Skip to content

[SimplifyCFG] Allow hoisting in the presence of pseudoprobes#199753

Open
mustartt wants to merge 2 commits into
llvm:mainfrom
mustartt:pseudo-probe-fccmp-chain
Open

[SimplifyCFG] Allow hoisting in the presence of pseudoprobes#199753
mustartt wants to merge 2 commits into
llvm:mainfrom
mustartt:pseudo-probe-fccmp-chain

Conversation

@mustartt
Copy link
Copy Markdown
Member

@mustartt mustartt commented May 26, 2026

Fix regressions in the presence of pseudoprobes that prevents SimplifyCFG from hoisting instructions into the predecessor. Teach hoistCommonCodeFromSuccessors and foldBranchToCommonDest to ignore pseudo probes and drop them when the BB is eliminated.

The minor loss of profile quality for these cases are justified, as not performing these hoists degrades performance more and blocks downstream passes like loop-vectorize (can be upto 30% in 526.blender_r and 525.x264_r).

@mustartt mustartt requested review from ahatanak, htyu and wlei-llvm May 26, 2026 20:21
@mustartt mustartt marked this pull request as ready for review May 26, 2026 20:24
@llvmorg-github-actions
Copy link
Copy Markdown

@llvm/pr-subscribers-llvm-transforms

Author: Henry Jiang (mustartt)

Changes

Fix regressions in the presence of pseudoprobes that prevents SimplifyCFG from hoisting instructions into the predecessor. Teach hoistCommonCodeFromSuccessors and foldBranchToCommonDest to ignore pseudo probes and drop them when the BB is eliminated.

The minor loss of profile quality for these cases are justified, as not performing these hoists degrades performance more and blocks downstream passes like loop-vectorize (can be upto 30% in 526.blender_r and 525.x264_r).


Full diff: https://github.com/llvm/llvm-project/pull/199753.diff

3 Files Affected:

  • (modified) llvm/lib/Transforms/Utils/SimplifyCFG.cpp (+11)
  • (added) llvm/test/Transforms/SimplifyCFG/fold-branch-to-common-dest-pseudoprobe.ll (+102)
  • (added) llvm/test/Transforms/SimplifyCFG/hoist-common-skip-pseudoprobe.ll (+103)
diff --git a/llvm/lib/Transforms/Utils/SimplifyCFG.cpp b/llvm/lib/Transforms/Utils/SimplifyCFG.cpp
index f1d47abe79365..99954ee13a740 100644
--- a/llvm/lib/Transforms/Utils/SimplifyCFG.cpp
+++ b/llvm/lib/Transforms/Utils/SimplifyCFG.cpp
@@ -1169,6 +1169,11 @@ static void cloneInstructionsIntoPredecessorBlockAndUpdateSSAUses(
     if (BonusInst.isTerminator())
       continue;
 
+    // Skip cloning pseudo probes into the predecessor, as it would overcount
+    // otherwise.
+    if (isa<PseudoProbeInst>(BonusInst))
+      continue;
+
     Instruction *NewBonusInst = BonusInst.clone();
 
     if (!NewBonusInst->getDebugLoc().isSameSourceLocation(PTI->getDebugLoc())) {
@@ -1519,6 +1524,9 @@ enum SkipFlags {
 };
 
