reg,pass: refactor allocation of aliased registers (#121)

Issue #100 demonstrated that register allocation for aliased registers is
fundamentally broken. The root of the issue is that currently accesses to the
same virtual register with different masks are treated as different registers.
This PR takes a different approach:

* Liveness analysis is masked: we now properly consider which parts of a register are live
* Register allocation produces a mapping from virtual to physical ID, and aliasing is applied later

In addition, a new pass ZeroExtend32BitOutputs accounts for the fact that 32-bit writes in 64-bit mode should actually be treated as 64-bit writes (the result is zero-extended).

Closes #100
This commit is contained in:
Michael McLoughlin
2020-01-22 22:50:40 -08:00
committed by GitHub
parent 126469f13d
commit f40d602170
33 changed files with 1241 additions and 362 deletions

View File

@@ -0,0 +1,66 @@
// +build ignore
package main
import (
"strconv"
. "github.com/mmcloughlin/avo/build"
. "github.com/mmcloughlin/avo/operand"
. "github.com/mmcloughlin/avo/reg"
)
// The goal is to test for correct handling of 32-bit operands in 64-bit mode,
// specifically that writes are zero-extended to 64 bits. This test is
// constructed such that the register allocator would fail if this feature is
// not accounted for. It consists of multiple copies of a 32-bit write followed
// by a 64-bit read of the same register. Without special treatment liveness
// analysis would consider the upper 32 bits to still be live prior to the
// write. Therefore if we stack up enough copies of this, we could cause the
// register allocator to fail.
func main() {
const (
r = 14 // number of registers
m = 3 // number of iterations
n = r * m
)
TEXT("Upper32", NOSPLIT, "func() uint64")
Doc("Upper32 computes the sum 1+2+...+" + strconv.Itoa(n) + ".")
Comment("Initialize sum.")
s := GP64()
XORQ(s, s)
// Allocate n 64-bit registers and populate them.
Comment("Initialize registers.")
x := make([]GPVirtual, n)
for i := 0; i < n; i++ {
x[i] = GP64()
MOVQ(U64(0x9e77d78aacb8cbcc), x[i])
}
k := 0
for i := 0; i < m; i++ {
Commentf("Iteration %d.", i+1)
// Write to the 32-bit aliases of r registers.
for j := 0; j < r; j++ {
MOVL(U32(k+j+1), x[k+j].As32())
}
// Sum them up.
for j := 0; j < r; j++ {
ADDQ(x[k+j], s)
}
k += r
}
Comment("Store result and return.")
Store(s, ReturnIndex(0))
RET()
Generate()
}

View File

@@ -0,0 +1,2 @@
// Package upper32 tests liveness analysis of 32-bit operations on 64-bit registers.
package upper32

View File

@@ -0,0 +1,6 @@
// Code generated by command: go run asm.go -out upper32.s -stubs stub.go. DO NOT EDIT.
package upper32
// Upper32 computes the sum 1+2+...+42.
func Upper32() uint64

