Doomhowl-Interactive/llvm-project
0
0
Fork 0
Code Issues Pull Requests Actions 5 Packages Projects Releases Wiki Activity
Files
main
llvm-project/llvm/test/CodeGen/AMDGPU/dynamic_stackalloc.ll
Aaditya fbd2467796 [AMDGPU] DPP implementations for Wave Reduction (#185814) 
Adding DPP reduction support for i32 types.
Supported Ops: `umin`, `min`, `umax`, `max`,
`add`, `sub`, `and`, `or`, `xor`.
2026-03-18 13:11:53 +05:30

3448 lines
174 KiB
LLVM
Raw Permalink Blame History

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefix=GFX9-SDAG %s
; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefix=GFX9-GISEL %s
; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11-SDAG %s
; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11-GISEL %s
define amdgpu_kernel void @test_dynamic_stackalloc_kernel_uniform(i32 %n) {
; GFX9-SDAG-LABEL: test_dynamic_stackalloc_kernel_uniform:
; GFX9-SDAG: ; %bb.0:
; GFX9-SDAG-NEXT: s_load_dword s4, s[8:9], 0x0
; GFX9-SDAG-NEXT: s_add_u32 s0, s0, s17
; GFX9-SDAG-NEXT: s_addc_u32 s1, s1, 0
; GFX9-SDAG-NEXT: s_movk_i32 s32, 0x400
; GFX9-SDAG-NEXT: s_mov_b32 s5, s32
; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-SDAG-NEXT: s_lshl2_add_u32 s4, s4, 15
; GFX9-SDAG-NEXT: s_and_b32 s4, s4, -16
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX9-SDAG-NEXT: s_lshl_b32 s4, s4, 6
; GFX9-SDAG-NEXT: s_mov_b32 s33, 0
; GFX9-SDAG-NEXT: s_add_i32 s32, s5, s4
; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], s5
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_endpgm
;
; GFX9-GISEL-LABEL: test_dynamic_stackalloc_kernel_uniform:
; GFX9-GISEL: ; %bb.0:
; GFX9-GISEL-NEXT: s_load_dword s5, s[8:9], 0x0
; GFX9-GISEL-NEXT: s_add_u32 s0, s0, s17
; GFX9-GISEL-NEXT: s_movk_i32 s32, 0x400
; GFX9-GISEL-NEXT: s_addc_u32 s1, s1, 0
; GFX9-GISEL-NEXT: s_mov_b32 s4, s32
; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-GISEL-NEXT: s_lshl2_add_u32 s5, s5, 15
; GFX9-GISEL-NEXT: s_and_b32 s5, s5, -16
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX9-GISEL-NEXT: v_mov_b32_e32 v1, s4
; GFX9-GISEL-NEXT: s_lshl_b32 s5, s5, 6
; GFX9-GISEL-NEXT: s_mov_b32 s33, 0
; GFX9-GISEL-NEXT: s_add_u32 s32, s4, s5
; GFX9-GISEL-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_endpgm
;
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_kernel_uniform:
; GFX11-SDAG: ; %bb.0:
; GFX11-SDAG-NEXT: s_load_b32 s0, s[4:5], 0x0
; GFX11-SDAG-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX11-SDAG-NEXT: s_mov_b32 s32, 16
; GFX11-SDAG-NEXT: s_mov_b32 s33, 0
; GFX11-SDAG-NEXT: s_mov_b32 s1, s32
; GFX11-SDAG-NEXT: scratch_store_b32 off, v0, s1 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: s_lshl2_add_u32 s0, s0, 15
; GFX11-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: s_and_b32 s0, s0, -16
; GFX11-SDAG-NEXT: s_lshl_b32 s0, s0, 5
; GFX11-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: s_add_i32 s32, s1, s0
; GFX11-SDAG-NEXT: s_endpgm
;
; GFX11-GISEL-LABEL: test_dynamic_stackalloc_kernel_uniform:
; GFX11-GISEL: ; %bb.0:
; GFX11-GISEL-NEXT: s_load_b32 s1, s[4:5], 0x0
; GFX11-GISEL-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX11-GISEL-NEXT: s_mov_b32 s32, 16
; GFX11-GISEL-NEXT: s_mov_b32 s33, 0
; GFX11-GISEL-NEXT: s_mov_b32 s0, s32
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s0 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: s_lshl2_add_u32 s1, s1, 15
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: s_and_b32 s1, s1, -16
; GFX11-GISEL-NEXT: s_lshl_b32 s1, s1, 5
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: s_add_u32 s32, s0, s1
; GFX11-GISEL-NEXT: s_endpgm
%alloca = alloca i32, i32 %n, addrspace(5)
store volatile i32 123, ptr addrspace(5) %alloca
ret void
}
define amdgpu_kernel void @test_dynamic_stackalloc_kernel_uniform_over_aligned(i32 %n) {
; GFX9-SDAG-LABEL: test_dynamic_stackalloc_kernel_uniform_over_aligned:
; GFX9-SDAG: ; %bb.0:
; GFX9-SDAG-NEXT: s_load_dword s4, s[8:9], 0x0
; GFX9-SDAG-NEXT: s_movk_i32 s32, 0x2000
; GFX9-SDAG-NEXT: s_add_u32 s0, s0, s17
; GFX9-SDAG-NEXT: s_addc_u32 s1, s1, 0
; GFX9-SDAG-NEXT: s_add_i32 s5, s32, 0x1fff
; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-SDAG-NEXT: s_lshl2_add_u32 s4, s4, 15
; GFX9-SDAG-NEXT: s_and_b32 s5, s5, 0xffffe000
; GFX9-SDAG-NEXT: s_and_b32 s4, s4, -16
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 10
; GFX9-SDAG-NEXT: s_lshl_b32 s4, s4, 6
; GFX9-SDAG-NEXT: v_mov_b32_e32 v1, s5
; GFX9-SDAG-NEXT: s_mov_b32 s33, 0
; GFX9-SDAG-NEXT: s_add_i32 s32, s5, s4
; GFX9-SDAG-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_endpgm
;
; GFX9-GISEL-LABEL: test_dynamic_stackalloc_kernel_uniform_over_aligned:
; GFX9-GISEL: ; %bb.0:
; GFX9-GISEL-NEXT: s_load_dword s4, s[8:9], 0x0
; GFX9-GISEL-NEXT: s_movk_i32 s32, 0x2000
; GFX9-GISEL-NEXT: s_add_u32 s0, s0, s17
; GFX9-GISEL-NEXT: s_addc_u32 s1, s1, 0
; GFX9-GISEL-NEXT: s_add_u32 s5, s32, 0x1fff
; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-GISEL-NEXT: s_lshl2_add_u32 s4, s4, 15
; GFX9-GISEL-NEXT: s_and_b32 s5, s5, 0xffffe000
; GFX9-GISEL-NEXT: s_and_b32 s4, s4, -16
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 10
; GFX9-GISEL-NEXT: v_mov_b32_e32 v1, s5
; GFX9-GISEL-NEXT: s_lshl_b32 s4, s4, 6
; GFX9-GISEL-NEXT: s_mov_b32 s33, 0
; GFX9-GISEL-NEXT: s_add_u32 s32, s5, s4
; GFX9-GISEL-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_endpgm
;
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_kernel_uniform_over_aligned:
; GFX11-SDAG: ; %bb.0:
; GFX11-SDAG-NEXT: s_load_b32 s0, s[4:5], 0x0
; GFX11-SDAG-NEXT: s_movk_i32 s32, 0x80
; GFX11-SDAG-NEXT: v_mov_b32_e32 v0, 10
; GFX11-SDAG-NEXT: s_add_i32 s1, s32, 0xfff
; GFX11-SDAG-NEXT: s_mov_b32 s33, 0
; GFX11-SDAG-NEXT: s_and_b32 s1, s1, 0xfffff000
; GFX11-SDAG-NEXT: scratch_store_b32 off, v0, s1 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: s_lshl2_add_u32 s0, s0, 15
; GFX11-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: s_and_b32 s0, s0, -16
; GFX11-SDAG-NEXT: s_lshl_b32 s0, s0, 5
; GFX11-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: s_add_i32 s32, s1, s0
; GFX11-SDAG-NEXT: s_endpgm
;
; GFX11-GISEL-LABEL: test_dynamic_stackalloc_kernel_uniform_over_aligned:
; GFX11-GISEL: ; %bb.0:
; GFX11-GISEL-NEXT: s_load_b32 s0, s[4:5], 0x0
; GFX11-GISEL-NEXT: s_movk_i32 s32, 0x80
; GFX11-GISEL-NEXT: v_mov_b32_e32 v0, 10
; GFX11-GISEL-NEXT: s_add_u32 s1, s32, 0xfff
; GFX11-GISEL-NEXT: s_mov_b32 s33, 0
; GFX11-GISEL-NEXT: s_and_b32 s1, s1, 0xfffff000
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s1 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: s_lshl2_add_u32 s0, s0, 15
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: s_and_b32 s0, s0, -16
; GFX11-GISEL-NEXT: s_lshl_b32 s0, s0, 5
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: s_add_u32 s32, s1, s0
; GFX11-GISEL-NEXT: s_endpgm
%alloca = alloca i32, i32 %n, align 128, addrspace(5)
store volatile i32 10, ptr addrspace(5) %alloca
ret void
}
define amdgpu_kernel void @test_dynamic_stackalloc_kernel_uniform_under_aligned(i32 %n) {
; GFX9-SDAG-LABEL: test_dynamic_stackalloc_kernel_uniform_under_aligned:
; GFX9-SDAG: ; %bb.0:
; GFX9-SDAG-NEXT: s_load_dword s4, s[8:9], 0x0
; GFX9-SDAG-NEXT: s_add_u32 s0, s0, s17
; GFX9-SDAG-NEXT: s_addc_u32 s1, s1, 0
; GFX9-SDAG-NEXT: s_movk_i32 s32, 0x400
; GFX9-SDAG-NEXT: s_mov_b32 s5, s32
; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-SDAG-NEXT: s_lshl2_add_u32 s4, s4, 15
; GFX9-SDAG-NEXT: s_and_b32 s4, s4, -16
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 22
; GFX9-SDAG-NEXT: s_lshl_b32 s4, s4, 6
; GFX9-SDAG-NEXT: s_mov_b32 s33, 0
; GFX9-SDAG-NEXT: s_add_i32 s32, s5, s4
; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], s5
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_endpgm
;
; GFX9-GISEL-LABEL: test_dynamic_stackalloc_kernel_uniform_under_aligned:
; GFX9-GISEL: ; %bb.0:
; GFX9-GISEL-NEXT: s_load_dword s5, s[8:9], 0x0
; GFX9-GISEL-NEXT: s_add_u32 s0, s0, s17
; GFX9-GISEL-NEXT: s_movk_i32 s32, 0x400
; GFX9-GISEL-NEXT: s_addc_u32 s1, s1, 0
; GFX9-GISEL-NEXT: s_mov_b32 s4, s32
; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-GISEL-NEXT: s_lshl2_add_u32 s5, s5, 15
; GFX9-GISEL-NEXT: s_and_b32 s5, s5, -16
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 22
; GFX9-GISEL-NEXT: v_mov_b32_e32 v1, s4
; GFX9-GISEL-NEXT: s_lshl_b32 s5, s5, 6
; GFX9-GISEL-NEXT: s_mov_b32 s33, 0
; GFX9-GISEL-NEXT: s_add_u32 s32, s4, s5
; GFX9-GISEL-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_endpgm
;
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_kernel_uniform_under_aligned:
; GFX11-SDAG: ; %bb.0:
; GFX11-SDAG-NEXT: s_load_b32 s0, s[4:5], 0x0
; GFX11-SDAG-NEXT: v_mov_b32_e32 v0, 22
; GFX11-SDAG-NEXT: s_mov_b32 s32, 16
; GFX11-SDAG-NEXT: s_mov_b32 s33, 0
; GFX11-SDAG-NEXT: s_mov_b32 s1, s32
; GFX11-SDAG-NEXT: scratch_store_b32 off, v0, s1 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: s_lshl2_add_u32 s0, s0, 15
; GFX11-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: s_and_b32 s0, s0, -16
; GFX11-SDAG-NEXT: s_lshl_b32 s0, s0, 5
; GFX11-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: s_add_i32 s32, s1, s0
; GFX11-SDAG-NEXT: s_endpgm
;
; GFX11-GISEL-LABEL: test_dynamic_stackalloc_kernel_uniform_under_aligned:
; GFX11-GISEL: ; %bb.0:
; GFX11-GISEL-NEXT: s_load_b32 s1, s[4:5], 0x0
; GFX11-GISEL-NEXT: v_mov_b32_e32 v0, 22
; GFX11-GISEL-NEXT: s_mov_b32 s32, 16
; GFX11-GISEL-NEXT: s_mov_b32 s33, 0
; GFX11-GISEL-NEXT: s_mov_b32 s0, s32
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s0 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: s_lshl2_add_u32 s1, s1, 15
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: s_and_b32 s1, s1, -16
; GFX11-GISEL-NEXT: s_lshl_b32 s1, s1, 5
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: s_add_u32 s32, s0, s1
; GFX11-GISEL-NEXT: s_endpgm
%alloca = alloca i32, i32 %n, align 2, addrspace(5)
store volatile i32 22, ptr addrspace(5) %alloca
ret void
}
define amdgpu_kernel void @test_dynamic_stackalloc_kernel_divergent() {
; GFX9-SDAG-LABEL: test_dynamic_stackalloc_kernel_divergent:
; GFX9-SDAG: ; %bb.0:
; GFX9-SDAG-NEXT: s_add_u32 s0, s0, s17
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-SDAG-NEXT: s_mov_b32 s33, 0
; GFX9-SDAG-NEXT: s_movk_i32 s32, 0x400
; GFX9-SDAG-NEXT: s_addc_u32 s1, s1, 0
; GFX9-SDAG-NEXT: v_and_b32_e32 v0, 0x1ff0, v0
; GFX9-SDAG-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[4:5]
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: v_readlane_b32 s6, v1, 63
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s4, s32
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, s4
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, s6, 6, v0
; GFX9-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], s4
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_endpgm
;
; GFX9-GISEL-LABEL: test_dynamic_stackalloc_kernel_divergent:
; GFX9-GISEL: ; %bb.0:
; GFX9-GISEL-NEXT: s_movk_i32 s32, 0x400
; GFX9-GISEL-NEXT: s_add_u32 s0, s0, s17
; GFX9-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-GISEL-NEXT: s_mov_b32 s33, 0
; GFX9-GISEL-NEXT: s_addc_u32 s1, s1, 0
; GFX9-GISEL-NEXT: s_mov_b32 s6, s32
; GFX9-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-GISEL-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[4:5]
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: v_readlane_b32 s7, v1, 63
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_lshl_b32 s4, s7, 6
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX9-GISEL-NEXT: v_mov_b32_e32 v2, s6
; GFX9-GISEL-NEXT: s_add_u32 s32, s6, s4
; GFX9-GISEL-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_endpgm
;
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_kernel_divergent:
; GFX11-SDAG: ; %bb.0:
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, 0x3ff, v0
; GFX11-SDAG-NEXT: s_mov_b32 s33, 0
; GFX11-SDAG-NEXT: s_mov_b32 s32, 16
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, 0x1ff0, v0
; GFX11-SDAG-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s0
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_readlane_b32 s1, v1, 31
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s0, s32
; GFX11-SDAG-NEXT: v_mov_b32_e32 v3, 0x7b
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, s1, 5, s0
; GFX11-SDAG-NEXT: scratch_store_b32 off, v3, s0 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX11-SDAG-NEXT: s_endpgm
;
; GFX11-GISEL-LABEL: test_dynamic_stackalloc_kernel_divergent:
; GFX11-GISEL: ; %bb.0:
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, 0x3ff, v0
; GFX11-GISEL-NEXT: s_mov_b32 s32, 16
; GFX11-GISEL-NEXT: s_mov_b32 s33, 0
; GFX11-GISEL-NEXT: s_mov_b32 s0, s32
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-GISEL-NEXT: s_or_saveexec_b32 s1, -1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_readlane_b32 s2, v1, 31
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s1
; GFX11-GISEL-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX11-GISEL-NEXT: s_lshl_b32 s1, s2, 5
; GFX11-GISEL-NEXT: s_add_u32 s32, s0, s1
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s0 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_endpgm
%idx = call i32 @llvm.amdgcn.workitem.id.x()
%alloca = alloca float, i32 %idx, addrspace(5)
store volatile i32 123, ptr addrspace(5) %alloca
ret void
}
define amdgpu_kernel void @test_dynamic_stackalloc_kernel_divergent_over_aligned() {
; GFX9-SDAG-LABEL: test_dynamic_stackalloc_kernel_divergent_over_aligned:
; GFX9-SDAG: ; %bb.0:
; GFX9-SDAG-NEXT: s_movk_i32 s32, 0x2000
; GFX9-SDAG-NEXT: s_add_u32 s0, s0, s17
; GFX9-SDAG-NEXT: s_addc_u32 s1, s1, 0
; GFX9-SDAG-NEXT: s_add_i32 s4, s32, 0x1fff
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-SDAG-NEXT: s_mov_b32 s33, 0
; GFX9-SDAG-NEXT: s_and_b32 s6, s4, 0xffffe000
