[llvm] [AMDGPU] Handle uniform control flow inside divergent (PR #217977)
Jay Foad via llvm-commits
llvm-commits at lists.llvm.org
Fri Aug 21 09:54:43 PDT 2026
https://github.com/jayfoad created https://github.com/llvm/llvm-project/pull/217977
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.
>From 4d7aaf9cec8aa8e22a1dc72cf98f239d7086b495 Mon Sep 17 00:00:00 2001
From: Jay Foad <jay.foad at amd.com>
Date: Fri, 21 Aug 2026 17:35:50 +0100
Subject: [PATCH 1/2] Precommit test
---
.../AMDGPU/uniform-inside-divergent-cfg.ll | 162 ++++++++++++++++++
1 file changed, 162 insertions(+)
create mode 100644 llvm/test/CodeGen/AMDGPU/uniform-inside-divergent-cfg.ll
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..0ae8199e16bf5
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/uniform-inside-divergent-cfg.ll
@@ -0,0 +1,162 @@
+; 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_mov_b32 s2, 0
+; 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_and_saveexec_b32 s1, 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_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_mov_b32 s2, -1
+; CHECK-NEXT: s_branch .LBB0_8
+; CHECK-NEXT: .LBB0_7:
+; CHECK-NEXT: s_mov_b32 s2, exec_lo
+; CHECK-NEXT: v_mov_b32_e32 v1, 1
+; CHECK-NEXT: .LBB0_8: ; %join
+; CHECK-NEXT: s_or_b32 exec_lo, exec_lo, s1
+; CHECK-NEXT: s_and_saveexec_b32 s0, s2
+; 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_mov_b32 s3, -1
+; 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
+}
>From e622b1d2645f6d1e57dd2fcce44e81441f0db303 Mon Sep 17 00:00:00 2001
From: Jay Foad <jay.foad at amd.com>
Date: Fri, 21 Aug 2026 17:37:45 +0100
Subject: [PATCH 2/2] [AMDGPU] Handle uniform control flow inside divergent
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.
---
llvm/lib/Target/AMDGPU/SILowerI1Copies.cpp | 10 +++++++---
.../AMDGPU/uniform-inside-divergent-cfg.ll | 15 ++++++++-------
2 files changed, 15 insertions(+), 10 deletions(-)
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
index 0ae8199e16bf5..a2e272d9127e4 100644
--- a/llvm/test/CodeGen/AMDGPU/uniform-inside-divergent-cfg.ll
+++ b/llvm/test/CodeGen/AMDGPU/uniform-inside-divergent-cfg.ll
@@ -6,14 +6,15 @@ define amdgpu_kernel void @div_unif_div(ptr addrspace(1) %out, float %ubeta, i32
; 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_mov_b32 s2, 0
; 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_and_saveexec_b32 s1, vcc_lo
+; 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
@@ -29,14 +30,14 @@ define amdgpu_kernel void @div_unif_div(ptr addrspace(1) %out, float %ubeta, i32
; 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_mov_b32 s2, -1
+; 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: s_mov_b32 s2, exec_lo
; CHECK-NEXT: v_mov_b32_e32 v1, 1
; CHECK-NEXT: .LBB0_8: ; %join
-; CHECK-NEXT: s_or_b32 exec_lo, exec_lo, s1
-; CHECK-NEXT: s_and_saveexec_b32 s0, s2
+; 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
@@ -110,7 +111,7 @@ define amdgpu_kernel void @unif_div(ptr addrspace(1) %out, float %u1, float %u2,
; CHECK-NEXT: .LBB1_5: ; %D
; CHECK-NEXT: v_mul_u32_u24_e32 v1, 3, v0
; CHECK-NEXT: .LBB1_6: ; %join
-; CHECK-NEXT: s_mov_b32 s3, -1
+; 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
More information about the llvm-commits
mailing list