[llvm] [InstSimplify] Fold trunc to constant when all result bits are known (PR #212123)

Timur Golubovich via llvm-commits llvm-commits at lists.llvm.org
Tue Jul 28 17:56:59 PDT 2026


https://github.com/timurgol007 updated https://github.com/llvm/llvm-project/pull/212123

>From 6403747773e4a8c45a9404c882cd6c49508db6d0 Mon Sep 17 00:00:00 2001
From: Timur Golubovich <timur.golubovich at intel.com>
Date: Sat, 18 Jul 2026 16:09:20 +0200
Subject: [PATCH 1/2] [InstSimplify] Fold trunc to constant when all result
 bits are known

Teach simplifyCastInst to fold trunc to a constant via
computeKnownBits when all surviving bits are determined.
---
 llvm/lib/Analysis/InstructionSimplify.cpp     |  8 ++++
 .../AddressSanitizer/asan-funclet.ll          | 44 +++++++++----------
 .../Transforms/InstCombine/zext-or-icmp.ll    |  7 +--
 .../X86/epilog-vectorization-inductions.ll    |  6 +--
 .../X86/SROA-after-final-loop-unrolling-2.ll  |  3 +-
 5 files changed, 33 insertions(+), 35 deletions(-)

diff --git a/llvm/lib/Analysis/InstructionSimplify.cpp b/llvm/lib/Analysis/InstructionSimplify.cpp
index ddaa44f43e631..5090d51e6dbfa 100644
--- a/llvm/lib/Analysis/InstructionSimplify.cpp
+++ b/llvm/lib/Analysis/InstructionSimplify.cpp
@@ -5679,6 +5679,14 @@ static Value *simplifyCastInst(unsigned CastOpc, Value *Op, Type *Ty,
     }
   }
 
+  // trunc Op to Ty -> constant if all surviving bits are known
+  if (CastOpc == Instruction::Trunc && Ty->isIntegerTy()) {
+    unsigned DestWidth = Ty->getScalarSizeInBits();
+    KnownBits Known = computeKnownBits(Op, Q);
+    if ((Known.Zero | Known.One).countr_one() >= DestWidth)
+      return ConstantInt::get(Ty, Known.One.trunc(DestWidth));
+  }
+
   return nullptr;
 }
 
diff --git a/llvm/test/Instrumentation/AddressSanitizer/asan-funclet.ll b/llvm/test/Instrumentation/AddressSanitizer/asan-funclet.ll
index 2a7216ff70cda..ee3a08e85e4b6 100644
--- a/llvm/test/Instrumentation/AddressSanitizer/asan-funclet.ll
+++ b/llvm/test/Instrumentation/AddressSanitizer/asan-funclet.ll
@@ -23,7 +23,7 @@ declare i32 @dummyPersonality(...)
 