; GFX9-SDAG-NEXT: v_and_b32_e32 v0, 0x1ff0, v0
; GFX9-SDAG-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[4:5]
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: v_readlane_b32 s7, v1, 63
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, s6
; GFX9-SDAG-NEXT: v_lshl_add_u32 v2, s7, 6, v0
; GFX9-SDAG-NEXT: v_readfirstlane_b32 s32, v2
; GFX9-SDAG-NEXT: v_mov_b32_e32 v2, 0x1bc
; GFX9-SDAG-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_endpgm
;
; GFX9-GISEL-LABEL: test_dynamic_stackalloc_kernel_divergent_over_aligned:
; GFX9-GISEL: ; %bb.0:
; GFX9-GISEL-NEXT: s_add_u32 s0, s0, s17
; GFX9-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-GISEL-NEXT: s_mov_b32 s33, 0
; GFX9-GISEL-NEXT: s_movk_i32 s32, 0x2000
; GFX9-GISEL-NEXT: s_addc_u32 s1, s1, 0
; GFX9-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-GISEL-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[4:5]
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: v_readlane_b32 s6, v1, 63
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_add_u32 s5, s32, 0x1fff
; GFX9-GISEL-NEXT: s_and_b32 s5, s5, 0xffffe000
; GFX9-GISEL-NEXT: s_lshl_b32 s4, s6, 6
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 0x1bc
; GFX9-GISEL-NEXT: v_mov_b32_e32 v2, s5
; GFX9-GISEL-NEXT: s_add_u32 s32, s5, s4
; GFX9-GISEL-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_endpgm
;
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_kernel_divergent_over_aligned:
; GFX11-SDAG: ; %bb.0:
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, 0x3ff, v0
; GFX11-SDAG-NEXT: s_mov_b32 s33, 0
; GFX11-SDAG-NEXT: s_movk_i32 s32, 0x80
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, 0x1ff0, v0
; GFX11-SDAG-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s0
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_readlane_b32 s1, v1, 31
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_add_i32 s0, s32, 0xfff
; GFX11-SDAG-NEXT: v_mov_b32_e32 v3, 0x1bc
; GFX11-SDAG-NEXT: s_and_b32 s0, s0, 0xfffff000
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, s1, 5, s0
; GFX11-SDAG-NEXT: scratch_store_b32 off, v3, s0 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX11-SDAG-NEXT: s_endpgm
;
; GFX11-GISEL-LABEL: test_dynamic_stackalloc_kernel_divergent_over_aligned:
; GFX11-GISEL: ; %bb.0:
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, 0x3ff, v0
; GFX11-GISEL-NEXT: s_mov_b32 s33, 0
; GFX11-GISEL-NEXT: s_movk_i32 s32, 0x80
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-GISEL-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s0
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_readlane_b32 s1, v1, 31
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: v_mov_b32_e32 v0, 0x1bc
; GFX11-GISEL-NEXT: s_add_u32 s0, s32, 0xfff
; GFX11-GISEL-NEXT: s_lshl_b32 s1, s1, 5
; GFX11-GISEL-NEXT: s_and_b32 s0, s0, 0xfffff000
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: s_add_u32 s32, s0, s1
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s0 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_endpgm
%idx = call i32 @llvm.amdgcn.workitem.id.x()
%alloca = alloca i32, i32 %idx, align 128, addrspace(5)
store volatile i32 444, ptr addrspace(5) %alloca
ret void
}
define amdgpu_kernel void @test_dynamic_stackalloc_kernel_divergent_under_aligned() {
; GFX9-SDAG-LABEL: test_dynamic_stackalloc_kernel_divergent_under_aligned:
; GFX9-SDAG: ; %bb.0:
; GFX9-SDAG-NEXT: s_add_u32 s0, s0, s17
; GFX9-SDAG-NEXT: s_mov_b32 s33, 0
; GFX9-SDAG-NEXT: s_movk_i32 s32, 0x400
; GFX9-SDAG-NEXT: s_addc_u32 s1, s1, 0
; GFX9-SDAG-NEXT: v_lshlrev_b32_e32 v0, 4, v0
; GFX9-SDAG-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[4:5]
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: v_readlane_b32 s6, v1, 63
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s4, s32
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, s4
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, s6, 6, v0
; GFX9-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 0x29a
; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], s4
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_endpgm
;
; GFX9-GISEL-LABEL: test_dynamic_stackalloc_kernel_divergent_under_aligned:
; GFX9-GISEL: ; %bb.0:
; GFX9-GISEL-NEXT: s_movk_i32 s32, 0x400
; GFX9-GISEL-NEXT: s_add_u32 s0, s0, s17
; GFX9-GISEL-NEXT: v_lshl_add_u32 v0, v0, 4, 15
; GFX9-GISEL-NEXT: s_mov_b32 s33, 0
; GFX9-GISEL-NEXT: s_addc_u32 s1, s1, 0
; GFX9-GISEL-NEXT: s_mov_b32 s6, s32
; GFX9-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-GISEL-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[4:5]
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: v_readlane_b32 s7, v1, 63
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_lshl_b32 s4, s7, 6
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 0x29a
; GFX9-GISEL-NEXT: v_mov_b32_e32 v2, s6
; GFX9-GISEL-NEXT: s_add_u32 s32, s6, s4
; GFX9-GISEL-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_endpgm
;
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_kernel_divergent_under_aligned:
; GFX11-SDAG: ; %bb.0:
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, 0x3ff, v0
; GFX11-SDAG-NEXT: s_mov_b32 s33, 0
; GFX11-SDAG-NEXT: s_mov_b32 s32, 16
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_lshlrev_b32_e32 v0, 4, v0
; GFX11-SDAG-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s0
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_readlane_b32 s1, v1, 31
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s0, s32
; GFX11-SDAG-NEXT: v_mov_b32_e32 v3, 0x29a
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, s1, 5, s0
; GFX11-SDAG-NEXT: scratch_store_b32 off, v3, s0 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX11-SDAG-NEXT: s_endpgm
;
; GFX11-GISEL-LABEL: test_dynamic_stackalloc_kernel_divergent_under_aligned:
; GFX11-GISEL: ; %bb.0:
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, 0x3ff, v0
; GFX11-GISEL-NEXT: s_mov_b32 s32, 16
; GFX11-GISEL-NEXT: s_mov_b32 s33, 0
; GFX11-GISEL-NEXT: s_mov_b32 s0, s32
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_lshl_add_u32 v0, v0, 4, 15
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-GISEL-NEXT: s_or_saveexec_b32 s1, -1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_readlane_b32 s2, v1, 31
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s1
; GFX11-GISEL-NEXT: v_mov_b32_e32 v0, 0x29a
; GFX11-GISEL-NEXT: s_lshl_b32 s1, s2, 5
; GFX11-GISEL-NEXT: s_add_u32 s32, s0, s1
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s0 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_endpgm
%idx = call i32 @llvm.amdgcn.workitem.id.x()
%alloca = alloca i128, i32 %idx, align 2, addrspace(5)
store volatile i32 666, ptr addrspace(5) %alloca
ret void
}
define amdgpu_kernel void @test_dynamic_stackalloc_kernel_multiple_allocas(i32 %n, i32 %m) {
; GFX9-SDAG-LABEL: test_dynamic_stackalloc_kernel_multiple_allocas:
; GFX9-SDAG: ; %bb.0: ; %entry
; GFX9-SDAG-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
; GFX9-SDAG-NEXT: s_add_u32 s0, s0, s17
; GFX9-SDAG-NEXT: s_addc_u32 s1, s1, 0
; GFX9-SDAG-NEXT: s_mov_b32 s33, 0
; GFX9-SDAG-NEXT: s_movk_i32 s32, 0x2000
; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-SDAG-NEXT: s_cmp_lg_u32 s4, 0
; GFX9-SDAG-NEXT: s_cbranch_scc1 .LBB6_2
; GFX9-SDAG-NEXT: ; %bb.1: ; %bb.0
; GFX9-SDAG-NEXT: s_lshl2_add_u32 s5, s5, 15
; GFX9-SDAG-NEXT: s_add_i32 s6, s32, 0xfff
; GFX9-SDAG-NEXT: s_and_b32 s5, s5, -16
; GFX9-SDAG-NEXT: s_and_b32 s8, s6, 0xfffff000
; GFX9-SDAG-NEXT: s_lshl_b32 s5, s5, 6
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-SDAG-NEXT: s_add_i32 s32, s8, s5
; GFX9-SDAG-NEXT: v_and_b32_e32 v0, 0x1ff0, v0
; GFX9-SDAG-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[6:7]
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: v_readlane_b32 s5, v1, 63
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-SDAG-NEXT: s_mov_b32 s6, s32
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, s6
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, s5, 6, v0
; GFX9-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 3
; GFX9-SDAG-NEXT: v_mov_b32_e32 v2, s8
; GFX9-SDAG-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 4
; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], s6
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: .LBB6_2: ; %bb.1
; GFX9-SDAG-NEXT: s_lshl2_add_u32 s4, s4, 15
; GFX9-SDAG-NEXT: s_and_b32 s4, s4, -16
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 1
; GFX9-SDAG-NEXT: s_lshl_b32 s4, s4, 6
; GFX9-SDAG-NEXT: s_mov_b32 s5, s32
; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], s33
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 2
; GFX9-SDAG-NEXT: s_add_i32 s32, s5, s4
; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], s5
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_endpgm
;
; GFX9-GISEL-LABEL: test_dynamic_stackalloc_kernel_multiple_allocas:
; GFX9-GISEL: ; %bb.0: ; %entry
; GFX9-GISEL-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
; GFX9-GISEL-NEXT: s_add_u32 s0, s0, s17
; GFX9-GISEL-NEXT: s_addc_u32 s1, s1, 0
; GFX9-GISEL-NEXT: s_mov_b32 s33, 0
; GFX9-GISEL-NEXT: s_movk_i32 s32, 0x2000
; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-GISEL-NEXT: s_cmp_lg_u32 s4, 0
; GFX9-GISEL-NEXT: s_cbranch_scc1 .LBB6_2
; GFX9-GISEL-NEXT: ; %bb.1: ; %bb.0
; GFX9-GISEL-NEXT: s_lshl2_add_u32 s5, s5, 15
; GFX9-GISEL-NEXT: s_and_b32 s5, s5, -16
; GFX9-GISEL-NEXT: s_add_u32 s6, s32, 0xfff
; GFX9-GISEL-NEXT: s_lshl_b32 s5, s5, 6
; GFX9-GISEL-NEXT: s_and_b32 s8, s6, 0xfffff000
; GFX9-GISEL-NEXT: s_add_u32 s32, s8, s5
; GFX9-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-GISEL-NEXT: s_mov_b32 s5, s32
; GFX9-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-GISEL-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[6:7]
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: v_readlane_b32 s9, v1, 63
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 3
; GFX9-GISEL-NEXT: v_mov_b32_e32 v2, s8
; GFX9-GISEL-NEXT: s_lshl_b32 s6, s9, 6
; GFX9-GISEL-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 4
; GFX9-GISEL-NEXT: v_mov_b32_e32 v2, s5
; GFX9-GISEL-NEXT: s_add_u32 s32, s5, s6
; GFX9-GISEL-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: .LBB6_2: ; %bb.1
; GFX9-GISEL-NEXT: s_lshl2_add_u32 s4, s4, 15
; GFX9-GISEL-NEXT: s_mov_b32 s5, s32
; GFX9-GISEL-NEXT: s_and_b32 s4, s4, -16
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 1
; GFX9-GISEL-NEXT: s_lshl_b32 s4, s4, 6
; GFX9-GISEL-NEXT: buffer_store_dword v0, off, s[0:3], s33
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 2
; GFX9-GISEL-NEXT: v_mov_b32_e32 v2, s5
; GFX9-GISEL-NEXT: s_add_u32 s32, s5, s4
; GFX9-GISEL-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_endpgm
;
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_kernel_multiple_allocas:
; GFX11-SDAG: ; %bb.0: ; %entry
; GFX11-SDAG-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
; GFX11-SDAG-NEXT: s_mov_b32 s33, 0
; GFX11-SDAG-NEXT: s_movk_i32 s32, 0x80
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: s_cmp_lg_u32 s0, 0
; GFX11-SDAG-NEXT: s_cbranch_scc1 .LBB6_2
; GFX11-SDAG-NEXT: ; %bb.1: ; %bb.0
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, 0x3ff, v0
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, 0x1ff0, v0
; GFX11-SDAG-NEXT: s_or_saveexec_b32 s2, -1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s2
; GFX11-SDAG-NEXT: s_lshl2_add_u32 s1, s1, 15
; GFX11-SDAG-NEXT: s_add_i32 s3, s32, 0x7ff
; GFX11-SDAG-NEXT: s_and_b32 s1, s1, -16
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_readlane_b32 s4, v1, 31
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s2
; GFX11-SDAG-NEXT: s_and_b32 s2, s3, 0xfffff800
; GFX11-SDAG-NEXT: s_lshl_b32 s1, s1, 5
; GFX11-SDAG-NEXT: v_dual_mov_b32 v3, 3 :: v_dual_mov_b32 v4, 4
; GFX11-SDAG-NEXT: s_add_i32 s32, s2, s1
; GFX11-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: s_mov_b32 s1, s32
; GFX11-SDAG-NEXT: scratch_store_b32 off, v3, s2 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: scratch_store_b32 off, v4, s1 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, s4, 5, s1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX11-SDAG-NEXT: .LBB6_2: ; %bb.1
; GFX11-SDAG-NEXT: s_lshl2_add_u32 s0, s0, 15
; GFX11-SDAG-NEXT: v_dual_mov_b32 v0, 1 :: v_dual_mov_b32 v3, 2
; GFX11-SDAG-NEXT: s_and_b32 s0, s0, -16
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-SDAG-NEXT: s_mov_b32 s1, s32
; GFX11-SDAG-NEXT: s_lshl_b32 s0, s0, 5
; GFX11-SDAG-NEXT: scratch_store_b32 off, v0, s33 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: scratch_store_b32 off, v3, s1 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: s_add_i32 s32, s1, s0
; GFX11-SDAG-NEXT: s_endpgm
;
; GFX11-GISEL-LABEL: test_dynamic_stackalloc_kernel_multiple_allocas:
; GFX11-GISEL: ; %bb.0: ; %entry
; GFX11-GISEL-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
; GFX11-GISEL-NEXT: s_mov_b32 s33, 0
; GFX11-GISEL-NEXT: s_movk_i32 s32, 0x80
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: s_cmp_lg_u32 s0, 0
; GFX11-GISEL-NEXT: s_cbranch_scc1 .LBB6_2
; GFX11-GISEL-NEXT: ; %bb.1: ; %bb.0
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, 0x3ff, v0
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-GISEL-NEXT: s_or_saveexec_b32 s2, -1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s2
; GFX11-GISEL-NEXT: s_lshl2_add_u32 s1, s1, 15
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_readlane_b32 s3, v1, 31
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s2
; GFX11-GISEL-NEXT: s_add_u32 s2, s32, 0x7ff
; GFX11-GISEL-NEXT: s_and_b32 s1, s1, -16
; GFX11-GISEL-NEXT: v_dual_mov_b32 v0, 3 :: v_dual_mov_b32 v3, 4
; GFX11-GISEL-NEXT: s_and_b32 s2, s2, 0xfffff800
; GFX11-GISEL-NEXT: s_lshl_b32 s1, s1, 5
