llvm-6502/test/CodeGen/AArch64/func-argpassing.ll
Tim Northover 1dad6937c5 AArch64: don't be too greedy when folding :lo12: accesses into mem ops.
This frequently leads to cases like:
   ldr xD, [xN, :lo12:var]
   add xA, xN, :lo12:var
   ldr xD, [xA, #8]

where the ADD would have been needed anyway, and the two distinct addressing
modes can prevent the formation of an ldp. Because of how we handle ADRP
(aggressively forming an ADRP/ADD pseudo-inst at ISel time), this pattern also
results in duplicated ADRP instructions (one on its own to cover the ldr, and
one combined with the add).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@223172 91177308-0d34-0410-b5e6-96231b3b80d8
2014-12-02 23:13:39 +00:00

207 lines
7.0 KiB
LLVM

; RUN: llc -verify-machineinstrs < %s -mtriple=aarch64-none-linux-gnu | FileCheck --check-prefix=CHECK %s
; RUN: llc -verify-machineinstrs < %s -mtriple=aarch64-none-linux-gnu -mattr=-fp-armv8 | FileCheck --check-prefix=CHECK-NOFP %s
%myStruct = type { i64 , i8, i32 }
@var8 = global i8 0
@var32 = global i32 0
@var64 = global i64 0
@var128 = global i128 0
@varfloat = global float 0.0
@vardouble = global double 0.0
@varstruct = global %myStruct zeroinitializer
define void @take_i8s(i8 %val1, i8 %val2) {
; CHECK-LABEL: take_i8s:
store i8 %val2, i8* @var8
; Not using w1 may be technically allowed, but it would indicate a
; problem in itself.
; CHECK: strb w1, [{{x[0-9]+}}, {{#?}}:lo12:var8]
ret void
}
define void @add_floats(float %val1, float %val2) {
; CHECK-LABEL: add_floats:
%newval = fadd float %val1, %val2
; CHECK: fadd [[ADDRES:s[0-9]+]], s0, s1
; CHECK-NOFP-NOT: fadd
store float %newval, float* @varfloat
; CHECK: str [[ADDRES]], [{{x[0-9]+}}, {{#?}}:lo12:varfloat]
ret void
}
; byval pointers should be allocated to the stack and copied as if
; with memcpy.
define void @take_struct(%myStruct* byval %structval) {
; CHECK-LABEL: take_struct:
%addr0 = getelementptr %myStruct* %structval, i64 0, i32 2
%addr1 = getelementptr %myStruct* %structval, i64 0, i32 0
%val0 = load volatile i32* %addr0
; Some weird move means x0 is used for one access
; CHECK: ldr [[REG32:w[0-9]+]], [{{x[0-9]+|sp}}, #12]
store volatile i32 %val0, i32* @var32
; CHECK: str [[REG32]], [{{x[0-9]+}}, {{#?}}:lo12:var32]
%val1 = load volatile i64* %addr1
; CHECK: ldr [[REG64:x[0-9]+]], [{{x[0-9]+|sp}}]
store volatile i64 %val1, i64* @var64
; CHECK: str [[REG64]], [{{x[0-9]+}}, {{#?}}:lo12:var64]
ret void
}
; %structval should be at sp + 16
define void @check_byval_align(i32* byval %ignore, %myStruct* byval align 16 %structval) {
; CHECK-LABEL: check_byval_align:
%addr0 = getelementptr %myStruct* %structval, i64 0, i32 2
%addr1 = getelementptr %myStruct* %structval, i64 0, i32 0
%val0 = load volatile i32* %addr0
; Some weird move means x0 is used for one access
; CHECK: ldr [[REG32:w[0-9]+]], [sp, #28]
store i32 %val0, i32* @var32
; CHECK: str [[REG32]], [{{x[0-9]+}}, {{#?}}:lo12:var32]
%val1 = load volatile i64* %addr1
; CHECK: ldr [[REG64:x[0-9]+]], [sp, #16]
store i64 %val1, i64* @var64
; CHECK: str [[REG64]], [{{x[0-9]+}}, {{#?}}:lo12:var64]
ret void
}
define i32 @return_int() {
; CHECK-LABEL: return_int:
%val = load i32* @var32
ret i32 %val
; CHECK: ldr w0, [{{x[0-9]+}}, {{#?}}:lo12:var32]
; Make sure epilogue follows
; CHECK-NEXT: ret
}
define double @return_double() {
; CHECK-LABEL: return_double:
ret double 3.14
; CHECK: ldr d0, [{{x[0-9]+}}, {{#?}}:lo12:.LCPI
; CHECK-NOFP-NOT: ldr d0,
}
; This is the kind of IR clang will produce for returning a struct
; small enough to go into registers. Not all that pretty, but it
; works.
define [2 x i64] @return_struct() {
; CHECK-LABEL: return_struct:
%addr = bitcast %myStruct* @varstruct to [2 x i64]*