 define void @FuncletPersonality(ptr %ptrParam) sanitize_address personality ptr @__CxxFrameHandler3 {
 ; CHECK-INLINE-LABEL: define void @FuncletPersonality(
-; CHECK-INLINE-SAME: ptr [[PTRPARAM:%.*]]) #[[ATTR2:[0-9]+]] personality ptr @__CxxFrameHandler3 {
+; CHECK-INLINE-SAME: ptr [[PTRPARAM:%.*]]) #[[ATTR3:[0-9]+]] personality ptr @__CxxFrameHandler3 {
 ; CHECK-INLINE-NEXT:  entry:
 ; CHECK-INLINE-NEXT:    [[TMP0:%.*]] = alloca i64, align 32
 ; CHECK-INLINE-NEXT:    store i64 0, ptr [[TMP0]], align 8
@@ -146,13 +146,12 @@ define void @FuncletPersonality(ptr %ptrParam) sanitize_address personality ptr
 ; CHECK-INLINE-NEXT:    br i1 [[TMP84]], label [[TMP85:%.*]], label [[TMP90:%.*]], !prof [[PROF1]]
 ; CHECK-INLINE:       79:
 ; CHECK-INLINE-NEXT:    [[TMP86:%.*]] = and i64 [[TMP77]], 7
-; CHECK-INLINE-NEXT:    [[TMP87:%.*]] = trunc i64 [[TMP86]] to i8
-; CHECK-INLINE-NEXT:    [[TMP88:%.*]] = icmp sge i8 [[TMP87]], [[TMP83]]
+; CHECK-INLINE-NEXT:    [[TMP88:%.*]] = icmp sge i8 0, [[TMP83]]
 ; CHECK-INLINE-NEXT:    br i1 [[TMP88]], label [[TMP89:%.*]], label [[TMP90]]
-; CHECK-INLINE:       83:
+; CHECK-INLINE:       82:
 ; CHECK-INLINE-NEXT:    call void @__asan_report_store1(i64 [[TMP77]]) #[[ATTR7]]
 ; CHECK-INLINE-NEXT:    unreachable
-; CHECK-INLINE:       84:
+; CHECK-INLINE:       83:
 ; CHECK-INLINE-NEXT:    store volatile i8 0, ptr [[TMP79]], align 1
 ; CHECK-INLINE-NEXT:    invoke void @MayThrowFunc()
 ; CHECK-INLINE-NEXT:            to label [[INVOKE_CONT:%.*]] unwind label [[EHCLEANUP:%.*]]
@@ -164,15 +163,15 @@ define void @FuncletPersonality(ptr %ptrParam) sanitize_address personality ptr
 ; CHECK-INLINE-NEXT:    store i64 1172321806, ptr [[TMP23]], align 8
 ; CHECK-INLINE-NEXT:    [[TMP93:%.*]] = icmp ne i64 [[TMP7]], 0
 ; CHECK-INLINE-NEXT:    br i1 [[TMP93]], label [[TMP94:%.*]], label [[TMP95:%.*]]
-; CHECK-INLINE:       88:
+; CHECK-INLINE:       87:
 ; CHECK-INLINE-NEXT:    call void @__asan_stack_free_8(i64 [[TMP7]], i64 8544)
 ; CHECK-INLINE-NEXT:    br label [[TMP97:%.*]]
-; CHECK-INLINE:       89:
+; CHECK-INLINE:       88:
 ; CHECK-INLINE-NEXT:    call void @__asan_set_shadow_00(i64 [[TMP29]], i64 4)
 ; CHECK-INLINE-NEXT:    [[TMP96:%.*]] = add i64 [[TMP29]], 1028
 ; CHECK-INLINE-NEXT:    call void @__asan_set_shadow_00(i64 [[TMP96]], i64 40)
 ; CHECK-INLINE-NEXT:    br label [[TMP97]]
-; CHECK-INLINE:       91:
+; CHECK-INLINE:       90:
 ; CHECK-INLINE-NEXT:    ret void
 ; CHECK-INLINE:       ehcleanup:
 ; CHECK-INLINE-NEXT:    [[TMP98:%.*]] = cleanuppad within none []
@@ -183,15 +182,14 @@ define void @FuncletPersonality(ptr %ptrParam) sanitize_address personality ptr
 ; CHECK-INLINE-NEXT:    [[TMP102:%.*]] = load i8, ptr [[TMP101]], align 1
 ; CHECK-INLINE-NEXT:    [[TMP103:%.*]] = icmp ne i8 [[TMP102]], 0
 ; CHECK-INLINE-NEXT:    br i1 [[TMP103]], label [[TMP104:%.*]], label [[TMP109:%.*]], !prof [[PROF1]]
-; CHECK-INLINE:       98:
+; CHECK-INLINE:       97:
