[llvm] fbec8f0 - [SimplifyCFG] Do not hoist/sink convergent function calls

Yaxun Liu via llvm-commits llvm-commits at lists.llvm.org
Tue Feb 28 09:42:24 PST 2023


Author: Yaxun (Sam) Liu
Date: 2023-02-28T12:41:56-05:00
New Revision: fbec8f04ab39211f01366d8d255e942d7b76354b

URL: https://github.com/llvm/llvm-project/commit/fbec8f04ab39211f01366d8d255e942d7b76354b
DIFF: https://github.com/llvm/llvm-project/commit/fbec8f04ab39211f01366d8d255e942d7b76354b.diff

LOG: [SimplifyCFG] Do not hoist/sink convergent function calls

Currently SimplifyCFG hoists/sink common instructions in then/else basic blocks
when certain options are enabled, which is the case for default clang optimization
pipelines for -O3. It tries to hoist/sink convergent function calls in divergent
control flow, which causes incorrect ISA generated for GPU, e.g.
https://github.com/ROCm-Developer-Tools/HIP/issues/3172

This patch fixes that by conservatively disable hoisting/sinking common
convergent function calls in then/else blocks.

Reviewed by: Artem Belevich

Differential Revision: https://reviews.llvm.org/D144756

Added: 
    

Modified: 
    llvm/lib/Transforms/Utils/SimplifyCFG.cpp
    llvm/test/Transforms/SimplifyCFG/convergent.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/Utils/SimplifyCFG.cpp b/llvm/lib/Transforms/Utils/SimplifyCFG.cpp
index 5f43f61f1abcb..747db1f5c63f5 100644
--- a/llvm/lib/Transforms/Utils/SimplifyCFG.cpp
+++ b/llvm/lib/Transforms/Utils/SimplifyCFG.cpp
@@ -1595,12 +1595,13 @@ bool SimplifyCFGOpt::HoistThenElseCodeToIf(BranchInst *BI,
       if (!TTI.isProfitableToHoist(I1) || !TTI.isProfitableToHoist(I2))
         return Changed;
 
-      // If any of the two call sites has nomerge attribute, stop hoisting.
+      // If any of the two call sites has nomerge or convergent attribute, stop
+      // hoisting.
       if (const auto *CB1 = dyn_cast<CallBase>(I1))
-        if (CB1->cannotMerge())
+        if (CB1->cannotMerge() || CB1->isConvergent())
           return Changed;
       if (const auto *CB2 = dyn_cast<CallBase>(I2))
-        if (CB2->cannotMerge())
+        if (CB2->cannotMerge() || CB2->isConvergent())
           return Changed;
 
       if (isa<DbgInfoIntrinsic>(I1) || isa<DbgInfoIntrinsic>(I2)) {
@@ -1808,9 +1809,9 @@ static bool canSinkInstructions(
     // Conservatively return false if I is an inline-asm instruction. Sinking
     // and merging inline-asm instructions can potentially create arguments
     // that cannot satisfy the inline-asm constraints.
-    // If the instruction has nomerge attribute, return false.
+    // If the instruction has nomerge or convergent attribute, return false.
     if (const auto *C = dyn_cast<CallBase>(I))
-      if (C->isInlineAsm() || C->cannotMerge())
+      if (C->isInlineAsm() || C->cannotMerge() || C->isConvergent())
         return false;
 
     // Each instruction must have zero or one use.

diff  --git a/llvm/test/Transforms/SimplifyCFG/convergent.ll b/llvm/test/Transforms/SimplifyCFG/convergent.ll
index 91e25b94d16b1..bbdba53524806 100644
--- a/llvm/test/Transforms/SimplifyCFG/convergent.ll
+++ b/llvm/test/Transforms/SimplifyCFG/convergent.ll
@@ -1,8 +1,12 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
-; RUN: opt -S -passes=simplifycfg -simplifycfg-require-and-preserve-domtree=1 < %s | FileCheck %s
-; RUN: opt -S -passes=simplifycfg < %s | FileCheck %s
+; RUN: opt -S -passes=simplifycfg -simplifycfg-require-and-preserve-domtree=1 < %s | FileCheck -check-prefixes=CHECK,NOSINK %s
+; RUN: opt -S -passes=simplifycfg < %s | FileCheck -check-prefixes=CHECK,NOSINK %s
+; RUN: opt -S -passes='simplifycfg<hoist-common-insts;sink-common-insts>' < %s | FileCheck -check-prefixes=CHECK,SINK %s
 
