bd87a2449d
This function was added for ARM targets, but aligning global/stack pointer arguments passed to memcpy/memmove/memset can improve code size and performance for all targets that don't have fast unaligned accesses. This adds a generic implementation that adjusts the alignment to pointer size if unaligned accesses are slow. Review D134168 suggests that this significantly improves performance on synthetic benchmarks such as Dhrystone on RV32 as it avoids memcpy() calls. Reviewed By: efriedma Differential Revision: https://reviews.llvm.org/D134282
201 lines
7.4 KiB
LLVM
201 lines
7.4 KiB
LLVM
; RUN: llc < %s -asm-verbose=false -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mcpu=mvp -mattr=+bulk-memory | FileCheck %s --check-prefixes CHECK,BULK-MEM
|
|
; RUN: llc < %s -asm-verbose=false -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mcpu=mvp -mattr=-bulk-memory | FileCheck %s --check-prefixes CHECK,NO-BULK-MEM
|
|
|
|
; Test that basic bulk memory codegen works correctly
|
|
|
|
target triple = "wasm32-unknown-unknown"
|
|
|
|
declare void @llvm.memcpy.p0.p0.i8(ptr, ptr, i8, i1)
|
|
declare void @llvm.memcpy.p0.p0.i32(ptr, ptr, i32, i1)
|
|
|
|
declare void @llvm.memmove.p0.p0.i8(ptr, ptr, i8, i1)
|
|
declare void @llvm.memmove.p0.p0.i32(ptr, ptr, i32, i1)
|
|
|
|
declare void @llvm.memset.p0.i8(ptr, i8, i8, i1)
|
|
declare void @llvm.memset.p0.i32(ptr, i8, i32, i1)
|
|
|
|
; CHECK-LABEL: memcpy_i8:
|
|
; NO-BULK-MEM-NOT: memory.copy
|
|
; BULK-MEM-NEXT: .functype memcpy_i8 (i32, i32, i32) -> ()
|
|
; BULK-MEM-NEXT: memory.copy 0, 0, $0, $1, $2
|
|
; BULK-MEM-NEXT: return
|
|
define void @memcpy_i8(ptr %dest, ptr %src, i8 zeroext %len) {
|
|
call void @llvm.memcpy.p0.p0.i8(ptr %dest, ptr %src, i8 %len, i1 0)
|
|
ret void
|
|
}
|
|
|
|
; CHECK-LABEL: memmove_i8:
|
|
; NO-BULK-MEM-NOT: memory.copy
|
|
; BULK-MEM-NEXT: .functype memmove_i8 (i32, i32, i32) -> ()
|
|
; BULK-MEM-NEXT: memory.copy 0, 0, $0, $1, $2
|
|
; BULK-MEM-NEXT: return
|
|
define void @memmove_i8(ptr %dest, ptr %src, i8 zeroext %len) {
|
|
call void @llvm.memmove.p0.p0.i8(ptr %dest, ptr %src, i8 %len, i1 0)
|
|
ret void
|
|
}
|
|
|
|
; CHECK-LABEL: memset_i8:
|
|
; NO-BULK-MEM-NOT: memory.fill
|
|
; BULK-MEM-NEXT: .functype memset_i8 (i32, i32, i32) -> ()
|
|
; BULK-MEM-NEXT: memory.fill 0, $0, $1, $2
|
|
; BULK-MEM-NEXT: return
|
|
define void @memset_i8(ptr %dest, i8 %val, i8 zeroext %len) {
|
|
call void @llvm.memset.p0.i8(ptr %dest, i8 %val, i8 %len, i1 0)
|
|
ret void
|
|
}
|
|
|
|
; CHECK-LABEL: memcpy_i32:
|
|
; NO-BULK-MEM-NOT: memory.copy
|
|
; BULK-MEM-NEXT: .functype memcpy_i32 (i32, i32, i32) -> ()
|
|
; BULK-MEM-NEXT: memory.copy 0, 0, $0, $1, $2
|
|
; BULK-MEM-NEXT: return
|
|
define void @memcpy_i32(ptr %dest, ptr %src, i32 %len) {
|
|
call void @llvm.memcpy.p0.p0.i32(ptr %dest, ptr %src, i32 %len, i1 0)
|
|
ret void
|
|
}
|
|
|
|
; CHECK-LABEL: memmove_i32:
|
|
; NO-BULK-MEM-NOT: memory.copy
|
|
; BULK-MEM-NEXT: .functype memmove_i32 (i32, i32, i32) -> ()
|
|
; BULK-MEM-NEXT: memory.copy 0, 0, $0, $1, $2
|
|
; BULK-MEM-NEXT: return
|
|
