Skip to content

Commit e5dbe7b

Browse files
folkertdevfolkertdev
committed
Update and add X86 musttail tests
1 parent 08a4817 commit e5dbe7b

File tree

4 files changed

+109
-14
lines changed

4 files changed

+109
-14
lines changed

‎llvm/test/CodeGen/X86/musttail-inalloca.ll‎

Lines changed: 0 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -18,15 +18,13 @@ define dso_local x86_thiscallcc void @methodWithVtorDisp_thunk(ptr %0, ptr inall
1818
; CHECK-LABEL: methodWithVtorDisp_thunk:
1919
; CHECK: # %bb.0:
2020
; CHECK-NEXT: pushl %esi
21-
; CHECK-NEXT: subl $20, %esp
2221
; CHECK-NEXT: movl %ecx, %esi
2322
; CHECK-NEXT: subl -4(%ecx), %esi
2423
; CHECK-NEXT: pushl {{[0-9]+}}(%esp)
2524
; CHECK-NEXT: pushl $_methodWithVtorDisp_thunk
2625
; CHECK-NEXT: calll ___cyg_profile_func_exit
2726
; CHECK-NEXT: addl $8, %esp
2827
; CHECK-NEXT: movl %esi, %ecx
29-
; CHECK-NEXT: addl $20, %esp
3028
; CHECK-NEXT: popl %esi
3129
; CHECK-NEXT: jmp _methodWithVtorDisp # TAILCALL
3230
%3 = getelementptr inbounds i8, ptr %0, i32 -4

‎llvm/test/CodeGen/X86/musttail-struct.ll‎

Lines changed: 107 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -1,6 +1,5 @@
11
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
22
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -x86-asm-syntax=intel | FileCheck %s
3-
; ; RUN: llc < %s -mtriple=i686-unknown-unknown | FileCheck %s
43