 ; CHECK-INLINE-NEXT:    [[TMP105:%.*]] = and i64 [[TMP54]], 7
-; CHECK-INLINE-NEXT:    [[TMP106:%.*]] = trunc i64 [[TMP105]] to i8
-; CHECK-INLINE-NEXT:    [[TMP107:%.*]] = icmp sge i8 [[TMP106]], [[TMP102]]
+; CHECK-INLINE-NEXT:    [[TMP107:%.*]] = icmp sge i8 0, [[TMP102]]
 ; CHECK-INLINE-NEXT:    br i1 [[TMP107]], label [[TMP108:%.*]], label [[TMP109]]
-; CHECK-INLINE:       102:
+; CHECK-INLINE:       100:
 ; CHECK-INLINE-NEXT:    call void @__asan_report_store1(i64 [[TMP54]]) #[[ATTR7]] [ "funclet"(token [[TMP98]]) ]
 ; CHECK-INLINE-NEXT:    unreachable
-; CHECK-INLINE:       103:
+; CHECK-INLINE:       101:
 ; CHECK-INLINE-NEXT:    store volatile i8 0, ptr [[TMP56]], align 1
 ; CHECK-INLINE-NEXT:    call void @__asan_poison_stack_memory(i64 [[TMP54]], i64 4) [ "funclet"(token [[TMP98]]) ]
 ; CHECK-INLINE-NEXT:    call void @DeInit(ptr [[TMP14]]) [ "funclet"(token [[TMP98]]) ]
@@ -214,27 +212,27 @@ define void @FuncletPersonality(ptr %ptrParam) sanitize_address personality ptr
 ; CHECK-INLINE-NEXT:    [[TMP120:%.*]] = load i8, ptr [[TMP119]], align 1
 ; CHECK-INLINE-NEXT:    [[TMP121:%.*]] = icmp ne i8 [[TMP120]], 0
 ; CHECK-INLINE-NEXT:    br i1 [[TMP121]], label [[TMP122:%.*]], label [[TMP127:%.*]], !prof [[PROF1]]
-; CHECK-INLINE:       118:
+; CHECK-INLINE:       116:
 ; CHECK-INLINE-NEXT:    [[TMP123:%.*]] = and i64 [[TMP116]], 7
 ; CHECK-INLINE-NEXT:    [[TMP124:%.*]] = trunc i64 [[TMP123]] to i8
 ; CHECK-INLINE-NEXT:    [[TMP125:%.*]] = icmp sge i8 [[TMP124]], [[TMP120]]
 ; CHECK-INLINE-NEXT:    br i1 [[TMP125]], label [[TMP126:%.*]], label [[TMP127]]
-; CHECK-INLINE:       122:
+; CHECK-INLINE:       120:
 ; CHECK-INLINE-NEXT:    call void @__asan_report_store_n(i64 [[TMP116]], i64 8) #[[ATTR7]] [ "funclet"(token [[TMP98]]) ]
 ; CHECK-INLINE-NEXT:    unreachable
-; CHECK-INLINE:       123:
+; CHECK-INLINE:       121:
 ; CHECK-INLINE-NEXT:    [[TMP128:%.*]] = lshr i64 [[TMP114]], 3
 ; CHECK-INLINE-NEXT:    [[TMP129:%.*]] = add i64 [[TMP128]], [[TMP1]]
 ; CHECK-INLINE-NEXT:    [[TMP130:%.*]] = inttoptr i64 [[TMP129]] to ptr
 ; CHECK-INLINE-NEXT:    [[TMP131:%.*]] = load i8, ptr [[TMP130]], align 1
 ; CHECK-INLINE-NEXT:    [[TMP132:%.*]] = icmp ne i8 [[TMP131]], 0
 ; CHECK-INLINE-NEXT:    br i1 [[TMP132]], label [[TMP133:%.*]], label [[EHEXIT:%.*]], !prof [[PROF1]]
-; CHECK-INLINE:       129:
+; CHECK-INLINE:       127:
 ; CHECK-INLINE-NEXT:    [[TMP134:%.*]] = and i64 [[TMP114]], 7
 ; CHECK-INLINE-NEXT:    [[TMP135:%.*]] = trunc i64 [[TMP134]] to i8
 ; CHECK-INLINE-NEXT:    [[TMP136:%.*]] = icmp sge i8 [[TMP135]], [[TMP131]]
 ; CHECK-INLINE-NEXT:    br i1 [[TMP136]], label [[TMP137:%.*]], label [[EHEXIT]]
-; CHECK-INLINE:       133:
+; CHECK-INLINE:       131:
 ; CHECK-INLINE-NEXT:    call void @__asan_report_store_n(i64 [[TMP114]], i64 8) #[[ATTR7]] [ "funclet"(token [[TMP98]]) ]
 ; CHECK-INLINE-NEXT:    unreachable
 ; CHECK-INLINE:       ehexit:
@@ -247,19 +245,19 @@ define void @FuncletPersonality(ptr %ptrParam) sanitize_address personality ptr