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: s_add_u32 s32, s2, s1
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s2 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_mov_b32 s1, s32
; GFX11-GISEL-NEXT: s_lshl_b32 s2, s3, 5
; GFX11-GISEL-NEXT: scratch_store_b32 off, v3, s1 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_add_u32 s32, s1, s2
; GFX11-GISEL-NEXT: .LBB6_2: ; %bb.1
; GFX11-GISEL-NEXT: s_lshl2_add_u32 s0, s0, 15
; GFX11-GISEL-NEXT: v_dual_mov_b32 v0, 1 :: v_dual_mov_b32 v3, 2
; GFX11-GISEL-NEXT: s_and_b32 s0, s0, -16
; GFX11-GISEL-NEXT: s_mov_b32 s1, s32
; GFX11-GISEL-NEXT: s_lshl_b32 s0, s0, 5
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s33 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: scratch_store_b32 off, v3, s1 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_add_u32 s32, s1, s0
; GFX11-GISEL-NEXT: s_endpgm
entry:
%cond = icmp eq i32 %n, 0
%alloca1 = alloca i32, i32 8, addrspace(5)
%alloca2 = alloca i17, i32 %n, addrspace(5)
br i1 %cond, label %bb.0, label %bb.1
bb.0:
%idx = call i32 @llvm.amdgcn.workitem.id.x()
%alloca3 = alloca i32, i32 %m, align 64, addrspace(5)
%alloca4 = alloca i32, i32 %idx, align 4, addrspace(5)
store volatile i32 3, ptr addrspace(5) %alloca3
store volatile i32 4, ptr addrspace(5) %alloca4
br label %bb.1
bb.1:
store volatile i32 1, ptr addrspace(5) %alloca1
store volatile i32 2, ptr addrspace(5) %alloca2
ret void
}
define amdgpu_kernel void @test_dynamic_stackalloc_kernel_control_flow(i32 %n, i32 %m) {
; GFX9-SDAG-LABEL: test_dynamic_stackalloc_kernel_control_flow:
; GFX9-SDAG: ; %bb.0: ; %entry
; GFX9-SDAG-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
; GFX9-SDAG-NEXT: s_add_u32 s0, s0, s17
; GFX9-SDAG-NEXT: s_addc_u32 s1, s1, 0
; GFX9-SDAG-NEXT: s_mov_b32 s33, 0
; GFX9-SDAG-NEXT: s_movk_i32 s32, 0x1000
; GFX9-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-SDAG-NEXT: s_cmp_lg_u32 s4, 0
; GFX9-SDAG-NEXT: s_cbranch_scc0 .LBB7_4
; GFX9-SDAG-NEXT: ; %bb.1: ; %bb.1
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-SDAG-NEXT: v_and_b32_e32 v0, 0x1ff0, v0
; GFX9-SDAG-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[6:7]
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: v_readlane_b32 s4, v1, 63
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-SDAG-NEXT: s_mov_b32 s6, s32
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, s6
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, s4, 6, v0
; GFX9-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 1
; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], s6
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_cbranch_execnz .LBB7_3
; GFX9-SDAG-NEXT: .LBB7_2: ; %bb.0
; GFX9-SDAG-NEXT: s_add_i32 s4, s32, 0xfff
; GFX9-SDAG-NEXT: s_lshl2_add_u32 s5, s5, 15
; GFX9-SDAG-NEXT: s_and_b32 s4, s4, 0xfffff000
; GFX9-SDAG-NEXT: s_and_b32 s5, s5, -16
; GFX9-SDAG-NEXT: s_lshl_b32 s5, s5, 6
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 2
; GFX9-SDAG-NEXT: v_mov_b32_e32 v2, s4
; GFX9-SDAG-NEXT: s_add_i32 s32, s4, s5
; GFX9-SDAG-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: .LBB7_3: ; %bb.2
; GFX9-SDAG-NEXT: s_endpgm
; GFX9-SDAG-NEXT: .LBB7_4:
; GFX9-SDAG-NEXT: s_branch .LBB7_2
;
; GFX9-GISEL-LABEL: test_dynamic_stackalloc_kernel_control_flow:
; GFX9-GISEL: ; %bb.0: ; %entry
; GFX9-GISEL-NEXT: s_load_dwordx2 s[4:5], s[8:9], 0x0
; GFX9-GISEL-NEXT: s_add_u32 s0, s0, s17
; GFX9-GISEL-NEXT: s_addc_u32 s1, s1, 0
; GFX9-GISEL-NEXT: s_mov_b32 s33, 0
; GFX9-GISEL-NEXT: s_movk_i32 s32, 0x1000
; GFX9-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-GISEL-NEXT: s_cmp_lg_u32 s4, 0
; GFX9-GISEL-NEXT: s_mov_b32 s4, 1
; GFX9-GISEL-NEXT: s_cbranch_scc0 .LBB7_2
; GFX9-GISEL-NEXT: ; %bb.1: ; %bb.1
; GFX9-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-GISEL-NEXT: s_mov_b32 s8, s32
; GFX9-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-GISEL-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[6:7]
; GFX9-GISEL-NEXT: s_mov_b32 s4, 0
; GFX9-GISEL-NEXT: s_nop 0
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: v_readlane_b32 s9, v1, 63
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-GISEL-NEXT: s_lshl_b32 s6, s9, 6
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 1
; GFX9-GISEL-NEXT: v_mov_b32_e32 v2, s8
; GFX9-GISEL-NEXT: s_add_u32 s32, s8, s6
; GFX9-GISEL-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: .LBB7_2: ; %Flow
; GFX9-GISEL-NEXT: s_xor_b32 s4, s4, 1
; GFX9-GISEL-NEXT: s_cmp_lg_u32 s4, 0
; GFX9-GISEL-NEXT: s_cbranch_scc1 .LBB7_4
; GFX9-GISEL-NEXT: ; %bb.3: ; %bb.0
; GFX9-GISEL-NEXT: s_lshl2_add_u32 s4, s5, 15
; GFX9-GISEL-NEXT: s_add_u32 s5, s32, 0xfff
; GFX9-GISEL-NEXT: s_and_b32 s4, s4, -16
; GFX9-GISEL-NEXT: s_and_b32 s5, s5, 0xfffff000
; GFX9-GISEL-NEXT: s_lshl_b32 s4, s4, 6
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 2
; GFX9-GISEL-NEXT: v_mov_b32_e32 v2, s5
; GFX9-GISEL-NEXT: s_add_u32 s32, s5, s4
; GFX9-GISEL-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: .LBB7_4: ; %bb.2
; GFX9-GISEL-NEXT: s_endpgm
;
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_kernel_control_flow:
; GFX11-SDAG: ; %bb.0: ; %entry
; GFX11-SDAG-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
; GFX11-SDAG-NEXT: s_mov_b32 s33, 0
; GFX11-SDAG-NEXT: s_mov_b32 s32, 64
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: s_cmp_lg_u32 s0, 0
; GFX11-SDAG-NEXT: s_cbranch_scc0 .LBB7_4
; GFX11-SDAG-NEXT: ; %bb.1: ; %bb.1
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, 0x3ff, v0
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, 0x1ff0, v0
; GFX11-SDAG-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s0
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_readlane_b32 s2, v1, 31
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s3, s32
; GFX11-SDAG-NEXT: v_mov_b32_e32 v3, 1
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, s2, 5, s3
; GFX11-SDAG-NEXT: scratch_store_b32 off, v3, s3 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX11-SDAG-NEXT: s_cbranch_execnz .LBB7_3
; GFX11-SDAG-NEXT: .LBB7_2: ; %bb.0
; GFX11-SDAG-NEXT: s_lshl2_add_u32 s1, s1, 15
; GFX11-SDAG-NEXT: v_mov_b32_e32 v0, 2
; GFX11-SDAG-NEXT: s_add_i32 s0, s32, 0x7ff
; GFX11-SDAG-NEXT: s_and_b32 s1, s1, -16
; GFX11-SDAG-NEXT: s_and_b32 s0, s0, 0xfffff800
; GFX11-SDAG-NEXT: s_lshl_b32 s1, s1, 5
; GFX11-SDAG-NEXT: scratch_store_b32 off, v0, s0 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: s_add_i32 s32, s0, s1
; GFX11-SDAG-NEXT: .LBB7_3: ; %bb.2
; GFX11-SDAG-NEXT: s_endpgm
; GFX11-SDAG-NEXT: .LBB7_4:
; GFX11-SDAG-NEXT: s_branch .LBB7_2
;
; GFX11-GISEL-LABEL: test_dynamic_stackalloc_kernel_control_flow:
; GFX11-GISEL: ; %bb.0: ; %entry
; GFX11-GISEL-NEXT: s_load_b64 s[0:1], s[4:5], 0x0
; GFX11-GISEL-NEXT: s_mov_b32 s33, 0
; GFX11-GISEL-NEXT: s_mov_b32 s32, 64
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: s_cmp_lg_u32 s0, 0
; GFX11-GISEL-NEXT: s_mov_b32 s0, 1
; GFX11-GISEL-NEXT: s_cbranch_scc0 .LBB7_2
; GFX11-GISEL-NEXT: ; %bb.1: ; %bb.1
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, 0x3ff, v0
; GFX11-GISEL-NEXT: s_mov_b32 s2, s32
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-GISEL-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s0
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_readlane_b32 s3, v1, 31
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: v_mov_b32_e32 v0, 1
; GFX11-GISEL-NEXT: s_lshl_b32 s3, s3, 5
; GFX11-GISEL-NEXT: s_mov_b32 s0, 0
; GFX11-GISEL-NEXT: s_add_u32 s32, s2, s3
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s2 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: .LBB7_2: ; %Flow
; GFX11-GISEL-NEXT: s_xor_b32 s0, s0, 1
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: s_cmp_lg_u32 s0, 0
; GFX11-GISEL-NEXT: s_cbranch_scc1 .LBB7_4
; GFX11-GISEL-NEXT: ; %bb.3: ; %bb.0
; GFX11-GISEL-NEXT: s_lshl2_add_u32 s0, s1, 15
; GFX11-GISEL-NEXT: v_mov_b32_e32 v0, 2
; GFX11-GISEL-NEXT: s_add_u32 s1, s32, 0x7ff
; GFX11-GISEL-NEXT: s_and_b32 s0, s0, -16
; GFX11-GISEL-NEXT: s_and_b32 s1, s1, 0xfffff800
; GFX11-GISEL-NEXT: s_lshl_b32 s0, s0, 5
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s1 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_add_u32 s32, s1, s0
; GFX11-GISEL-NEXT: .LBB7_4: ; %bb.2
; GFX11-GISEL-NEXT: s_endpgm
entry:
%cond = icmp eq i32 %n, 0
br i1 %cond, label %bb.0, label %bb.1
bb.0:
%alloca2 = alloca i32, i32 %m, align 64, addrspace(5)
store volatile i32 2, ptr addrspace(5) %alloca2
br label %bb.2
bb.1:
%idx = call i32 @llvm.amdgcn.workitem.id.x()
%alloca1 = alloca i32, i32 %idx, align 4, addrspace(5)
store volatile i32 1, ptr addrspace(5) %alloca1
br label %bb.2
bb.2:
ret void
}
define void @test_dynamic_stackalloc_device_uniform(i32 %n) {
; GFX9-SDAG-LABEL: test_dynamic_stackalloc_device_uniform:
; GFX9-SDAG: ; %bb.0:
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s7, s33
; GFX9-SDAG-NEXT: s_mov_b32 s33, s32
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_store_dword v1, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-SDAG-NEXT: s_addk_i32 s32, 0x400
; GFX9-SDAG-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-SDAG-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[4:5]
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: v_readlane_b32 s6, v1, 63
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s4, s32
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, s4
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, s6, 6, v0
; GFX9-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], s4
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s32, s33
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_load_dword v1, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s33, s7
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-GISEL-LABEL: test_dynamic_stackalloc_device_uniform:
; GFX9-GISEL: ; %bb.0:
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s8, s33
; GFX9-GISEL-NEXT: s_mov_b32 s33, s32
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_store_dword v1, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_addk_i32 s32, 0x400
; GFX9-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-GISEL-NEXT: s_mov_b32 s6, s32
; GFX9-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-GISEL-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[4:5]
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: v_readlane_b32 s7, v1, 63
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_lshl_b32 s4, s7, 6
; GFX9-GISEL-NEXT: s_add_u32 s32, s6, s4
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX9-GISEL-NEXT: v_mov_b32_e32 v2, s6
; GFX9-GISEL-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s32, s33
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_load_dword v1, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_mov_b32 s33, s8
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_device_uniform:
; GFX11-SDAG: ; %bb.0:
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-SDAG-NEXT: s_mov_b32 s2, s33
; GFX11-SDAG-NEXT: s_mov_b32 s33, s32
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x1 ; 8-byte Folded Spill
; GFX11-SDAG-NEXT: scratch_store_b32 off, v1, s33
; GFX11-SDAG-NEXT: scratch_store_b32 off, v2, s33 offset:4
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-SDAG-NEXT: s_add_i32 s32, s32, 16
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-SDAG-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s0
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_readlane_b32 s1, v1, 31
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s0, s32
; GFX11-SDAG-NEXT: v_mov_b32_e32 v3, 0x7b
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, s1, 5, s0
; GFX11-SDAG-NEXT: scratch_store_b32 off, v3, s0 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX11-SDAG-NEXT: s_mov_b32 s32, s33
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x1 ; 8-byte Folded Reload
; GFX11-SDAG-NEXT: scratch_load_b32 v1, off, s33
; GFX11-SDAG-NEXT: scratch_load_b32 v2, off, s33 offset:4
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s33, s2
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX11-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-GISEL-LABEL: test_dynamic_stackalloc_device_uniform:
; GFX11-GISEL: ; %bb.0:
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-GISEL-NEXT: s_mov_b32 s3, s33
; GFX11-GISEL-NEXT: s_mov_b32 s33, s32
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x1 ; 8-byte Folded Spill
; GFX11-GISEL-NEXT: scratch_store_b32 off, v1, s33
; GFX11-GISEL-NEXT: scratch_store_b32 off, v2, s33 offset:4
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-GISEL-NEXT: s_add_i32 s32, s32, 16
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: s_mov_b32 s0, s32
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-GISEL-NEXT: s_or_saveexec_b32 s1, -1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_readlane_b32 s2, v1, 31
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s1
; GFX11-GISEL-NEXT: v_mov_b32_e32 v0, 0x7b
; GFX11-GISEL-NEXT: s_lshl_b32 s1, s2, 5
; GFX11-GISEL-NEXT: s_add_u32 s32, s0, s1
; GFX11-GISEL-NEXT: s_mov_b32 s32, s33
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s0 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x1 ; 8-byte Folded Reload
; GFX11-GISEL-NEXT: scratch_load_b32 v1, off, s33
; GFX11-GISEL-NEXT: scratch_load_b32 v2, off, s33 offset:4
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: s_mov_b32 s33, s3
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31]
%alloca = alloca i32, i32 %n, addrspace(5)
store volatile i32 123, ptr addrspace(5) %alloca
ret void
}
define void @test_dynamic_stackalloc_device_uniform_over_aligned(i32 %n) {
; GFX9-SDAG-LABEL: test_dynamic_stackalloc_device_uniform_over_aligned:
