratch_load_b32 v165, off, s33 offset:408 +; DAGISEL64-NEXT: scratch_load_b32 v166, off, s33 offset:412 +; DAGISEL64-NEXT: scratch_load_b32 v167, off, s33 offset:416 +; DAGISEL64-NEXT: scratch_load_b32 v176, off, s33 offset:420 +; DAGISEL64-NEXT: scratch_load_b32 v177, off, s33 offset:424 +; DAGISEL64-NEXT: scratch_load_b32 v178, off, s33 offset:428 +; DAGISEL64-NEXT: scratch_load_b32 v179, off, s33 offset:432 +; DAGISEL64-NEXT: scratch_load_b32 v180, off, s33 offset:436 +; DAGISEL64-NEXT: scratch_load_b32 v181, off, s33 offset:440 +; DAGISEL64-NEXT: scratch_load_b32 v182, off, s33 offset:444 +; DAGISEL64-NEXT: scratch_load_b32 v183, off, s33 offset:448 +; DAGISEL64-NEXT: scratch_load_b32 v192, off, s33 offset:452 +; DAGISEL64-NEXT: scratch_load_b32 v193, off, s33 offset:456 +; DAGISEL64-NEXT: scratch_load_b32 v194, off, s33 offset:460 +; DAGISEL64-NEXT: scratch_load_b32 v195, off, s33 offset:464 +; DAGISEL64-NEXT: scratch_load_b32 v196, off, s33 offset:468 +; DAGISEL64-NEXT: scratch_load_b32 v197, off, s33 offset:472 +; DAGISEL64-NEXT: scratch_load_b32 v198, off, s33 offset:476 +; DAGISEL64-NEXT: scratch_load_b32 v199, off, s33 offset:480 +; DAGISEL64-NEXT: scratch_load_b32 v208, off, s33 offset:484 +; DAGISEL64-NEXT: scratch_load_b32 v209, off, s33 offset:488 +; DAGISEL64-NEXT: scratch_load_b32 v210, off, s33 offset:492 +; DAGISEL64-NEXT: scratch_load_b32 v211, off, s33 offset:496 +; DAGISEL64-NEXT: scratch_load_b32 v212, off, s33 offset:500 +; DAGISEL64-NEXT: scratch_load_b32 v213, off, s33 offset:504 +; DAGISEL64-NEXT: scratch_load_b32 v214, off, s33 offset:508 +; DAGISEL64-NEXT: scratch_load_b32 v215, off, s33 offset:512 +; DAGISEL64-NEXT: s_clause 0xf +; DAGISEL64-NEXT: scratch_load_b32 v224, off, s33 offset:516 +; DAGISEL64-NEXT: scratch_load_b32 v225, off, s33 offset:520 +; DAGISEL64-NEXT: scratch_load_b32 v226, off, s33 offset:524 +; DAGISEL64-NEXT: scratch_load_b32 v227, off, s33 offset:528 +; DAGISEL64-NEXT: scratch_load_b32 v228, off, s33 offset:532 +; DAGISEL64-NEXT: scratch_load_b32 v229, off, s33 offset:536 +; DAGISEL64-NEXT: scratch_load_b32 v230, off, s33 offset:540 +; DAGISEL64-NEXT: scratch_load_b32 v231, off, s33 offset:544 +; DAGISEL64-NEXT: scratch_load_b32 v240, off, s33 offset:548 +; DAGISEL64-NEXT: scratch_load_b32 v241, off, s33 offset:552 +; DAGISEL64-NEXT: scratch_load_b32 v242, off, s33 offset:556 +; DAGISEL64-NEXT: scratch_load_b32 v243, off, s33 offset:560 +; DAGISEL64-NEXT: scratch_load_b32 v244, off, s33 offset:564 +; DAGISEL64-NEXT: scratch_load_b32 v245, off, s33 offset:568 +; DAGISEL64-NEXT: scratch_load_b32 v246, off, s33 offset:572 +; DAGISEL64-NEXT: scratch_load_b32 v247, off, s33 offset:576 +; DAGISEL64-NEXT: s_mov_b64 exec, s[4:5] +; DAGISEL64-NEXT: s_mov_b32 s33, s0 +; DAGISEL64-NEXT: