[InstCombine] Refine nuw propagation in OptimizePointerDifference (#147059)

After #146100, the offset may
be generated by a previous call to `EmitGEPOffsets`, causing the nuw
flag on shl to be lost. This patch handles the `shl+ptradd` case as
well. It also fixes a miscompilation in the case of `mul + ptradd`.
Alive2: https://alive2.llvm.org/ce/z/BeaNzE

This patch removes many unnecessary masking operations in Rust programs
with the `ptr_offset_from_unsigned` intrinsic :
https://github.com/dtcxzyw/llvm-opt-benchmark/pull/2538/files

Author: dtcxzyw

llvm/lib/Transforms/InstCombine/InstCombineAddSub.cpp

@@ -2164,10 +2164,14 @@ Value *InstCombinerImpl::OptimizePointerDifference(Value *LHS, Value *RHS,
 
   // If this is a single inbounds GEP and the original sub was nuw,
   // then the final multiplication is also nuw.
-  if (auto *I = dyn_cast<Instruction>(Result))
+  if (auto *I = dyn_cast<OverflowingBinaryOperator>(Result))
     if (IsNUW && match(Offset2, m_Zero()) && Base.LHSNW.isInBounds() &&
-        I->getOpcode() == Instruction::Mul)
-      I->setHasNoUnsignedWrap();
+        (I->use_empty() || I->hasOneUse()) && I->hasNoSignedWrap() &&
+        !I->hasNoUnsignedWrap() &&
+        ((I->getOpcode() == Instruction::Mul &&
+          match(I->getOperand(1), m_NonNegative())) ||
+         I->getOpcode() == Instruction::Shl))
+      cast<Instruction>(I)->setHasNoUnsignedWrap();
 
   // If we have a 2nd GEP of the same base pointer, subtract the offsets.
   // If both GEPs are inbounds, then the subtract does not have signed overflow.

llvm/test/Transforms/InstCombine/sub-gep.ll

@@ -1089,3 +1089,72 @@ define <2 x i64> @splat_geps_multiple(ptr %base, i64 %idx0, <2 x i64> %idx1, <2
   %d = sub <2 x i64> %gep2.int, %gep1.int
   ret <2 x i64> %d
 }
+
+define i64 @nuw_ptrdiff_shl_nsw(ptr %base, i64 %idx) {
+; CHECK-LABEL: @nuw_ptrdiff_shl_nsw(
+; CHECK-NEXT:    [[OFFSET:%.*]] = shl nuw nsw i64 [[IDX:%.*]], 3
+; CHECK-NEXT:    ret i64 [[OFFSET]]
+;
+  %offset = shl nsw i64 %idx, 3
+  %gep = getelementptr inbounds i8, ptr %base, i64 %offset
+  %lhs = ptrtoint ptr %gep to i64
+  %rhs = ptrtoint ptr %base to i64
+  %diff = sub nuw i64 %lhs, %rhs
+  ret i64 %diff
+}
+
+define i64 @nuw_ptrdiff_shl_nonsw(ptr %base, i64 %idx) {
+; CHECK-LABEL: @nuw_ptrdiff_shl_nonsw(
+; CHECK-NEXT:    [[OFFSET:%.*]] = shl i64 [[IDX:%.*]], 3
+; CHECK-NEXT:    ret i64 [[OFFSET]]
+;
+  %offset = shl i64 %idx, 3
+  %gep = getelementptr inbounds i8, ptr %base, i64 %offset
+  %lhs = ptrtoint ptr %gep to i64
+  %rhs = ptrtoint ptr %base to i64
+  %diff = sub nuw i64 %lhs, %rhs
+  ret i64 %diff
+}
+
+define i64 @nuw_ptrdiff_mul_nsw_nneg_scale(ptr %base, i64 %idx) {
+; CHECK-LABEL: @nuw_ptrdiff_mul_nsw_nneg_scale(
+; CHECK-NEXT:    [[OFFSET:%.*]] = mul nuw nsw i64 [[IDX:%.*]], 3
+; CHECK-NEXT:    ret i64 [[OFFSET]]
+;
+  %offset = mul nsw i64 %idx, 3
+  %gep = getelementptr inbounds i8, ptr %base, i64 %offset
+  %lhs = ptrtoint ptr %gep to i64
+  %rhs = ptrtoint ptr %base to i64
+  %diff = sub nuw i64 %lhs, %rhs
+  ret i64 %diff
+}
+
+define i64 @nuw_ptrdiff_mul_nsw_unknown_scale(ptr %base, i64 %idx, i64 %scale) {
+; CHECK-LABEL: @nuw_ptrdiff_mul_nsw_unknown_scale(
+; CHECK-NEXT:    [[OFFSET:%.*]] = mul nsw i64 [[IDX:%.*]], [[SCALE:%.*]]
+; CHECK-NEXT:    ret i64 [[OFFSET]]
+;
+  %offset = mul nsw i64 %idx, %scale
+  %gep = getelementptr inbounds i8, ptr %base, i64 %offset
+  %lhs = ptrtoint ptr %gep to i64
+  %rhs = ptrtoint ptr %base to i64
+  %diff = sub nuw i64 %lhs, %rhs
+  ret i64 %diff
+}
+
+declare void @usei64(i64)
+
+define i64 @nuw_ptrdiff_mul_nsw_nneg_scale_multiuse(ptr %base, i64 %idx) {
+; CHECK-LABEL: @nuw_ptrdiff_mul_nsw_nneg_scale_multiuse(
+; CHECK-NEXT:    [[OFFSET:%.*]] = mul nsw i64 [[IDX:%.*]], 3
+; CHECK-NEXT:    call void @usei64(i64 [[OFFSET]])
+; CHECK-NEXT:    ret i64 [[OFFSET]]
+;
+  %offset = mul nsw i64 %idx, 3
+  call void @usei64(i64 %offset)
+  %gep = getelementptr inbounds i8, ptr %base, i64 %offset
+  %lhs = ptrtoint ptr %gep to i64
+  %rhs = ptrtoint ptr %base to i64
+  %diff = sub nuw i64 %lhs, %rhs
+  ret i64 %diff
+}