; GFX9-SDAG: ; %bb.0:
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s8, s33
; GFX9-SDAG-NEXT: s_add_i32 s33, s32, 0x1fc0
; GFX9-SDAG-NEXT: s_and_b32 s33, s33, 0xffffe000
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_store_dword v1, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s9, s34
; GFX9-SDAG-NEXT: s_mov_b32 s34, s32
; GFX9-SDAG-NEXT: s_addk_i32 s32, 0x4000
; GFX9-SDAG-NEXT: s_add_i32 s4, s32, 0x1fff
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-SDAG-NEXT: s_and_b32 s6, s4, 0xffffe000
; GFX9-SDAG-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-SDAG-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[4:5]
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: v_readlane_b32 s7, v1, 63
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, s6
; GFX9-SDAG-NEXT: v_lshl_add_u32 v2, s7, 6, v0
; GFX9-SDAG-NEXT: v_readfirstlane_b32 s32, v2
; GFX9-SDAG-NEXT: v_mov_b32_e32 v2, 10
; GFX9-SDAG-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s32, s34
; GFX9-SDAG-NEXT: s_mov_b32 s34, s9
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_load_dword v1, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s33, s8
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-GISEL-LABEL: test_dynamic_stackalloc_device_uniform_over_aligned:
; GFX9-GISEL: ; %bb.0:
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s7, s33
; GFX9-GISEL-NEXT: s_add_i32 s33, s32, 0x1fc0
; GFX9-GISEL-NEXT: s_and_b32 s33, s33, 0xffffe000
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_store_dword v1, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-GISEL-NEXT: s_mov_b32 s8, s34
; GFX9-GISEL-NEXT: s_mov_b32 s34, s32
; GFX9-GISEL-NEXT: s_addk_i32 s32, 0x4000
; GFX9-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-GISEL-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[4:5]
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: v_readlane_b32 s6, v1, 63
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_add_u32 s5, s32, 0x1fff
; GFX9-GISEL-NEXT: s_lshl_b32 s4, s6, 6
; GFX9-GISEL-NEXT: s_and_b32 s5, s5, 0xffffe000
; GFX9-GISEL-NEXT: s_add_u32 s32, s5, s4
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 10
; GFX9-GISEL-NEXT: v_mov_b32_e32 v2, s5
; GFX9-GISEL-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s32, s34
; GFX9-GISEL-NEXT: s_mov_b32 s34, s8
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_load_dword v1, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_mov_b32 s33, s7
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_device_uniform_over_aligned:
; GFX11-SDAG: ; %bb.0:
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-SDAG-NEXT: s_mov_b32 s2, s33
; GFX11-SDAG-NEXT: s_add_i32 s33, s32, 0x7f
; GFX11-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: s_and_b32 s33, s33, 0xffffff80
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x1 ; 8-byte Folded Spill
; GFX11-SDAG-NEXT: scratch_store_b32 off, v1, s33
; GFX11-SDAG-NEXT: scratch_store_b32 off, v2, s33 offset:4
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-SDAG-NEXT: s_mov_b32 s3, s34
; GFX11-SDAG-NEXT: s_mov_b32 s34, s32
; GFX11-SDAG-NEXT: s_addk_i32 s32, 0x100
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-SDAG-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s0
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_readlane_b32 s1, v1, 31
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_add_i32 s0, s32, 0xfff
; GFX11-SDAG-NEXT: v_mov_b32_e32 v3, 10
; GFX11-SDAG-NEXT: s_and_b32 s0, s0, 0xfffff000
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, s1, 5, s0
; GFX11-SDAG-NEXT: scratch_store_b32 off, v3, s0 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX11-SDAG-NEXT: s_mov_b32 s32, s34
; GFX11-SDAG-NEXT: s_mov_b32 s34, s3
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x1 ; 8-byte Folded Reload
; GFX11-SDAG-NEXT: scratch_load_b32 v1, off, s33
; GFX11-SDAG-NEXT: scratch_load_b32 v2, off, s33 offset:4
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s33, s2
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX11-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-GISEL-LABEL: test_dynamic_stackalloc_device_uniform_over_aligned:
; GFX11-GISEL: ; %bb.0:
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-GISEL-NEXT: s_mov_b32 s2, s33
; GFX11-GISEL-NEXT: s_add_i32 s33, s32, 0x7f
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: s_and_b32 s33, s33, 0xffffff80
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x1 ; 8-byte Folded Spill
; GFX11-GISEL-NEXT: scratch_store_b32 off, v1, s33
; GFX11-GISEL-NEXT: scratch_store_b32 off, v2, s33 offset:4
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-GISEL-NEXT: s_mov_b32 s3, s34
; GFX11-GISEL-NEXT: s_mov_b32 s34, s32
; GFX11-GISEL-NEXT: s_addk_i32 s32, 0x100
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-GISEL-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s0
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_readlane_b32 s1, v1, 31
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: s_add_u32 s0, s32, 0xfff
; GFX11-GISEL-NEXT: v_mov_b32_e32 v0, 10
; GFX11-GISEL-NEXT: s_and_b32 s0, s0, 0xfffff000
; GFX11-GISEL-NEXT: s_lshl_b32 s1, s1, 5
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: s_add_u32 s32, s0, s1
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s0 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_mov_b32 s32, s34
; GFX11-GISEL-NEXT: s_mov_b32 s34, s3
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x1 ; 8-byte Folded Reload
; GFX11-GISEL-NEXT: scratch_load_b32 v1, off, s33
; GFX11-GISEL-NEXT: scratch_load_b32 v2, off, s33 offset:4
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: s_mov_b32 s33, s2
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31]
%alloca = alloca i32, i32 %n, align 128, addrspace(5)
store volatile i32 10, ptr addrspace(5) %alloca
ret void
}
define void @test_dynamic_stackalloc_device_uniform_under_aligned(i32 %n) {
; GFX9-SDAG-LABEL: test_dynamic_stackalloc_device_uniform_under_aligned:
; GFX9-SDAG: ; %bb.0:
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s7, s33
; GFX9-SDAG-NEXT: s_mov_b32 s33, s32
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_store_dword v1, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-SDAG-NEXT: s_addk_i32 s32, 0x400
; GFX9-SDAG-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-SDAG-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[4:5]
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: v_readlane_b32 s6, v1, 63
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s4, s32
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, s4
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, s6, 6, v0
; GFX9-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 22
; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], s4
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s32, s33
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_load_dword v1, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s33, s7
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-GISEL-LABEL: test_dynamic_stackalloc_device_uniform_under_aligned:
; GFX9-GISEL: ; %bb.0:
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s8, s33
; GFX9-GISEL-NEXT: s_mov_b32 s33, s32
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_store_dword v1, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_addk_i32 s32, 0x400
; GFX9-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-GISEL-NEXT: s_mov_b32 s6, s32
; GFX9-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-GISEL-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[4:5]
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: v_readlane_b32 s7, v1, 63
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_lshl_b32 s4, s7, 6
; GFX9-GISEL-NEXT: s_add_u32 s32, s6, s4
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 22
; GFX9-GISEL-NEXT: v_mov_b32_e32 v2, s6
; GFX9-GISEL-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s32, s33
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_load_dword v1, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_mov_b32 s33, s8
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_device_uniform_under_aligned:
; GFX11-SDAG: ; %bb.0:
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-SDAG-NEXT: s_mov_b32 s2, s33
; GFX11-SDAG-NEXT: s_mov_b32 s33, s32
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x1 ; 8-byte Folded Spill
; GFX11-SDAG-NEXT: scratch_store_b32 off, v1, s33
; GFX11-SDAG-NEXT: scratch_store_b32 off, v2, s33 offset:4
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-SDAG-NEXT: s_add_i32 s32, s32, 16
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-SDAG-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s0
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_readlane_b32 s1, v1, 31
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s0, s32
; GFX11-SDAG-NEXT: v_mov_b32_e32 v3, 22
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, s1, 5, s0
; GFX11-SDAG-NEXT: scratch_store_b32 off, v3, s0 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX11-SDAG-NEXT: s_mov_b32 s32, s33
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x1 ; 8-byte Folded Reload
; GFX11-SDAG-NEXT: scratch_load_b32 v1, off, s33
; GFX11-SDAG-NEXT: scratch_load_b32 v2, off, s33 offset:4
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s33, s2
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX11-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-GISEL-LABEL: test_dynamic_stackalloc_device_uniform_under_aligned:
; GFX11-GISEL: ; %bb.0:
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-GISEL-NEXT: s_mov_b32 s3, s33
; GFX11-GISEL-NEXT: s_mov_b32 s33, s32
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x1 ; 8-byte Folded Spill
; GFX11-GISEL-NEXT: scratch_store_b32 off, v1, s33
; GFX11-GISEL-NEXT: scratch_store_b32 off, v2, s33 offset:4
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-GISEL-NEXT: s_add_i32 s32, s32, 16
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: s_mov_b32 s0, s32
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-GISEL-NEXT: s_or_saveexec_b32 s1, -1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_readlane_b32 s2, v1, 31
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s1
; GFX11-GISEL-NEXT: v_mov_b32_e32 v0, 22
; GFX11-GISEL-NEXT: s_lshl_b32 s1, s2, 5
; GFX11-GISEL-NEXT: s_add_u32 s32, s0, s1
; GFX11-GISEL-NEXT: s_mov_b32 s32, s33
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s0 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x1 ; 8-byte Folded Reload
; GFX11-GISEL-NEXT: scratch_load_b32 v1, off, s33
; GFX11-GISEL-NEXT: scratch_load_b32 v2, off, s33 offset:4
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: s_mov_b32 s33, s3
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31]
%alloca = alloca i32, i32 %n, align 2, addrspace(5)
store volatile i32 22, ptr addrspace(5) %alloca
ret void
}
define void @test_dynamic_stackalloc_device_divergent() {
; GFX9-SDAG-LABEL: test_dynamic_stackalloc_device_divergent:
; GFX9-SDAG: ; %bb.0:
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s7, s33
; GFX9-SDAG-NEXT: s_mov_b32 s33, s32
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: v_and_b32_e32 v1, 0x3ff, v31
; GFX9-SDAG-NEXT: v_lshl_add_u32 v1, v1, 2, 15
; GFX9-SDAG-NEXT: s_addk_i32 s32, 0x400
; GFX9-SDAG-NEXT: v_and_b32_e32 v1, 0x1ff0, v1
; GFX9-SDAG-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, v1, s[4:5]
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: v_readlane_b32 s6, v0, 63
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s4, s32
; GFX9-SDAG-NEXT: v_mov_b32_e32 v1, s4
; GFX9-SDAG-NEXT: v_lshl_add_u32 v1, s6, 6, v1
; GFX9-SDAG-NEXT: v_readfirstlane_b32 s32, v1
; GFX9-SDAG-NEXT: v_mov_b32_e32 v1, 0x7b
; GFX9-SDAG-NEXT: buffer_store_dword v1, off, s[0:3], s4
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s32, s33
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_load_dword v0, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s33, s7
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-GISEL-LABEL: test_dynamic_stackalloc_device_divergent:
; GFX9-GISEL: ; %bb.0:
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s8, s33
; GFX9-GISEL-NEXT: s_mov_b32 s33, s32
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_store_dword v0, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: v_and_b32_e32 v1, 0x3ff, v31
; GFX9-GISEL-NEXT: s_addk_i32 s32, 0x400
; GFX9-GISEL-NEXT: v_lshl_add_u32 v1, v1, 2, 15
; GFX9-GISEL-NEXT: s_mov_b32 s6, s32
; GFX9-GISEL-NEXT: v_and_b32_e32 v1, -16, v1
; GFX9-GISEL-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v1, s[4:5]
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: v_readlane_b32 s7, v0, 63
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_lshl_b32 s4, s7, 6
; GFX9-GISEL-NEXT: s_add_u32 s32, s6, s4
; GFX9-GISEL-NEXT: v_mov_b32_e32 v1, 0x7b
; GFX9-GISEL-NEXT: v_mov_b32_e32 v2, s6
; GFX9-GISEL-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s32, s33
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_load_dword v0, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_mov_b32 s33, s8
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_device_divergent:
; GFX11-SDAG: ; %bb.0:
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-SDAG-NEXT: s_mov_b32 s2, s33
; GFX11-SDAG-NEXT: s_mov_b32 s33, s32
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x1 ; 8-byte Folded Spill
; GFX11-SDAG-NEXT: scratch_store_b32 off, v0, s33
; GFX11-SDAG-NEXT: scratch_store_b32 off, v1, s33 offset:4
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: v_and_b32_e32 v2, 0x3ff, v31
; GFX11-SDAG-NEXT: s_add_i32 s32, s32, 16
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_lshl_add_u32 v2, v2, 2, 15