 ; CHECK-INLINE-NEXT:    store i64 1172321806, ptr [[TMP23]], align 8
 ; CHECK-INLINE-NEXT:    [[TMP142:%.*]] = icmp ne i64 [[TMP7]], 0
 ; CHECK-INLINE-NEXT:    br i1 [[TMP142]], label [[TMP143:%.*]], label [[TMP144:%.*]]
-; CHECK-INLINE:       139:
+; CHECK-INLINE:       137:
 ; CHECK-INLINE-NEXT:    call void @__asan_stack_free_8(i64 [[TMP7]], i64 8544) [ "funclet"(token [[TMP98]]) ]
 ; CHECK-INLINE-NEXT:    br label [[TMP146:%.*]]
-; CHECK-INLINE:       140:
+; CHECK-INLINE:       138:
 ; CHECK-INLINE-NEXT:    call void @__asan_set_shadow_00(i64 [[TMP29]], i64 4) [ "funclet"(token [[TMP98]]) ]
 ; CHECK-INLINE-NEXT:    [[TMP145:%.*]] = add i64 [[TMP29]], 1028
 ; CHECK-INLINE-NEXT:    call void @__asan_set_shadow_00(i64 [[TMP145]], i64 40) [ "funclet"(token [[TMP98]]) ]
 ; CHECK-INLINE-NEXT:    br label [[TMP146]]
-; CHECK-INLINE:       142:
+; CHECK-INLINE:       140:
 ; CHECK-INLINE-NEXT:    cleanupret from [[TMP98]] unwind to caller
 ;
 ; CHECK-OUTLINE-LABEL: define void @FuncletPersonality(
-; CHECK-OUTLINE-SAME: ptr [[PTRPARAM:%.*]]) #[[ATTR2:[0-9]+]] personality ptr @__CxxFrameHandler3 {
+; CHECK-OUTLINE-SAME: ptr [[PTRPARAM:%.*]]) #[[ATTR3:[0-9]+]] personality ptr @__CxxFrameHandler3 {
 ; CHECK-OUTLINE-NEXT:  entry:
 ; CHECK-OUTLINE-NEXT:    [[TMP0:%.*]] = alloca i64, align 32
 ; CHECK-OUTLINE-NEXT:    store i64 0, ptr [[TMP0]], align 8
@@ -483,7 +481,7 @@ nopredecessor:
 ; Non-Windows personality, ensure no funclet gets attached to asan runtime call.
 define void @OtherPersonality(ptr %ptrParam) sanitize_address personality ptr @dummyPersonality {
 ; CHECK-LABEL: define void @OtherPersonality(
-; CHECK-SAME: ptr [[PTRPARAM:%.*]]) #[[ATTR2:[0-9]+]] personality ptr @dummyPersonality {
+; CHECK-SAME: ptr [[PTRPARAM:%.*]]) #[[ATTR3:[0-9]+]] personality ptr @dummyPersonality {
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[TMP0:%.*]] = load i64, ptr @__asan_shadow_memory_dynamic_address, align 8
 ; CHECK-NEXT:    [[ASAN_LOCAL_STACK_BASE:%.*]] = alloca i64, align 8
diff --git a/llvm/test/Transforms/InstCombine/zext-or-icmp.ll b/llvm/test/Transforms/InstCombine/zext-or-icmp.ll
index feb4be9e37050..fc77acfb0b35b 100644
--- a/llvm/test/Transforms/InstCombine/zext-or-icmp.ll
+++ b/llvm/test/Transforms/InstCombine/zext-or-icmp.ll
@@ -179,12 +179,9 @@ define i8 @PR49475_infloop(i32 %t0, i16 %insert, i64 %e, i8 %i162) "instcombine-
 ; CHECK-NEXT:    [[SUB17:%.*]] = sub i64 [[CONV16]], [[E:%.*]]
 ; CHECK-NEXT:    [[SEXT:%.*]] = shl i64 [[SUB17]], 32
 ; CHECK-NEXT:    [[CONV18:%.*]] = ashr exact i64 [[SEXT]], 32
-; CHECK-NEXT:    [[CMP:%.*]] = icmp sge i64 [[XOR]], [[CONV18]]
-; CHECK-NEXT:    [[TRUNC44:%.*]] = zext i1 [[CMP]] to i8
-; CHECK-NEXT:    [[INC:%.*]] = add i8 [[I162]], [[TRUNC44]]
-; CHECK-NEXT:    [[TOBOOL23_NOT:%.*]] = xor i1 [[CMP]], true
+; CHECK-NEXT:    [[TOBOOL23_NOT:%.*]] = icmp slt i64 [[XOR]], [[CONV18]]
 ; CHECK-NEXT:    call void @llvm.assume(i1 [[TOBOOL23_NOT]])