54
; Test correct handling of a musttail call with a byval struct argument.
65

@@ -72,17 +71,117 @@ entry:
7271
}
7372

7473
define dso_local i32 @swapByValArguments(ptr byval(%struct.1xi32) %0, ptr byval(%struct.1xi32) %1) {
75-
; CHECK-LABEL: swapArguments:
74+
; CHECK-LABEL: swapByValArguments:
7675
; CHECK: # %bb.0:
77-
7876
; CHECK-NEXT: mov eax, dword ptr [rsp + 8]
7977
; CHECK-NEXT: mov dword ptr [rsp - 16], eax
80-
; CHECK-NEXT: mov ecx, dword ptr [rsp + 16]
81-
; CHECK-NEXT: mov dword ptr [rsp - 8], ecx
82-
83-
; CHECK-NEXT: mov dword ptr [rsp + 8], ecx
84-
; CHECK-NEXT: mov dword ptr [rsp + 16], eax
78+
; CHECK-NEXT: mov eax, dword ptr [rsp + 16]
79+
; CHECK-NEXT: mov dword ptr [rsp - 8], eax
8580
; CHECK-NEXT: jmp swap # TAILCALL
81+
82+
8683
%r = musttail call i32 @swap(ptr byval(%struct.1xi32) %1, ptr byval(%struct.1xi32) %0)
8784
ret i32 %r
8885
}
86+
87+
; Clang only uses byval for arguments of 65 bytes or larger, but e.g. rustc
88+
; does use byval for smaller types. Here we use a 20 byte struct to keep
89+
; the tests more readable.
90+
%twenty_bytes = type { [5 x i32] }
91+
declare void @large_callee(%twenty_bytes* byval(%twenty_bytes) align 4)
92+
93+
; Functions with byval parameters can be tail-called, because the value is
94+
; actually passed in registers and the stack in the same way for the caller and
95+
; callee. On x86 byval arguments are never (partially) passed via registers.
96+
define void @large_caller(%twenty_bytes* byval(%twenty_bytes) align 4 %a) {
97+
; CHECK-LABEL: large_caller:
98+
; CHECK: # %bb.0: # %entry
99+
; CHECK-NEXT: jmp large_callee@PLT # TAILCALL
100+
entry:
101+
musttail call void @large_callee(%twenty_bytes* byval(%twenty_bytes) align 4 %a)
102+
ret void
103+
}
104+
105+
; The IR for this one looks dodgy, because it has an alloca passed to a
106+
; musttail function, but it is passed as a byval argument, so will be copied
107+
; into the stack space allocated by @large_caller_new_value's caller, so is
108+
; valid.
109+
define void @large_caller_new_value(%twenty_bytes* byval(%twenty_bytes) align 4 %a) {
110+
; CHECK-LABEL: large_caller_new_value:
111+
; CHECK: # %bb.0: # %entry
112+
; CHECK-NEXT: movabs rax, 4294967296
113+
; CHECK-NEXT: mov qword ptr [rsp - 20], rax
114+
; CHECK-NEXT: movabs rcx, 12884901890
115+
; CHECK-NEXT: mov qword ptr [rsp - 12], rcx
116+
; CHECK-NEXT: mov dword ptr [rsp - 4], 4
117+
; CHECK-NEXT: mov qword ptr [rsp + 8], rax
118+
; CHECK-NEXT: mov qword ptr [rsp + 16], rcx
119+
; CHECK-NEXT: mov dword ptr [rsp + 24], 4
120+
; CHECK-NEXT: jmp large_callee@PLT # TAILCALL
121+
entry:
122+
%y = alloca %twenty_bytes, align 4
123+
store i32 0, ptr %y, align 4
124+
%0 = getelementptr inbounds i8, ptr %y, i32 4
125+
store i32 1, ptr %0, align 4
126+
%1 = getelementptr inbounds i8, ptr %y, i32 8
127+
store i32 2, ptr %1, align 4
128+
%2 = getelementptr inbounds i8, ptr %y, i32 12
129+
store i32 3, ptr %2, align 4
130+
%3 = getelementptr inbounds i8, ptr %y, i32 16
131+
store i32 4, ptr %3, align 4
132+
musttail call void @large_callee(%twenty_bytes* byval(%twenty_bytes) align 4 %y)
133+
ret void
134+
}
135+
136+
declare void @two_byvals_callee(%twenty_bytes* byval(%twenty_bytes) align 4, %twenty_bytes* byval(%twenty_bytes) align 4)
137+
define void @swap_byvals(%twenty_bytes* byval(%twenty_bytes) align 4 %a, %twenty_bytes* byval(%twenty_bytes) align 4 %b) {
138+
; CHECK-LABEL: swap_byvals:
139+
; CHECK: # %bb.0: # %entry
140+
; CHECK-NEXT: mov eax, dword ptr [rsp + 24]
141+
; CHECK-NEXT: mov dword ptr [rsp - 8], eax
142+
; CHECK-NEXT: movaps xmm0, xmmword ptr [rsp + 8]
143+
; CHECK-NEXT: movaps xmmword ptr [rsp - 24], xmm0
144+
; CHECK-NEXT: mov eax, dword ptr [rsp + 48]
145+
; CHECK-NEXT: mov dword ptr [rsp - 32], eax
146+
; CHECK-NEXT: mov rax, qword ptr [rsp + 32]
147+
; CHECK-NEXT: mov rcx, qword ptr [rsp + 40]
148+
; CHECK-NEXT: mov qword ptr [rsp - 40], rcx
149+
; CHECK-NEXT: mov qword ptr [rsp - 48], rax
150+
; CHECK-NEXT: jmp two_byvals_callee@PLT # TAILCALL
151+
entry:
152+
musttail call void @two_byvals_callee(%twenty_bytes* byval(%twenty_bytes) align 4 %b, %twenty_bytes* byval(%twenty_bytes) align 4 %a)
153+
ret void
154+
}
155+
156+
; A forwarded byval arg, but at a different argument position. Because
157+
; x86 does not (partially) pass byval arguments in registers, the byval
158+
; arg is in the correct position already, so this is not a sibcall but
159+
; can be tail-call optimized.
160+
declare void @shift_byval_callee(%twenty_bytes* byval(%twenty_bytes) align 4)
161+
define void @shift_byval(i32 %a, %twenty_bytes* byval(%twenty_bytes) align 4 %b) {
162+
; CHECK-LABEL: shift_byval:
163+
; CHECK: # %bb.0: # %entry
164+
; CHECK-NEXT: jmp shift_byval_callee@PLT # TAILCALL
165+
entry:
166+
tail call void @shift_byval_callee(%twenty_bytes* byval(%twenty_bytes) align 4 %b)
167+
ret void
168+
}
169+
170+
; A global object passed to a byval argument, so it must be copied, but doesn't
171+
; need a stack temporary.
172+
@large_global = external global %twenty_bytes
173+
define void @large_caller_from_global(%twenty_bytes* byval(%twenty_bytes) align 4 %a) {
174+
; CHECK-LABEL: large_caller_from_global:
175+
; CHECK: # %bb.0: # %entry
176+
; CHECK-NEXT: mov rax, qword ptr [rip + large_global@GOTPCREL]
177+
; CHECK-NEXT: mov ecx, dword ptr [rax + 16]
178+
; CHECK-NEXT: mov dword ptr [rsp + 24], ecx
179+
; CHECK-NEXT: mov rcx, qword ptr [rax]
180+
; CHECK-NEXT: mov rax, qword ptr [rax + 8]
181+
; CHECK-NEXT: mov qword ptr [rsp + 16], rax
182+
; CHECK-NEXT: mov qword ptr [rsp + 8], rcx
183+
; CHECK-NEXT: jmp large_callee@PLT # TAILCALL
184+
entry:
185+
musttail call void @large_callee(%twenty_bytes* byval(%twenty_bytes) align 4 @large_global)
186+
ret void
187+
}

‎llvm/test/CodeGen/X86/tailcallbyval64.ll‎

Lines changed: 1 addition & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -5,8 +5,7 @@
55
; Expect the entry point.
66
; CHECK-LABEL: tailcaller:
77

8-
; Expect 2 rep;movs because of tail call byval lowering.
9-
; CHECK: rep;
8+
; Expect 1 rep;movs because of tail call stack argument lowering.
109
; CHECK: rep;
1110

1211
; A sequence of copyto/copyfrom virtual registers is used to deal with byval

‎llvm/test/CodeGen/X86/tailccbyval64.ll‎

Lines changed: 1 addition & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -5,8 +5,7 @@
55
; Expect the entry point.
66
; CHECK-LABEL: tailcaller:
77

8-
; Expect 2 rep;movs because of tail call byval lowering.
9-
; CHECK: rep;
8+
; Expect 1 rep;movs because of tail call stack argument lowering.
109
; CHECK: rep;
1110

1211
; A sequence of copyto/copyfrom virtual registers is used to deal with byval

0 commit comments

Comments
 (0)