View File

@@ -0,0 +1,146 @@
// Code generated by command: go run asm.go -out upper32.s -stubs stub.go. DO NOT EDIT.
#include "textflag.h"
// func Upper32() uint64
TEXT ·Upper32(SB), NOSPLIT, $0-8
// Initialize sum.
XORQ AX, AX
// Initialize registers.
MOVQ $0x9e77d78aacb8cbcc, CX
MOVQ $0x9e77d78aacb8cbcc, DX
MOVQ $0x9e77d78aacb8cbcc, BX
MOVQ $0x9e77d78aacb8cbcc, BP
MOVQ $0x9e77d78aacb8cbcc, SI
MOVQ $0x9e77d78aacb8cbcc, DI
MOVQ $0x9e77d78aacb8cbcc, R8
MOVQ $0x9e77d78aacb8cbcc, R9
MOVQ $0x9e77d78aacb8cbcc, R10
MOVQ $0x9e77d78aacb8cbcc, R11
MOVQ $0x9e77d78aacb8cbcc, R12
MOVQ $0x9e77d78aacb8cbcc, R13
MOVQ $0x9e77d78aacb8cbcc, R14
MOVQ $0x9e77d78aacb8cbcc, R15
MOVQ $0x9e77d78aacb8cbcc, CX
MOVQ $0x9e77d78aacb8cbcc, DX
MOVQ $0x9e77d78aacb8cbcc, BX
MOVQ $0x9e77d78aacb8cbcc, BP
MOVQ $0x9e77d78aacb8cbcc, SI
MOVQ $0x9e77d78aacb8cbcc, DI
MOVQ $0x9e77d78aacb8cbcc, R8
MOVQ $0x9e77d78aacb8cbcc, R9
MOVQ $0x9e77d78aacb8cbcc, R10
MOVQ $0x9e77d78aacb8cbcc, R11
MOVQ $0x9e77d78aacb8cbcc, R12
MOVQ $0x9e77d78aacb8cbcc, R13
MOVQ $0x9e77d78aacb8cbcc, R14
MOVQ $0x9e77d78aacb8cbcc, R15
MOVQ $0x9e77d78aacb8cbcc, CX
MOVQ $0x9e77d78aacb8cbcc, DX
MOVQ $0x9e77d78aacb8cbcc, BX
MOVQ $0x9e77d78aacb8cbcc, BP
MOVQ $0x9e77d78aacb8cbcc, SI
MOVQ $0x9e77d78aacb8cbcc, DI
MOVQ $0x9e77d78aacb8cbcc, R8
MOVQ $0x9e77d78aacb8cbcc, R9
MOVQ $0x9e77d78aacb8cbcc, R10
MOVQ $0x9e77d78aacb8cbcc, R11
MOVQ $0x9e77d78aacb8cbcc, R12
MOVQ $0x9e77d78aacb8cbcc, R13
MOVQ $0x9e77d78aacb8cbcc, R14
MOVQ $0x9e77d78aacb8cbcc, R15
// Iteration 1.
MOVL $0x00000001, CX
MOVL $0x00000002, DX
MOVL $0x00000003, BX
MOVL $0x00000004, BP
MOVL $0x00000005, SI
MOVL $0x00000006, DI
MOVL $0x00000007, R8
MOVL $0x00000008, R9
MOVL $0x00000009, R10
MOVL $0x0000000a, R11
MOVL $0x0000000b, R12
MOVL $0x0000000c, R13
MOVL $0x0000000d, R14
MOVL $0x0000000e, R15
ADDQ CX, AX
ADDQ DX, AX
ADDQ BX, AX
ADDQ BP, AX
ADDQ SI, AX
ADDQ DI, AX
ADDQ R8, AX
ADDQ R9, AX
ADDQ R10, AX
ADDQ R11, AX
ADDQ R12, AX
ADDQ R13, AX
ADDQ R14, AX
ADDQ R15, AX
// Iteration 2.
MOVL $0x0000000f, CX
MOVL $0x00000010, DX
MOVL $0x00000011, BX
MOVL $0x00000012, BP
MOVL $0x00000013, SI
MOVL $0x00000014, DI
MOVL $0x00000015, R8
MOVL $0x00000016, R9
MOVL $0x00000017, R10
MOVL $0x00000018, R11
MOVL $0x00000019, R12
MOVL $0x0000001a, R13
MOVL $0x0000001b, R14
MOVL $0x0000001c, R15
ADDQ CX, AX
ADDQ DX, AX
ADDQ BX, AX
ADDQ BP, AX
ADDQ SI, AX
ADDQ DI, AX
ADDQ R8, AX
ADDQ R9, AX
ADDQ R10, AX
ADDQ R11, AX
ADDQ R12, AX
ADDQ R13, AX
ADDQ R14, AX
ADDQ R15, AX
// Iteration 3.
MOVL $0x0000001d, CX
MOVL $0x0000001e, DX
MOVL $0x0000001f, BX
MOVL $0x00000020, BP
MOVL $0x00000021, SI
MOVL $0x00000022, DI
MOVL $0x00000023, R8
MOVL $0x00000024, R9
MOVL $0x00000025, R10
MOVL $0x00000026, R11
MOVL $0x00000027, R12
MOVL $0x00000028, R13
MOVL $0x00000029, R14
MOVL $0x0000002a, R15
ADDQ CX, AX
ADDQ DX, AX
ADDQ BX, AX
ADDQ BP, AX
ADDQ SI, AX
ADDQ DI, AX
ADDQ R8, AX
ADDQ R9, AX
ADDQ R10, AX
ADDQ R11, AX
ADDQ R12, AX
ADDQ R13, AX
ADDQ R14, AX
ADDQ R15, AX
// Store result and return.
MOVQ AX, ret+0(FP)
RET

View File

@@ -0,0 +1,15 @@
package upper32
import (
"testing"
)
//go:generate go run asm.go -out upper32.s -stubs stub.go
func TestUpper32(t *testing.T) {
const n = 14 * 3
const expect = n * (n + 1) / 2
if got := Upper32(); got != expect {
t.Fatalf("Upper32() = %v; expect %v", got, expect)
}
}