; GFX11-SDAG-NEXT: v_and_b32_e32 v2, 0x1ff0, v2
; GFX11-SDAG-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, v2, s0
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: ds_swizzle_b32 v1, v0 offset:swizzle(BROADCAST,32,15)
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: v_max_u32_e32 v0, v0, v1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_readlane_b32 s1, v0, 31
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s0, s32
; GFX11-SDAG-NEXT: v_mov_b32_e32 v3, 0x7b
; GFX11-SDAG-NEXT: v_lshl_add_u32 v2, s1, 5, s0
; GFX11-SDAG-NEXT: scratch_store_b32 off, v3, s0 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: v_readfirstlane_b32 s32, v2
; GFX11-SDAG-NEXT: s_mov_b32 s32, s33
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x1 ; 8-byte Folded Reload
; GFX11-SDAG-NEXT: scratch_load_b32 v0, off, s33
; GFX11-SDAG-NEXT: scratch_load_b32 v1, off, s33 offset:4
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s33, s2
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX11-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-GISEL-LABEL: test_dynamic_stackalloc_device_divergent:
; GFX11-GISEL: ; %bb.0:
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-GISEL-NEXT: s_mov_b32 s3, s33
; GFX11-GISEL-NEXT: s_mov_b32 s33, s32
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x1 ; 8-byte Folded Spill
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s33
; GFX11-GISEL-NEXT: scratch_store_b32 off, v1, s33 offset:4
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: v_and_b32_e32 v2, 0x3ff, v31
; GFX11-GISEL-NEXT: s_add_i32 s32, s32, 16
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: s_mov_b32 s0, s32
; GFX11-GISEL-NEXT: v_lshl_add_u32 v2, v2, 2, 15
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_and_b32_e32 v2, -16, v2
; GFX11-GISEL-NEXT: s_or_saveexec_b32 s1, -1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: ds_swizzle_b32 v1, v0 offset:swizzle(BROADCAST,32,15)
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: v_max_u32_e32 v0, v0, v1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_readlane_b32 s2, v0, 31
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s1
; GFX11-GISEL-NEXT: v_mov_b32_e32 v2, 0x7b
; GFX11-GISEL-NEXT: s_lshl_b32 s1, s2, 5
; GFX11-GISEL-NEXT: s_add_u32 s32, s0, s1
; GFX11-GISEL-NEXT: s_mov_b32 s32, s33
; GFX11-GISEL-NEXT: scratch_store_b32 off, v2, s0 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x1 ; 8-byte Folded Reload
; GFX11-GISEL-NEXT: scratch_load_b32 v0, off, s33
; GFX11-GISEL-NEXT: scratch_load_b32 v1, off, s33 offset:4
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: s_mov_b32 s33, s3
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31]
%idx = call i32 @llvm.amdgcn.workitem.id.x()
%alloca = alloca i32, i32 %idx, addrspace(5)
store volatile i32 123, ptr addrspace(5) %alloca
ret void
}
define void @test_dynamic_stackalloc_device_divergent_over_aligned() {
; GFX9-SDAG-LABEL: test_dynamic_stackalloc_device_divergent_over_aligned:
; GFX9-SDAG: ; %bb.0:
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s8, s33
; GFX9-SDAG-NEXT: s_add_i32 s33, s32, 0x1fc0
; GFX9-SDAG-NEXT: s_and_b32 s33, s33, 0xffffe000
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s9, s34
; GFX9-SDAG-NEXT: s_mov_b32 s34, s32
; GFX9-SDAG-NEXT: s_addk_i32 s32, 0x4000
; GFX9-SDAG-NEXT: v_and_b32_e32 v1, 0x3ff, v31
; GFX9-SDAG-NEXT: s_add_i32 s4, s32, 0x1fff
; GFX9-SDAG-NEXT: v_lshl_add_u32 v1, v1, 2, 15
; GFX9-SDAG-NEXT: s_and_b32 s6, s4, 0xffffe000
; GFX9-SDAG-NEXT: v_and_b32_e32 v1, 0x1ff0, v1
; GFX9-SDAG-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, v1, s[4:5]
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: v_readlane_b32 s7, v0, 63
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: v_mov_b32_e32 v1, s6
; GFX9-SDAG-NEXT: v_lshl_add_u32 v2, s7, 6, v1
; GFX9-SDAG-NEXT: v_readfirstlane_b32 s32, v2
; GFX9-SDAG-NEXT: v_mov_b32_e32 v2, 0x1bc
; GFX9-SDAG-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s32, s34
; GFX9-SDAG-NEXT: s_mov_b32 s34, s9
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_load_dword v0, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s33, s8
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-GISEL-LABEL: test_dynamic_stackalloc_device_divergent_over_aligned:
; GFX9-GISEL: ; %bb.0:
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s7, s33
; GFX9-GISEL-NEXT: s_add_i32 s33, s32, 0x1fc0
; GFX9-GISEL-NEXT: s_and_b32 s33, s33, 0xffffe000
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_store_dword v0, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: v_and_b32_e32 v1, 0x3ff, v31
; GFX9-GISEL-NEXT: v_lshl_add_u32 v1, v1, 2, 15
; GFX9-GISEL-NEXT: s_mov_b32 s8, s34
; GFX9-GISEL-NEXT: s_mov_b32 s34, s32
; GFX9-GISEL-NEXT: s_addk_i32 s32, 0x4000
; GFX9-GISEL-NEXT: v_and_b32_e32 v1, -16, v1
; GFX9-GISEL-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v1, s[4:5]
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: v_readlane_b32 s6, v0, 63
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_add_u32 s5, s32, 0x1fff
; GFX9-GISEL-NEXT: s_lshl_b32 s4, s6, 6
; GFX9-GISEL-NEXT: s_and_b32 s5, s5, 0xffffe000
; GFX9-GISEL-NEXT: s_add_u32 s32, s5, s4
; GFX9-GISEL-NEXT: v_mov_b32_e32 v1, 0x1bc
; GFX9-GISEL-NEXT: v_mov_b32_e32 v2, s5
; GFX9-GISEL-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s32, s34
; GFX9-GISEL-NEXT: s_mov_b32 s34, s8
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_load_dword v0, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_mov_b32 s33, s7
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_device_divergent_over_aligned:
; GFX11-SDAG: ; %bb.0:
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-SDAG-NEXT: s_mov_b32 s2, s33
; GFX11-SDAG-NEXT: s_add_i32 s33, s32, 0x7f
; GFX11-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: s_and_b32 s33, s33, 0xffffff80
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x1 ; 8-byte Folded Spill
; GFX11-SDAG-NEXT: scratch_store_b32 off, v0, s33
; GFX11-SDAG-NEXT: scratch_store_b32 off, v1, s33 offset:4
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: v_and_b32_e32 v2, 0x3ff, v31
; GFX11-SDAG-NEXT: s_mov_b32 s3, s34
; GFX11-SDAG-NEXT: s_mov_b32 s34, s32
; GFX11-SDAG-NEXT: s_addk_i32 s32, 0x100
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_lshl_add_u32 v2, v2, 2, 15
; GFX11-SDAG-NEXT: v_and_b32_e32 v2, 0x1ff0, v2
; GFX11-SDAG-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, v2, s0
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: ds_swizzle_b32 v1, v0 offset:swizzle(BROADCAST,32,15)
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: v_max_u32_e32 v0, v0, v1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_readlane_b32 s1, v0, 31
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_add_i32 s0, s32, 0xfff
; GFX11-SDAG-NEXT: v_mov_b32_e32 v3, 0x1bc
; GFX11-SDAG-NEXT: s_and_b32 s0, s0, 0xfffff000
; GFX11-SDAG-NEXT: v_lshl_add_u32 v2, s1, 5, s0
; GFX11-SDAG-NEXT: scratch_store_b32 off, v3, s0 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: v_readfirstlane_b32 s32, v2
; GFX11-SDAG-NEXT: s_mov_b32 s32, s34
; GFX11-SDAG-NEXT: s_mov_b32 s34, s3
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x1 ; 8-byte Folded Reload
; GFX11-SDAG-NEXT: scratch_load_b32 v0, off, s33
; GFX11-SDAG-NEXT: scratch_load_b32 v1, off, s33 offset:4
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s33, s2
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX11-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-GISEL-LABEL: test_dynamic_stackalloc_device_divergent_over_aligned:
; GFX11-GISEL: ; %bb.0:
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-GISEL-NEXT: s_mov_b32 s2, s33
; GFX11-GISEL-NEXT: s_add_i32 s33, s32, 0x7f
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: s_and_b32 s33, s33, 0xffffff80
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x1 ; 8-byte Folded Spill
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s33
; GFX11-GISEL-NEXT: scratch_store_b32 off, v1, s33 offset:4
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: v_and_b32_e32 v2, 0x3ff, v31
; GFX11-GISEL-NEXT: s_mov_b32 s3, s34
; GFX11-GISEL-NEXT: s_mov_b32 s34, s32
; GFX11-GISEL-NEXT: s_addk_i32 s32, 0x100
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_lshl_add_u32 v2, v2, 2, 15
; GFX11-GISEL-NEXT: v_and_b32_e32 v2, -16, v2
; GFX11-GISEL-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s0
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: ds_swizzle_b32 v1, v0 offset:swizzle(BROADCAST,32,15)
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: v_max_u32_e32 v0, v0, v1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_readlane_b32 s1, v0, 31
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: s_add_u32 s0, s32, 0xfff
; GFX11-GISEL-NEXT: v_mov_b32_e32 v2, 0x1bc
; GFX11-GISEL-NEXT: s_and_b32 s0, s0, 0xfffff000
; GFX11-GISEL-NEXT: s_lshl_b32 s1, s1, 5
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: s_add_u32 s32, s0, s1
; GFX11-GISEL-NEXT: scratch_store_b32 off, v2, s0 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_mov_b32 s32, s34
; GFX11-GISEL-NEXT: s_mov_b32 s34, s3
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x1 ; 8-byte Folded Reload
; GFX11-GISEL-NEXT: scratch_load_b32 v0, off, s33
; GFX11-GISEL-NEXT: scratch_load_b32 v1, off, s33 offset:4
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: s_mov_b32 s33, s2
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31]
%idx = call i32 @llvm.amdgcn.workitem.id.x()
%alloca = alloca i32, i32 %idx, align 128, addrspace(5)
store volatile i32 444, ptr addrspace(5) %alloca
ret void
}
define void @test_dynamic_stackalloc_device_divergent_under_aligned() {
; GFX9-SDAG-LABEL: test_dynamic_stackalloc_device_divergent_under_aligned:
; GFX9-SDAG: ; %bb.0:
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s7, s33
; GFX9-SDAG-NEXT: s_mov_b32 s33, s32
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: v_and_b32_e32 v1, 0x3ff, v31
; GFX9-SDAG-NEXT: v_lshl_add_u32 v1, v1, 2, 15
; GFX9-SDAG-NEXT: s_addk_i32 s32, 0x400
; GFX9-SDAG-NEXT: v_and_b32_e32 v1, 0x1ff0, v1
; GFX9-SDAG-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, v1, s[4:5]
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: v_readlane_b32 s6, v0, 63
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s4, s32
; GFX9-SDAG-NEXT: v_mov_b32_e32 v1, s4
; GFX9-SDAG-NEXT: v_lshl_add_u32 v1, s6, 6, v1
; GFX9-SDAG-NEXT: v_readfirstlane_b32 s32, v1
; GFX9-SDAG-NEXT: v_mov_b32_e32 v1, 0x29a
; GFX9-SDAG-NEXT: buffer_store_dword v1, off, s[0:3], s4
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s32, s33
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_load_dword v0, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s33, s7
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-GISEL-LABEL: test_dynamic_stackalloc_device_divergent_under_aligned:
; GFX9-GISEL: ; %bb.0:
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s8, s33
; GFX9-GISEL-NEXT: s_mov_b32 s33, s32
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_store_dword v0, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: v_and_b32_e32 v1, 0x3ff, v31
; GFX9-GISEL-NEXT: s_addk_i32 s32, 0x400
; GFX9-GISEL-NEXT: v_lshl_add_u32 v1, v1, 2, 15
; GFX9-GISEL-NEXT: s_mov_b32 s6, s32
; GFX9-GISEL-NEXT: v_and_b32_e32 v1, -16, v1
; GFX9-GISEL-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v1, s[4:5]
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: v_readlane_b32 s7, v0, 63
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_lshl_b32 s4, s7, 6
; GFX9-GISEL-NEXT: s_add_u32 s32, s6, s4
; GFX9-GISEL-NEXT: v_mov_b32_e32 v1, 0x29a
; GFX9-GISEL-NEXT: v_mov_b32_e32 v2, s6
; GFX9-GISEL-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s32, s33
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_load_dword v0, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_mov_b32 s33, s8
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_device_divergent_under_aligned:
; GFX11-SDAG: ; %bb.0:
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-SDAG-NEXT: s_mov_b32 s2, s33
; GFX11-SDAG-NEXT: s_mov_b32 s33, s32
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x1 ; 8-byte Folded Spill
; GFX11-SDAG-NEXT: scratch_store_b32 off, v0, s33
; GFX11-SDAG-NEXT: scratch_store_b32 off, v1, s33 offset:4
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: v_and_b32_e32 v2, 0x3ff, v31
; GFX11-SDAG-NEXT: s_add_i32 s32, s32, 16
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_lshl_add_u32 v2, v2, 2, 15
; GFX11-SDAG-NEXT: v_and_b32_e32 v2, 0x1ff0, v2
; GFX11-SDAG-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v0, 0, v2, s0
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: ds_swizzle_b32 v1, v0 offset:swizzle(BROADCAST,32,15)
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: v_max_u32_e32 v0, v0, v1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_readlane_b32 s1, v0, 31
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s0, s32
; GFX11-SDAG-NEXT: v_mov_b32_e32 v3, 0x29a
; GFX11-SDAG-NEXT: v_lshl_add_u32 v2, s1, 5, s0
; GFX11-SDAG-NEXT: scratch_store_b32 off, v3, s0 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: v_readfirstlane_b32 s32, v2
; GFX11-SDAG-NEXT: s_mov_b32 s32, s33
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x1 ; 8-byte Folded Reload
; GFX11-SDAG-NEXT: scratch_load_b32 v0, off, s33
; GFX11-SDAG-NEXT: scratch_load_b32 v1, off, s33 offset:4
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s33, s2
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX11-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-GISEL-LABEL: test_dynamic_stackalloc_device_divergent_under_aligned:
; GFX11-GISEL: ; %bb.0:
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-GISEL-NEXT: s_mov_b32 s3, s33