 static unsigned skippedInstrFlags(Instruction *I) {
+  // Pseudo probes don't constrain reordering of other instructions.
+  if (isa<PseudoProbeInst>(I))
+    return 0;
   unsigned Flags = 0;
   if (I->mayReadFromMemory())
     Flags |= SkipReadMem;
@@ -4147,6 +4155,9 @@ bool llvm::foldBranchToCommonDest(CondBrInst *BI, DomTreeUpdater *DTU,
     // Ignore the terminator.
     if (isa<UncondBrInst, CondBrInst>(I))
       continue;
+    // Pseudo probes aren't speculatable but can be dropped on fold.
+    if (isa<PseudoProbeInst>(I))
+      continue;
     // I must be safe to execute unconditionally.
     if (!isSafeToSpeculativelyExecute(&I))
       return false;
diff --git a/llvm/test/Transforms/SimplifyCFG/fold-branch-to-common-dest-pseudoprobe.ll b/llvm/test/Transforms/SimplifyCFG/fold-branch-to-common-dest-pseudoprobe.ll
new file mode 100644
index 0000000000000..c3fafd5806823
--- /dev/null
+++ b/llvm/test/Transforms/SimplifyCFG/fold-branch-to-common-dest-pseudoprobe.ll
@@ -0,0 +1,102 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -passes=simplifycfg %s | FileCheck %s
+
+declare void @llvm.pseudoprobe(i64, i64, i32, i64)
+
+; Check that probe 2 does not prevent folding and is dropped
+define i32 @chained_cmp(float %v, ptr %out) {
+; CHECK-LABEL: @chained_cmp(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[CMP1:%.*]] = fcmp olt float [[V:%.*]], f0x35B28000
+; CHECK-NEXT:    call void @llvm.pseudoprobe(i64 -503749374543619030, i64 1, i32 0, i64 -1)
+; CHECK-NEXT:    [[CMP2:%.*]] = fcmp ogt float [[V]], f0xBF800001
+; CHECK-NEXT:    [[OR_COND:%.*]] = and i1 [[CMP1]], [[CMP2]]
+; CHECK-NEXT:    br i1 [[OR_COND]], label [[IF_THEN:%.*]], label [[COMMON_RET:%.*]]
+; CHECK:       common.ret:
+; CHECK-NEXT:    [[COMMON_RET_OP:%.*]] = phi i32 [ 1, [[IF_THEN]] ], [ 0, [[ENTRY:%.*]] ]
+; CHECK-NEXT:    ret i32 [[COMMON_RET_OP]]
+; CHECK:       if.then:
+; CHECK-NEXT:    store float [[V]], ptr [[OUT:%.*]], align 4
+; CHECK-NEXT:    br label [[COMMON_RET]]
+;
+entry:
+  %cmp1 = fcmp olt float %v, 0x3EB6500000000000
+  call void @llvm.pseudoprobe(i64 -503749374543619030, i64 1, i32 0, i64 -1)
+  br i1 %cmp1, label %land.lhs.true, label %ret_zero
+
+land.lhs.true:
+  %cmp2 = fcmp ogt float %v, 0xBFF0000020000000
+  call void @llvm.pseudoprobe(i64 -503749374543619030, i64 2, i32 0, i64 -1)
+  br i1 %cmp2, label %if.then, label %ret_zero
+
+if.then:
+  store float %v, ptr %out, align 4
+  ret i32 1
+
+ret_zero:
+  ret i32 0
+}
+
+; Probe should not enable folding when another bonus instruction is not
+; speculatable like udiv
+define i32 @no_fold_with_unsafe_bonus(i32 %x, i32 %y, i32 %d) {
+; CHECK-LABEL: @no_fold_with_unsafe_bonus(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[C0:%.*]] = icmp eq i32 [[X:%.*]], 0
+; CHECK-NEXT:    br i1 [[C0]], label [[BB:%.*]], label [[COMMON:%.*]]
+; CHECK:       bb:
+; CHECK-NEXT:    call void @llvm.pseudoprobe(i64 100, i64 1, i32 0, i64 -1)
+; CHECK-NEXT:    [[DIV:%.*]] = udiv i32 [[Y:%.*]], [[D:%.*]]
