llvm-6502/test/CodeGen/AArch64/arm64-stp.ll
Chad Rosier e2e26b486d [AArch64] Remove an overly conservative check when generating store pairs.
Store instructions do not modify register values and therefore it's safe
to form a store pair even if the source register has been read in between
the two store instructions.

Previously, the read of w1 (see below) prevented the formation of a stp.

        str      w0, [x2]
        ldr     w8, [x2, #8]
        add      w0, w8, w1
        str     w1, [x2, #4]
        ret

We now generate the following code.

        stp      w0, w1, [x2]
        ldr     w8, [x2, #8]
        add      w0, w8, w1
        ret

All correctness tests with -Ofast on A57 with Spec200x and EEMBC pass.
Performance results for SPEC2K were within noise.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@239432 91177308-0d34-0410-b5e6-96231b3b80d8
2015-06-09 20:59:41 +00:00

134 lines
4.4 KiB
LLVM

; RUN: llc < %s -march=arm64 -aarch64-stp-suppress=false -verify-machineinstrs -mcpu=cyclone | FileCheck %s
; RUN: llc < %s -march=arm64 -aarch64-unscaled-mem-op=true\
; RUN: -verify-machineinstrs -mcpu=cyclone | FileCheck -check-prefix=STUR_CHK %s
; CHECK: stp_int
; CHECK: stp w0, w1, [x2]
define void @stp_int(i32 %a, i32 %b, i32* nocapture %p) nounwind {
store i32 %a, i32* %p, align 4
%add.ptr = getelementptr inbounds i32, i32* %p, i64 1
store i32 %b, i32* %add.ptr, align 4
ret void
}
; CHECK: stp_long
; CHECK: stp x0, x1, [x2]
define void @stp_long(i64 %a, i64 %b, i64* nocapture %p) nounwind {
store i64 %a, i64* %p, align 8
%add.ptr = getelementptr inbounds i64, i64* %p, i64 1
store i64 %b, i64* %add.ptr, align 8
ret void
}
; CHECK: stp_float
; CHECK: stp s0, s1, [x0]
define void @stp_float(float %a, float %b, float* nocapture %p) nounwind {
store float %a, float* %p, align 4
%add.ptr = getelementptr inbounds float, float* %p, i64 1
store float %b, float* %add.ptr, align 4
ret void
}
; CHECK: stp_double
; CHECK: stp d0, d1, [x0]
define void @stp_double(double %a, double %b, double* nocapture %p) nounwind {
store double %a, double* %p, align 8
%add.ptr = getelementptr inbounds double, double* %p, i64 1
store double %b, double* %add.ptr, align 8
ret void
}
; Test the load/store optimizer---combine ldurs into a ldp, if appropriate
define void @stur_int(i32 %a, i32 %b, i32* nocapture %p) nounwind {
; STUR_CHK: stur_int
; STUR_CHK: stp w{{[0-9]+}}, {{w[0-9]+}}, [x{{[0-9]+}}, #-8]
; STUR_CHK-NEXT: ret
%p1 = getelementptr inbounds i32, i32* %p, i32 -1
store i32 %a, i32* %p1, align 2
%p2 = getelementptr inbounds i32, i32* %p, i32 -2
store i32 %b, i32* %p2, align 2
ret void
}
define void @stur_long(i64 %a, i64 %b, i64* nocapture %p) nounwind {
; STUR_CHK: stur_long
; STUR_CHK: stp x{{[0-9]+}}, {{x[0-9]+}}, [x{{[0-9]+}}, #-16]
; STUR_CHK-NEXT: ret
%p1 = getelementptr inbounds i64, i64* %p, i32 -1
store i64 %a, i64* %p1, align 2
%p2 = getelementptr inbounds i64, i64* %p, i32 -2
store i64 %b, i64* %p2, align 2
ret void
}
define void @stur_float(float %a, float %b, float* nocapture %p) nounwind {
; STUR_CHK: stur_float
; STUR_CHK: stp s{{[0-9]+}}, {{s[0-9]+}}, [x{{[0-9]+}}, #-8]
; STUR_CHK-NEXT: ret
%p1 = getelementptr inbounds float, float* %p, i32 -1
store float %a, float* %p1, align 2
%p2 = getelementptr inbounds float, float* %p, i32 -2
store float %b, float* %p2, align 2
ret void
}
define void @stur_double(double %a, double %b, double* nocapture %p) nounwind {
; STUR_CHK: stur_double
; STUR_CHK: stp d{{[0-9]+}}, {{d[0-9]+}}, [x{{[0-9]+}}, #-16]
; STUR_CHK-NEXT: ret
%p1 = getelementptr inbounds double, double* %p, i32 -1
store double %a, double* %p1, align 2
%p2 = getelementptr inbounds double, double* %p, i32 -2
store double %b, double* %p2, align 2
ret void
}
define void @splat_v4i32(i32 %v, i32 *%p) {
entry:
; CHECK-LABEL: splat_v4i32
; CHECK-DAG: stp w0, w0, [x1]
; CHECK-DAG: stp w0, w0, [x1, #8]
; CHECK: ret
%p17 = insertelement <4 x i32> undef, i32 %v, i32 0
%p18 = insertelement <4 x i32> %p17, i32 %v, i32 1
%p19 = insertelement <4 x i32> %p18, i32 %v, i32 2
%p20 = insertelement <4 x i32> %p19, i32 %v, i32 3
%p21 = bitcast i32* %p to <4 x i32>*
store <4 x i32> %p20, <4 x i32>* %p21, align 4
ret void
}
; Read of %b to compute %tmp2 shouldn't prevent formation of stp
; CHECK-LABEL: stp_int_rar_hazard
; CHECK: stp w0, w1, [x2]
; CHECK: ldr [[REG:w[0-9]+]], [x2, #8]
; CHECK: add w0, [[REG]], w1
; CHECK: ret
define i32 @stp_int_rar_hazard(i32 %a, i32 %b, i32* nocapture %p) nounwind {
store i32 %a, i32* %p, align 4
%ld.ptr = getelementptr inbounds i32, i32* %p, i64 2
%tmp = load i32, i32* %ld.ptr, align 4
%tmp2 = add i32 %tmp, %b
%add.ptr = getelementptr inbounds i32, i32* %p, i64 1
store i32 %b, i32* %add.ptr, align 4
ret i32 %tmp2
}
; Read of %b to compute %tmp2 shouldn't prevent formation of stp
; CHECK-LABEL: stp_int_rar_hazard_after
; CHECK: ldr [[REG:w[0-9]+]], [x3, #4]
; CHECK: add w0, [[REG]], w2
; CHECK: stp w1, w2, [x3]
; CHECK: ret
define i32 @stp_int_rar_hazard_after(i32 %w0, i32 %a, i32 %b, i32* nocapture %p) nounwind {
store i32 %a, i32* %p, align 4
%ld.ptr = getelementptr inbounds i32, i32* %p, i64 1
%tmp = load i32, i32* %ld.ptr, align 4
%tmp2 = add i32 %tmp, %b
%add.ptr = getelementptr inbounds i32, i32* %p, i64 1
store i32 %b, i32* %add.ptr, align 4
ret i32 %tmp2
}