s_wait_loadcnt 0x0 +; DAGISEL64-NEXT: s_wait_alu 0xfffe +; DAGISEL64-NEXT: s_setpc_b64 s[30:31] +; +; GISEL64-LABEL: call_gfx_from_whole_wave: +; GISEL64: ; %bb.0: +; GISEL64-NEXT: s_wait_loadcnt_dscnt 0x0 +; GISEL64-NEXT: s_wait_expcnt 0x0 +; GISEL64-NEXT: s_wait_samplecnt 0x0 +; GISEL64-NEXT: s_wait_bvhcnt 0x0 +; GISEL64-NEXT: s_wait_kmcnt 0x0 +; GISEL64-NEXT: s_mov_b32 s0, s33 +; GISEL64-NEXT: s_mov_b32 s33, s32 +; GISEL64-NEXT: s_xor_saveexec_b64 s[4:5], -1 +; GISEL64-NEXT: s_clause 0x1f +; GISEL64-NEXT: scratch_store_b32 off, v0, s33 offset:4 +; GISEL64-NEXT: scratch_store_b32 off, v1, s33 offset:8 +; GISEL64-NEXT: scratch_store_b32 off, v2, s33 offset:12 +; GISEL64-NEXT: scratch_store_b32 off, v3, s33 offset:16 +; GISEL64-NEXT: scratch_store_b32 off, v4, s33 offset:20 +; GISEL64-NEXT: scratch_store_b32 off, v5, s33 offset:24 +; GISEL64-NEXT: scratch_store_b32 off, v6, s33 offset:28 +; GISEL64-NEXT: scratch_store_b32 off, v7, s33 offset:32 +; GISEL64-NEXT: scratch_store_b32 off, v8, s33 offset:36 +; GISEL64-NEXT: scratch_store_b32 off, v9, s33 offset:40 +; GISEL64-NEXT: scratch_store_b32 off, v10, s33 offset:44 +; GISEL64-NEXT: scratch_store_b32 off, v11, s33 offset:48 +; GISEL64-NEXT: scratch_store_b32 off, v12, s33 offset:52 +; GISEL64-NEXT: scratch_store_b32 off, v13, s33 offset:56 +; GISEL64-NEXT: scratch_store_b32 off, v14, s33 offset:60 +; GISEL64-NEXT: scratch_store_b32 off, v15, s33 offset:64 +; GISEL64-NEXT: scratch_store_b32 off, v16, s33 offset:68 +; GISEL64-NEXT: scratch_store_b32 off, v17, s33 offset:72 +; GISEL64-NEXT: scratch_store_b32 off, v18, s33 offset:76 +; GISEL64-NEXT: scratch_store_b32 off, v19, s33 offset:80 +; GISEL64-NEXT: scratch_store_b32 off, v20, s33 offset:84 +; GISEL64-NEXT: scratch_store_b32 off, v21, s33 offset:88 +; GISEL64-NEXT: scratch_store_b32 off, v22, s33 offset:92 +; GISEL64-NEXT: scratch_store_b32 off, v23, s33 offset:96 +; GISEL64-NEXT: scratch_store_b32 off, v24, s33 offset:100 +; GISEL64-NEXT: scratch_store_b32 off, v25, s33 offset:104 +; GISEL64-NEXT: scratch_store_b32 off, v26, s33 offset:108 +; GISEL64-NEXT: scratch_store_b32 off, v27, s33 offset:112 +; GISEL64-NEXT: scratch_store_b32 off, v28, s33 offset:116 +; GISEL64-NEXT: scratch_store_b32 off, v29, s33 offset:120 +; GISEL64-NEXT: scratch_store_b32 off, v30, s33 offset:124 +; GISEL64-NEXT: scratch_store_b32 off, v31, s33 offset:128 +; GISEL64-NEXT: s_clause 0x1f +; GISEL64-NEXT: scratch_store_b32 off, v32, s33 offset:132 +; GISEL64-NEXT: scratch_store_b32 off, v33, s33 offset:136 +; GISEL64-NEXT: scratch_store_b32 off, v34, s33 offset:140 +; GISEL64-NEXT: scratch_store_b32 off, v35, s33 offset:144 +; GISEL64-NEXT: scratch_store_b32 off, v36, s33 offset:148 +; GISEL64-NEXT: scratch_store_b32 off, v37, s33 offset:152 +; GISEL64-NEXT: scratch_store_b32 off, v38, s33 offset:156 +; GISEL64-NEXT: scratch_store_b32 off, v39, s33 offset:160 +; GISEL64-NEXT: scratch_store_b32 off, v48, s33 offset:164 +; GISEL64-NEXT: scratch_store_b32 off, v49, s33 offset:168 +; GISEL64-NEXT: scratch_store_b32 off, v50, s33 offset:172 +; GISEL64-NEXT: scratch_store_b32 off, v51, s33 offset:176 +; GISEL64-NEXT: scratch_store_b32 off, v52, s33 offset:180 +; GISEL64-NEXT: scratch_store_b32 off, v53, s33 offset:184 +; GISEL64-NEXT: scratch_store_b32 off, v54, s33 offset:188 +; GISEL64-NEXT: scratch_store_b32 off, v55, s33 offset:192 +; GISEL64-NEXT: scratch_store_b32 off, v64, s33 offset:196 +; GISEL64-NEXT: scratch_store_b32 off, v65, s33 offset:200 +; GISEL64-NEXT: scratch_store_b32 off, v66, s33 offset:204 +; GISEL64-NEXT: scratch_store_b32 off, v67, s33 offset:208 +; GISEL64-NEXT: scratch_store_b32 off, v68, s33 offset:212 +; GISEL64-NEXT: scratch_store_b32 off, v69, s33 offset:216 +; GISEL64-NEXT: scratch_store_b32 off, v70, s33 offset:220 +; GISEL64-NEXT: scratch_store_b32 off, v71, s33 offset:224 +; GISEL64-NEXT: scratch_store_b32 off, v80, s33 offset:228 +; GISEL64-NEXT: scratch_store_b32 off, v81, s33 offset:232 +; GISEL64-NEXT: scratch_store_b32 off, v82, s33 offset:236 +; GISEL64-NEXT: scratch_store_b32 off, v83, s33 offset:240 +; GISEL64-NEXT: scratch_store_b32 off, v84, s33 offset:244 +; GISEL64-NEXT: scratch_store_b32 off, v85, s33 offset:248 +; GISEL64-NEXT: scratch_store_b32 off, v86, s33 offset:252 +; GISEL64-NEXT: scratch_store_b32 off, v87, s33 offset:256 +; GISEL64-NEXT: s_clause 0x1f +; GISEL64-NEXT: scratch_store_b32 off, v96, s33 offset:260 +; GISEL64-NEXT: scratch_store_b32 off, v97, s33 offset:264 +; GISEL64-NEXT: scratch_store_b32 off, v98, s33 offset:268 +; GISEL64-NEXT: scratch_store_b32 off, v99, s33 offset:272 +; GISEL64-NEXT: scratch_store_b32 off, v100, s33 offset:276 +; GISEL64-NEXT: scratch_store_b32 off, v101, s33 offset:280 +; GISEL64-NEXT: scratch_store_b32 off, v102, s33 offset:284 +; GISEL64-NEXT: scratch_store_b32 off, v103, s33 offset:288 +; GISEL64-NEXT: scratch_store_b32 off, v112, s33 offset:292 +; GISEL64-NEXT: scratch_store_b32 off, v113, s33 offset:296 +; GISEL64-NEXT: scratch_store_b32 off, v114, s33 offset:300 +; GISEL64-NEXT: scratch_store_b32 off, v115, s33 offset:304 +; GISEL64-NEXT: scratch_store_b32 off, v116, s33 offset:308 +; GISEL64-NEXT: scratch_store_b32 off, v117, s33 offset:312 +; GISEL64-NEXT: scratch_store_b32 off, v118, s33 offset:316 +; GISEL64-NEXT: scratch_store_b32 off, v119, s33 offset:320 +; GISEL64-NEXT: scratch_store_b32 off, v128, s33 offset:324 +; GISEL64-NEXT: scratch_store_b32 off, v129, s33 offset:328 +; GISEL64-NEXT: scratch_store_b32 off, v130, s33 offset:332 +; GISEL64-NEXT: scratch_store_b32 off, v131, s33 offset:336 +; GISEL64-NEXT: scratch_store_b32 off, v132, s33 offset:340 +; GISEL64-NEXT: scratch_store_b32 off, v133, s33 offset:344 +; GISEL64-NEXT: scratch_store_b32 off, v134, s33 offset:348 +; GISEL64-NEXT: scratch_store_b32 off, v135, s33 offset:352 +; GISEL64-NEXT: scratch_store_b32 off, v144, s33 offset:356 +; GISEL64-NEXT: scratch_store_b32 off, v145, s33 offset:360 +; GISEL64-NEXT: scratch_store_b32 off, v146, s33 offset:364 +; GISEL64-NEXT: scratch_store_b32 off, v147, s33 offset:368 +; GISEL64-NEXT: scratch_store_b32 off, v148, s33 offset:372 +; GISEL64-NEXT: scratch_store_b32 off, v149, s33 offset:376 +; GISEL64-NEXT: scratch_store_b32 off, v150, s33 offset:380 +; GISEL64-NEXT: scratch_store_b32 off, v151, s33 offset:384 +; GISEL64-NEXT: s_clause 0x1f +; GISEL64-NEXT: scratch_store_b32 off, v160, s33 offset:388 +; GISEL64-NEXT: scratch_store_b32 off, v161, s33 offset:392 +; GISEL64-NEXT: scratch_store_b32 off, v162, s33 offset:396 +; GISEL64-NEXT: scratch_store_b32 off, v163, s33 offset:400 +; GISEL64-NEXT: scratch_store_b32 off, v164, s33 offset:404 +; GISEL64-NEXT: scratch_store_b32 off, v165, s33 offset:408 +; GISEL64-NEXT: scratch_store_b32 off, v166, s33 offset:412 +; GISEL64-NEXT: scratch_store_b32 off, v167, s33 offset:416 +; GISEL64-NEXT: scratch_store_b32 off, v176, s33 offset:420 +; GISEL64-NEXT: scratch_store_b32 off, v177, s33 offset:424 +; GISEL64-NEXT: scratch_store_b32 off, v178, s33 offset:428 +; GISEL64-NEXT: scratch_store_b32 off, v179, s33 offset:432 +; GISEL64-NEXT: scratch_store_b32 off, v180, s33 offset:436 +; GISEL64-NEXT: scratch_store_b32 off, v181, s33 offset:440 +; GISEL64-NEXT: scratch_store_b32 off, v182, s33 offset:444 +; GISEL64-NEXT: scratch_store_b32 off, v183, s33 offset:448 +; GISEL64-NEXT: scratch_store_b32 off, v192, s33 offset:452 +; GISEL64-NEXT: scratch_store_b32 off, v193, s33 offset:456 +; GISEL64-NEXT: scratch_store_b32 off, v194, s33 offset:460 +; GISEL64-NEXT: scratch_store_b32 off, v195, s33 offset:464 +; GISEL64-NEXT: scratch_store_b32 off, v196, s33 offset:468 +; GISEL64-NEXT: scratch_store_b32 off, v197, s33 offset:472 +; GISEL64-NEXT: scratch_store_b32 off, v198, s33 offset:476 +; GISEL64-NEXT: scratch_store_b32 off, v199, s33 offset:480 +; GISEL64-NEXT: scratch_store_b32 off, v208, s33 offset:484 +; GISEL64-NEXT: scratch_store_b32 off, v209, s33 offset:488 +; GISEL64-NEXT: scratch_store_b32 off, v210, s33 offset:492 +; GISEL64-NEXT: scratch_store_b32 off, v211, s33 offset:496 +; GISEL64-NEXT: scratch_store_b32 off, v212, s33 offset:500 +; GISEL64-NEXT: scratch_store_b32 off, v213, s33 offset:504 +; GISEL64-NEXT: scratch_store_b32 off, v214, s33 offset:508 +; GISEL64-NEXT: scratch_store_b32 off, v215, s33 offset:512 +; GISEL64-NEXT: s_clause 0xf +; GISEL64-NEXT: scratch_store_b32 off, v224, s33 offset:516 +; GISEL64-NEXT: scratch_store_b32 off, v225, s33 offset:520 +; GISEL64-NEXT: scratch_store_b32 off, v226, s33 offset:524 +; GISEL64-NEXT: scratch_store_b32 off, v227, s33 offset:528 +; GISEL64-NEXT: scratch_store_b32 off, v228, s33 offset:532 +; GISEL64-NEXT: scratch_store_b32 off, v229, s33 offset:536 +; GISEL64-NEXT: scratch_store_b32 off, v230, s33 offset:540 +; GISEL64-NEXT: scratch_store_b32 off, v231, s33 offset:544 +; GISEL64-NEXT: scratch_store_b32 off, v240, s33 offset:548 +; GISEL64-NEXT: scratch_store_b32 off, v241, s33 offset:552 +; GISEL64-NEXT: scratch_store_b32 off, v242, s33 offset:556 +; GISEL64-NEXT: scratch_store_b32 off, v243, s33 offset:560 +; GISEL64-NEXT: scratch_store_b32 off, v244, s33 offset:564 +; GISEL64-NEXT: scratch_store_b32 off, v245, s33 offset:568 +; GISEL64-NEXT: scratch_store_b32 off, v246, s33 offset:572 +; GISEL64-NEXT: scratch_store_b32 off, v247, s33 offset:576 +; GISEL64-NEXT: s_mov_b64 exec, -1 +; GISEL64-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill +; GISEL64-NEXT: s_wait_alu 0xfffe +; GISEL64-NEXT: v_writelane_b32 v40, s0, 4 +; GISEL64-NEXT: v_mov_b32_e32 v2, v0 +; GISEL64-NEXT: v_swap_b32 v0, v1 +; GISEL64-NEXT: s_mov_b32 s0, gfx_callee@abs32@lo +; GISEL64-NEXT: v_writelane_b32 v40, s4, 0 +; GISEL64-NEXT: s_mov_b32 s1, gfx_callee@abs32@hi +; GISEL64-NEXT: s_addk_co_i32 s32, 0x250 +; GISEL64-NEXT: v_writelane_b32 v40, s5, 1 +; GISEL64-NEXT: v_writelane_b32 v40, s30, 2 +; GISEL64-NEXT: v_writelane_b32 v40, s31, 3 +; GISEL64-NEXT: s_wait_alu 0xfffe +; GISEL64-NEXT: s_swappc_b64 s[30:31], s[0:1] +; GISEL64-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GISEL64-NEXT: v_readlane_b32 s31, v40, 3 +; GISEL64-NEXT: v_readlane_b32 s30, v40, 2 +; GISEL64-NEXT: v_readlane_b32 s5, v40, 1 +; GISEL64-NEXT: v_readlane_b32 s4, v40, 0 +; GISEL64-NEXT: v_readlane_b32 s0, v40, 4 +; GISEL64-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload +; GISEL64-NEXT: s_mov_b32 s32, s33 +; GISEL64-NEXT: s_xor_b64 exec, s[4:5], -1 +; GISEL64-NEXT: s_clause 0x1f +; GISEL64-NEXT: scratch_load_b32 v0, off, s33 offset:4 +; GISEL64-NEXT: scratch_load_b32 v1, off, s33 offset:8 +; GISEL64-NEXT: scratch_load_b32 v2, off, s33 offset:12 +; GISEL64-NEXT: scratch_load_b32 v3, off, s33 offset:16 +; GISEL64-NEXT: scratch_load_b32 v4, off, s33 offset:20 +; GISEL64-NEXT: scratch_load_b32 v5, off, s33 offset:24 +; GISEL64-NEXT: scratch_load_b32 v6, off, s33 offset:28 +; GISEL64-NEXT: scratch_load_b32 v7, off, s33 offset:32 +; GISEL64-NEXT: scratch_load_b32 v8, off, s33 offset:36 +; GISEL64-NEXT: scratch_load_b32 v9, off, s33 offset:40 +; GISEL64-NEXT: scratch_load_b32 v10, off, s33 offset:44 +; GISEL64-NEXT: scratch_load_b32 