; GFX11-GISEL-NEXT: s_mov_b32 s33, s32
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x1 ; 8-byte Folded Spill
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s33
; GFX11-GISEL-NEXT: scratch_store_b32 off, v1, s33 offset:4
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: v_and_b32_e32 v2, 0x3ff, v31
; GFX11-GISEL-NEXT: s_add_i32 s32, s32, 16
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: s_mov_b32 s0, s32
; GFX11-GISEL-NEXT: v_lshl_add_u32 v2, v2, 2, 15
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_and_b32_e32 v2, -16, v2
; GFX11-GISEL-NEXT: s_or_saveexec_b32 s1, -1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v0, 0, v2, s1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v0, v0, v0 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: ds_swizzle_b32 v1, v0 offset:swizzle(BROADCAST,32,15)
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: v_max_u32_e32 v0, v0, v1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_readlane_b32 s2, v0, 31
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s1
; GFX11-GISEL-NEXT: v_mov_b32_e32 v2, 0x29a
; GFX11-GISEL-NEXT: s_lshl_b32 s1, s2, 5
; GFX11-GISEL-NEXT: s_add_u32 s32, s0, s1
; GFX11-GISEL-NEXT: s_mov_b32 s32, s33
; GFX11-GISEL-NEXT: scratch_store_b32 off, v2, s0 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x1 ; 8-byte Folded Reload
; GFX11-GISEL-NEXT: scratch_load_b32 v0, off, s33
; GFX11-GISEL-NEXT: scratch_load_b32 v1, off, s33 offset:4
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: s_mov_b32 s33, s3
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31]
%idx = call i32 @llvm.amdgcn.workitem.id.x()
%alloca = alloca i32, i32 %idx, align 2, addrspace(5)
store volatile i32 666, ptr addrspace(5) %alloca
ret void
}
define void @test_dynamic_stackalloc_device_multiple_allocas(i32 %n, i32 %m) {
; GFX9-SDAG-LABEL: test_dynamic_stackalloc_device_multiple_allocas:
; GFX9-SDAG: ; %bb.0: ; %entry
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s10, s33
; GFX9-SDAG-NEXT: s_add_i32 s33, s32, 0xfc0
; GFX9-SDAG-NEXT: s_and_b32 s33, s33, 0xfffff000
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_store_dword v2, off, s[0:3], s33 offset:64 ; 4-byte Folded Spill
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s11, s34
; GFX9-SDAG-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX9-SDAG-NEXT: s_mov_b32 s34, s32
; GFX9-SDAG-NEXT: s_addk_i32 s32, 0x3000
; GFX9-SDAG-NEXT: s_and_saveexec_b64 s[4:5], vcc
; GFX9-SDAG-NEXT: s_cbranch_execz .LBB14_2
; GFX9-SDAG-NEXT: ; %bb.1: ; %bb.0
; GFX9-SDAG-NEXT: s_add_i32 s6, s32, 0xfff
; GFX9-SDAG-NEXT: v_lshl_add_u32 v1, v1, 2, 15
; GFX9-SDAG-NEXT: s_and_b32 s8, s6, 0xfffff000
; GFX9-SDAG-NEXT: v_and_b32_e32 v1, -16, v1
; GFX9-SDAG-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v2, 0, v1, s[6:7]
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: v_readlane_b32 s9, v2, 63
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-SDAG-NEXT: v_mov_b32_e32 v1, s8
; GFX9-SDAG-NEXT: v_lshl_add_u32 v3, s9, 6, v1
; GFX9-SDAG-NEXT: v_readfirstlane_b32 s32, v3
; GFX9-SDAG-NEXT: v_and_b32_e32 v3, 0x3ff, v31
; GFX9-SDAG-NEXT: v_lshl_add_u32 v3, v3, 2, 15
; GFX9-SDAG-NEXT: v_and_b32_e32 v3, 0x1ff0, v3
; GFX9-SDAG-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v2, 0, v3, s[6:7]
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: v_readlane_b32 s8, v2, 63
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-SDAG-NEXT: s_mov_b32 s6, s32
; GFX9-SDAG-NEXT: v_mov_b32_e32 v3, s6
; GFX9-SDAG-NEXT: v_lshl_add_u32 v3, s8, 6, v3
; GFX9-SDAG-NEXT: v_readfirstlane_b32 s32, v3
; GFX9-SDAG-NEXT: v_mov_b32_e32 v3, 3
; GFX9-SDAG-NEXT: buffer_store_dword v3, v1, s[0:3], 0 offen
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: v_mov_b32_e32 v1, 4
; GFX9-SDAG-NEXT: buffer_store_dword v1, off, s[0:3], s6
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: .LBB14_2: ; %bb.1
; GFX9-SDAG-NEXT: s_or_b64 exec, exec, s[4:5]
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-SDAG-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-SDAG-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v2, 0, v0, s[4:5]
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: v_readlane_b32 s6, v2, 63
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s4, s32
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, s4
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, s6, 6, v0
; GFX9-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 1
; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], s33
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 2
; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], s4
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s32, s34
; GFX9-SDAG-NEXT: s_mov_b32 s34, s11
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:64 ; 4-byte Folded Reload
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s33, s10
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-GISEL-LABEL: test_dynamic_stackalloc_device_multiple_allocas:
; GFX9-GISEL: ; %bb.0: ; %entry
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s11, s33
; GFX9-GISEL-NEXT: s_add_i32 s33, s32, 0xfc0
; GFX9-GISEL-NEXT: s_and_b32 s33, s33, 0xfffff000
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_store_dword v2, off, s[0:3], s33 offset:64 ; 4-byte Folded Spill
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_mov_b32 s12, s34
; GFX9-GISEL-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX9-GISEL-NEXT: s_mov_b32 s34, s32
; GFX9-GISEL-NEXT: s_addk_i32 s32, 0x3000
; GFX9-GISEL-NEXT: s_and_saveexec_b64 s[4:5], vcc
; GFX9-GISEL-NEXT: s_cbranch_execz .LBB14_2
; GFX9-GISEL-NEXT: ; %bb.1: ; %bb.0
; GFX9-GISEL-NEXT: v_lshl_add_u32 v1, v1, 2, 15
; GFX9-GISEL-NEXT: v_and_b32_e32 v3, 0x3ff, v31
; GFX9-GISEL-NEXT: v_and_b32_e32 v1, -16, v1
; GFX9-GISEL-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v1, s[6:7]
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: v_readlane_b32 s8, v2, 63
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-GISEL-NEXT: s_add_u32 s7, s32, 0xfff
; GFX9-GISEL-NEXT: s_lshl_b32 s6, s8, 6
; GFX9-GISEL-NEXT: s_and_b32 s8, s7, 0xfffff000
; GFX9-GISEL-NEXT: s_add_u32 s32, s8, s6
; GFX9-GISEL-NEXT: v_lshl_add_u32 v1, v3, 2, 15
; GFX9-GISEL-NEXT: s_mov_b32 s9, s32
; GFX9-GISEL-NEXT: v_and_b32_e32 v1, -16, v1
; GFX9-GISEL-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v1, s[6:7]
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: v_readlane_b32 s10, v2, 63
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-GISEL-NEXT: v_mov_b32_e32 v1, 3
; GFX9-GISEL-NEXT: v_mov_b32_e32 v3, s8
; GFX9-GISEL-NEXT: s_lshl_b32 s6, s10, 6
; GFX9-GISEL-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: v_mov_b32_e32 v1, 4
; GFX9-GISEL-NEXT: v_mov_b32_e32 v3, s9
; GFX9-GISEL-NEXT: s_add_u32 s32, s9, s6
; GFX9-GISEL-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: .LBB14_2: ; %bb.1
; GFX9-GISEL-NEXT: s_or_b64 exec, exec, s[4:5]
; GFX9-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-GISEL-NEXT: s_mov_b32 s6, s32
; GFX9-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-GISEL-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v0, s[4:5]
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: v_readlane_b32 s7, v2, 63
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_lshl_b32 s4, s7, 6
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 1
; GFX9-GISEL-NEXT: s_add_u32 s32, s6, s4
; GFX9-GISEL-NEXT: buffer_store_dword v0, off, s[0:3], s33
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 2
; GFX9-GISEL-NEXT: v_mov_b32_e32 v1, s6
; GFX9-GISEL-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s32, s34
; GFX9-GISEL-NEXT: s_mov_b32 s34, s12
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_load_dword v2, off, s[0:3], s33 offset:64 ; 4-byte Folded Reload
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_mov_b32 s33, s11
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_device_multiple_allocas:
; GFX11-SDAG: ; %bb.0: ; %entry
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-SDAG-NEXT: s_mov_b32 s5, s33
; GFX11-SDAG-NEXT: s_add_i32 s33, s32, 63
; GFX11-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: s_and_not1_b32 s33, s33, 63
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x3 ; 16-byte Folded Spill
; GFX11-SDAG-NEXT: scratch_store_b32 off, v2, s33 offset:64
; GFX11-SDAG-NEXT: scratch_store_b32 off, v3, s33 offset:68
; GFX11-SDAG-NEXT: scratch_store_b32 off, v4, s33 offset:72
; GFX11-SDAG-NEXT: scratch_store_b32 off, v5, s33 offset:76
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s6, s34
; GFX11-SDAG-NEXT: s_mov_b32 s0, exec_lo
; GFX11-SDAG-NEXT: s_mov_b32 s34, s32
; GFX11-SDAG-NEXT: s_addk_i32 s32, 0xc0
; GFX11-SDAG-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX11-SDAG-NEXT: s_cbranch_execz .LBB14_2
; GFX11-SDAG-NEXT: ; %bb.1: ; %bb.0
; GFX11-SDAG-NEXT: v_lshl_add_u32 v1, v1, 2, 15
; GFX11-SDAG-NEXT: v_and_b32_e32 v6, 0x3ff, v31
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-SDAG-NEXT: v_and_b32_e32 v1, -16, v1
; GFX11-SDAG-NEXT: v_lshl_add_u32 v6, v6, 2, 15
; GFX11-SDAG-NEXT: s_or_saveexec_b32 s1, -1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v2, 0, v1, s1
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-SDAG-NEXT: v_and_b32_e32 v1, 0x1ff0, v6
; GFX11-SDAG-NEXT: s_or_saveexec_b32 s1, -1
; GFX11-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v3, 0, v1, s1
; GFX11-SDAG-NEXT: s_add_i32 s3, s32, 0x7ff
; GFX11-SDAG-NEXT: s_and_b32 s3, s3, 0xfffff800
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v3, v3, v3 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v3, v3, v3 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v3, v3, v3 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: ds_swizzle_b32 v4, v2 offset:swizzle(BROADCAST,32,15)
; GFX11-SDAG-NEXT: v_max_u32_dpp v3, v3, v3 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: ds_swizzle_b32 v5, v3 offset:swizzle(BROADCAST,32,15)
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(1)
; GFX11-SDAG-NEXT: v_max_u32_e32 v2, v2, v4
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_readlane_b32 s2, v2, 31
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: v_max_u32_e32 v2, v3, v5
; GFX11-SDAG-NEXT: v_readlane_b32 s4, v2, 31
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s1
; GFX11-SDAG-NEXT: v_lshl_add_u32 v1, s2, 5, s3
; GFX11-SDAG-NEXT: v_dual_mov_b32 v6, 3 :: v_dual_mov_b32 v7, 4
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-SDAG-NEXT: v_readfirstlane_b32 s32, v1
; GFX11-SDAG-NEXT: s_mov_b32 s1, s32
; GFX11-SDAG-NEXT: scratch_store_b32 off, v6, s3 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: scratch_store_b32 off, v7, s1 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: v_lshl_add_u32 v1, s4, 5, s1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_readfirstlane_b32 s32, v1
; GFX11-SDAG-NEXT: .LBB14_2: ; %bb.1
; GFX11-SDAG-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-SDAG-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v2, 0, v0, s0
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: ds_swizzle_b32 v3, v2 offset:swizzle(BROADCAST,32,15)
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: v_max_u32_e32 v2, v2, v3
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_readlane_b32 s1, v2, 31
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s0, s32
; GFX11-SDAG-NEXT: v_dual_mov_b32 v1, 1 :: v_dual_mov_b32 v6, 2
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, s1, 5, s0
; GFX11-SDAG-NEXT: scratch_store_b32 off, v1, s33 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: scratch_store_b32 off, v6, s0 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX11-SDAG-NEXT: s_mov_b32 s32, s34
; GFX11-SDAG-NEXT: s_mov_b32 s34, s6
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x3 ; 16-byte Folded Reload
; GFX11-SDAG-NEXT: scratch_load_b32 v2, off, s33 offset:64
; GFX11-SDAG-NEXT: scratch_load_b32 v3, off, s33 offset:68
; GFX11-SDAG-NEXT: scratch_load_b32 v4, off, s33 offset:72
; GFX11-SDAG-NEXT: scratch_load_b32 v5, off, s33 offset:76
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s33, s5
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX11-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-GISEL-LABEL: test_dynamic_stackalloc_device_multiple_allocas:
; GFX11-GISEL: ; %bb.0: ; %entry
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-GISEL-NEXT: s_mov_b32 s4, s33
; GFX11-GISEL-NEXT: s_add_i32 s33, s32, 63
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: s_and_not1_b32 s33, s33, 63
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x3 ; 16-byte Folded Spill
; GFX11-GISEL-NEXT: scratch_store_b32 off, v2, s33 offset:64
; GFX11-GISEL-NEXT: scratch_store_b32 off, v3, s33 offset:68
; GFX11-GISEL-NEXT: scratch_store_b32 off, v4, s33 offset:72
; GFX11-GISEL-NEXT: scratch_store_b32 off, v5, s33 offset:76
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: s_mov_b32 s5, s34
; GFX11-GISEL-NEXT: s_mov_b32 s0, exec_lo
; GFX11-GISEL-NEXT: s_mov_b32 s34, s32
; GFX11-GISEL-NEXT: s_addk_i32 s32, 0xc0
; GFX11-GISEL-NEXT: v_cmpx_eq_u32_e32 0, v0
; GFX11-GISEL-NEXT: s_cbranch_execz .LBB14_2
; GFX11-GISEL-NEXT: ; %bb.1: ; %bb.0
; GFX11-GISEL-NEXT: v_and_b32_e32 v6, 0x3ff, v31
; GFX11-GISEL-NEXT: v_lshl_add_u32 v1, v1, 2, 15
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-GISEL-NEXT: v_lshl_add_u32 v6, v6, 2, 15
; GFX11-GISEL-NEXT: v_and_b32_e32 v1, -16, v1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-GISEL-NEXT: v_and_b32_e32 v6, -16, v6
; GFX11-GISEL-NEXT: s_or_saveexec_b32 s1, -1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_2) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v1, s1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v3, 0, v6, s1