-; CHECK-NEXT:    ret i8 [[INC]]
+; CHECK-NEXT:    ret i8 [[I162]]
 ;
   %b = icmp eq i32 %t0, 0
   %b2 = icmp eq i16 %insert, 0
diff --git a/llvm/test/Transforms/LoopVectorize/X86/epilog-vectorization-inductions.ll b/llvm/test/Transforms/LoopVectorize/X86/epilog-vectorization-inductions.ll
index 4f6c9a01786a0..2abba6ff12637 100644
--- a/llvm/test/Transforms/LoopVectorize/X86/epilog-vectorization-inductions.ll
+++ b/llvm/test/Transforms/LoopVectorize/X86/epilog-vectorization-inductions.ll
@@ -265,14 +265,10 @@ define i8 @multiple_inductions_start_at_0() {
 ; CHECK-NEXT:    br i1 false, label [[VEC_EPILOG_SCALAR_PH]], label [[VEC_EPILOG_PH]], !prof [[PROF11:![0-9]+]]
 ; CHECK:       vec.epilog.ph:
 ; CHECK-NEXT:    [[VEC_EPILOG_RESUME_VAL:%.*]] = phi i32 [ 1024, [[VEC_EPILOG_ITER_CHECK]] ], [ 0, [[VECTOR_MAIN_LOOP_ITER_CHECK]] ]
-; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ 0, [[VEC_EPILOG_ITER_CHECK]] ], [ 0, [[VECTOR_MAIN_LOOP_ITER_CHECK]] ]
-; CHECK-NEXT:    [[TMP2:%.*]] = trunc i32 [[BC_RESUME_VAL]] to i8
-; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i8> poison, i8 [[TMP2]], i64 0
-; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i8> [[BROADCAST_SPLATINSERT]], <4 x i8> poison, <4 x i32> zeroinitializer
 ; CHECK-NEXT:    br label [[VEC_EPILOG_VECTOR_BODY:%.*]]
 ; CHECK:       vec.epilog.vector.body:
 ; CHECK-NEXT:    [[INDEX1:%.*]] = phi i32 [ [[VEC_EPILOG_RESUME_VAL]], [[VEC_EPILOG_PH]] ], [ [[INDEX_NEXT3:%.*]], [[VEC_EPILOG_VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_IND2:%.*]] = phi <4 x i8> [ [[BROADCAST_SPLAT]], [[VEC_EPILOG_PH]] ], [ [[VEC_IND2]], [[VEC_EPILOG_VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_IND2:%.*]] = phi <4 x i8> [ zeroinitializer, [[VEC_EPILOG_PH]] ], [ [[VEC_IND2]], [[VEC_EPILOG_VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[INDEX_NEXT3]] = add nuw i32 [[INDEX1]], 4
 ; CHECK-NEXT:    [[TMP3:%.*]] = icmp eq i32 [[INDEX_NEXT3]], 1052
 ; CHECK-NEXT:    br i1 [[TMP3]], label [[VEC_EPILOG_MIDDLE_BLOCK:%.*]], label [[VEC_EPILOG_VECTOR_BODY]], !llvm.loop [[LOOP12:![0-9]+]]
diff --git a/llvm/test/Transforms/PhaseOrdering/X86/SROA-after-final-loop-unrolling-2.ll b/llvm/test/Transforms/PhaseOrdering/X86/SROA-after-final-loop-unrolling-2.ll
index 51ea03f465c7e..30565aaddc5b7 100644
--- a/llvm/test/Transforms/PhaseOrdering/X86/SROA-after-final-loop-unrolling-2.ll
+++ b/llvm/test/Transforms/PhaseOrdering/X86/SROA-after-final-loop-unrolling-2.ll
@@ -171,7 +171,6 @@ attributes #3 = { nounwind }
 ; CHECK: [[CHAR_TBAA5]] = !{[[META6:![0-9]+]], [[META6]], i64 0}
 ; CHECK: [[META6]] = !{!"omnipotent char", [[META7:![0-9]+]], i64 0}
 ; CHECK: [[META7]] = !{!"Simple C++ TBAA"}
-; CHECK: [[LOOP8]] = distinct !{[[LOOP8]], [[META9:![0-9]+]], [[META10:![0-9]+]]}
+; CHECK: [[LOOP8]] = distinct !{[[LOOP8]], [[META9:![0-9]+]]}
 ; CHECK: [[META9]] = !{!"llvm.loop.mustprogress"}
-; CHECK: [[META10]] = !{!"llvm.loop.isvectorized", i32 1}
 ;.