 declare void @foo() convergent
+declare i32 @tid()
+declare i32 @mbcnt(i32 %a, i32 %b) convergent
+declare i32 @bpermute(i32 %a, i32 %b) convergent
 
 define i32 @test_01(i32 %a) {
 ; CHECK-LABEL: @test_01(
@@ -40,3 +44,98 @@ if.false.2:
 exit:
   ret i32 %a
 }
+
+define void @test_02(ptr %y.coerce) convergent {
+; NOSINK-LABEL: @test_02(
+; NOSINK-NEXT:  entry:
+; NOSINK-NEXT:    [[TMP0:%.*]] = tail call i32 @tid()
+; NOSINK-NEXT:    [[REM:%.*]] = and i32 [[TMP0]], 1
+; NOSINK-NEXT:    [[CMP_NOT:%.*]] = icmp eq i32 [[REM]], 0
+; NOSINK-NEXT:    br i1 [[CMP_NOT]], label [[IF_ELSE:%.*]], label [[IF_THEN:%.*]]
+; NOSINK:       if.then:
+; NOSINK-NEXT:    [[TMP1:%.*]] = tail call i32 @mbcnt(i32 -1, i32 0)
+; NOSINK-NEXT:    [[TMP2:%.*]] = tail call i32 @mbcnt(i32 -1, i32 [[TMP1]])
+; NOSINK-NEXT:    [[AND_I:%.*]] = shl i32 [[TMP2]], 2
+; NOSINK-NEXT:    [[ADD_I:%.*]] = and i32 [[AND_I]], -256
+; NOSINK-NEXT:    [[SHL_I:%.*]] = or i32 [[ADD_I]], 16
+; NOSINK-NEXT:    [[TMP3:%.*]] = tail call i32 @bpermute(i32 [[SHL_I]], i32 [[TMP0]])
+; NOSINK-NEXT:    [[IDXPROM:%.*]] = zext i32 [[TMP0]] to i64
+; NOSINK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[Y_COERCE:%.*]], i64 [[IDXPROM]]
+; NOSINK-NEXT:    store i32 [[TMP3]], ptr [[ARRAYIDX]], align 4
+; NOSINK-NEXT:    br label [[IF_END:%.*]]
+; NOSINK:       if.else:
+; NOSINK-NEXT:    [[TMP4:%.*]] = tail call i32 @mbcnt(i32 -1, i32 0)
+; NOSINK-NEXT:    [[TMP5:%.*]] = tail call i32 @mbcnt(i32 -1, i32 [[TMP4]])
+; NOSINK-NEXT:    [[AND_I11:%.*]] = shl i32 [[TMP5]], 2
+; NOSINK-NEXT:    [[ADD_I12:%.*]] = and i32 [[AND_I11]], -256
+; NOSINK-NEXT:    [[SHL_I13:%.*]] = or i32 [[ADD_I12]], 8
+; NOSINK-NEXT:    [[TMP6:%.*]] = tail call i32 @bpermute(i32 [[SHL_I13]], i32 [[TMP0]])
+; NOSINK-NEXT:    [[IDXPROM4:%.*]] = zext i32 [[TMP0]] to i64
+; NOSINK-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i32, ptr [[Y_COERCE]], i64 [[IDXPROM4]]
+; NOSINK-NEXT:    store i32 [[TMP6]], ptr [[ARRAYIDX5]], align 4
+; NOSINK-NEXT:    br label [[IF_END]]
+; NOSINK:       if.end:
+; NOSINK-NEXT:    ret void
+;
+; SINK-LABEL: @test_02(
+; SINK-NEXT:  entry:
+; SINK-NEXT:    [[TMP0:%.*]] = tail call i32 @tid()
+; SINK-NEXT:    [[REM:%.*]] = and i32 [[TMP0]], 1
+; SINK-NEXT:    [[CMP_NOT:%.*]] = icmp eq i32 [[REM]], 0
+; SINK-NEXT:    br i1 [[CMP_NOT]], label [[IF_ELSE:%.*]], label [[IF_THEN:%.*]]
