shylie/llvm-project
1
0
Fork 0
Code Issues Pull Requests Actions 6 Packages Projects Releases Wiki Activity
llvm-project/llvm/test/CodeGen/AMDGPU/spill-offset-calculation.ll
Matt Arsenault 31973062ec AMDGPU: Fix clobbering SCC when expanding large offset spill pseudos 
If we had a large offset which required materializing in a register,
we would emit an s_add_i32, clobbering SCC. Start checking if SCC is
live, and instead use a VGPR offset. For MUBUF, we switch to using
offen. We would do this anyway in a normal load/store with a frame
index, but not for spills.

The same problem still exists in other contexts where we expand frame
indices.

The nasty edge case is when SGPRs are spilled to memory at a large
frame offset where SCC is also clobbered. This requires a second
scavenging index, and also required several patches in the scavenger
to correctly handle multiple recursive scavenge indexes.

An even nastier edge case we still don't support is if we don't have
any free SGPRs. If SCC is live and we don't have any free SGPRs to
save exec, we have no way of flipping exec back and forth without also
clobbering SCC.

Fixes: SWDEV-309419
2022-02-07 10:02:03 -05:00

341 lines
16 KiB
LLVM
Raw Blame History

; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -enable-misched=0 -post-RA-scheduler=0 -stress-regalloc=8 < %s | FileCheck -check-prefixes=GCN,MUBUF %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -enable-misched=0 -post-RA-scheduler=0 -stress-regalloc=8 -amdgpu-enable-flat-scratch < %s | FileCheck -check-prefixes=GCN,FLATSCR %s
; Test that the VGPR spiller correctly switches to SGPR offsets when the
; instruction offset field would overflow, and that it accounts for memory
; swizzling.
; GCN-LABEL: test_inst_offset_kernel
define amdgpu_kernel void @test_inst_offset_kernel() {
entry:
; Occupy 4092 bytes of scratch, so the offset of the spill of %a just fits in
; the instruction offset field.
%alloca = alloca i8, i32 4088, align 4, addrspace(5)
%buf = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
%aptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
; MUBUF: buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], 0 offset:4092 ; 4-byte Folded Spill
; FLATSCR: scratch_store_dword off, v{{[0-9]+}}, s{{[0-9]+}} ; 4-byte Folded Spill
%a = load volatile i32, i32 addrspace(5)* %aptr
; Force %a to spill.
call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
%outptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
store volatile i32 %a, i32 addrspace(5)* %outptr
ret void
}
; GCN-LABEL: test_sgpr_offset_kernel
define amdgpu_kernel void @test_sgpr_offset_kernel() {
entry:
; Occupy 4096 bytes of scratch, so the offset of the spill of %a does not
; fit in the instruction, and has to live in the SGPR offset.
%alloca = alloca i8, i32 4092, align 4, addrspace(5)
%buf = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
%aptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
; 0x40000 / 64 = 4096 (for wave64)
; MUBUF: s_mov_b32 s4, 0x40000
; MUBUF: buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s4 ; 4-byte Folded Spill
; FLATSCR: s_movk_i32 s2, 0x1000
; FLATSCR: scratch_store_dword off, v{{[0-9]+}}, s2 ; 4-byte Folded Spill
%a = load volatile i32, i32 addrspace(5)* %aptr
; Force %a to spill
call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
%outptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
store volatile i32 %a, i32 addrspace(5)* %outptr
ret void
}
; FIXME: If we fail to scavenge an SGPR in a kernel we don't have a stack
; pointer to temporarily update, so we just crash.
; GCN-LABEL: test_sgpr_offset_function_scavenge_fail_func
define void @test_sgpr_offset_function_scavenge_fail_func() #2 {
entry:
; Occupy 4096 bytes of scratch, so the offset of the spill of %a does not
; fit in the instruction, and has to live in the SGPR offset.
