[Regression 9] i1 is not treated as 8bit in store combine #42791
Description
| Bugzilla Link | 43446 |
| Resolution | FIXED |
| Resolved on | Feb 26, 2020 06:32 |
| Version | 9.0 |
| OS | Linux |
| Blocks | #30613 #43900 |
| CC | @topperc,@zmodem,@LebedevRI,@RKSimon,@slacka,@Kojoley,@rotateright |
| Fixed by commit(s) | d20907d |
Extended Description
On LLVM 9 we have some code that breaks when LLVM optimizes it. We have been unable to produce a minimum example in our frontend, but we've narrowed down a small LLVM IR snippet that seems to have the same problem. This was not an issue on LLVM 8.0.0.
testing.ll:
; ModuleID = 'testing-not-working'
source_filename = "testing-not-working"
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"
; Function Attrs: nobuiltin nounwind
define i64 @​testing() local_unnamed_addr #​1 {
Entry:
%hasher = alloca i64
%0 = tail call i64 asm sideeffect "syscall", "={rax},{rax},{rdi},{rsi},{rdx},{r10},{r8},{r9},~{rcx},~{r11},~{memory}"(i64 9, i64 0, i64 1, i64 3, i64 34, i64 -1, i64 0) #​6
%1 = inttoptr i64 %0 to i8*
store i8 -2, i8* %1
%2 = inttoptr i64 %0 to i1*
store i1 true, i1* %2
%3 = bitcast i64* %hasher to i8*
store i64 0, i64* %hasher
%4 = load i1, i1* %2
%5 = zext i1 %4 to i64
ret i64 %5
}
attributes #​0 = { cold noreturn nounwind }
attributes #​1 = { nobuiltin nounwind }
attributes #​2 = { argmemonly nounwind }
attributes #​3 = { naked nobuiltin noreturn nounwind }
attributes #​4 = { nobuiltin noreturn nounwind }
attributes #​5 = { nobuiltin noinline nounwind }
attributes #​6 = { nounwind }
attributes #​7 = { noinline }
test.c:
#include <inttypes.h>
#include <stdio.h>
uint64_t testing();
void main() {
printf("%" PRId64 "", testing());
}path/to/llc testing.ll -filetype=obj && gcc test.c testing.o && ./a.out
Compiling this code and running it prints 255. This is very unexpected as we are returning the result of a zext instruction.
%5 = zext i1 %4 to i64
ret i64 %5
The docs for llvm say:
When zero extending from i1, the result will always be either 0 or 1.
https://llvm.org/docs/LangRef.html#zext-to-instruction
So no matter what the code above is doing, zext should ALWAYS give 0 or 1. But this is clearly not the case as testing returns 255. Compiling with llc from llvm8 and testing returns 1 as expected.
❯ ~/repo/llvm-cfe-install/bin/llc --version | grep LLVM
LLVM (http://llvm.org/):
LLVM version 9.0.0
❯ llc --version | grep LLVM
LLVM (http://llvm.org/):
LLVM version 8.0.1
❯ llc testing.ll -filetype=obj && gcc test.c testing.o && ./a.out
1
❯ ~/repo/llvm-cfe-install/bin/llc testing.ll -filetype=obj && gcc test.c testing.o && ./a.out
255Thank-you.
Credit to Jimmi Holst Christensen [email protected]