189900eb14
This stops reporting CostPerUse 1 for `R8`-`R15` and `XMM8`-`XMM31`.
This was previously done because instruction encoding require a REX
prefix when using them resulting in longer instruction encodings. I
found that this regresses the quality of the register allocation as the
costs impose an ordering on eviction candidates. I also feel that there
is a bit of an impedance mismatch as the actual costs occure when
encoding instructions using those registers, but the order of VReg
assignments is not primarily ordered by number of Defs+Uses.
I did extensive measurements with the llvm-test-suite wiht SPEC2006 +
SPEC2017 included, internal services showed similar patterns. Generally
there are a log of improvements but also a lot of regression. But on
average the allocation quality seems to improve at a small code size
regression.
Results for measuring static and dynamic instruction counts:
Dynamic Counts (scaled by execution frequency) / Optimization Remarks:
Spills+FoldedSpills -5.6%
Reloads+FoldedReloads -4.2%
Copies -0.1%
Static / LLVM Statistics:
regalloc.NumSpills mean -1.6%, geomean -2.8%
regalloc.NumReloads mean -1.7%, geomean -3.1%
size..text mean +0.4%, geomean +0.4%
Static / LLVM Statistics:
mean -2.2%, geomean -3.1%) regalloc.NumSpills
mean -2.6%, geomean -3.9%) regalloc.NumReloads
mean +0.6%, geomean +0.6%) size..text
Static / LLVM Statistics:
regalloc.NumSpills mean -3.0%
regalloc.NumReloads mean -3.3%
size..text mean +0.3%, geomean +0.3%
Differential Revision: https://reviews.llvm.org/D133902
135 lines
8.2 KiB
LLVM
135 lines
8.2 KiB
LLVM
; RUN: llc -verify-machineinstrs -stack-symbol-ordering=0 < %s | FileCheck %s
|
|
|
|
target datalayout = "e-i64:64-f80:128-n8:16:32:64-S128"
|
|
target triple = "x86_64-pc-linux-gnu"
|
|
|
|
; This test is checking to make sure that we reuse the same stack slots
|
|
; for GC values spilled over two different call sites. Since the order
|
|
; of GC arguments differ, niave lowering code would insert loads and
|
|
; stores to rearrange items on the stack. We need to make sure (for
|
|
; performance) that this doesn't happen.
|
|
define i32 @back_to_back_calls(ptr addrspace(1) %a, ptr addrspace(1) %b, ptr addrspace(1) %c) #1 gc "statepoint-example" {
|
|
; CHECK-LABEL: back_to_back_calls
|
|
; The exact stores don't matter, but there need to be three stack slots created
|
|
; CHECK-DAG: movq %rdi, {{[0-9]*}}(%rsp)
|
|
; CHECK-DAG: movq %rdx, {{[0-9]*}}(%rsp)
|
|
; CHECK-DAG: movq %rsi, {{[0-9]*}}(%rsp)
|
|
; There should be no more than three moves
|
|
; CHECK-NOT: movq
|
|
%safepoint_token = tail call token (i64, i32, ptr, i32, i32, ...) @llvm.experimental.gc.statepoint.p0(i64 0, i32 0, ptr elementtype(void ()) undef, i32 0, i32 0, i32 0, i32 0) ["gc-live" (ptr addrspace(1) %a, ptr addrspace(1) %b, ptr addrspace(1) %c), "deopt" (i32 0, i32 -1, i32 0, i32 0, i32 0)]
|
|
%a1 = tail call coldcc ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token, i32 0, i32 0)
|
|
%b1 = tail call coldcc ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token, i32 0, i32 1)
|
|
%c1 = tail call coldcc ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token, i32 0, i32 2)
|
|
; CHECK: callq
|
|
; This is the key check. There should NOT be any memory moves here
|
|
; CHECK-NOT: movq
|
|
%safepoint_token2 = tail call token (i64, i32, ptr, i32, i32, ...) @llvm.experimental.gc.statepoint.p0(i64 0, i32 0, ptr elementtype(void ()) undef, i32 0, i32 0, i32 0, i32 0) ["gc-live" (ptr addrspace(1) %c1, ptr addrspace(1) %b1, ptr addrspace(1) %a1), "deopt" (i32 0, i32 -1, i32 0, i32 0, i32 0)]
|
|
%a2 = tail call coldcc ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token2, i32 0, i32 2)
|
|
%b2 = tail call coldcc ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token2, i32 0, i32 1)
|
|
%c2 = tail call coldcc ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token2, i32 0, i32 0)
|
|
; CHECK: callq
|
|
ret i32 1
|
|
}
|
|
|
|
; This test simply checks that minor changes in vm state don't prevent slots
|
|
; being reused for gc values.