%alloca = alloca i8, i32 4096, align 4, addrspace(5)
%buf = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
%aptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
%asm.0 = call { i32, i32, i32, i32, i32, i32, i32, i32 } asm sideeffect "", "=s,=s,=s,=s,=s,=s,=s,=s"()
%asm0.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 0
%asm1.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 1
%asm2.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 2
%asm3.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 3
%asm4.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 4
%asm5.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 5
%asm6.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 6
%asm7.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 7
; 0x40000 / 64 = 4096 (for wave64)
%a = load volatile i32, i32 addrspace(5)* %aptr
; MUBUF: v_mov_b32_e32 [[OFFSET:v[0-9]+]], 0x1004
; MUBUF-NEXT: buffer_store_dword v{{[0-9]+}}, [[OFFSET]], s[{{[0-9]+:[0-9]+}}], s32 offen ; 4-byte Folded Spill
; FLATSCR: s_add_i32 [[SOFF:s[0-9]+]], s32, 0x1004
; FLATSCR: scratch_store_dword off, v{{[0-9]+}}, [[SOFF]] ; 4-byte Folded Spill
call void asm sideeffect "", "s,s,s,s,s,s,s,s,v"(i32 %asm0.0, i32 %asm1.0, i32 %asm2.0, i32 %asm3.0, i32 %asm4.0, i32 %asm5.0, i32 %asm6.0, i32 %asm7.0, i32 %a)
%asm = call { i32, i32, i32, i32, i32, i32, i32, i32 } asm sideeffect "", "=s,=s,=s,=s,=s,=s,=s,=s"()
%asm0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 0
%asm1 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 1
%asm2 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 2
%asm3 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 3
%asm4 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 4
%asm5 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 5
%asm6 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 6
%asm7 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 7
call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}"() #0
; MUBUF: v_mov_b32_e32 [[OFFSET:v[0-9]+]], 0x1004
; MUBUF-NEXT: buffer_load_dword v{{[0-9]+}}, [[OFFSET]], s[{{[0-9]+:[0-9]+}}], s32 offen ; 4-byte Folded Reload
; FLATSCR: s_add_i32 [[SOFF:s[0-9]+]], s32, 0x1004
; FLATSCR: scratch_load_dword v{{[0-9]+}}, off, [[SOFF]] ; 4-byte Folded Reload
; Force %a to spill with no free SGPRs
call void asm sideeffect "", "s,s,s,s,s,s,s,s,v"(i32 %asm0, i32 %asm1, i32 %asm2, i32 %asm3, i32 %asm4, i32 %asm5, i32 %asm6, i32 %asm7, i32 %a)
ret void
}
define amdgpu_kernel void @test_sgpr_offset_function_scavenge_fail_kernel() #3 {
entry:
; Occupy 4096 bytes of scratch, so the offset of the spill of %a does not
; fit in the instruction, and has to live in the SGPR offset.
%alloca = alloca i8, i32 4096, align 4, addrspace(5)
%buf = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
%aptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
%asm.0 = call { i32, i32, i32, i32, i32, i32, i32, i32 } asm sideeffect "", "=s,=s,=s,=s,=s,=s,=s,=s"()
%asm0.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 0
%asm1.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 1
%asm2.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 2
%asm3.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 3
%asm4.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 4
%asm5.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 5
%asm6.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 6
%asm7.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 7
; 0x40000 / 64 = 4096 (for wave64)
%a = load volatile i32, i32 addrspace(5)* %aptr
; MUBUF: v_mov_b32_e32 [[OFFSET:v[0-9]+]], 0x1004
; MUBUF: buffer_store_dword v{{[0-9]+}}, [[OFFSET]], s[{{[0-9]+:[0-9]+}}], 0 offen ; 4-byte Folded Spill
; FLATSCR: s_movk_i32 [[SOFF:s[0-9]+]], 0x1004
; FLATSCR: scratch_store_dword off, v{{[0-9]+}}, [[SOFF]] ; 4-byte Folded Spill
call void asm sideeffect "", "s,s,s,s,s,s,s,s,v"(i32 %asm0.0, i32 %asm1.0, i32 %asm2.0, i32 %asm3.0, i32 %asm4.0, i32 %asm5.0, i32 %asm6.0, i32 %asm7.0, i32 %a)
%asm = call { i32, i32, i32, i32, i32, i32, i32, i32 } asm sideeffect "", "=s,=s,=s,=s,=s,=s,=s,=s"()
%asm0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 0
%asm1 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 1
%asm2 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 2
%asm3 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 3
%asm4 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 4
%asm5 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 5
%asm6 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 6
%asm7 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 7
call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}"() #0
; MUBUF: v_mov_b32_e32 [[OFFSET:v[0-9]+]], 0x1004
; MUBUF: buffer_load_dword v{{[0-9]+}}, [[OFFSET]], s[{{[0-9]+:[0-9]+}}], 0 offen ; 4-byte Folded Reload
; FLATSCR: s_movk_i32 [[SOFF:s[0-9]+]], 0x1004
; FLATSCR: scratch_load_dword v{{[0-9]+}}, off, [[SOFF]] ; 4-byte Folded Reload
; Force %a to spill with no free SGPRs
call void asm sideeffect "", "s,s,s,s,s,s,s,s,v"(i32 %asm0, i32 %asm1, i32 %asm2, i32 %asm3, i32 %asm4, i32 %asm5, i32 %asm6, i32 %asm7, i32 %a)
ret void
}
; GCN-LABEL: test_sgpr_offset_subregs_kernel
define amdgpu_kernel void @test_sgpr_offset_subregs_kernel() {
entry:
; Occupy 4088 bytes of scratch, so that the spill of the last subreg of %a
; still fits below offset 4096 (4088 + 8 - 4 = 4092), and can be placed in
; the instruction offset field.