v11, off, s33 offset:48 +; GISEL64-NEXT: scratch_load_b32 v12, off, s33 offset:52 +; GISEL64-NEXT: scratch_load_b32 v13, off, s33 offset:56 +; GISEL64-NEXT: scratch_load_b32 v14, off, s33 offset:60 +; GISEL64-NEXT: scratch_load_b32 v15, off, s33 offset:64 +; GISEL64-NEXT: scratch_load_b32 v16, off, s33 offset:68 +; GISEL64-NEXT: scratch_load_b32 v17, off, s33 offset:72 +; GISEL64-NEXT: scratch_load_b32 v18, off, s33 offset:76 +; GISEL64-NEXT: scratch_load_b32 v19, off, s33 offset:80 +; GISEL64-NEXT: scratch_load_b32 v20, off, s33 offset:84 +; GISEL64-NEXT: scratch_load_b32 v21, off, s33 offset:88 +; GISEL64-NEXT: scratch_load_b32 v22, off, s33 offset:92 +; GISEL64-NEXT: scratch_load_b32 v23, off, s33 offset:96 +; GISEL64-NEXT: scratch_load_b32 v24, off, s33 offset:100 +; GISEL64-NEXT: scratch_load_b32 v25, off, s33 offset:104 +; GISEL64-NEXT: scratch_load_b32 v26, off, s33 offset:108 +; GISEL64-NEXT: scratch_load_b32 v27, off, s33 offset:112 +; GISEL64-NEXT: scratch_load_b32 v28, off, s33 offset:116 +; GISEL64-NEXT: scratch_load_b32 v29, off, s33 offset:120 +; GISEL64-NEXT: scratch_load_b32 v30, off, s33 offset:124 +; GISEL64-NEXT: scratch_load_b32 v31, off, s33 offset:128 +; GISEL64-NEXT: s_clause 0x1f +; GISEL64-NEXT: scratch_load_b32 v32, off, s33 offset:132 +; GISEL64-NEXT: scratch_load_b32 v33, off, s33 offset:136 +; GISEL64-NEXT: scratch_load_b32 v34, off, s33 offset:140 +; GISEL64-NEXT: scratch_load_b32 v35, off, s33 offset:144 +; GISEL64-NEXT: scratch_load_b32 v36, off, s33 offset:148 +; GISEL64-NEXT: scratch_load_b32 v37, off, s33 offset:152 +; GISEL64-NEXT: scratch_load_b32 v38, off, s33 offset:156 +; GISEL64-NEXT: scratch_load_b32 v39, off, s33 offset:160 +; GISEL64-NEXT: scratch_load_b32 v48, off, s33 offset:164 +; GISEL64-NEXT: scratch_load_b32 v49, off, s33 offset:168 +; GISEL64-NEXT: scratch_load_b32 v50, off, s33 offset:172 +; GISEL64-NEXT: scratch_load_b32 v51, off, s33 offset:176 +; GISEL64-NEXT: scratch_load_b32 v52, off, s33 offset:180 +; GISEL64-NEXT: scratch_load_b32 v53, off, s33 offset:184 +; GISEL64-NEXT: scratch_load_b32 v54, off, s33 offset:188 +; GISEL64-NEXT: scratch_load_b32 v55, off, s33 offset:192 +; GISEL64-NEXT: scratch_load_b32 v64, off, s33 offset:196 +; GISEL64-NEXT: scratch_load_b32 v65, off, s33 offset:200 +; GISEL64-NEXT: scratch_load_b32 v66, off, s33 offset:204 +; GISEL64-NEXT: scratch_load_b32 v67, off, s33 offset:208 +; GISEL64-NEXT: scratch_load_b32 v68, off, s33 offset:212 +; GISEL64-NEXT: scratch_load_b32 v69, off, s33 offset:216 +; GISEL64-NEXT: scratch_load_b32 v70, off, s33 offset:220 +; GISEL64-NEXT: scratch_load_b32 v71, off, s33 offset:224 +; GISEL64-NEXT: scratch_load_b32 v80, off, s33 offset:228 +; GISEL64-NEXT: scratch_load_b32 v81, off, s33 offset:232 +; GISEL64-NEXT: scratch_load_b32 v82, off, s33 offset:236 +; GISEL64-NEXT: scratch_load_b32 v83, off, s33 offset:240 +; GISEL64-NEXT: scratch_load_b32 v84, off, s33 offset:244 +; GISEL64-NEXT: scratch_load_b32 v85, off, s33 offset:248 +; GISEL64-NEXT: scratch_load_b32 v86, off, s33 offset:252 +; GISEL64-NEXT: scratch_load_b32 v87, off, s33 offset:256 +; GISEL64-NEXT: s_clause 0x1f +; GISEL64-NEXT: scratch_load_b32 v96, off, s33 offset:260 +; GISEL64-NEXT: scratch_load_b32 v97, off, s33 offset:264 +; GISEL64-NEXT: scratch_load_b32 v98, off, s33 offset:268 +; GISEL64-NEXT: scratch_load_b32 v99, off, s33 offset:272 +; GISEL64-NEXT: scratch_load_b32 v100, off, s33 offset:276 +; GISEL64-NEXT: scratch_load_b32 v101, off, s33 offset:280 +; GISEL64-NEXT: scratch_load_b32 v102, off, s33 offset:284 +; GISEL64-NEXT: scratch_load_b32 v103, off, s33 offset:288 +; GISEL64-NEXT: scratch_load_b32 v112, off, s33 offset:292 +; GISEL64-NEXT: scratch_load_b32 v113, off, s33 offset:296 +; GISEL64-NEXT: scratch_load_b32 v114, off, s33 offset:300 +; GISEL64-NEXT: scratch_load_b32 v115, off, s33 offset:304 +; GISEL64-NEXT: scratch_load_b32 v116, off, s33 offset:308 +; GISEL64-NEXT: scratch_load_b32 v117, off, s33 offset:312 +; GISEL64-NEXT: scratch_load_b32 v118, off, s33 offset:316 +; GISEL64-NEXT: scratch_load_b32 v119, off, s33 offset:320 +; GISEL64-NEXT: scratch_load_b32 v128, off, s33 offset:324 +; GISEL64-NEXT: scratch_load_b32 v129, off, s33 offset:328 +; GISEL64-NEXT: scratch_load_b32 v130, off, s33 offset:332 +; GISEL64-NEXT: scratch_load_b32 v131, off, s33 offset:336 +; GISEL64-NEXT: scratch_load_b32 v132, off, s33 offset:340 +; GISEL64-NEXT: scratch_load_b32 v133, off, s33 offset:344 +; GISEL64-NEXT: scratch_load_b32 v134, off, s33 offset:348 +; GISEL64-NEXT: scratch_load_b32 v135, off, s33 offset:352 +; GISEL64-NEXT: scratch_load_b32 v144, off, s33 offset:356 +; GISEL64-NEXT: scratch_load_b32 v145, off, s33 offset:360 +; GISEL64-NEXT: scratch_load_b32 v146, off, s33 offset:364 +; GISEL64-NEXT: scratch_load_b32 v147, off, s33 offset:368 +; GISEL64-NEXT: scratch_load_b32 v148, off, s33 offset:372 +; GISEL64-NEXT: scratch_load_b32 v149, off, s33 offset:376 +; GISEL64-NEXT: scratch_load_b32 v150, off, s33 offset:380 +; GISEL64-NEXT: scratch_load_b32 v151, off, s33 offset:384 +; GISEL64-NEXT: s_clause 0x1f +; GISEL64-NEXT: scratch_load_b32 v160, off, s33 offset:388 +; GISEL64-NEXT: scratch_load_b32 v161, off, s33 offset:392 +; GISEL64-NEXT: scratch_load_b32 v162, off, s33 offset:396 +; GISEL64-NEXT: scratch_load_b32 v163, off, s33 offset:400 +; GISEL64-NEXT: scratch_load_b32 v164, off, s33 offset:404 +; GISEL64-NEXT: scratch_load_b32 v165, off, s33 offset:408 +; GISEL64-NEXT: scratch_load_b32 v166, off, s33 offset:412 +; GISEL64-NEXT: scratch_load_b32 