%val = load [2 x i64]* %addr
ret [2 x i64] %val
; CHECK: add x[[VARSTRUCT:[0-9]+]], {{x[0-9]+}}, :lo12:varstruct
; CHECK: ldp x0, x1, [x[[VARSTRUCT]]]
; Make sure epilogue immediately follows
; CHECK-NEXT: ret
}
; Large structs are passed by reference (storage allocated by caller
; to preserve value semantics) in x8. Strictly this only applies to
; structs larger than 16 bytes, but C semantics can still be provided
; if LLVM does it to %myStruct too. So this is the simplest check
define void @return_large_struct(%myStruct* sret %retval) {
; CHECK-LABEL: return_large_struct:
%addr0 = getelementptr %myStruct* %retval, i64 0, i32 0
%addr1 = getelementptr %myStruct* %retval, i64 0, i32 1
%addr2 = getelementptr %myStruct* %retval, i64 0, i32 2
store i64 42, i64* %addr0
store i8 2, i8* %addr1
store i32 9, i32* %addr2
; CHECK: str {{x[0-9]+}}, [x8]
; CHECK: strb {{w[0-9]+}}, [x8, #8]
; CHECK: str {{w[0-9]+}}, [x8, #12]
ret void
}
; This struct is just too far along to go into registers: (only x7 is
; available, but it needs two). Also make sure that %stacked doesn't
; sneak into x7 behind.
define i32 @struct_on_stack(i8 %var0, i16 %var1, i32 %var2, i64 %var3, i128 %var45,
i32* %var6, %myStruct* byval %struct, i32* byval %stacked,
double %notstacked) {
; CHECK-LABEL: struct_on_stack:
%addr = getelementptr %myStruct* %struct, i64 0, i32 0
%val64 = load volatile i64* %addr
store volatile i64 %val64, i64* @var64
; Currently nothing on local stack, so struct should be at sp
; CHECK: ldr [[VAL64:x[0-9]+]], [sp]
; CHECK: str [[VAL64]], [{{x[0-9]+}}, {{#?}}:lo12:var64]
store volatile double %notstacked, double* @vardouble
; CHECK-NOT: ldr d0
; CHECK: str d0, [{{x[0-9]+}}, {{#?}}:lo12:vardouble
; CHECK-NOFP-NOT: str d0,
%retval = load volatile i32* %stacked
ret i32 %retval
; CHECK-LE: ldr w0, [sp, #16]
}
define void @stacked_fpu(float %var0, double %var1, float %var2, float %var3,
float %var4, float %var5, float %var6, float %var7,
float %var8) {
; CHECK-LABEL: stacked_fpu:
store float %var8, float* @varfloat
; Beware as above: the offset would be different on big-endian
; machines if the first ldr were changed to use s-registers.
; CHECK: ldr {{[ds]}}[[VALFLOAT:[0-9]+]], [sp]
; CHECK: str s[[VALFLOAT]], [{{x[0-9]+}}, {{#?}}:lo12:varfloat]
ret void
}
; 128-bit integer types should be passed in xEVEN, xODD rather than
; the reverse. In this case x2 and x3. Nothing should use x1.
define i64 @check_i128_regalign(i32 %val0, i128 %val1, i64 %val2) {
; CHECK-LABEL: check_i128_regalign
store i128 %val1, i128* @var128
; CHECK: add x[[VAR128:[0-9]+]], {{x[0-9]+}}, :lo12:var128
; CHECK-DAG: stp x2, x3, [x[[VAR128]]]
ret i64 %val2
; CHECK: mov x0, x4
}
define void @check_i128_stackalign(i32 %val0, i32 %val1, i32 %val2, i32 %val3,
i32 %val4, i32 %val5, i32 %val6, i32 %val7,
i32 %stack1, i128 %stack2) {
; CHECK-LABEL: check_i128_stackalign
store i128 %stack2, i128* @var128
; Nothing local on stack in current codegen, so first stack is 16 away
; CHECK-LE: add x[[REG:[0-9]+]], sp, #16
; CHECK-LE: ldr {{x[0-9]+}}, [x[[REG]], #8]
; Important point is that we address sp+24 for second dword
; CHECK: ldp {{x[0-9]+}}, {{x[0-9]+}}, [sp, #16]
ret void
}
declare void @llvm.memcpy.p0i8.p0i8.i32(i8*, i8*, i32, i32, i1)
define i32 @test_extern() {
; CHECK-LABEL: test_extern:
call void @llvm.memcpy.p0i8.p0i8.i32(i8* undef, i8* undef, i32 undef, i32 4, i1 0)
; CHECK: bl memcpy
ret i32 0
}
; A sub-i32 stack argument must be loaded on big endian with ldr{h,b}, not just
; implicitly extended to a 32-bit load.
define i16 @stacked_i16(i32 %val0, i32 %val1, i32 %val2, i32 %val3,
i32 %val4, i32 %val5, i32 %val6, i32 %val7,
i16 %stack1) {
; CHECK-LABEL: stacked_i16
ret i16 %stack1
}