%alloca = alloca i8, i32 4084, align 4, addrspace(5)
%bufv1 = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
%bufv2 = bitcast i8 addrspace(5)* %alloca to <2 x i32> addrspace(5)*
; MUBUF: buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], 0 offset:4088 ; 4-byte Folded Spill
; MUBUF: buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], 0 offset:4092 ; 4-byte Folded Spill
; FLATSCR: s_movk_i32 [[SOFF:s[0-9]+]], 0xff8
; FLATSCR: scratch_store_dwordx2 off, v[{{[0-9:]+}}], [[SOFF]] ; 8-byte Folded Spill
%aptr = getelementptr <2 x i32>, <2 x i32> addrspace(5)* %bufv2, i32 1
%a = load volatile <2 x i32>, <2 x i32> addrspace(5)* %aptr
; Force %a to spill.
call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
; Ensure the alloca sticks around.
%bptr = getelementptr i32, i32 addrspace(5)* %bufv1, i32 1
%b = load volatile i32, i32 addrspace(5)* %bptr
; Ensure the spill is of the full super-reg.
call void asm sideeffect "; $0", "r"(<2 x i32> %a)
ret void
}
; GCN-LABEL: test_inst_offset_subregs_kernel
define amdgpu_kernel void @test_inst_offset_subregs_kernel() {
entry:
; Occupy 4092 bytes of scratch, so that the spill of the last subreg of %a
; does not fit below offset 4096 (4092 + 8 - 4 = 4096), and has to live
; in the SGPR offset.
%alloca = alloca i8, i32 4088, align 4, addrspace(5)
%bufv1 = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
%bufv2 = bitcast i8 addrspace(5)* %alloca to <2 x i32> addrspace(5)*
; 0x3ff00 / 64 = 4092 (for wave64)
; MUBUF: s_mov_b32 s4, 0x3ff00
; MUBUF: buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s4 ; 4-byte Folded Spill
; MUBUF: buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s4 offset:4 ; 4-byte Folded Spill
; FLATSCR: s_movk_i32 [[SOFF:s[0-9]+]], 0xffc
; FLATSCR: scratch_store_dwordx2 off, v[{{[0-9:]+}}], [[SOFF]] ; 8-byte Folded Spill
%aptr = getelementptr <2 x i32>, <2 x i32> addrspace(5)* %bufv2, i32 1
%a = load volatile <2 x i32>, <2 x i32> addrspace(5)* %aptr
; Force %a to spill.
call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
; Ensure the alloca sticks around.
%bptr = getelementptr i32, i32 addrspace(5)* %bufv1, i32 1
%b = load volatile i32, i32 addrspace(5)* %bptr
; Ensure the spill is of the full super-reg.
call void asm sideeffect "; $0", "r"(<2 x i32> %a)
ret void
}
; GCN-LABEL: test_inst_offset_function
define void @test_inst_offset_function() {
entry:
; Occupy enough bytes of scratch, so the offset of the spill of %a
; just fits in the instruction offset field when the emergency stack
; slot is added. It's hard to hit the actual limit since we're also
; going to insert the emergency stack slot for large frames.
%alloca = alloca i8, i32 4088, align 4, addrspace(5)
%buf = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
%aptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
; MUBUF: buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offset:4088 ; 4-byte Folded Spill
; FLATSCR: scratch_store_dword off, v{{[0-9]+}}, s{{[0-9]+}} offset:4088 ; 4-byte Folded Spill
%a = load volatile i32, i32 addrspace(5)* %aptr
; Force %a to spill.
call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
%outptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
store volatile i32 %a, i32 addrspace(5)* %outptr
ret void
}
; GCN-LABEL: test_sgpr_offset_function
define void @test_sgpr_offset_function() {
entry:
; Occupy 4096 bytes of scratch, so the offset of the spill of %a does not
; fit in the instruction, and has to live in the SGPR offset.
