mirror of
https://github.com/RPCS3/llvm-mirror.git
synced 2024-10-19 11:02:59 +02:00
7bae0ef103
Currently LLVM assumes that a pointer addrspacecasted to a different addr space is equivalent to trunc or zext bitwise, which is not true. For example, in amdgcn target, when a null pointer is addrspacecasted from addr space 4 to 0, its value is changed from i64 0 to i32 -1. This patch teaches LLVM not to assume known bits of addrspacecast instruction to its operand. Differential Revision: https://reviews.llvm.org/D26803 llvm-svn: 287545
25 lines
688 B
LLVM
25 lines
688 B
LLVM
; RUN: opt -instcombine -S < %s | FileCheck %s
|
|
|
|
; When a pointer is addrspacecasted to a another addr space, we cannot assume
|
|
; anything about the new bits.
|
|
|
|
target datalayout = "p:32:32-p3:32:32-p4:64:64"
|
|
|
|
; CHECK-LABEL: @test_shift
|
|
; CHECK-NOT: ret i64 0
|
|
define i64 @test_shift(i8* %p) {
|
|
%g = addrspacecast i8* %p to i8 addrspace(4)*
|
|
%i = ptrtoint i8 addrspace(4)* %g to i64
|
|
%shift = lshr i64 %i, 32
|
|
ret i64 %shift
|
|
}
|
|
|
|
; CHECK-LABEL: @test_null
|
|
; A null pointer casted to another addr space may no longer have null value.
|
|
; CHECK-NOT: ret i32 0
|
|
define i32 @test_null() {
|
|
%g = addrspacecast i8* null to i8 addrspace(3)*
|
|
%i = ptrtoint i8 addrspace(3)* %g to i32
|
|
ret i32 %i
|
|
}
|