; GFX11-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-GISEL-NEXT: v_max_u32_dpp v3, v3, v3 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-GISEL-NEXT: v_max_u32_dpp v3, v3, v3 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-GISEL-NEXT: v_max_u32_dpp v3, v3, v3 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-GISEL-NEXT: v_max_u32_dpp v3, v3, v3 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: ds_swizzle_b32 v4, v2 offset:swizzle(BROADCAST,32,15)
; GFX11-GISEL-NEXT: ds_swizzle_b32 v5, v3 offset:swizzle(BROADCAST,32,15)
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(1)
; GFX11-GISEL-NEXT: v_max_u32_e32 v2, v2, v4
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: v_max_u32_e32 v3, v3, v5
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-GISEL-NEXT: v_readlane_b32 s2, v2, 31
; GFX11-GISEL-NEXT: v_readlane_b32 s3, v3, 31
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s1
; GFX11-GISEL-NEXT: s_add_u32 s1, s32, 0x7ff
; GFX11-GISEL-NEXT: v_dual_mov_b32 v1, 3 :: v_dual_mov_b32 v6, 4
; GFX11-GISEL-NEXT: s_and_b32 s1, s1, 0xfffff800
; GFX11-GISEL-NEXT: s_lshl_b32 s2, s2, 5
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: s_add_u32 s32, s1, s2
; GFX11-GISEL-NEXT: scratch_store_b32 off, v1, s1 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_mov_b32 s1, s32
; GFX11-GISEL-NEXT: s_lshl_b32 s2, s3, 5
; GFX11-GISEL-NEXT: scratch_store_b32 off, v6, s1 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_add_u32 s32, s1, s2
; GFX11-GISEL-NEXT: .LBB14_2: ; %bb.1
; GFX11-GISEL-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-GISEL-NEXT: s_mov_b32 s0, s32
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-GISEL-NEXT: s_or_saveexec_b32 s1, -1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v0, s1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: ds_swizzle_b32 v3, v2 offset:swizzle(BROADCAST,32,15)
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: v_max_u32_e32 v2, v2, v3
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_readlane_b32 s2, v2, 31
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s1
; GFX11-GISEL-NEXT: v_dual_mov_b32 v0, 1 :: v_dual_mov_b32 v1, 2
; GFX11-GISEL-NEXT: s_lshl_b32 s1, s2, 5
; GFX11-GISEL-NEXT: s_add_u32 s32, s0, s1
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s33 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: scratch_store_b32 off, v1, s0 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_mov_b32 s32, s34
; GFX11-GISEL-NEXT: s_mov_b32 s34, s5
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x3 ; 16-byte Folded Reload
; GFX11-GISEL-NEXT: scratch_load_b32 v2, off, s33 offset:64
; GFX11-GISEL-NEXT: scratch_load_b32 v3, off, s33 offset:68
; GFX11-GISEL-NEXT: scratch_load_b32 v4, off, s33 offset:72
; GFX11-GISEL-NEXT: scratch_load_b32 v5, off, s33 offset:76
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: s_mov_b32 s33, s4
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31]
entry:
%cond = icmp eq i32 %n, 0
%alloca1 = alloca i32, i32 8, addrspace(5)
%alloca2 = alloca i32, i32 %n, addrspace(5)
br i1 %cond, label %bb.0, label %bb.1
bb.0:
%idx = call i32 @llvm.amdgcn.workitem.id.x()
%alloca3 = alloca i32, i32 %m, align 64, addrspace(5)
%alloca4 = alloca i32, i32 %idx, align 4, addrspace(5)
store volatile i32 3, ptr addrspace(5) %alloca3
store volatile i32 4, ptr addrspace(5) %alloca4
br label %bb.1
bb.1:
store volatile i32 1, ptr addrspace(5) %alloca1
store volatile i32 2, ptr addrspace(5) %alloca2
ret void
}
define void @test_dynamic_stackalloc_device_control_flow(i32 %n, i32 %m) {
; GFX9-SDAG-LABEL: test_dynamic_stackalloc_device_control_flow:
; GFX9-SDAG: ; %bb.0: ; %entry
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s10, s33
; GFX9-SDAG-NEXT: s_add_i32 s33, s32, 0xfc0
; GFX9-SDAG-NEXT: s_and_b32 s33, s33, 0xfffff000
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s11, s34
; GFX9-SDAG-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
; GFX9-SDAG-NEXT: s_mov_b32 s34, s32
; GFX9-SDAG-NEXT: s_addk_i32 s32, 0x2000
; GFX9-SDAG-NEXT: s_and_saveexec_b64 s[4:5], vcc
; GFX9-SDAG-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-SDAG-NEXT: s_cbranch_execz .LBB15_2
; GFX9-SDAG-NEXT: ; %bb.1: ; %bb.1
; GFX9-SDAG-NEXT: s_add_i32 s6, s32, 0xfff
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, v1, 2, 15
; GFX9-SDAG-NEXT: s_and_b32 s8, s6, 0xfffff000
; GFX9-SDAG-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-SDAG-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v2, 0, v0, s[6:7]
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: v_readlane_b32 s9, v2, 63
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, s8
; GFX9-SDAG-NEXT: v_lshl_add_u32 v1, s9, 6, v0
; GFX9-SDAG-NEXT: v_readfirstlane_b32 s32, v1
; GFX9-SDAG-NEXT: v_mov_b32_e32 v1, 2
; GFX9-SDAG-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: ; implicit-def: $vgpr31
; GFX9-SDAG-NEXT: .LBB15_2: ; %Flow
; GFX9-SDAG-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-SDAG-NEXT: s_cbranch_execz .LBB15_4
; GFX9-SDAG-NEXT: ; %bb.3: ; %bb.0
; GFX9-SDAG-NEXT: v_and_b32_e32 v0, 0x3ff, v31
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-SDAG-NEXT: v_and_b32_e32 v0, 0x1ff0, v0
; GFX9-SDAG-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v2, 0, v0, s[6:7]
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: v_readlane_b32 s8, v2, 63
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-SDAG-NEXT: s_mov_b32 s6, s32
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, s6
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, s8, 6, v0
; GFX9-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 1
; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], s6
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: .LBB15_4: ; %bb.2
; GFX9-SDAG-NEXT: s_or_b64 exec, exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s32, s34
; GFX9-SDAG-NEXT: s_mov_b32 s34, s11
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s33, s10
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-GISEL-LABEL: test_dynamic_stackalloc_device_control_flow:
; GFX9-GISEL: ; %bb.0: ; %entry
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s10, s33
; GFX9-GISEL-NEXT: s_add_i32 s33, s32, 0xfc0
; GFX9-GISEL-NEXT: s_and_b32 s33, s33, 0xfffff000
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_mov_b32 s11, s34
; GFX9-GISEL-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0
; GFX9-GISEL-NEXT: s_mov_b32 s34, s32
; GFX9-GISEL-NEXT: s_addk_i32 s32, 0x2000
; GFX9-GISEL-NEXT: s_and_saveexec_b64 s[4:5], vcc
; GFX9-GISEL-NEXT: s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-GISEL-NEXT: s_cbranch_execz .LBB15_2
; GFX9-GISEL-NEXT: ; %bb.1: ; %bb.1
; GFX9-GISEL-NEXT: v_lshl_add_u32 v0, v1, 2, 15
; GFX9-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-GISEL-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v0, s[6:7]
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: v_readlane_b32 s8, v2, 63
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-GISEL-NEXT: s_add_u32 s7, s32, 0xfff
; GFX9-GISEL-NEXT: s_and_b32 s7, s7, 0xfffff000
; GFX9-GISEL-NEXT: s_lshl_b32 s6, s8, 6
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 2
; GFX9-GISEL-NEXT: v_mov_b32_e32 v1, s7
; GFX9-GISEL-NEXT: s_add_u32 s32, s7, s6
; GFX9-GISEL-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: ; implicit-def: $vgpr31
; GFX9-GISEL-NEXT: .LBB15_2: ; %Flow
; GFX9-GISEL-NEXT: s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-GISEL-NEXT: s_cbranch_execz .LBB15_4
; GFX9-GISEL-NEXT: ; %bb.3: ; %bb.0
; GFX9-GISEL-NEXT: v_and_b32_e32 v0, 0x3ff, v31
; GFX9-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-GISEL-NEXT: s_mov_b32 s8, s32
; GFX9-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-GISEL-NEXT: s_or_saveexec_b64 s[6:7], -1
; GFX9-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v0, s[6:7]
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: v_readlane_b32 s9, v2, 63
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[6:7]
; GFX9-GISEL-NEXT: s_lshl_b32 s6, s9, 6
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 1
; GFX9-GISEL-NEXT: v_mov_b32_e32 v1, s8
; GFX9-GISEL-NEXT: s_add_u32 s32, s8, s6
; GFX9-GISEL-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: .LBB15_4: ; %bb.2
; GFX9-GISEL-NEXT: s_or_b64 exec, exec, s[4:5]
; GFX9-GISEL-NEXT: s_mov_b32 s32, s34
; GFX9-GISEL-NEXT: s_mov_b32 s34, s11
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_mov_b32 s33, s10
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_device_control_flow:
; GFX11-SDAG: ; %bb.0: ; %entry
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-SDAG-NEXT: s_mov_b32 s3, s33
; GFX11-SDAG-NEXT: s_add_i32 s33, s32, 63
; GFX11-SDAG-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: s_and_not1_b32 s33, s33, 63
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x1 ; 8-byte Folded Spill
; GFX11-SDAG-NEXT: scratch_store_b32 off, v2, s33
; GFX11-SDAG-NEXT: scratch_store_b32 off, v3, s33 offset:4
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s4, s34
; GFX11-SDAG-NEXT: s_mov_b32 s0, exec_lo
; GFX11-SDAG-NEXT: s_mov_b32 s34, s32
; GFX11-SDAG-NEXT: s_addk_i32 s32, 0x80
; GFX11-SDAG-NEXT: v_cmpx_ne_u32_e32 0, v0
; GFX11-SDAG-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-SDAG-NEXT: s_cbranch_execz .LBB15_2
; GFX11-SDAG-NEXT: ; %bb.1: ; %bb.1
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, v1, 2, 15
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-SDAG-NEXT: s_or_saveexec_b32 s1, -1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v2, 0, v0, s1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: ds_swizzle_b32 v3, v2 offset:swizzle(BROADCAST,32,15)
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: v_max_u32_e32 v2, v2, v3
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_readlane_b32 s2, v2, 31
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s1
; GFX11-SDAG-NEXT: s_add_i32 s1, s32, 0x7ff
; GFX11-SDAG-NEXT: v_mov_b32_e32 v1, 2
; GFX11-SDAG-NEXT: s_and_b32 s1, s1, 0xfffff800
; GFX11-SDAG-NEXT: ; implicit-def: $vgpr31
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, s2, 5, s1
; GFX11-SDAG-NEXT: scratch_store_b32 off, v1, s1 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX11-SDAG-NEXT: .LBB15_2: ; %Flow
; GFX11-SDAG-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-SDAG-NEXT: s_cbranch_execz .LBB15_4
; GFX11-SDAG-NEXT: ; %bb.3: ; %bb.0
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, 0x3ff, v31
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, 0x1ff0, v0
; GFX11-SDAG-NEXT: s_or_saveexec_b32 s1, -1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v2, 0, v0, s1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: ds_swizzle_b32 v3, v2 offset:swizzle(BROADCAST,32,15)
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: v_max_u32_e32 v2, v2, v3
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_readlane_b32 s2, v2, 31
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s1
; GFX11-SDAG-NEXT: s_mov_b32 s1, s32
; GFX11-SDAG-NEXT: v_mov_b32_e32 v1, 1
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, s2, 5, s1
; GFX11-SDAG-NEXT: scratch_store_b32 off, v1, s1 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX11-SDAG-NEXT: .LBB15_4: ; %bb.2
; GFX11-SDAG-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s32, s34
; GFX11-SDAG-NEXT: s_mov_b32 s34, s4
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x1 ; 8-byte Folded Reload
; GFX11-SDAG-NEXT: scratch_load_b32 v2, off, s33
; GFX11-SDAG-NEXT: scratch_load_b32 v3, off, s33 offset:4
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s33, s3
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX11-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-GISEL-LABEL: test_dynamic_stackalloc_device_control_flow:
; GFX11-GISEL: ; %bb.0: ; %entry
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-GISEL-NEXT: s_mov_b32 s4, s33
; GFX11-GISEL-NEXT: s_add_i32 s33, s32, 63
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: s_and_not1_b32 s33, s33, 63
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x1 ; 8-byte Folded Spill
; GFX11-GISEL-NEXT: scratch_store_b32 off, v2, s33
; GFX11-GISEL-NEXT: scratch_store_b32 off, v3, s33 offset:4
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: s_mov_b32 s5, s34
; GFX11-GISEL-NEXT: s_mov_b32 s0, exec_lo
; GFX11-GISEL-NEXT: s_mov_b32 s34, s32
; GFX11-GISEL-NEXT: s_addk_i32 s32, 0x80
; GFX11-GISEL-NEXT: v_cmpx_ne_u32_e32 0, v0
; GFX11-GISEL-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-GISEL-NEXT: s_cbranch_execz .LBB15_2
; GFX11-GISEL-NEXT: ; %bb.1: ; %bb.1
; GFX11-GISEL-NEXT: v_lshl_add_u32 v0, v1, 2, 15
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-GISEL-NEXT: s_or_saveexec_b32 s1, -1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v0, s1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: ds_swizzle_b32 v3, v2 offset:swizzle(BROADCAST,32,15)
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: v_max_u32_e32 v2, v2, v3
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_readlane_b32 s2, v2, 31
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s1
; GFX11-GISEL-NEXT: v_mov_b32_e32 v0, 2
; GFX11-GISEL-NEXT: s_add_u32 s1, s32, 0x7ff
; GFX11-GISEL-NEXT: s_lshl_b32 s2, s2, 5
; GFX11-GISEL-NEXT: s_and_b32 s1, s1, 0xfffff800
; GFX11-GISEL-NEXT: ; implicit-def: $vgpr31
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: s_add_u32 s32, s1, s2
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s1 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: .LBB15_2: ; %Flow
; GFX11-GISEL-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-GISEL-NEXT: s_cbranch_execz .LBB15_4
; GFX11-GISEL-NEXT: ; %bb.3: ; %bb.0
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, 0x3ff, v31
