[llvm] GlobalISel lane masks merging (PR #73337)
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Thu Dec 7 06:23:43 PST 2023
================
@@ -0,0 +1,338 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
+; RUN: llc -global-isel -amdgpu-global-isel-risky-select -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
+
+; Divergent phis that don't require lowering using lane mask merging
+
+; - divergent phi that has divergent incoming value (this makes it divergent)
+; but is reachable through only one path - branch instruction that chooses
+; path is uniform
+
+; - divergent phi that is used only inside the loop and has incoming from
+; previous iteration. After phi-elimination (rewrite lane mask in phi def with
+; lane mask value from previous iteration), phi will hold lane mask valid for
+; current iteration which is fine since it is not used outside of the loop.
+
+; And one more that is tricky (is branch divergent or not ?)
+; "amdgpu-flat-work-group-size"="1,1" aka single lane execution does not stop
+; shader from activating multiple lanes by using some intrinsic (entering wwm
+; and using dpp instructions)
+; - there are cases with single lane execution where branch instructions are not
+; lowered to si_if (or other intrinsic branches) - with intention to use
+; uniform branch after instruction selection?
+; PhiIncomingAnalysis does not recognize G_BRCOND as divergent branch and does
+; not perform lane mask merging
+
+
+
+define amdgpu_ps void @divergent_i1_phi_uniform_branch(i32 addrspace(1)* %out, i32 %tid, i32 inreg %cond, i32 addrspace(1)* %dummyaddr) {
+; GFX10-LABEL: divergent_i1_phi_uniform_branch:
+; GFX10: ; %bb.0: ; %A
+; GFX10-NEXT: s_cmp_lg_u32 s0, 0
+; GFX10-NEXT: s_cbranch_scc0 .LBB0_2
+; GFX10-NEXT: ; %bb.1:
+; GFX10-NEXT: v_cmp_le_u32_e64 s0, 6, v2
+; GFX10-NEXT: s_branch .LBB0_3
+; GFX10-NEXT: .LBB0_2: ; %dummy
+; GFX10-NEXT: v_mov_b32_e32 v5, 0x7b
+; GFX10-NEXT: v_cmp_gt_u32_e64 s0, 1, v2
+; GFX10-NEXT: global_store_dword v[3:4], v5, off
+; GFX10-NEXT: .LBB0_3: ; %exit
+; GFX10-NEXT: v_cndmask_b32_e64 v2, 2, 1, s0
+; GFX10-NEXT: global_store_dword v[0:1], v2, off
+; GFX10-NEXT: s_endpgm
+A:
+ %val_A = icmp uge i32 %tid, 6
+ %cmp = icmp eq i32 %cond, 0
+ br i1 %cmp, label %dummy, label %exit
+
+dummy:
+ store i32 123, i32 addrspace(1)* %dummyaddr
+ br label %B
+
+B:
+ %val_B = icmp ult i32 %tid, 1
+ br label %exit
+
+exit:
+ %phi = phi i1 [ %val_A, %A ], [ %val_B, %B ]
+ %sel = select i1 %phi, i32 1, i32 2
+ store i32 %sel, i32 addrspace(1)* %out
+ ret void
+}
+
+; Fix me - there is no need to merge lane masks here
+define amdgpu_ps void @divergent_i1_phi_uniform_branch_simple(i32 addrspace(1)* %out, i32 %tid, i32 inreg %cond) {
+; GFX10-LABEL: divergent_i1_phi_uniform_branch_simple:
+; GFX10: ; %bb.0: ; %A
+; GFX10-NEXT: s_cmp_lg_u32 s0, 0
+; GFX10-NEXT: s_cbranch_scc0 .LBB1_2
+; GFX10-NEXT: ; %bb.1:
+; GFX10-NEXT: v_cmp_le_u32_e64 s0, 6, v2
+; GFX10-NEXT: s_branch .LBB1_3
+; GFX10-NEXT: .LBB1_2: ; %B
+; GFX10-NEXT: v_cmp_gt_u32_e64 s0, 1, v2
+; GFX10-NEXT: .LBB1_3: ; %exit
+; GFX10-NEXT: v_cndmask_b32_e64 v2, 2, 1, s0
+; GFX10-NEXT: global_store_dword v[0:1], v2, off
+; GFX10-NEXT: s_endpgm
+A:
+ %val_A = icmp uge i32 %tid, 6
+ %cmp = icmp eq i32 %cond, 0
+ br i1 %cmp, label %B, label %exit
+
+B:
+ %val_B = icmp ult i32 %tid, 1
+ br label %exit
+
+exit:
+ %phi = phi i1 [ %val_A, %A ], [ %val_B, %B ]
+ %sel = select i1 %phi, i32 1, i32 2
+ store i32 %sel, i32 addrspace(1)* %out
+ ret void
+}
+
+
+; Divergent i1 phi that uses value from previous iteration.