+; CHECK-NEXT:    [[C1:%.*]] = icmp eq i32 [[DIV]], 0
+; CHECK-NEXT:    br i1 [[C1]], label [[COMMON_RET:%.*]], label [[COMMON]]
+; CHECK:       common.ret:
+; CHECK-NEXT:    [[COMMON_RET_OP:%.*]] = phi i32 [ 0, [[COMMON]] ], [ 1, [[BB]] ]
+; CHECK-NEXT:    ret i32 [[COMMON_RET_OP]]
+; CHECK:       common:
+; CHECK-NEXT:    br label [[COMMON_RET]]
+;
+entry:
+  %c0 = icmp eq i32 %x, 0
+  br i1 %c0, label %bb, label %common
+
+bb:
+  call void @llvm.pseudoprobe(i64 100, i64 1, i32 0, i64 -1)
+  %div = udiv i32 %y, %d
+  %c1 = icmp eq i32 %div, 0
+  br i1 %c1, label %if.then, label %common
+
+if.then:
+  ret i32 1
+
+common:
+  ret i32 0
+}
+
+; Multiple probes in BB are all dropped on fold and predecessor probe is preserved.
+define i32 @multiple_probes_in_bb(i32 %x, i32 %y) {
+; CHECK-LABEL: @multiple_probes_in_bb(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    call void @llvm.pseudoprobe(i64 200, i64 1, i32 0, i64 -1)
+; CHECK-NEXT:    [[C0:%.*]] = icmp eq i32 [[X:%.*]], 0
+; CHECK-NEXT:    [[C1:%.*]] = icmp eq i32 [[Y:%.*]], 0
+; CHECK-NEXT:    [[OR_COND:%.*]] = select i1 [[C0]], i1 [[C1]], i1 false
+; CHECK-NEXT:    [[COMMON_RET_OP:%.*]] = select i1 [[OR_COND]], i32 1, i32 0
+; CHECK-NEXT:    ret i32 [[COMMON_RET_OP]]
+;
+entry:
+  call void @llvm.pseudoprobe(i64 200, i64 1, i32 0, i64 -1)
+  %c0 = icmp eq i32 %x, 0
+  br i1 %c0, label %bb, label %common
+
+bb:
+  call void @llvm.pseudoprobe(i64 200, i64 2, i32 0, i64 -1)
+  %c1 = icmp eq i32 %y, 0
+  call void @llvm.pseudoprobe(i64 200, i64 3, i32 0, i64 -1)
+  br i1 %c1, label %if.then, label %common
+
+if.then:
+  ret i32 1
+
+common:
+  ret i32 0
+}
diff --git a/llvm/test/Transforms/SimplifyCFG/hoist-common-skip-pseudoprobe.ll b/llvm/test/Transforms/SimplifyCFG/hoist-common-skip-pseudoprobe.ll
new file mode 100644
index 0000000000000..9b5e8798072c2
--- /dev/null
+++ b/llvm/test/Transforms/SimplifyCFG/hoist-common-skip-pseudoprobe.ll
@@ -0,0 +1,103 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S --passes='simplifycfg<hoist-common-insts>' %s | FileCheck %s
+
+declare void @llvm.pseudoprobe(i64, i64, i32, i64)
+
+; Check that instructions past the pseudoprobe instrs are still hoisted into the predecessor.
+define void @hoist_loads_past_pseudoprobe(i1 %c, ptr %d, ptr %m, ptr %b, i32 %v) {
+; CHECK-LABEL: @hoist_loads_past_pseudoprobe(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP1:%.*]] = load i16, ptr [[M:%.*]], align 2
+; CHECK-NEXT:    [[CONV0:%.*]] = zext i16 [[TMP1]] to i32
+; CHECK-NEXT:    [[TMP3:%.*]] = load i16, ptr [[M1:%.*]], align 2
+; CHECK-NEXT:    [[CONV1:%.*]] = zext i16 [[TMP3]] to i32
+; CHECK-NEXT:    br i1 [[C:%.*]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK:       if.then:
+; CHECK-NEXT:    call void @llvm.pseudoprobe(i64 1, i64 1, i32 0, i64 -1)
+; CHECK-NEXT:    [[ADD:%.*]] = add i32 [[CONV0]], [[V:%.*]]
+; CHECK-NEXT:    [[MUL:%.*]] = mul i32 [[ADD]], [[CONV1]]
+; CHECK-NEXT:    [[SHR:%.*]] = lshr i32 [[MUL]], 16
