Extra check when rejoining branches. #167799
Description
https://godbolt.org/z/E4GWGTjf6
Code
#include <utility> __attribute__((cold)) std::pair<const char*, long> cold(); inline std::pair<const char*, long> sub(const char* ptr) { if (*ptr < 5) { return cold(); } return {ptr, (long)*ptr}; } long rejoin(const char* ptr) { auto result = sub(ptr); auto rptr = std::get<0>(result); if (!rptr) { while (true) {} } return std::get<1>(result); } inline std::pair<const char*, long> sub_fix(const char* ptr) { if (*ptr < 5) { auto value = cold(); auto rptr = std::get<0>(value); if (rptr) { return {ptr + (rptr - ptr), std::get<1>(value)}; } else { return {nullptr, std::get<1>(value)}; } } return {ptr, (long)*ptr}; } long rejoin_fix(const char* ptr) { auto result = sub_fix(ptr); auto rptr = std::get<0>(result); if (!rptr) { while (true) {} } return std::get<1>(result); }Assembly
rejoin(char const*): movsx rax, byte ptr [rdi] cmp rax, 4 jle .LBB0_1 test rdi, rdi ; <--- je .LBB0_3 .LBB0_4: ret .LBB0_1: push rax call cold()@PLT mov rdi, rax mov rax, rdx add rsp, 8 test rdi, rdi jne .LBB0_4 .LBB0_3: jmp .LBB0_3 rejoin_fix(char const*): movsx rax, byte ptr [rdi] cmp rax, 4 jle .LBB1_1 ret .LBB1_1: push rax call cold()@PLT test rax, rax lea rsp, [rsp + 8] je .LBB1_4 mov rax, rdx ret .LBB1_4: jmp .LBB1_4IR
define dso_local noundef i64 @rejoin(char const*)(ptr noundef %ptr) local_unnamed_addr { entry: %0 = load i8, ptr %ptr, align 1 %cmp.i = icmp slt i8 %0, 5 br i1 %cmp.i, label %if.then.i, label %if.end.i if.then.i: %call.i = tail call { ptr, i64 } @cold()() #2 br label %sub(char const*).exit if.end.i: %conv1.i = zext nneg i8 %0 to i64 %1 = insertvalue { ptr, i64 } poison, ptr %ptr, 0 %2 = insertvalue { ptr, i64 } %1, i64 %conv1.i, 1 br label %sub(char const*).exit sub(char const*).exit: // <--- branches merge %.fca.1.insert.merged.i = phi { ptr, i64 } [ %call.i, %if.then.i ], [ %2, %if.end.i ] %3 = extractvalue { ptr, i64 } %.fca.1.insert.merged.i, 0 %tobool.not = icmp eq ptr %3, null // <--- not optimized out br i1 %tobool.not, label %while.cond, label %if.end while.cond: br label %while.cond if.end: %4 = extractvalue { ptr, i64 } %.fca.1.insert.merged.i, 1 ret i64 %4 }