>From 8c384714b4954e412797e404fae5d1370580e525 Mon Sep 17 00:00:00 2001
From: Timur Golubovich <timur.golubovich at intel.com>
Date: Mon, 27 Jul 2026 12:38:35 +0200
Subject: [PATCH 2/2] finished idea

---
 .../InstCombine/InstCombineCasts.cpp          | 12 +++-
 .../InstCombine/trunc-multiuse-phi.ll         | 65 +++++++++++++++++++
 2 files changed, 75 insertions(+), 2 deletions(-)
 create mode 100644 llvm/test/Transforms/InstCombine/trunc-multiuse-phi.ll

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
index 2db6396b9d661..1a9879882125e 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp
@@ -270,9 +270,17 @@ Instruction *InstCombinerImpl::commonCastTransforms(CastInst &CI) {
     // Don't do this if it would create a PHI node with an illegal type from a
     // legal type.
     if (!Src->getType()->isIntegerTy() || !CI.getType()->isIntegerTy() ||
-        shouldChangeType(CI.getSrcTy(), CI.getType()))
-      if (Instruction *NV = foldOpIntoPhi(CI, PN))
+        shouldChangeType(CI.getSrcTy(), CI.getType())) {
+      bool AllConst =
+          !PN->hasOneUse() && all_of(PN->blocks(), [&](BasicBlock *InBB) {
+            Value *V = simplifyCastInst(
+                CI.getOpcode(), PN->getIncomingValueForBlock(InBB),
+                CI.getType(), SQ.getWithInstruction(InBB->getTerminator()));
+            return V && isa<Constant>(V);
+          });
+      if (Instruction *NV = foldOpIntoPhi(CI, PN, AllConst))
         return NV;
+    }
   }
 