; GFX11-GISEL-NEXT: s_mov_b32 s1, s32
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-GISEL-NEXT: s_or_saveexec_b32 s2, -1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v2, 0, v0, s2
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v2, v2, v2 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: ds_swizzle_b32 v3, v2 offset:swizzle(BROADCAST,32,15)
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: v_max_u32_e32 v2, v2, v3
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_readlane_b32 s3, v2, 31
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s2
; GFX11-GISEL-NEXT: v_mov_b32_e32 v0, 1
; GFX11-GISEL-NEXT: s_lshl_b32 s2, s3, 5
; GFX11-GISEL-NEXT: s_add_u32 s32, s1, s2
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s1 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: .LBB15_4: ; %bb.2
; GFX11-GISEL-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-GISEL-NEXT: s_mov_b32 s32, s34
; GFX11-GISEL-NEXT: s_mov_b32 s34, s5
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x1 ; 8-byte Folded Reload
; GFX11-GISEL-NEXT: scratch_load_b32 v2, off, s33
; GFX11-GISEL-NEXT: scratch_load_b32 v3, off, s33 offset:4
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: s_mov_b32 s33, s4
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31]
entry:
%cond = icmp eq i32 %n, 0
br i1 %cond, label %bb.0, label %bb.1
bb.0:
%idx = call i32 @llvm.amdgcn.workitem.id.x()
%alloca1 = alloca i32, i32 %idx, align 4, addrspace(5)
store volatile i32 1, ptr addrspace(5) %alloca1
br label %bb.2
bb.1:
%alloca2 = alloca i32, i32 %m, align 64, addrspace(5)
store volatile i32 2, ptr addrspace(5) %alloca2
br label %bb.2
bb.2:
ret void
}
define void @test_dynamic_stackalloc_device_divergent_non_standard_size_i16(i16 %n) {
; GFX9-SDAG-LABEL: test_dynamic_stackalloc_device_divergent_non_standard_size_i16:
; GFX9-SDAG: ; %bb.0:
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s7, s33
; GFX9-SDAG-NEXT: s_mov_b32 s33, s32
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_store_dword v1, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-SDAG-NEXT: s_addk_i32 s32, 0x400
; GFX9-SDAG-NEXT: v_and_b32_e32 v0, 0x7fff0, v0
; GFX9-SDAG-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[4:5]
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: v_readlane_b32 s6, v1, 63
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s4, s32
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, s4
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, s6, 6, v0
; GFX9-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 0x29a
; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], s4
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s32, s33
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_load_dword v1, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s33, s7
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-GISEL-LABEL: test_dynamic_stackalloc_device_divergent_non_standard_size_i16:
; GFX9-GISEL: ; %bb.0:
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s8, s33
; GFX9-GISEL-NEXT: s_mov_b32 s33, s32
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_store_dword v1, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX9-GISEL-NEXT: s_addk_i32 s32, 0x400
; GFX9-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-GISEL-NEXT: s_mov_b32 s6, s32
; GFX9-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-GISEL-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[4:5]
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: v_readlane_b32 s7, v1, 63
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_lshl_b32 s4, s7, 6
; GFX9-GISEL-NEXT: s_add_u32 s32, s6, s4
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 0x29a
; GFX9-GISEL-NEXT: v_mov_b32_e32 v2, s6
; GFX9-GISEL-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s32, s33
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_load_dword v1, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_mov_b32 s33, s8
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_device_divergent_non_standard_size_i16:
; GFX11-SDAG: ; %bb.0:
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-SDAG-NEXT: s_mov_b32 s2, s33
; GFX11-SDAG-NEXT: s_mov_b32 s33, s32
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x1 ; 8-byte Folded Spill
; GFX11-SDAG-NEXT: scratch_store_b32 off, v1, s33
; GFX11-SDAG-NEXT: scratch_store_b32 off, v2, s33 offset:4
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: v_mov_b16_e32 v3.h, 0
; GFX11-SDAG-NEXT: v_mov_b16_e32 v3.l, v0.l
; GFX11-SDAG-NEXT: s_add_i32 s32, s32, 16
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, v3, 2, 15
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, 0x7fff0, v0
; GFX11-SDAG-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s0
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_readlane_b32 s1, v1, 31
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s0, s32
; GFX11-SDAG-NEXT: v_mov_b32_e32 v3, 0x29a
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, s1, 5, s0
; GFX11-SDAG-NEXT: scratch_store_b32 off, v3, s0 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX11-SDAG-NEXT: s_mov_b32 s32, s33
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x1 ; 8-byte Folded Reload
; GFX11-SDAG-NEXT: scratch_load_b32 v1, off, s33
; GFX11-SDAG-NEXT: scratch_load_b32 v2, off, s33 offset:4
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s33, s2
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX11-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-GISEL-LABEL: test_dynamic_stackalloc_device_divergent_non_standard_size_i16:
; GFX11-GISEL: ; %bb.0:
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-GISEL-NEXT: s_mov_b32 s3, s33
; GFX11-GISEL-NEXT: s_mov_b32 s33, s32
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x1 ; 8-byte Folded Spill
; GFX11-GISEL-NEXT: scratch_store_b32 off, v1, s33
; GFX11-GISEL-NEXT: scratch_store_b32 off, v2, s33 offset:4
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX11-GISEL-NEXT: s_add_i32 s32, s32, 16
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: s_mov_b32 s0, s32
; GFX11-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-GISEL-NEXT: s_or_saveexec_b32 s1, -1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_readlane_b32 s2, v1, 31
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s1
; GFX11-GISEL-NEXT: v_mov_b32_e32 v0, 0x29a
; GFX11-GISEL-NEXT: s_lshl_b32 s1, s2, 5
; GFX11-GISEL-NEXT: s_add_u32 s32, s0, s1
; GFX11-GISEL-NEXT: s_mov_b32 s32, s33
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s0 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x1 ; 8-byte Folded Reload
; GFX11-GISEL-NEXT: scratch_load_b32 v1, off, s33
; GFX11-GISEL-NEXT: scratch_load_b32 v2, off, s33 offset:4
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: s_mov_b32 s33, s3
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31]
%alloca = alloca i32, i16 %n, align 2, addrspace(5)
store volatile i32 666, ptr addrspace(5) %alloca
ret void
}
define void @test_dynamic_stackalloc_device_divergent_non_standard_size_i64(i64 %n) {
; GFX9-SDAG-LABEL: test_dynamic_stackalloc_device_divergent_non_standard_size_i64:
; GFX9-SDAG: ; %bb.0:
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s7, s33
; GFX9-SDAG-NEXT: s_mov_b32 s33, s32
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_store_dword v1, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-SDAG-NEXT: s_addk_i32 s32, 0x400
; GFX9-SDAG-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-SDAG-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[4:5]
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: s_nop 1
; GFX9-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-SDAG-NEXT: v_readlane_b32 s6, v1, 63
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s4, s32
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, s4
; GFX9-SDAG-NEXT: v_lshl_add_u32 v0, s6, 6, v0
; GFX9-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX9-SDAG-NEXT: v_mov_b32_e32 v0, 0x29a
; GFX9-SDAG-NEXT: buffer_store_dword v0, off, s[0:3], s4
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_mov_b32 s32, s33
; GFX9-SDAG-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-SDAG-NEXT: buffer_load_dword v1, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-SDAG-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-SDAG-NEXT: s_mov_b32 s33, s7
; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-GISEL-LABEL: test_dynamic_stackalloc_device_divergent_non_standard_size_i64:
; GFX9-GISEL: ; %bb.0:
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s8, s33
; GFX9-GISEL-NEXT: s_mov_b32 s33, s32
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_store_dword v1, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_addk_i32 s32, 0x400
; GFX9-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX9-GISEL-NEXT: s_mov_b32 s6, s32
; GFX9-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX9-GISEL-NEXT: s_or_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s[4:5]
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:15 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: s_nop 1
; GFX9-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_bcast:31 row_mask:0xf bank_mask:0xf
; GFX9-GISEL-NEXT: v_readlane_b32 s7, v1, 63
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_lshl_b32 s4, s7, 6
; GFX9-GISEL-NEXT: s_add_u32 s32, s6, s4
; GFX9-GISEL-NEXT: v_mov_b32_e32 v0, 0x29a
; GFX9-GISEL-NEXT: v_mov_b32_e32 v2, s6
; GFX9-GISEL-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_mov_b32 s32, s33
; GFX9-GISEL-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-GISEL-NEXT: buffer_load_dword v1, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-GISEL-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-GISEL-NEXT: s_mov_b32 s33, s8
; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-SDAG-LABEL: test_dynamic_stackalloc_device_divergent_non_standard_size_i64:
; GFX11-SDAG: ; %bb.0:
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-SDAG-NEXT: s_mov_b32 s2, s33
; GFX11-SDAG-NEXT: s_mov_b32 s33, s32
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x1 ; 8-byte Folded Spill
; GFX11-SDAG-NEXT: scratch_store_b32 off, v1, s33
; GFX11-SDAG-NEXT: scratch_store_b32 off, v2, s33 offset:4
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-SDAG-NEXT: s_add_i32 s32, s32, 16
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-SDAG-NEXT: s_or_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, v0, s0
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-SDAG-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-SDAG-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-SDAG-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-SDAG-NEXT: v_readlane_b32 s1, v1, 31
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s0, s32
; GFX11-SDAG-NEXT: v_mov_b32_e32 v3, 0x29a
; GFX11-SDAG-NEXT: v_lshl_add_u32 v0, s1, 5, s0
; GFX11-SDAG-NEXT: scratch_store_b32 off, v3, s0 dlc
; GFX11-SDAG-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-SDAG-NEXT: v_readfirstlane_b32 s32, v0
; GFX11-SDAG-NEXT: s_mov_b32 s32, s33
; GFX11-SDAG-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-SDAG-NEXT: s_clause 0x1 ; 8-byte Folded Reload
; GFX11-SDAG-NEXT: scratch_load_b32 v1, off, s33
; GFX11-SDAG-NEXT: scratch_load_b32 v2, off, s33 offset:4
; GFX11-SDAG-NEXT: s_mov_b32 exec_lo, s0
; GFX11-SDAG-NEXT: s_mov_b32 s33, s2
; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0)
; GFX11-SDAG-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-GISEL-LABEL: test_dynamic_stackalloc_device_divergent_non_standard_size_i64:
; GFX11-GISEL: ; %bb.0:
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-GISEL-NEXT: s_mov_b32 s3, s33
; GFX11-GISEL-NEXT: s_mov_b32 s33, s32
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x1 ; 8-byte Folded Spill
; GFX11-GISEL-NEXT: scratch_store_b32 off, v1, s33
; GFX11-GISEL-NEXT: scratch_store_b32 off, v2, s33 offset:4
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: v_lshl_add_u32 v0, v0, 2, 15
; GFX11-GISEL-NEXT: s_add_i32 s32, s32, 16
; GFX11-GISEL-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: s_mov_b32 s0, s32
; GFX11-GISEL-NEXT: v_and_b32_e32 v0, -16, v0
; GFX11-GISEL-NEXT: s_or_saveexec_b32 s1, -1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, v0, s1
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:1 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:2 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:4 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: v_max_u32_dpp v1, v1, v1 row_shr:8 row_mask:0xf bank_mask:0xf
; GFX11-GISEL-NEXT: ds_swizzle_b32 v2, v1 offset:swizzle(BROADCAST,32,15)
; GFX11-GISEL-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT: v_max_u32_e32 v1, v1, v2
; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT: v_readlane_b32 s2, v1, 31
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s1
; GFX11-GISEL-NEXT: v_mov_b32_e32 v0, 0x29a
; GFX11-GISEL-NEXT: s_lshl_b32 s1, s2, 5
; GFX11-GISEL-NEXT: s_add_u32 s32, s0, s1
; GFX11-GISEL-NEXT: s_mov_b32 s32, s33
; GFX11-GISEL-NEXT: scratch_store_b32 off, v0, s0 dlc
; GFX11-GISEL-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-GISEL-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-GISEL-NEXT: s_clause 0x1 ; 8-byte Folded Reload
; GFX11-GISEL-NEXT: scratch_load_b32 v1, off, s33
; GFX11-GISEL-NEXT: scratch_load_b32 v2, off, s33 offset:4
; GFX11-GISEL-NEXT: s_mov_b32 exec_lo, s0
; GFX11-GISEL-NEXT: s_mov_b32 s33, s3
; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0)
; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31]
%alloca = alloca i32, i64 %n, align 2, addrspace(5)
store volatile i32 666, ptr addrspace(5) %alloca
ret void
}
Reference in New Issue View Git Blame Copy Permalink