define void @memmove_i32(ptr %dest, ptr %src, i32 %len) {
|
|
call void @llvm.memmove.p0.p0.i32(ptr %dest, ptr %src, i32 %len, i1 0)
|
|
ret void
|
|
}
|
|
|
|
; CHECK-LABEL: memset_i32:
|
|
; NO-BULK-MEM-NOT: memory.fill
|
|
; BULK-MEM-NEXT: .functype memset_i32 (i32, i32, i32) -> ()
|
|
; BULK-MEM-NEXT: memory.fill 0, $0, $1, $2
|
|
; BULK-MEM-NEXT: return
|
|
define void @memset_i32(ptr %dest, i8 %val, i32 %len) {
|
|
call void @llvm.memset.p0.i32(ptr %dest, i8 %val, i32 %len, i1 0)
|
|
ret void
|
|
}
|
|
|
|
; CHECK-LABEL: memcpy_1:
|
|
; CHECK-NEXT: .functype memcpy_1 (i32, i32) -> ()
|
|
; CHECK-NEXT: i32.load8_u $push[[L0:[0-9]+]]=, 0($1)
|
|
; CHECK-NEXT: i32.store8 0($0), $pop[[L0]]
|
|
; CHECK-NEXT: return
|
|
define void @memcpy_1(ptr %dest, ptr %src) {
|
|
call void @llvm.memcpy.p0.p0.i32(ptr %dest, ptr %src, i32 1, i1 0)
|
|
ret void
|
|
}
|
|
|
|
; CHECK-LABEL: memmove_1:
|
|
; CHECK-NEXT: .functype memmove_1 (i32, i32) -> ()
|
|
; CHECK-NEXT: i32.load8_u $push[[L0:[0-9]+]]=, 0($1)
|
|
; CHECK-NEXT: i32.store8 0($0), $pop[[L0]]
|
|
; CHECK-NEXT: return
|
|
define void @memmove_1(ptr %dest, ptr %src) {
|
|
call void @llvm.memmove.p0.p0.i32(ptr %dest, ptr %src, i32 1, i1 0)
|
|
ret void
|
|
}
|
|
|
|
; CHECK-LABEL: memset_1:
|
|
; NO-BULK-MEM-NOT: memory.fill
|
|
; BULK-MEM-NEXT: .functype memset_1 (i32, i32) -> ()
|
|
; BULK-MEM-NEXT: i32.store8 0($0), $1
|
|
; BULK-MEM-NEXT: return
|
|
define void @memset_1(ptr %dest, i8 %val) {
|
|
call void @llvm.memset.p0.i32(ptr %dest, i8 %val, i32 1, i1 0)
|
|
ret void
|
|
}
|
|
|
|
; CHECK-LABEL: memcpy_1024:
|
|
; NO-BULK-MEM-NOT: memory.copy
|
|
; BULK-MEM-NEXT: .functype memcpy_1024 (i32, i32) -> ()
|
|
; BULK-MEM-NEXT: i32.const $push[[L0:[0-9]+]]=, 1024
|
|
; BULK-MEM-NEXT: memory.copy 0, 0, $0, $1, $pop[[L0]]
|
|
; BULK-MEM-NEXT: return
|
|
define void @memcpy_1024(ptr %dest, ptr %src) {
|
|
call void @llvm.memcpy.p0.p0.i32(ptr %dest, ptr %src, i32 1024, i1 0)
|
|
ret void
|
|
}
|
|
|
|
; CHECK-LABEL: memmove_1024:
|
|
; NO-BULK-MEM-NOT: memory.copy
|
|
; BULK-MEM-NEXT: .functype memmove_1024 (i32, i32) -> ()
|
|
; BULK-MEM-NEXT: i32.const $push[[L0:[0-9]+]]=, 1024
|
|
; BULK-MEM-NEXT: memory.copy 0, 0, $0, $1, $pop[[L0]]
|
|
; BULK-MEM-NEXT: return
|
|
define void @memmove_1024(ptr %dest, ptr %src) {
|
|
call void @llvm.memmove.p0.p0.i32(ptr %dest, ptr %src, i32 1024, i1 0)
|
|
ret void
|
|
}
|
|
|
|
; CHECK-LABEL: memset_1024:
|
|
; NO-BULK-MEM-NOT: memory.fill
|
|
; BULK-MEM-NEXT: .functype memset_1024 (i32, i32) -> ()
|
|
; BULK-MEM-NEXT: i32.const $push[[L0:[0-9]+]]=, 1024
|
|
; BULK-MEM-NEXT: memory.fill 0, $0, $1, $pop[[L0]]
|
|
; BULK-MEM-NEXT: return
|
|
define void @memset_1024(ptr %dest, i8 %val) {
|
|
call void @llvm.memset.p0.i32(ptr %dest, i8 %val, i32 1024, i1 0)
|
|
ret void
|
|
}
|
|
|
|
; The following tests check that frame index elimination works for
|
|
; bulk memory instructions. The stack pointer is bumped by 112 instead
|
|
; of 100 because the stack pointer in WebAssembly is currently always
|
|
; 16-byte aligned, even in leaf functions, although it is not written
|
|
; back to the global in this case.