|
|
define i32 @reserve_first(ptr addrspace(1) %a, ptr addrspace(1) %b, ptr addrspace(1) %c) #1 gc "statepoint-example" {
|
|
; CHECK-LABEL: reserve_first
|
|
; The exact stores don't matter, but there need to be three stack slots created
|
|
; CHECK-DAG: movq %rdi, {{[0-9]*}}(%rsp)
|
|
; CHECK-DAG: movq %rdx, {{[0-9]*}}(%rsp)
|
|
; CHECK-DAG: movq %rsi, {{[0-9]*}}(%rsp)
|
|
%safepoint_token = tail call token (i64, i32, ptr, i32, i32, ...) @llvm.experimental.gc.statepoint.p0(i64 0, i32 0, ptr elementtype(void ()) undef, i32 0, i32 0, i32 0, i32 0) ["gc-live" (ptr addrspace(1) %a, ptr addrspace(1) %b, ptr addrspace(1) %c), "deopt" (i32 0, i32 -1, i32 0, i32 0, i32 0)]
|
|
%a1 = tail call coldcc ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token, i32 0, i32 0)
|
|
%b1 = tail call coldcc ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token, i32 0, i32 1)
|
|
%c1 = tail call coldcc ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token, i32 0, i32 2)
|
|
; CHECK: callq
|
|
; This is the key check. There should NOT be any memory moves here
|
|
; CHECK-NOT: movq
|
|
%safepoint_token2 = tail call token (i64, i32, ptr, i32, i32, ...) @llvm.experimental.gc.statepoint.p0(i64 0, i32 0, ptr elementtype(void ()) undef, i32 0, i32 0, i32 0, i32 0) ["gc-live" (ptr addrspace(1) %c1, ptr addrspace(1) %b1, ptr addrspace(1) %a1), "deopt" (ptr addrspace(1) %a1, i32 0, ptr addrspace(1) %c1, i32 0, i32 0)]
|
|
%a2 = tail call coldcc ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token2, i32 0, i32 2)
|
|
%b2 = tail call coldcc ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token2, i32 0, i32 1)
|
|
%c2 = tail call coldcc ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token2, i32 0, i32 0)
|
|
; CHECK: callq
|
|
ret i32 1
|
|
}
|
|
|
|
; Check that we reuse the same stack slot across multiple calls. The use of
|
|
; more than two calls here is critical. We've had a bug which allowed reuse
|
|
; exactly once which went undetected for a long time.
|
|
define i32 @back_to_back_deopt(i32 %a, i32 %b, i32 %c) #1
|
|
gc "statepoint-example" {
|
|
; CHECK-LABEL: back_to_back_deopt
|
|
; The exact stores don't matter, but there need to be three stack slots created
|
|
; CHECK-DAG: movl %edi, 12(%rsp)
|
|
; CHECK-DAG: movl %esi, 8(%rsp)
|
|
; CHECK-DAG: movl %edx, 4(%rsp)
|
|
; CHECK: callq
|
|
; CHECK-DAG: movl %r14d, 12(%rsp)
|
|
; CHECK-DAG: movl %ebp, 8(%rsp)
|
|
; CHECK-DAG: movl %ebx, 4(%rsp)
|
|
; CHECK: callq
|
|
; CHECK-DAG: movl %r14d, 12(%rsp)
|
|
; CHECK-DAG: movl %ebp, 8(%rsp)
|
|
; CHECK-DAG: movl %ebx, 4(%rsp)
|
|
; CHECK: callq
|
|
; CHECK-DAG: movl %r14d, 12(%rsp)
|
|
; CHECK-DAG: movl %ebp, 8(%rsp)
|
|
; CHECK-DAG: movl %ebx, 4(%rsp)
|
|
; CHECK: callq
|
|
call token (i64, i32, ptr, i32, i32, ...) @llvm.experimental.gc.statepoint.p0(i64 0, i32 0, ptr elementtype(void ()) undef, i32 0, i32 0, i32 0, i32 0) ["deopt" (i32 %a, i32 %b, i32 %c)]
|
|
call token (i64, i32, ptr, i32, i32, ...) @llvm.experimental.gc.statepoint.p0(i64 0, i32 0, ptr elementtype(void ()) undef, i32 0, i32 0, i32 0, i32 0) ["deopt" (i32 %a, i32 %b, i32 %c)]
