1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
|
; RUN: llc -mtriple=armv7s-apple-ios7.0 -show-mc-encoding %s -o - | FileCheck %s --check-prefix=CHECK-ARM
; RUN: llc -mtriple=thumbv7s-apple-ios7.0 -show-mc-encoding %s -o - | FileCheck %s --check-prefix=CHECK-THUMB
; RUN: llc -mtriple=thumbv7m-none-eabi -show-mc-encoding %s -o - | FileCheck %s --check-prefix=CHECK-THUMB
; In the ARM backend, most compares are glued to their uses so CPSR can't
; escape. However, for long ADCS chains (and last ditch fallback) the dependency
; is carried in the DAG because duplicating them can be more expensive than
; copying CPSR.
; Crafting a test for this was a little tricky, in case it breaks here are some
; notes on what I was tring to achieve:
; + We want 2 long ADCS chains
; + We want them to split after an initial common prefix (so that a single
; CPSR is used twice).
; + We want both chains to write CPSR post-split (so that the copy can't be
; elided).
; + We want the chains to be long enough that duplicating them is expensive.
define void @test_copy_cpsr(i128 %lhs, i128 %rhs, i128* %addr) {
; CHECK-ARM: test_copy_cpsr:
; CHECK-THUMB: test_copy_cpsr:
; CHECK-ARM: mrs [[TMP:r[0-9]+]], apsr @ encoding: [0x00,0x{{[0-9a-f]}}0,0x0f,0xe1]
; CHECK-ARM: msr APSR_nzcvq, [[TMP]] @ encoding: [0x0{{[0-9a-f]}},0xf0,0x28,0xe1]
; In Thumb mode v7M and v7AR have different MRS/MSR instructions that happen
; to overlap for the apsr case, so it's definitely worth checking both.
; CHECK-THUMB: mrs [[TMP:r[0-9]+]], apsr @ encoding: [0xef,0xf3,0x00,0x8{{[0-9a-f]}}]
; CHECK-THUMB: msr {{APSR|apsr}}_nzcvq, [[TMP]] @ encoding: [0x8{{[0-9a-f]}},0xf3,0x00,0x88]
%sum = add i128 %lhs, %rhs
store volatile i128 %sum, i128* %addr
%rhs2.tmp1 = trunc i128 %rhs to i64
%rhs2 = zext i64 %rhs2.tmp1 to i128
%sum2 = add i128 %lhs, %rhs2
store volatile i128 %sum2, i128* %addr
ret void
}
|
