File: spill-sgpr-stack-no-sgpr.ll

package info (click to toggle)
llvm-toolchain-21 1%3A21.1.4-5
  • links: PTS, VCS
  • area: main
  • in suites: sid
  • size: 2,236,516 kB
  • sloc: cpp: 7,619,569; ansic: 1,433,956; asm: 1,058,748; python: 252,181; f90: 94,671; objc: 70,753; lisp: 42,813; pascal: 18,401; sh: 8,601; ml: 5,111; perl: 4,720; makefile: 3,585; awk: 3,523; javascript: 2,272; xml: 892; fortran: 770
file content (44 lines) | stat: -rw-r--r-- 1,807 bytes parent folder | download | duplicates (18)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -mattr=+wavefrontsize64 < %s | FileCheck -check-prefix=GFX10 %s

; The test was originally written to spill an SGPR to scratch without
; having spare SGPRs available to save exec. This scenario won't be
; true anymore as we reserve SGPR(s) upfront for saving exec.

define amdgpu_kernel void @test() #1 {
; GFX10-LABEL: test:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_mov_b32 s8, SCRATCH_RSRC_DWORD0
; GFX10-NEXT:    s_mov_b32 s9, SCRATCH_RSRC_DWORD1
; GFX10-NEXT:    s_mov_b32 s10, -1
; GFX10-NEXT:    s_mov_b32 s11, 0x31e16000
; GFX10-NEXT:    s_add_u32 s8, s8, s7
; GFX10-NEXT:    s_addc_u32 s9, s9, 0
; GFX10-NEXT:    ;;#ASMSTART
; GFX10-NEXT:    ; def s[0:7]
; GFX10-NEXT:    ;;#ASMEND
; GFX10-NEXT:    ;;#ASMSTART
; GFX10-NEXT:    ; def s[8:12]
; GFX10-NEXT:    ;;#ASMEND
; GFX10-NEXT:    ;;#ASMSTART
; GFX10-NEXT:    ;;#ASMEND
; GFX10-NEXT:    ;;#ASMSTART
; GFX10-NEXT:    ; use s[0:7]
; GFX10-NEXT:    ;;#ASMEND
; GFX10-NEXT:    ;;#ASMSTART
; GFX10-NEXT:    ; use s[8:12]
; GFX10-NEXT:    ;;#ASMEND
; GFX10-NEXT:    s_endpgm
  %wide.sgpr0 = call <8 x i32> asm sideeffect "; def $0", "={s[0:7]}" () #0
  %wide.sgpr2 = call <5 x i32> asm sideeffect "; def $0", "={s[8:12]}" () #0
  call void asm sideeffect "", "~{v[0:7]}" () #0
  call void asm sideeffect "; use $0", "s"(<8 x i32> %wide.sgpr0) #0
  call void asm sideeffect "; use $0", "s"(<5 x i32> %wide.sgpr2) #0
  ret void
}

; FIXME: amdgpu-no attributese are a workaround for cases where the
; number of incoming arguments is larger than the number of permitted
; registers.
attributes #0 = { nounwind }
attributes #1 = { nounwind "amdgpu-num-sgpr"="16" "amdgpu-num-vgpr"="8" "amdgpu-no-queue-ptr" "amdgpu-no-dispatch-id" }