Skip to content

AMDGPU: Remove arbitrary SCC liveness scan threshold #94097

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Jun 1, 2024

Conversation

arsenm
Copy link
Contributor

@arsenm arsenm commented Jun 1, 2024

When checking if SCC needs to be saved and restored, we need an accurate liveness query. Fixes verifier error from inserting an undef use of SCC.

Fixes #92561

When checking if SCC needs to be saved and restored, we need an
accurate liveness query. Fixes verifier error from inserting an
undef use of SCC.

Fixes llvm#92561
@llvmbot
Copy link
Member

llvmbot commented Jun 1, 2024

@llvm/pr-subscribers-backend-amdgpu

Author: Matt Arsenault (arsenm)

Changes

When checking if SCC needs to be saved and restored, we need an accurate liveness query. Fixes verifier error from inserting an undef use of SCC.

Fixes #92561


Full diff: https://github.com/llvm/llvm-project/pull/94097.diff

2 Files Affected:

  • (modified) llvm/lib/Target/AMDGPU/SIInstrInfo.cpp (+7-2)
  • (added) llvm/test/CodeGen/AMDGPU/issue92561-restore-undef-scc-verifier-error.ll (+160)
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 779f4c853326d..d8e21da8019a7 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -6437,8 +6437,13 @@ loadMBUFScalarOperandsFromVGPR(const SIInstrInfo &TII, MachineInstr &MI,
 
