https://llvm.org/bugs/show_bug.cgi?id=30505
Bug ID: 30505
Summary: !range metadata pessimizes generated assembly
Product: libraries
Version: 3.9
Hardware: PC
OS: Linux
Status: NEW
Severity: normal
Priority: P
Component: Transformation Utilities
Assignee: unassignedb...@nondot.org
Reporter: simonas+llvm....@kazlauskas.me
CC: llvm-bugs@lists.llvm.org
Classification: Unclassified
Created attachment 17340
--> https://llvm.org/bugs/attachment.cgi?id=17340&action=edit
Case where pessimisation happens
In the attached test case the instruction
%1 = load i8, i8* %0, !range !1
has range metadata attached to it. In LLVM 3.9 Combine Redundant instructions
pass replaces this load with
%1 = load i8, i8* %0, align 1, !range !1
%trunc = trunc i8 %1 to i2
The truncation ends up never getting removed and results in extra and
unnecessary `andb $3, %dil` instruction in the optimised assembly. This extra
instruction is causing some noticeable overhead in a tight loop.
---
LLVM 3.8 does not apply the transformation detailed above. Therefore it does
not emit the superfluous `andb` instruction either.
As far as fixes are concerned, perhaps it would make sense to only `trunc` to
the bitwidths supported by the target machine?
--
You are receiving this mail because:
You are on the CC list for the bug.
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
