[DAGCombiner] Fix miscompile bug in combineShiftOfShiftedLogic

[bjope]

Ensure that the sum of the shift amounts does not overflow the
shift amount type when combining shifts in combineShiftOfShiftedLogic.

Solves a miscompile bug found when testing the C23 BitInt feature.

Targets like X86 that only use an i8 for shift amounts after
legalization seems to be extra susceptible for bugs like this as it
isn't legal to shift more than 255 steps.

[llvmbot]

@llvm/pr-subscribers-llvm-selectiondag

@llvm/pr-subscribers-backend-x86

Author: Björn Pettersson (bjope)

Changes

Ensure that the sum of the shift amounts does not overflow the
shift amount type when combining shifts in combineShiftOfShiftedLogic.

Solves a miscompile bug found when testing the C23 BitInt feature.

Targets like X86 that only use an i8 for shift amounts after
legalization seems to be extra susceptible for bugs like this as it
isn't legal to shift more than 255 steps.

---

Full diff: https://github.com/llvm/llvm-project/pull/89616.diff

2 Files Affected:

• (modified) llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (+8-1)
• (modified) llvm/test/CodeGen/X86/shift-combine.ll (+35)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index e6e0a1fc7d8277..fd265b12d73ca4 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -9503,8 +9503,15 @@ static SDValue combineShiftOfShiftedLogic(SDNode *Shift, SelectionDAG &DAG) {
     if (ShiftAmtVal->getBitWidth() != C1Val.getBitWidth())
       return false;
 
+    // The fold is not valid if the sum of the shift values doesn't fit in the
+    // given shift amount type.
+    bool Overflow = false;
+    APInt NewShiftAmt = C1Val.uadd_ov(*ShiftAmtVal, Overflow);
+    if (Overflow)
+      return false;
+
     // The fold is not valid if the sum of the shift values exceeds bitwidth.
-    if ((*ShiftAmtVal + C1Val).uge(V.getScalarValueSizeInBits()))
+    if (NewShiftAmt.uge(V.getScalarValueSizeInBits()))
       return false;
 
     return true;
diff --git a/llvm/test/CodeGen/X86/shift-combine.ll b/llvm/test/CodeGen/X86/shift-combine.ll
index bb0fd9c68afbaf..ca3898d358e221 100644
--- a/llvm/test/CodeGen/X86/shift-combine.ll
+++ b/llvm/test/CodeGen/X86/shift-combine.ll
@@ -787,3 +787,38 @@ define <4 x i32> @or_tree_with_mismatching_shifts_vec_i32(<4 x i32> %a, <4 x i32
   %r = or <4 x i32> %or.ab, %or.cd
   ret <4 x i32> %r
 }
+
+; FIXME: Reproducer for a DAGCombiner::combineShiftOfShiftedLogic
+; bug. DAGCombiner need to check that the sum of the shift amounts fits in i8,
+; which is the legal type used to described X86 shift amounts. Verify that we
+; do not try to create a shift with 140+120 as shift amount, and verify that
+; the stored value do not depend on %a1.
+define void @combineShiftOfShiftedLogic(i128 %a1, i32 %a2, ptr %p) {
+; X86-LABEL: combineShiftOfShiftedLogic:
+; X86:       # %bb.0:
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    movl %eax, 20(%ecx)
+; X86-NEXT:    movl $0, 16(%ecx)
+; X86-NEXT:    movl $0, 12(%ecx)
+; X86-NEXT:    movl $0, 8(%ecx)
+; X86-NEXT:    movl $0, 4(%ecx)
+; X86-NEXT:    movl $0, (%ecx)
+; X86-NEXT:    retl
+;
+; X64-LABEL: combineShiftOfShiftedLogic:
+; X64:       # %bb.0:
+; X64-NEXT:    # kill: def $edx killed $edx def $rdx
+; X64-NEXT:    shlq $32, %rdx
+; X64-NEXT:    movq %rdx, 16(%rcx)
+; X64-NEXT:    movq $0, 8(%rcx)
+; X64-NEXT:    movq $0, (%rcx)
+; X64-NEXT:    retq
+  %zext1 = zext i128 %a1 to i192
+  %zext2 = zext i32 %a2 to i192
+  %shl = shl i192 %zext1, 130
+  %or = or i192 %shl, %zext2
+  %res = shl i192 %or, 160
+  store i192 %res, ptr %p, align 8
+  ret void
+}

[nikic]

Leftover FIXME

[bjope]

Thanks! (I actually realized that I had forgotten to remove this when I woke up this morning, but was not quick enough to fix it before you found it.)

[jayfoad]

Should be 130+160?

[jayfoad]

LGTM

[AreaZR]

/cherry-pick 5fd9bbd f9b419b

[llvmbot]

/cherry-pick 5fd9bbd f9b419b

Error: Command failed due to missing milestone.