   // Save SCC. Waterfall Loop may overwrite SCC.
   Register SaveSCCReg;
-  bool SCCNotDead = (MBB.computeRegisterLiveness(TRI, AMDGPU::SCC, MI, 30) !=
-                     MachineBasicBlock::LQR_Dead);
+
+  // FIXME: We should maintain SCC liveness while doing the FixSGPRCopies walk
+  // rather than unlimited scan everywhere
+  bool SCCNotDead =
+      MBB.computeRegisterLiveness(TRI, AMDGPU::SCC, MI,
+                                  std::numeric_limits<unsigned>::max()) !=
+      MachineBasicBlock::LQR_Dead;
   if (SCCNotDead) {
     SaveSCCReg = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
     BuildMI(MBB, Begin, DL, TII.get(AMDGPU::S_CSELECT_B32), SaveSCCReg)
diff --git a/llvm/test/CodeGen/AMDGPU/issue92561-restore-undef-scc-verifier-error.ll b/llvm/test/CodeGen/AMDGPU/issue92561-restore-undef-scc-verifier-error.ll
new file mode 100644
index 0000000000000..0adce2b84aa0d
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/issue92561-restore-undef-scc-verifier-error.ll
@@ -0,0 +1,160 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck -check-prefix=SDAG %s
+; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck -check-prefix=GISEL %s
+
+; Check for verifier error due to trying to save and restore SCC
+; around a waterfall looop when it was never defined. We have to get
+; an accurate liveness at the use point and cannot rely on an
+; imprecise maybe-live query.
+
+define void @issue92561(ptr addrspace(1) %arg) {
+; SDAG-LABEL: issue92561:
+; SDAG:       ; %bb.0: ; %bb
+; SDAG-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SDAG-NEXT:    s_clause 0x1
+; SDAG-NEXT:    global_load_b128 v[4:7], v[0:1], off offset:16
+; SDAG-NEXT:    global_load_b128 v[0:3], v[0:1], off
+; SDAG-NEXT:    v_mov_b32_e32 v8, 0
+; SDAG-NEXT:    s_mov_b32 s12, 0
+; SDAG-NEXT:    s_mov_b32 s3, exec_lo
+; SDAG-NEXT:    s_mov_b32 s13, s12
+; SDAG-NEXT:    s_mov_b32 s14, s12
+; SDAG-NEXT:    s_mov_b32 s15, s12
+; SDAG-NEXT:    s_waitcnt vmcnt(0)
+; SDAG-NEXT:  .LBB0_1: ; =>This Inner Loop Header: Depth=1
+; SDAG-NEXT:    v_readfirstlane_b32 s4, v0
+; SDAG-NEXT:    v_readfirstlane_b32 s5, v1
+; SDAG-NEXT:    v_readfirstlane_b32 s6, v2
+; SDAG-NEXT:    v_readfirstlane_b32 s7, v3
+; SDAG-NEXT:    v_readfirstlane_b32 s8, v4
+; SDAG-NEXT:    v_readfirstlane_b32 s9, v5
+; SDAG-NEXT:    v_readfirstlane_b32 s10, v6
+; SDAG-NEXT:    v_readfirstlane_b32 s11, v7
+; SDAG-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
+; SDAG-NEXT:    v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
+; SDAG-NEXT:    v_cmp_eq_u64_e64 s1, s[8:9], v[4:5]
+; SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
+; SDAG-NEXT:    v_cmp_eq_u64_e64 s2, s[10:11], v[6:7]
+; SDAG-NEXT:    s_and_b32 s0, vcc_lo, s0
+; SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instid1(SALU_CYCLE_1)
+; SDAG-NEXT:    s_and_b32 s0, s0, s1
+; SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; SDAG-NEXT:    s_and_b32 s0, s0, s2
+; SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; SDAG-NEXT:    s_and_saveexec_b32 s0, s0
+; SDAG-NEXT:    image_sample_c_lz v9, [v8, v8, v8, v8], s[4:11], s[12:15] dmask:0x1 dim:SQ_RSRC_IMG_2D_ARRAY
+; SDAG-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
+; SDAG-NEXT:    ; implicit-def: $vgpr8
+; SDAG-NEXT:    s_xor_b32 exec_lo, exec_lo, s0
