This patch improves ValueTracking's impliesPoison(V1, V2) to do this reasoning:
%res = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 %a, i64 %b) %overflow = extractvalue { i64, i1 } %res, 1 %mul = extractvalue { i64, i1 } %res, 0 ; If %mul is poison, %overflow is also poison, and vice versa.
This improvement leads to supporting this optimization under -instcombine-unsafe-select-transform=0:
define i1 @test2_logical(i64 %a, i64 %b, i64* %ptr) { ; CHECK-LABEL: @test2_logical( ; CHECK-NEXT: [[MUL:%.*]] = mul i64 [[A:%.*]], [[B:%.*]] ; CHECK-NEXT: [[TMP1:%.*]] = icmp ne i64 [[A]], 0 ; CHECK-NEXT: [[TMP2:%.*]] = icmp ne i64 [[B]], 0 ; CHECK-NEXT: [[OVERFLOW_1:%.*]] = and i1 [[TMP1]], [[TMP2]] ; CHECK-NEXT: [[NEG:%.*]] = sub i64 0, [[MUL]] ; CHECK-NEXT: store i64 [[NEG]], i64* [[PTR:%.*]], align 8 ; CHECK-NEXT: ret i1 [[OVERFLOW_1]] ; %res = tail call { i64, i1 } @llvm.umul.with.overflow.i64(i64 %a, i64 %b) %overflow = extractvalue { i64, i1 } %res, 1 %mul = extractvalue { i64, i1 } %res, 0 %cmp = icmp ne i64 %mul, 0 %overflow.1 = select i1 %overflow, i1 true, i1 %cmp %neg = sub i64 0, %mul store i64 %neg, i64* %ptr, align 8 ret i1 %overflow.1 }
Previously, this didn't happen because the flag prevented select i1 %overflow, i1 true, i1 %cmp from being or i1 %overflow, %cmp.
Note that the select -> or conversion happens only when impliesPoison(%cmp, %overflow) returns true.
This improvement allows impliesPoison to do the reasoning.
I chose to not make propagatesPoison return true on extractvalue because otherwise we lose this nice property:
"If propagatesPoison(op x y) is true, and the result is non-poison, x and y are also non-poison"
If extractvalue is added, this doesn't hold any more; imagine