v167, off, s33 offset:416 +; GISEL64-NEXT: scratch_load_b32 v176, off, s33 offset:420 +; GISEL64-NEXT: scratch_load_b32 v177, off, s33 offset:424 +; GISEL64-NEXT: scratch_load_b32 v178, off, s33 offset:428 +; GISEL64-NEXT: scratch_load_b32 v179, off, s33 offset:432 +; GISEL64-NEXT: scratch_load_b32 v180, off, s33 offset:436 +; GISEL64-NEXT: scratch_load_b32 v181, off, s33 offset:440 +; GISEL64-NEXT: scratch_load_b32 v182, off, s33 offset:444 +; GISEL64-NEXT: scratch_load_b32 v183, off, s33 offset:448 +; GISEL64-NEXT: scratch_load_b32 v192, off, s33 offset:452 +; GISEL64-NEXT: scratch_load_b32 v193, off, s33 offset:456 +; GISEL64-NEXT: scratch_load_b32 v194, off, s33 offset:460 +; GISEL64-NEXT: scratch_load_b32 v195, off, s33 offset:464 +; GISEL64-NEXT: scratch_load_b32 v196, off, s33 offset:468 +; GISEL64-NEXT: scratch_load_b32 v197, off, s33 offset:472 +; GISEL64-NEXT: scratch_load_b32 v198, off, s33 offset:476 +; GISEL64-NEXT: scratch_load_b32 v199, off, s33 offset:480 +; GISEL64-NEXT: scratch_load_b32 v208, off, s33 offset:484 +; GISEL64-NEXT: scratch_load_b32 v209, off, s33 offset:488 +; GISEL64-NEXT: scratch_load_b32 v210, off, s33 offset:492 +; GISEL64-NEXT: scratch_load_b32 v211, off, s33 offset:496 +; GISEL64-NEXT: scratch_load_b32 v212, off, s33 offset:500 +; GISEL64-NEXT: scratch_load_b32 v213, off, s33 offset:504 +; GISEL64-NEXT: scratch_load_b32 v214, off, s33 offset:508 +; GISEL64-NEXT: scratch_load_b32 v215, off, s33 offset:512 +; GISEL64-NEXT: s_clause 0xf +; GISEL64-NEXT: scratch_load_b32 v224, off, s33 offset:516 +; GISEL64-NEXT: scratch_load_b32 v225, off, s33 offset:520 +; GISEL64-NEXT: scratch_load_b32 v226, off, s33 offset:524 +; GISEL64-NEXT: scratch_load_b32 v227, off, s33 offset:528 +; GISEL64-NEXT: scratch_load_b32 v228, off, s33 offset:532 +; GISEL64-NEXT: scratch_load_b32 v229, off, s33 offset:536 +; GISEL64-NEXT: scratch_load_b32 v230, off, s33 offset:540 +; GISEL64-NEXT: scratch_load_b32 v231, off, s33 offset:544 +; GISEL64-NEXT: scratch_load_b32 v240, off, s33 offset:548 +; GISEL64-NEXT: scratch_load_b32 v241, off, s33 offset:552 +; GISEL64-NEXT: scratch_load_b32 v242, off, s33 offset:556 +; GISEL64-NEXT: scratch_load_b32 v243, off, s33 offset:560 +; GISEL64-NEXT: scratch_load_b32 v244, off, s33 offset:564 +; GISEL64-NEXT: scratch_load_b32 v245, off, s33 offset:568 +; GISEL64-NEXT: scratch_load_b32 v246, off, s33 offset:572 +; GISEL64-NEXT: scratch_load_b32 v247, off, s33 offset:576 +; GISEL64-NEXT: s_mov_b64 exec, s[4:5] +; GISEL64-NEXT: s_mov_b32 s33, s0 +; GISEL64-NEXT: s_wait_loadcnt 0x0 +; GISEL64-NEXT: s_wait_alu 0xfffe +; GISEL64-NEXT: s_setpc_b64 s[30:31] + %ret = call amdgpu_gfx <2 x half>(<2 x half>, <2 x half>) @gfx_callee(<2 x half> %y, <2 x half> %x) convergent + ret <2 x half> %ret +} |