[AMDGPU] Handle uniform control flow inside divergent - #217977
Open
jayfoad wants to merge 2 commits into
Open
Conversation
Fix PhiIncomingAnalysis::analyze. When it detected an incoming block that started divergent control flow ending at DefBlock, it tried to add all blocks in that divergent region to ReachableMap. But the propagation could be stopped by other blocks (those that did not start divergent control flow) that were already in ReachableMap. Fix this by doing the propagation first, before adding the remaining incoming blocks.
This was referenced Aug 21, 2026
|
@llvm/pr-subscribers-backend-amdgpu Author: Jay Foad (jayfoad) ChangesFix PhiIncomingAnalysis::analyze. When it detected an incoming block Full diff: https://github.com/llvm/llvm-project/pull/217977.diff 2 Files Affected:
diff --git a/llvm/lib/Target/AMDGPU/SILowerI1Copies.cpp b/llvm/lib/Target/AMDGPU/SILowerI1Copies.cpp
index 62b3638812ab9..c49270e34bf55 100644
--- a/llvm/lib/Target/AMDGPU/SILowerI1Copies.cpp
+++ b/llvm/lib/Target/AMDGPU/SILowerI1Copies.cpp
@@ -147,12 +147,10 @@ class PhiIncomingAnalysis {
continue;
}
- ReachableMap.try_emplace(MBB, false);
-
// If this block has a divergent terminator and the def block is its
// post-dominator, the wave may first visit the other successors.
if (TII->hasDivergentBranch(MBB) && PDT.dominates(&DefBlock, MBB))
- append_range(Stack, MBB->successors());
+ Stack.push_back(MBB);
}
while (!Stack.empty()) {
@@ -161,6 +159,12 @@ class PhiIncomingAnalysis {
append_range(Stack, MBB->successors());
}
+ // Insert remaining incoming blocks.
+ for (auto Incoming : Incomings) {
+ MachineBasicBlock *MBB = Incoming.Block;
+ ReachableMap.try_emplace(MBB, false);
+ }
+
for (auto &[MBB, IsSource] : ReachableMap) {
bool HaveReachablePred = false;
for (MachineBasicBlock *Pred : MBB->predecessors()) {
diff --git a/llvm/test/CodeGen/AMDGPU/uniform-inside-divergent-cfg.ll b/llvm/test/CodeGen/AMDGPU/uniform-inside-divergent-cfg.ll
new file mode 100644
index 0000000000000..a2e272d9127e4
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/uniform-inside-divergent-cfg.ll
@@ -0,0 +1,163 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
+; RUN: llc -mtriple=amdgpu11.00 -structurizecfg-skip-uniform-regions < %s | FileCheck %s
+
+define amdgpu_kernel void @div_unif_div(ptr addrspace(1) %out, float %ubeta, i32 %m) {
+; CHECK-LABEL: div_unif_div:
+; CHECK: ; %bb.0: ; %entry
+; CHECK-NEXT: s_load_b64 s[0:1], s[4:5], 0x2c
+; CHECK-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_and_b32 v0, 0x3ff, v0
+; CHECK-NEXT: s_waitcnt lgkmcnt(0)
+; CHECK-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; CHECK-NEXT: v_cmp_gt_u32_e32 vcc_lo, s1, v0
+; CHECK-NEXT: s_mov_b32 s1, 0
+; CHECK-NEXT: s_and_saveexec_b32 s2, vcc_lo
+; CHECK-NEXT: s_cbranch_execz .LBB0_8
+; CHECK-NEXT: ; %bb.1: ; %A
+; CHECK-NEXT: v_cmp_eq_f32_e64 s0, s0, 0