   // Canonicalize a unary shuffle after the cast if neither operation changes
diff --git a/llvm/test/Transforms/InstCombine/trunc-multiuse-phi.ll b/llvm/test/Transforms/InstCombine/trunc-multiuse-phi.ll
new file mode 100644
index 0000000000000..dd3021e39f6c8
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/trunc-multiuse-phi.ll
@@ -0,0 +1,65 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -passes=instcombine -S | FileCheck %s
+
+; Reduced from libc++ basic_string short-string optimization.
+; The low bit of the header word distinguishes heap-allocated (bit0=1) from
+; inline (bit0=0). The trunc to i1 checks whether to deallocate, and the phi
+; also feeds the capacity mask (and i64 %hdr, -2). Because the phi has multiple
+; uses, InstCombine does not push the trunc through it, so it misses that
+; the trunc result is simply !%is.short (the branch condition).
+
+declare void @_ZdlPvm(ptr, i64)
+declare noalias nonnull ptr @_Znwm(i64)
+
+define void @trunc_multiuse_phi(i64 %len, ptr %buf, ptr %src) {
+; CHECK-LABEL: @trunc_multiuse_phi(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[IS_SHORT:%.*]] = icmp ugt i64 [[LEN:%.*]], 22
+; CHECK-NEXT:    br i1 [[IS_SHORT]], label [[SHORT:%.*]], label [[LONG:%.*]]
+; CHECK:       short:
+; CHECK-NEXT:    [[HDR_SHORT:%.*]] = shl nuw nsw i64 [[LEN]], 1
+; CHECK-NEXT:    br label [[MERGE:%.*]]
+; CHECK:       long:
+; CHECK-NEXT:    [[CAP_RAW:%.*]] = and i64 [[LEN]], -16
+; CHECK-NEXT:    [[CAP:%.*]] = add nuw i64 [[CAP_RAW]], 16
+; CHECK-NEXT:    [[HEAP:%.*]] = call noalias ptr @_Znwm(i64 [[CAP]])
+; CHECK-NEXT:    br label [[MERGE]]
+; CHECK:       merge:
+; CHECK-NEXT:    [[CAP_MASK:%.*]] = phi i64 [ [[HDR_SHORT]], [[LONG]] ], [ [[CAP]], [[SHORT]] ]
+; CHECK-NEXT:    [[DATA:%.*]] = phi ptr [ [[BUF:%.*]], [[LONG]] ], [ [[HEAP]], [[SHORT]] ]
+; CHECK-NEXT:    br i1 [[IS_SHORT]], label [[DO_FREE:%.*]], label [[DONE:%.*]]
+; CHECK:       do.free:
+; CHECK-NEXT:    call void @_ZdlPvm(ptr [[DATA]], i64 [[CAP_MASK]])
+; CHECK-NEXT:    br label [[DONE]]
+; CHECK:       done:
+; CHECK-NEXT:    ret void
+;
+entry:
+  %is.short = icmp ult i64 %len, 23
+  br i1 %is.short, label %short, label %long
+
+short:
+  %hdr.short = shl nuw nsw i64 %len, 1
+  br label %merge
+
+long:
+  %cap.raw = and i64 %len, -16
+  %cap = add nuw i64 %cap.raw, 16
+  %heap = call noalias ptr @_Znwm(i64 %cap)
+  %hdr.long = or disjoint i64 %cap, 1
+  br label %merge
+
+merge:
+  %hdr = phi i64 [ %hdr.short, %short ], [ %hdr.long, %long ]
+  %data = phi ptr [ %buf, %short ], [ %heap, %long ]
+  %is.heap = trunc i64 %hdr to i1
+  br i1 %is.heap, label %do.free, label %done
+
+do.free:
+  %cap.mask = and i64 %hdr, -2
+  call void @_ZdlPvm(ptr %data, i64 %cap.mask)
+  br label %done
+
+done:
+  ret void
+}



More information about the llvm-commits mailing list