+; SDAG-NEXT:    s_cbranch_execnz .LBB0_1
+; SDAG-NEXT:  ; %bb.2:
+; SDAG-NEXT:    s_mov_b32 exec_lo, s3
+; SDAG-NEXT:    v_dual_mov_b32 v0, 0x7fc00000 :: v_dual_mov_b32 v1, 0
+; SDAG-NEXT:    v_mov_b32_e32 v2, 1.0
+; SDAG-NEXT:    s_mov_b32 s0, s12
+; SDAG-NEXT:    s_mov_b32 s1, s12
+; SDAG-NEXT:    s_mov_b32 s2, s12
+; SDAG-NEXT:    s_mov_b32 s3, s12
+; SDAG-NEXT:    s_mov_b32 s4, s12
+; SDAG-NEXT:    s_mov_b32 s5, s12
+; SDAG-NEXT:    s_mov_b32 s6, s12
+; SDAG-NEXT:    s_mov_b32 s7, s12
+; SDAG-NEXT:    s_clause 0x2
+; SDAG-NEXT:    image_sample_c_lz v0, [v1, v1, v0, v1], s[0:7], s[12:15] dmask:0x1 dim:SQ_RSRC_IMG_2D_ARRAY
+; SDAG-NEXT:    image_sample_c_lz v3, [v1, v1, v1, v1], s[0:7], s[12:15] dmask:0x1 dim:SQ_RSRC_IMG_2D_ARRAY
+; SDAG-NEXT:    image_sample_c_lz v2, [v1, v2, v1, v1], s[0:7], s[12:15] dmask:0x1 dim:SQ_RSRC_IMG_2D_ARRAY
+; SDAG-NEXT:    v_mov_b32_e32 v4, v1
+; SDAG-NEXT:    s_waitcnt vmcnt(2)
+; SDAG-NEXT:    v_add_f32_e32 v0, v9, v0
+; SDAG-NEXT:    s_waitcnt vmcnt(0)
+; SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; SDAG-NEXT:    v_add_f32_e32 v0, v2, v0
+; SDAG-NEXT:    v_mov_b32_e32 v2, v1
+; SDAG-NEXT:    v_dual_add_f32 v0, v3, v0 :: v_dual_mov_b32 v3, v1
+; SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; SDAG-NEXT:    v_mul_f32_e32 v0, 0x3e800000, v0
+; SDAG-NEXT:    image_store v[0:2], v[3:4], s[0:7] dim:SQ_RSRC_IMG_2D unorm
+; SDAG-NEXT:    s_setpc_b64 s[30:31]
+;
+; GISEL-LABEL: issue92561:
+; GISEL:       ; %bb.0: ; %bb
+; GISEL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GISEL-NEXT:    s_clause 0x1
+; GISEL-NEXT:    global_load_b128 v[2:5], v[0:1], off
+; GISEL-NEXT:    global_load_b128 v[6:9], v[0:1], off offset:16
+; GISEL-NEXT:    v_mov_b32_e32 v0, 0
+; GISEL-NEXT:    s_mov_b32 s20, 0
+; GISEL-NEXT:    s_mov_b32 s3, exec_lo
+; GISEL-NEXT:    s_mov_b32 s21, s20
+; GISEL-NEXT:    s_mov_b32 s22, s20
+; GISEL-NEXT:    s_mov_b32 s23, s20
+; GISEL-NEXT:    s_mov_b32 s4, s20
+; GISEL-NEXT:    s_mov_b32 s5, s20
+; GISEL-NEXT:    s_mov_b32 s6, s20
+; GISEL-NEXT:    s_mov_b32 s7, s20
+; GISEL-NEXT:    s_mov_b32 s8, s20
+; GISEL-NEXT:    s_mov_b32 s9, s20
+; GISEL-NEXT:    s_mov_b32 s10, s20
+; GISEL-NEXT:    s_mov_b32 s11, s20
+; GISEL-NEXT:    s_waitcnt vmcnt(0)
+; GISEL-NEXT:  .LBB0_1: ; =>This Inner Loop Header: Depth=1
+; GISEL-NEXT:    v_readfirstlane_b32 s12, v2
+; GISEL-NEXT:    v_readfirstlane_b32 s13, v3
+; GISEL-NEXT:    v_readfirstlane_b32 s14, v4
+; GISEL-NEXT:    v_readfirstlane_b32 s15, v5
+; GISEL-NEXT:    v_readfirstlane_b32 s16, v6
+; GISEL-NEXT:    v_readfirstlane_b32 s17, v7
+; GISEL-NEXT:    v_readfirstlane_b32 s18, v8
+; GISEL-NEXT:    v_readfirstlane_b32 s19, v9
+; GISEL-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[12:13], v[2:3]
+; GISEL-NEXT:    v_cmp_eq_u64_e64 s0, s[14:15], v[4:5]
+; GISEL-NEXT:    v_cmp_eq_u64_e64 s1, s[16:17], v[6:7]
+; GISEL-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GISEL-NEXT:    v_cmp_eq_u64_e64 s2, s[18:19], v[8:9]
+; GISEL-NEXT:    s_and_b32 s0, vcc_lo, s0
+; GISEL-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instid1(SALU_CYCLE_1)
+; GISEL-NEXT:    s_and_b32 s0, s0, s1
+; GISEL-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
+; GISEL-NEXT:    s_and_b32 s0, s0, s2
+; GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GISEL-NEXT:    s_and_saveexec_b32 s0, s0
+; GISEL-NEXT:    image_sample_c_lz v1, [v0, v0, v0, v0], s[12:19], s[20:23] dmask:0x1 dim:SQ_RSRC_IMG_2D_ARRAY
+; GISEL-NEXT:    ; implicit-def: $vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9
+; GISEL-NEXT:    ; implicit-def: $vgpr0
+; GISEL-NEXT:    s_xor_b32 exec_lo, exec_lo, s0
+; GISEL-NEXT:    s_cbranch_execnz .LBB0_1
+; GISEL-NEXT:  ; %bb.2:
+; GISEL-NEXT:    s_mov_b32 exec_lo, s3
+; GISEL-NEXT:    v_dual_mov_b32 v2, 0 :: v_dual_mov_b32 v3, 1.0
+; GISEL-NEXT:    v_mov_b32_e32 v0, 0x7fc00000
+; GISEL-NEXT:    s_clause 0x2
+; GISEL-NEXT:    image_sample_c_lz v0, [v2, v2, v0, v2], s[4:11], s[20:23] dmask:0x1 dim:SQ_RSRC_IMG_2D_ARRAY
+; GISEL-NEXT:    image_sample_c_lz v3, [v2, v3, v2, v2], s[4:11], s[20:23] dmask:0x1 dim:SQ_RSRC_IMG_2D_ARRAY
+; GISEL-NEXT:    image_sample_c_lz v4, [v2, v2, v2, v2], s[4:11], s[20:23] dmask:0x1 dim:SQ_RSRC_IMG_2D_ARRAY
+; GISEL-NEXT:    s_mov_b32 s21, s20
+; GISEL-NEXT:    s_waitcnt vmcnt(2)
+; GISEL-NEXT:    v_add_f32_e32 v0, v1, v0
+; GISEL-NEXT:    s_waitcnt vmcnt(1)
+; GISEL-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GISEL-NEXT:    v_dual_add_f32 v0, v3, v0 :: v_dual_mov_b32 v3, v2
+; GISEL-NEXT:    s_waitcnt vmcnt(0)
+; GISEL-NEXT:    v_add_f32_e32 v0, v4, v0
+; GISEL-NEXT:    v_dual_mov_b32 v4, s20 :: v_dual_mov_b32 v5, s21
+; GISEL-NEXT:    s_delay_alu instid0(VALU_DEP_2)
+; GISEL-NEXT:    v_mul_f32_e32 v1, 0x3e800000, v0
+; GISEL-NEXT:    image_store v[1:3], v[4:5], s[4:11] dim:SQ_RSRC_IMG_2D unorm
+; GISEL-NEXT:    s_setpc_b64 s[30:31]
+bb:
+  %descriptor = load <8 x i32>, ptr addrspace(1) %arg, align 32
+  %needs.waterfall = call float @llvm.amdgcn.image.sample.c.lz.2darray.f32.f32(i32 1, float 0.0, float 0.0, float 0.0, float 0.0, <8 x i32> %descriptor, <4 x i32> zeroinitializer, i1 false, i32 0, i32 0)
+  %i2 = call float @llvm.amdgcn.image.sample.c.lz.2darray.f32.f32(i32 1, float 0.0, float 0.0, float 0x7FF8000000000000, float 0.0, <8 x i32> zeroinitializer, <4 x i32> zeroinitializer, i1 false, i32 0, i32 0)
+  %i3 = fadd float %needs.waterfall, %i2
+  %i4 = call float @llvm.amdgcn.image.sample.c.lz.2darray.f32.f32(i32 1, float 0.0, float 1.000000e+00, float 0.0, float 0.0, <8 x i32> zeroinitializer, <4 x i32> zeroinitializer, i1 false, i32 0, i32 0)
+  %i5 = fadd float %i4, %i3
+  %i6 = call float @llvm.amdgcn.image.sample.c.lz.2darray.f32.f32(i32 1, float 0.0, float 0.0, float 0.0, float 0.0, <8 x i32> zeroinitializer, <4 x i32> zeroinitializer, i1 false, i32 0, i32 0)
+  %i7 = fadd float %i6, %i5
+  %i8 = fmul float %i7, 2.500000e-01
+  %i9 = insertelement <3 x float> zeroinitializer, float %i8, i64 0
+  call void @llvm.amdgcn.image.store.2d.v3f32.i32(<3 x float> %i9, i32 0, i32 0, i32 0, <8 x i32> zeroinitializer, i32 0, i32 0)
+  ret void
+}

Copy link
Collaborator

@rampitec rampitec left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I assume it may have terrible compile time consequences in extreme cases. But correctness first.

@arsenm arsenm merged commit b263033 into llvm:main Jun 1, 2024
9 checks passed
@arsenm arsenm deleted the issue92561 branch June 1, 2024 12:19
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

SIFixSGPRCopies inserts waterfall loop with undefined implicit use of scc
3 participants