+; CHECK-NEXT: s_mov_b32 s1, exec_lo
+; CHECK-NEXT: s_and_b32 vcc_lo, exec_lo, s0
+; CHECK-NEXT: s_cbranch_vccnz .LBB0_7
+; CHECK-NEXT: ; %bb.2: ; %B
+; CHECK-NEXT: s_mov_b32 s0, exec_lo
+; CHECK-NEXT: ; implicit-def: $vgpr1
+; CHECK-NEXT: v_cmpx_lt_u32_e32 2, v0
+; CHECK-NEXT: s_xor_b32 s0, exec_lo, s0
+; CHECK-NEXT: ; %bb.3: ; %B2
+; CHECK-NEXT: v_mul_u32_u24_e32 v1, 3, v0
+; CHECK-NEXT: ; %bb.4: ; %Flow
+; CHECK-NEXT: s_and_not1_saveexec_b32 s0, s0
+; CHECK-NEXT: ; %bb.5: ; %B1
+; CHECK-NEXT: v_add_nc_u32_e32 v1, 7, v0
+; CHECK-NEXT: ; %bb.6: ; %Flow3
+; CHECK-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; CHECK-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; CHECK-NEXT: s_or_b32 s1, s1, exec_lo
+; CHECK-NEXT: s_branch .LBB0_8
+; CHECK-NEXT: .LBB0_7:
+; CHECK-NEXT: v_mov_b32_e32 v1, 1
+; CHECK-NEXT: .LBB0_8: ; %join
+; CHECK-NEXT: s_or_b32 exec_lo, exec_lo, s2
+; CHECK-NEXT: s_and_saveexec_b32 s0, s1
+; CHECK-NEXT: s_cbranch_execz .LBB0_10
+; CHECK-NEXT: ; %bb.9: ; %do.store
+; CHECK-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; CHECK-NEXT: v_lshlrev_b32_e32 v0, 2, v0
+; CHECK-NEXT: s_waitcnt lgkmcnt(0)
+; CHECK-NEXT: global_store_b32 v0, v1, s[0:1]
+; CHECK-NEXT: .LBB0_10: ; %exit
+; CHECK-NEXT: s_endpgm
+entry:
+ %tid = call i32 @llvm.amdgcn.workitem.id.x()
+ %div = icmp ult i32 %tid, %m ; DIVERGENT -> becomes SI_IF
+ br i1 %div, label %A, label %join
+
+A: ; only lanes with tid < m
+ %unif = fcmp une float %ubeta, 0.0 ; UNIFORM -> region gets skipped
+ br i1 %unif, label %B, label %join
+
+B:
+ %div2 = icmp ult i32 %tid, 3 ; divergent inner if/else
+ br i1 %div2, label %B1, label %B2
+B1:
+ %v1 = add i32 %tid, 7
+ br label %C
+B2:
+ %v2 = mul i32 %tid, 3
+ br label %C
+C:
+ %v = phi i32 [ %v1, %B1 ], [ %v2, %B2 ]
+ br label %join
+
+join:
+ %val = phi i32 [ 0, %entry ], [ 1, %A ], [ %v, %C ]
+ %pred = phi i1 [ false, %entry ], [ true, %A ], [ true, %C ]
+ br i1 %pred, label %do.store, label %exit
+
+do.store:
+ %gep = getelementptr i32, ptr addrspace(1) %out, i32 %tid
+ store i32 %val, ptr addrspace(1) %gep
+ br label %exit
+
+exit:
+ ret void
+}
+
+define amdgpu_kernel void @unif_div(ptr addrspace(1) %out, float %u1, float %u2, i32 %m) {
+; CHECK-LABEL: unif_div:
+; CHECK: ; %bb.0: ; %entry
+; CHECK-NEXT: s_load_b128 s[0:3], s[4:5], 0x2c
+; CHECK-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_and_b32 v0, 0x3ff, v0
+; CHECK-NEXT: s_waitcnt lgkmcnt(0)
+; CHECK-NEXT: s_mov_b32 s3, 0
+; CHECK-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; CHECK-NEXT: v_cmp_gt_u32_e32 vcc_lo, s2, v0
+; CHECK-NEXT: s_and_saveexec_b32 s2, vcc_lo
+; CHECK-NEXT: s_cbranch_execz .LBB1_7
+; CHECK-NEXT: ; %bb.1: ; %A
+; CHECK-NEXT: v_cmp_eq_f32_e64 s0, s0, 0
+; CHECK-NEXT: s_and_b32 vcc_lo, exec_lo, s0
+; CHECK-NEXT: s_cbranch_vccnz .LBB1_4
+; CHECK-NEXT: ; %bb.2: ; %B