%alloca = alloca i8, i32 4096, align 4, addrspace(5)
%buf = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
%aptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
; 0x40000 / 64 = 4096 (for wave64)
; MUBUF: s_add_i32 s4, s32, 0x40100
; MUBUF: buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s4 ; 4-byte Folded Spill
; FLATSCR: s_add_i32 s0, s32, 0x1004
; FLATSCR: scratch_store_dword off, v{{[0-9]+}}, s0 ; 4-byte Folded Spill
%a = load volatile i32, i32 addrspace(5)* %aptr
; Force %a to spill
call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
%outptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
store volatile i32 %a, i32 addrspace(5)* %outptr
ret void
}
; GCN-LABEL: test_sgpr_offset_subregs_function
define void @test_sgpr_offset_subregs_function() {
entry:
; We want to test the spill of the last subreg of %a is the highest
; valid value for the immediate offset. We enable the emergency
; stack slot for large frames, so it's hard to get the frame layout
; exactly as we want to test it.
;
; Occupy 4084 bytes of scratch, so that the spill of the last subreg of %a
; still fits below offset 4096 (4084 + 8 - 4 = 4092), and can be placed in
; the instruction offset field.
%alloca = alloca i8, i32 4084, align 4, addrspace(5)
%bufv1 = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
%bufv2 = bitcast i8 addrspace(5)* %alloca to <2 x i32> addrspace(5)*
; MUBUF: buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offset:4084 ; 4-byte Folded Spill
; MUBUF: buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offset:4088 ; 4-byte Folded Spill
; FLATSCR: scratch_store_dwordx2 off, v[{{[0-9:]+}}], s32 offset:4084 ; 8-byte Folded Spill
%aptr = getelementptr <2 x i32>, <2 x i32> addrspace(5)* %bufv2, i32 1
%a = load volatile <2 x i32>, <2 x i32> addrspace(5)* %aptr
; Force %a to spill.
call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
; Ensure the alloca sticks around.
%bptr = getelementptr i32, i32 addrspace(5)* %bufv1, i32 1
%b = load volatile i32, i32 addrspace(5)* %bptr
; Ensure the spill is of the full super-reg.
call void asm sideeffect "; $0", "r"(<2 x i32> %a)
ret void
}
; GCN-LABEL: test_inst_offset_subregs_function
define void @test_inst_offset_subregs_function() {
entry:
; Occupy 4088 bytes of scratch, so that the spill of the last subreg of %a
; does not fit below offset 4096 (408 + 4 + 8 - 4 = 4096), and has to live
; in the SGPR offset.
%alloca = alloca i8, i32 4088, align 4, addrspace(5)
%bufv1 = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
%bufv2 = bitcast i8 addrspace(5)* %alloca to <2 x i32> addrspace(5)*
; 0x3ff0000 / 64 = 4092 (for wave64)
; MUBUF: s_add_i32 s4, s32, 0x3ff00
; MUBUF: buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s4 ; 4-byte Folded Spill
; MUBUF: buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s4 offset:4 ; 4-byte Folded Spill
; FLATSCR: scratch_store_dwordx2 off, v[{{[0-9:]+}}], s32 offset:4092 ; 8-byte Folded Spill
%aptr = getelementptr <2 x i32>, <2 x i32> addrspace(5)* %bufv2, i32 1
%a = load volatile <2 x i32>, <2 x i32> addrspace(5)* %aptr
; Force %a to spill.
call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
; Ensure the alloca sticks around.
%bptr = getelementptr i32, i32 addrspace(5)* %bufv1, i32 1
%b = load volatile i32, i32 addrspace(5)* %bptr
; Ensure the spill is of the full super-reg.
call void asm sideeffect "; $0", "r"(<2 x i32> %a)
ret void
}
attributes #0 = { nounwind }
attributes #1 = { nounwind "amdgpu-num-sgpr"="17" "amdgpu-num-vgpr"="8" }
attributes #2 = { nounwind "amdgpu-num-sgpr"="14" "amdgpu-num-vgpr"="8" }
attributes #3 = { nounwind "amdgpu-num-sgpr"="16" "amdgpu-num-vgpr"="8" }
Reference in New Issue View Git Blame Copy Permalink