+; SINK:       if.then:
+; SINK-NEXT:    [[TMP1:%.*]] = tail call i32 @mbcnt(i32 -1, i32 0)
+; SINK-NEXT:    [[TMP2:%.*]] = tail call i32 @mbcnt(i32 -1, i32 [[TMP1]])
+; SINK-NEXT:    [[AND_I:%.*]] = shl i32 [[TMP2]], 2
+; SINK-NEXT:    [[ADD_I:%.*]] = and i32 [[AND_I]], -256
+; SINK-NEXT:    [[SHL_I:%.*]] = or i32 [[ADD_I]], 16
+; SINK-NEXT:    [[TMP3:%.*]] = tail call i32 @bpermute(i32 [[SHL_I]], i32 [[TMP0]])
+; SINK-NEXT:    br label [[IF_END:%.*]]
+; SINK:       if.else:
+; SINK-NEXT:    [[TMP4:%.*]] = tail call i32 @mbcnt(i32 -1, i32 0)
+; SINK-NEXT:    [[TMP5:%.*]] = tail call i32 @mbcnt(i32 -1, i32 [[TMP4]])
+; SINK-NEXT:    [[AND_I11:%.*]] = shl i32 [[TMP5]], 2
+; SINK-NEXT:    [[ADD_I12:%.*]] = and i32 [[AND_I11]], -256
+; SINK-NEXT:    [[SHL_I13:%.*]] = or i32 [[ADD_I12]], 8
+; SINK-NEXT:    [[TMP6:%.*]] = tail call i32 @bpermute(i32 [[SHL_I13]], i32 [[TMP0]])
+; SINK-NEXT:    br label [[IF_END]]
+; SINK:       if.end:
+; SINK-NEXT:    [[DOTSINK:%.*]] = phi i32 [ [[TMP6]], [[IF_ELSE]] ], [ [[TMP3]], [[IF_THEN]] ]
+; SINK-NEXT:    [[IDXPROM4:%.*]] = zext i32 [[TMP0]] to i64
+; SINK-NEXT:    [[ARRAYIDX5:%.*]] = getelementptr inbounds i32, ptr [[Y_COERCE:%.*]], i64 [[IDXPROM4]]
+; SINK-NEXT:    store i32 [[DOTSINK]], ptr [[ARRAYIDX5]], align 4
+; SINK-NEXT:    ret void
+;
+entry:
+  %0 = tail call i32 @tid()
+  %rem = and i32 %0, 1
+  %cmp.not = icmp eq i32 %rem, 0
+  br i1 %cmp.not, label %if.else, label %if.then
+
+if.then:
+  %1 = tail call i32 @mbcnt(i32 -1, i32 0)
+  %2 = tail call i32 @mbcnt(i32 -1, i32 %1)
+  %and.i = shl i32 %2, 2
+  %add.i = and i32 %and.i, -256
+  %shl.i = or i32 %add.i, 16
+  %3 = tail call i32 @bpermute(i32 %shl.i, i32 %0)
+  %idxprom = zext i32 %0 to i64
+  %arrayidx = getelementptr inbounds i32, ptr %y.coerce, i64 %idxprom
+  store i32 %3, ptr %arrayidx
+  br label %if.end
+
+if.else:
+  %4 = tail call i32 @mbcnt(i32 -1, i32 0)
+  %5 = tail call i32 @mbcnt(i32 -1, i32 %4)
+  %and.i11 = shl i32 %5, 2
+  %add.i12 = and i32 %and.i11, -256
+  %shl.i13 = or i32 %add.i12, 8
+  %6 = tail call i32 @bpermute(i32 %shl.i13, i32 %0)
+  %idxprom4 = zext i32 %0 to i64
+  %arrayidx5 = getelementptr inbounds i32, ptr %y.coerce, i64 %idxprom4
+  store i32 %6, ptr %arrayidx5
+  br label %if.end
+
+if.end:
+  ret void
+}


        


More information about the llvm-commits mailing list