+; CHECK-NEXT:    [[CONV12:%.*]] = trunc i32 [[SHR]] to i16
+; CHECK-NEXT:    br label [[IF_END:%.*]]
+; CHECK:       if.else:
+; CHECK-NEXT:    call void @llvm.pseudoprobe(i64 1, i64 2, i32 0, i64 -1)
+; CHECK-NEXT:    [[SUB:%.*]] = sub i32 [[CONV0]], [[V]]
+; CHECK-NEXT:    [[MUL24:%.*]] = mul i32 [[SUB]], [[CONV1]]
+; CHECK-NEXT:    [[SHR25:%.*]] = lshr i32 [[MUL24]], 16
+; CHECK-NEXT:    [[TMP2:%.*]] = trunc i32 [[SHR25]] to i16
+; CHECK-NEXT:    [[CONV27:%.*]] = sub i16 0, [[TMP2]]
+; CHECK-NEXT:    br label [[IF_END]]
+; CHECK:       if.end:
+; CHECK-NEXT:    [[STOREMERGE:%.*]] = phi i16 [ [[CONV27]], [[IF_ELSE]] ], [ [[CONV12]], [[IF_THEN]] ]
+; CHECK-NEXT:    store i16 [[STOREMERGE]], ptr [[D:%.*]], align 2
+; CHECK-NEXT:    ret void
+;
+entry:
+  br i1 %c, label %if.then, label %if.else
+
+if.then:
+  call void @llvm.pseudoprobe(i64 1, i64 1, i32 0, i64 -1)
+  %0 = load i16, ptr %b, align 2
+  %conv0a = zext i16 %0 to i32
+  %add = add i32 %conv0a, %v
+  %1 = load i16, ptr %m, align 2
+  %conv1a = zext i16 %1 to i32
+  %mul = mul i32 %add, %conv1a
+  %shr = lshr i32 %mul, 16
+  %conv12 = trunc i32 %shr to i16
+  br label %if.end
+
+if.else:
+  call void @llvm.pseudoprobe(i64 1, i64 2, i32 0, i64 -1)
+  %2 = load i16, ptr %b, align 2
+  %conv0b = zext i16 %2 to i32
+  %sub = sub i32 %conv0b, %v
+  %3 = load i16, ptr %m, align 2
+  %conv1b = zext i16 %3 to i32
+  %mul24 = mul i32 %sub, %conv1b
+  %shr25 = lshr i32 %mul24, 16
+  %4 = trunc i32 %shr25 to i16
+  %conv27 = sub i16 0, %4
+  br label %if.end
+
+if.end:
+  %storemerge = phi i16 [ %conv27, %if.else ], [ %conv12, %if.then ]
+  store i16 %storemerge, ptr %d, align 2
+  ret void
+}
+
+; Stores past pseudo probes are also hoisted into the predecessor.
+define void @hoist_stores_past_pseudoprobe(i1 %c, ptr %d, ptr %d2, i32 %v) {
+; CHECK-LABEL: @hoist_stores_past_pseudoprobe(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    store i32 [[V:%.*]], ptr [[D:%.*]], align 4
+; CHECK-NEXT:    br i1 [[C:%.*]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]
+; CHECK:       if.then:
+; CHECK-NEXT:    call void @llvm.pseudoprobe(i64 2, i64 1, i32 0, i64 -1)
+; CHECK-NEXT:    store i32 0, ptr [[D2:%.*]], align 4
+; CHECK-NEXT:    br label [[IF_END:%.*]]
+; CHECK:       if.else:
+; CHECK-NEXT:    call void @llvm.pseudoprobe(i64 2, i64 2, i32 0, i64 -1)
+; CHECK-NEXT:    store i32 1, ptr [[D2]], align 4
+; CHECK-NEXT:    br label [[IF_END]]
+; CHECK:       if.end:
+; CHECK-NEXT:    ret void
+;
+entry:
+  br i1 %c, label %if.then, label %if.else
+
+if.then:
+  call void @llvm.pseudoprobe(i64 2, i64 1, i32 0, i64 -1)
+  store i32 %v, ptr %d, align 4
+  store i32 0, ptr %d2, align 4
+  br label %if.end
+
+if.else:
+  call void @llvm.pseudoprobe(i64 2, i64 2, i32 0, i64 -1)
+  store i32 %v, ptr %d, align 4
+  store i32 1, ptr %d2, align 4
+  br label %if.end
+
+if.end:
+  ret void
+}

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.

1 participant