+; Used only inside the loop (variable name is bool_counter)
+define void @divergent_i1_phi_used_inside_loop(float %val, ptr %addr) {
+; GFX10-LABEL: divergent_i1_phi_used_inside_loop:
+; GFX10: ; %bb.0: ; %entry
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_mov_b32_e32 v3, 1
+; GFX10-NEXT: v_mov_b32_e32 v4, s4
+; GFX10-NEXT: .LBB2_1: ; %loop
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_cvt_f32_u32_e32 v5, v4
+; GFX10-NEXT: v_xor_b32_e32 v3, 1, v3
+; GFX10-NEXT: v_add_nc_u32_e32 v4, 1, v4
+; GFX10-NEXT: v_cmp_gt_f32_e32 vcc_lo, v5, v0
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execnz .LBB2_1
+; GFX10-NEXT: ; %bb.2: ; %exit
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_and_b32_e32 v0, 1, v3
+; GFX10-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1.0, vcc_lo
+; GFX10-NEXT: flat_store_dword v[1:2], v0
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+entry:
+ br label %loop
+
+loop:
+ %counter = phi i32 [ 0, %entry ], [ %counterPlus1, %loop ]
+ %bool_counter = phi i1 [ true, %entry ], [ %neg_bool_counter, %loop ]
+ %neg_bool_counter = xor i1 %bool_counter, true
+ %fcounter = uitofp i32 %counter to float
+ %cond = fcmp ogt float %fcounter, %val
+ %counterPlus1 = add i32 %counter, 1
+ br i1 %cond, label %exit, label %loop
+
+exit:
+ %select = select i1 %neg_bool_counter, float 1.000000e+00, float 0.000000e+00
+ store float %select, ptr %addr
+ ret void
+}
+
+define void @divergent_i1_phi_used_inside_loop_bigger_loop_body(float %val, float %pre_cond_val, ptr %addr, ptr %addr_if, ptr %addr_else) {
+; GFX10-LABEL: divergent_i1_phi_used_inside_loop_bigger_loop_body:
+; GFX10: ; %bb.0: ; %entry
+; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT: v_cmp_lt_f32_e32 vcc_lo, 1.0, v1
+; GFX10-NEXT: s_mov_b32 s4, 0
+; GFX10-NEXT: v_mov_b32_e32 v8, 0x3e8
+; GFX10-NEXT: v_mov_b32_e32 v9, s4
+; GFX10-NEXT: v_cndmask_b32_e64 v1, 0, 1, vcc_lo
+; GFX10-NEXT: s_branch .LBB3_2
+; GFX10-NEXT: .LBB3_1: ; %loop_body
+; GFX10-NEXT: ; in Loop: Header=BB3_2 Depth=1
+; GFX10-NEXT: v_cvt_f32_u32_e32 v10, v9
+; GFX10-NEXT: v_xor_b32_e32 v1, 1, v1
+; GFX10-NEXT: v_add_nc_u32_e32 v9, 1, v9
+; GFX10-NEXT: v_cmp_gt_f32_e32 vcc_lo, v10, v0
+; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4
+; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: s_cbranch_execz .LBB3_6
+; GFX10-NEXT: .LBB3_2: ; %loop_start
+; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1
+; GFX10-NEXT: v_cmp_ge_i32_e32 vcc_lo, 0x3e8, v9
+; GFX10-NEXT: s_mov_b32 s5, 1
+; GFX10-NEXT: s_cbranch_vccz .LBB3_4
+; GFX10-NEXT: ; %bb.3: ; %else
+; GFX10-NEXT: ; in Loop: Header=BB3_2 Depth=1
+; GFX10-NEXT: s_mov_b32 s5, 0
+; GFX10-NEXT: flat_store_dword v[6:7], v8
+; GFX10-NEXT: .LBB3_4: ; %Flow
+; GFX10-NEXT: ; in Loop: Header=BB3_2 Depth=1
+; GFX10-NEXT: s_xor_b32 s5, s5, 1
+; GFX10-NEXT: s_and_b32 s5, s5, 1
+; GFX10-NEXT: s_cmp_lg_u32 s5, 0
+; GFX10-NEXT: s_cbranch_scc1 .LBB3_1
+; GFX10-NEXT: ; %bb.5: ; %if
+; GFX10-NEXT: ; in Loop: Header=BB3_2 Depth=1
+; GFX10-NEXT: flat_store_dword v[4:5], v8
+; GFX10-NEXT: s_branch .LBB3_1
+; GFX10-NEXT: .LBB3_6: ; %exit
+; GFX10-NEXT: s_or_b32 exec_lo, exec_lo, s4
+; GFX10-NEXT: v_and_b32_e32 v0, 1, v1
+; GFX10-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1.0, vcc_lo
+; GFX10-NEXT: flat_store_dword v[2:3], v0
+; GFX10-NEXT: s_waitcnt lgkmcnt(0)
+; GFX10-NEXT: s_setpc_b64 s[30:31]
+entry:
+ %pre_cond = fcmp ogt float %pre_cond_val, 1.0
+ br label %loop_start
+
+loop_start:
+ %counter = phi i32 [ 0, %entry ], [ %counterPlus1, %loop_body ]
+ %bool_counter = phi i1 [ %pre_cond, %entry ], [ %neg_bool_counter, %loop_body ]
+ %cond_break = icmp sgt i32 %counter, 1000
+ br i1 %cond_break, label %if, label %else
+
+if:
+ store i32 1000, ptr %addr_if
+ br label %loop_body
+
+else:
+ store i32 1000, ptr %addr_else
+ br label %loop_body
+
+loop_body:
+ %neg_bool_counter = xor i1 %bool_counter, true
+ %fcounter = uitofp i32 %counter to float
+ %cond = fcmp ogt float %fcounter, %val
+ %counterPlus1 = add i32 %counter, 1
+ br i1 %cond, label %exit, label %loop_start
+
+exit:
+ %select = select i1 %neg_bool_counter, float 1.000000e+00, float 0.000000e+00
+ store float %select, ptr %addr
+ ret void
+}
+
+; There is a divergent, according to machine uniformity info, g_brcond branch
+; here, not lowered to si_if because of "amdgpu-flat-work-group-size"="1,1".
+define amdgpu_cs void @_amdgpu_cs_main( i32 inreg noundef %.userdata0, <3 x i32> inreg noundef %.WorkgroupId, <3 x i32> noundef %.LocalInvocationId) #0 {
----------------
arsenm wrote:
Name the function to indicate it's the 1,1 case
https://github.com/llvm/llvm-project/pull/73337
More information about the llvm-commits
mailing list