|
|
|
|
; TODO: Change TransientStackAlignment to 1 to avoid this extra
|
|
; arithmetic. This will require forcing the use of StackAlignment in
|
|
; PrologEpilogEmitter.cpp when
|
|
; WebAssemblyFrameLowering::needsSPWriteback would be true.
|
|
|
|
; CHECK-LABEL: memcpy_alloca_src:
|
|
; NO-BULK-MEM-NOT: memory.copy
|
|
; BULK-MEM-NEXT: .functype memcpy_alloca_src (i32) -> ()
|
|
; BULK-MEM-NEXT: global.get $push[[L0:[0-9]+]]=, __stack_pointer
|
|
; BULK-MEM-NEXT: i32.const $push[[L1:[0-9]+]]=, 112
|
|
; BULK-MEM-NEXT: i32.sub $push[[L2:[0-9]+]]=, $pop[[L0]], $pop[[L1]]
|
|
; BULK-MEM-NEXT: i32.const $push[[L3:[0-9]+]]=, 8
|
|
; BULK-MEM-NEXT: i32.add $push[[L4:[0-9]+]]=, $pop[[L2]], $pop[[L3]]
|
|
; BULK-MEM-NEXT: i32.const $push[[L5:[0-9]+]]=, 100
|
|
; BULK-MEM-NEXT: memory.copy 0, 0, $0, $pop[[L4]], $pop[[L5]]
|
|
; BULK-MEM-NEXT: return
|
|
define void @memcpy_alloca_src(ptr %dst) {
|
|
%a = alloca [100 x i8]
|
|
call void @llvm.memcpy.p0.p0.i32(ptr %dst, ptr %a, i32 100, i1 false)
|
|
ret void
|
|
}
|
|
|
|
; CHECK-LABEL: memcpy_alloca_dst:
|
|
; NO-BULK-MEM-NOT: memory.copy
|
|
; BULK-MEM-NEXT: .functype memcpy_alloca_dst (i32) -> ()
|
|
; BULK-MEM-NEXT: global.get $push[[L0:[0-9]+]]=, __stack_pointer
|
|
; BULK-MEM-NEXT: i32.const $push[[L1:[0-9]+]]=, 112
|
|
; BULK-MEM-NEXT: i32.sub $push[[L2:[0-9]+]]=, $pop[[L0]], $pop[[L1]]
|
|
; BULK-MEM-NEXT: i32.const $push[[L3:[0-9]+]]=, 8
|
|
; BULK-MEM-NEXT: i32.add $push[[L4:[0-9]+]]=, $pop[[L2]], $pop[[L3]]
|
|
; BULK-MEM-NEXT: i32.const $push[[L5:[0-9]+]]=, 100
|
|
; BULK-MEM-NEXT: memory.copy 0, 0, $pop[[L4]], $0, $pop[[L5]]
|
|
; BULK-MEM-NEXT: return
|
|
define void @memcpy_alloca_dst(ptr %src) {
|
|
%a = alloca [100 x i8]
|
|
call void @llvm.memcpy.p0.p0.i32(ptr %a, ptr %src, i32 100, i1 false)
|
|
ret void
|
|
}
|
|
|
|
; CHECK-LABEL: memset_alloca:
|
|
; NO-BULK-MEM-NOT: memory.fill
|
|
; BULK-MEM-NEXT: .functype memset_alloca (i32) -> ()
|
|
; BULK-MEM-NEXT: global.get $push[[L0:[0-9]+]]=, __stack_pointer
|
|
; BULK-MEM-NEXT: i32.const $push[[L1:[0-9]+]]=, 112
|
|
; BULK-MEM-NEXT: i32.sub $push[[L2:[0-9]+]]=, $pop[[L0]], $pop[[L1]]
|
|
; BULK-MEM-NEXT: i32.const $push[[L3:[0-9]+]]=, 8
|
|
; BULK-MEM-NEXT: i32.add $push[[L4:[0-9]+]]=, $pop[[L2]], $pop[[L3]]
|
|
; BULK-MEM-NEXT: i32.const $push[[L5:[0-9]+]]=, 100
|
|
; BULK-MEM-NEXT: memory.fill 0, $pop[[L4]], $0, $pop[[L5]]
|
|
; BULK-MEM-NEXT: return
|
|
define void @memset_alloca(i8 %val) {
|
|
%a = alloca [100 x i8]
|
|
call void @llvm.memset.p0.i32(ptr %a, i8 %val, i32 100, i1 false)
|
|
ret void
|
|
}
|