|
|
call token (i64, i32, ptr, i32, i32, ...) @llvm.experimental.gc.statepoint.p0(i64 0, i32 0, ptr elementtype(void ()) undef, i32 0, i32 0, i32 0, i32 0) ["deopt" (i32 %a, i32 %b, i32 %c)]
|
|
call token (i64, i32, ptr, i32, i32, ...) @llvm.experimental.gc.statepoint.p0(i64 0, i32 0, ptr elementtype(void ()) undef, i32 0, i32 0, i32 0, i32 0) ["deopt" (i32 %a, i32 %b, i32 %c)]
|
|
ret i32 1
|
|
}
|
|
|
|
; Test that stack slots are reused for invokes
|
|
define i32 @back_to_back_invokes(ptr addrspace(1) %a, ptr addrspace(1) %b, ptr addrspace(1) %c) #1 gc "statepoint-example" personality ptr @"personality_function" {
|
|
; CHECK-LABEL: back_to_back_invokes
|
|
entry:
|
|
; The exact stores don't matter, but there need to be three stack slots created
|
|
; CHECK-DAG: movq %rdi, {{[0-9]*}}(%rsp)
|
|
; CHECK-DAG: movq %rdx, {{[0-9]*}}(%rsp)
|
|
; CHECK-DAG: movq %rsi, {{[0-9]*}}(%rsp)
|
|
; CHECK: callq
|
|
%safepoint_token = invoke token (i64, i32, ptr, i32, i32, ...) @llvm.experimental.gc.statepoint.p0(i64 0, i32 0, ptr elementtype(void ()) undef, i32 0, i32 0, i32 0, i32 0) ["gc-live" (ptr addrspace(1) %a, ptr addrspace(1) %b, ptr addrspace(1) %c), "deopt" (i32 0, i32 -1, i32 0, i32 0, i32 0)]
|
|
to label %normal_return unwind label %exceptional_return
|
|
|
|
normal_return:
|
|
%a1 = tail call coldcc ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token, i32 0, i32 0)
|
|
%b1 = tail call coldcc ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token, i32 0, i32 1)
|
|
%c1 = tail call coldcc ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token, i32 0, i32 2)
|
|
; Should work even through bitcasts
|
|
; This is the key check. There should NOT be any memory moves here
|
|
; CHECK-NOT: movq
|
|
; CHECK: callq
|
|
%safepoint_token2 = invoke token (i64, i32, ptr, i32, i32, ...) @llvm.experimental.gc.statepoint.p0(i64 0, i32 0, ptr elementtype(void ()) undef, i32 0, i32 0, i32 0, i32 0) ["gc-live" (ptr addrspace(1) %c1, ptr addrspace(1) %b1, ptr addrspace(1) %a1), "deopt" (i32 0, i32 -1, i32 0, i32 0, i32 0)]
|
|
to label %normal_return2 unwind label %exceptional_return2
|
|
|
|
normal_return2:
|
|
%a2 = tail call coldcc ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token2, i32 0, i32 2)
|
|
%b2 = tail call coldcc ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token2, i32 0, i32 1)
|
|
%c2 = tail call coldcc ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token %safepoint_token2, i32 0, i32 0)
|
|
ret i32 1
|
|
|
|
exceptional_return:
|
|
%landing_pad = landingpad { ptr, i32 }
|
|
cleanup
|
|
ret i32 0
|
|
|
|
exceptional_return2:
|
|
%landing_pad2 = landingpad { ptr, i32 }
|
|
cleanup
|
|
ret i32 0
|
|
}
|
|
|
|
; Function Attrs: nounwind
|
|
declare ptr addrspace(1) @llvm.experimental.gc.relocate.p1(token, i32, i32) #3
|
|
|
|
declare token @llvm.experimental.gc.statepoint.p0(i64, i32, ptr, i32, i32, ...)
|
|
|
|
declare i32 @"personality_function"()
|
|
|
|
attributes #1 = { uwtable }
|