+; CHECK-NEXT: v_cmp_eq_f32_e64 s0, s1, 0
+; CHECK-NEXT: s_and_b32 vcc_lo, exec_lo, s0
+; CHECK-NEXT: s_cbranch_vccnz .LBB1_5
+; CHECK-NEXT: ; %bb.3: ; %C
+; CHECK-NEXT: v_add_nc_u32_e32 v1, 7, v0
+; CHECK-NEXT: s_branch .LBB1_6
+; CHECK-NEXT: .LBB1_4:
+; CHECK-NEXT: s_mov_b32 s3, exec_lo
+; CHECK-NEXT: v_mov_b32_e32 v1, 1
+; CHECK-NEXT: s_branch .LBB1_7
+; CHECK-NEXT: .LBB1_5: ; %D
+; CHECK-NEXT: v_mul_u32_u24_e32 v1, 3, v0
+; CHECK-NEXT: .LBB1_6: ; %join
+; CHECK-NEXT: s_or_b32 s3, exec_lo, exec_lo
+; CHECK-NEXT: .LBB1_7: ; %join
+; CHECK-NEXT: s_or_b32 exec_lo, exec_lo, s2
+; CHECK-NEXT: s_and_saveexec_b32 s0, s3
+; CHECK-NEXT: s_cbranch_execz .LBB1_9
+; CHECK-NEXT: ; %bb.8: ; %do.store
+; CHECK-NEXT: s_load_b64 s[0:1], s[4:5], 0x24
+; CHECK-NEXT: v_lshlrev_b32_e32 v0, 2, v0
+; CHECK-NEXT: s_waitcnt lgkmcnt(0)
+; CHECK-NEXT: global_store_b32 v0, v1, s[0:1]
+; CHECK-NEXT: .LBB1_9: ; %exit
+; CHECK-NEXT: s_endpgm
+entry:
+ %tid = call i32 @llvm.amdgcn.workitem.id.x()
+ %div = icmp ult i32 %tid, %m ; DIVERGENT -> SI_IF
+ br i1 %div, label %A, label %join
+
+A: ; only lanes with tid < m
+ %c1 = fcmp une float %u1, 0.0 ; uniform
+ br i1 %c1, label %B, label %join
+
+B:
+ %c2 = fcmp une float %u2, 0.0 ; uniform
+ br i1 %c2, label %C, label %D
+
+C:
+ %vc = add i32 %tid, 7
+ br label %E
+
+D:
+ %vd = mul i32 %tid, 3
+ br label %E
+
+E: ; not a direct successor of A
+ %v = phi i32 [ %vc, %C ], [ %vd, %D ]
+ br label %join
+
+join:
+ %val = phi i32 [ 0, %entry ], [ 1, %A ], [ %v, %E ]
+ %pred = phi i1 [ false, %entry ], [ true, %A ], [ true, %E ]
+ br i1 %pred, label %do.store, label %exit
+
+do.store:
+ %gep = getelementptr i32, ptr addrspace(1) %out, i32 %tid
+ store i32 %val, ptr addrspace(1) %gep
+ br label %exit
+
+exit:
+ ret void
+}
|
arsenm
approved these changes
Aug 21, 2026
🪟 Windows x64 Test Results
Failed Tests(click on a test name to see its output) ClangClang.Modules/GH204633.cppmIf these failures are unrelated to your changes (for example tests are broken or flaky at HEAD), please open an issue at https://github.com/llvm/llvm-project/issues and add the |
🐧 Linux x64 Test Results
Failed Tests(click on a test name to see its output) ClangClang.Modules/GH204633.cppmIf these failures are unrelated to your changes (for example tests are broken or flaky at HEAD), please open an issue at https://github.com/llvm/llvm-project/issues and add the |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fix PhiIncomingAnalysis::analyze. When it detected an incoming block
that started divergent control flow ending at DefBlock, it tried to add
all blocks in that divergent region to ReachableMap. But the propagation
could be stopped by other blocks (those that did not start divergent
control flow) that were already in ReachableMap. Fix this by doing the
propagation first, before adding the remaining incoming blocks.