RISC-V: Improve li expansion for better code density.

li is a pseudo instruction in RISC-V, it might expand to more than one
instructions if the immediate value can't fit addi or lui, but the
assembler will always using 4-byte instructions during expansion.

For example:

 li a0, 0x12345001

will expand into

 12345537 lui a0,0x12345
 00150513 addi a0,a0,1

but addi could be compress into

 0505 addi a0,a0,1

It because load_const use macro_build to emit instructions,
and macro_build call append_insn, and expect it will compress
it if possible, but the fact is append_insn never compress anything,

So this patch redirect the li expansion flow to normal instruction
emission flow via md_assemble, added md_assemblef as an wrapper for
that for easier emit instruction with printf-style argument to build
instruction.

gas/ChangeLog:

	* tc-riscv.c (md_assemblef): New.
	(load_const) Use md_assemblef instead of macro_build to emit
	instructions.
	* testsuite/gas/riscv/li32.d: New.
	* testsuite/gas/riscv/li32.s: Ditto.
	* testsuite/gas/riscv/li64.d: Ditto.
	* testsuite/gas/riscv/li64.s: Ditto.
This commit is contained in:
Kito Cheng 2019-08-20 17:47:58 +08:00
parent 23c13d4299
commit db3b6ecc28
5 changed files with 108 additions and 5 deletions

View File

@ -928,6 +928,29 @@ macro_build (expressionS *ep, const char *name, const char *fmt, ...)
append_insn (&insn, ep, r);
}
/* Build an instruction created by a macro expansion. Like md_assemble but
accept a printf-style format string and arguments. */
static void
md_assemblef (const char *format, ...)
{
char *buf = NULL;
va_list ap;
int r;
va_start (ap, format);
r = vasprintf (&buf, format, ap);
if (r < 0)
as_fatal (_("internal error: vasprintf failed"));
md_assemble (buf);
free(buf);
va_end (ap);
}
/* Sign-extend 32-bit mode constants that have bit 31 set and all higher bits
unset. */
static void
@ -1013,6 +1036,7 @@ static void
load_const (int reg, expressionS *ep)
{
int shift = RISCV_IMM_BITS;
bfd_vma upper_imm;
expressionS upper = *ep, lower = *ep;
lower.X_add_number = (int32_t) ep->X_add_number << (32-shift) >> (32-shift);
upper.X_add_number -= lower.X_add_number;
@ -1032,9 +1056,10 @@ load_const (int reg, expressionS *ep)
upper.X_add_number = (int64_t) upper.X_add_number >> shift;
load_const (reg, &upper);
macro_build (NULL, "slli", "d,s,>", reg, reg, shift);
md_assemblef ("slli x%d, x%d, 0x%x", reg, reg, shift);
if (lower.X_add_number != 0)
macro_build (&lower, "addi", "d,s,j", reg, reg, BFD_RELOC_RISCV_LO12_I);
md_assemblef ("addi x%d, x%d, %" BFD_VMA_FMT "d", reg, reg,
lower.X_add_number);
}
else
{
@ -1043,13 +1068,16 @@ load_const (int reg, expressionS *ep)
if (upper.X_add_number != 0)
{
macro_build (ep, "lui", "d,u", reg, BFD_RELOC_RISCV_HI20);
/* Discard low part and zero-extend upper immediate. */
upper_imm = ((uint32_t)upper.X_add_number >> shift);
md_assemblef ("lui x%d, 0x%" BFD_VMA_FMT "x", reg, upper_imm);
hi_reg = reg;
}
if (lower.X_add_number != 0 || hi_reg == 0)
macro_build (ep, ADD32_INSN, "d,s,j", reg, hi_reg,
BFD_RELOC_RISCV_LO12_I);
md_assemblef ("%s x%d, x%d, %" BFD_VMA_FMT "d", ADD32_INSN, reg, hi_reg,
lower.X_add_number);
}
}

View File

@ -0,0 +1,17 @@
#as: -march=rv32ic -mabi=ilp32
#objdump: -dr
.*: file format elf32-littleriscv
Disassembly of section .text:
0+000 <target>:
[^:]+:[ ]+6521[ ]+lui[ ]+a0,0x8
[^:]+:[ ]+0505[ ]+addi[ ]+a0,a0,1
[^:]+:[ ]+6509[ ]+lui[ ]+a0,0x2
[^:]+:[ ]+f0150513[ ]+addi[ ]+a0,a0,-255 # .*
[^:]+:[ ]+12345537[ ]+lui[ ]+a0,0x12345
[^:]+:[ ]+0505[ ]+addi[ ]+a0,a0,1
[^:]+:[ ]+f2345537[ ]+lui[ ]+a0,0xf2345
[^:]+:[ ]+0505[ ]+addi[ ]+a0,a0,1

View File

@ -0,0 +1,5 @@
target:
li a0, 0x8001
li a0, 0x1f01
li a0, 0x12345001
li a0, 0xf2345001

View File

@ -0,0 +1,44 @@
#as: -march=rv64ic -mabi=lp64
#objdump: -dr
.*: file format elf64-littleriscv
Disassembly of section .text:
0000000000000000 <target>:
[^:]+:[ ]+6521[ ]+lui[ ]+a0,0x8
[^:]+:[ ]+2505[ ]+addiw[ ]+a0,a0,1
[^:]+:[ ]+6509[ ]+lui[ ]+a0,0x2
[^:]+:[ ]+f015051b[ ]+addiw[ ]+a0,a0,-255
[^:]+:[ ]+12345537[ ]+lui[ ]+a0,0x12345
[^:]+:[ ]+2505[ ]+addiw[ ]+a0,a0,1
[^:]+:[ ]+000f2537[ ]+lui[ ]+a0,0xf2
[^:]+:[ ]+3455051b[ ]+addiw[ ]+a0,a0,837
[^:]+:[ ]+0532[ ]+slli[ ]+a0,a0,0xc
[^:]+:[ ]+0505[ ]+addi[ ]+a0,a0,1
[^:]+:[ ]+00f12537[ ]+lui[ ]+a0,0xf12
[^:]+:[ ]+3455051b[ ]+addiw[ ]+a0,a0,837
[^:]+:[ ]+0532[ ]+slli[ ]+a0,a0,0xc
[^:]+:[ ]+0505[ ]+addi[ ]+a0,a0,1
[^:]+:[ ]+ff010537[ ]+lui[ ]+a0,0xff010
[^:]+:[ ]+f015051b[ ]+addiw[ ]+a0,a0,-255
[^:]+:[ ]+054e[ ]+slli[ ]+a0,a0,0x13
[^:]+:[ ]+80150513[ ]+addi[ ]+a0,a0,-2047 # .*
[^:]+:[ ]+0536[ ]+slli[ ]+a0,a0,0xd
[^:]+:[ ]+f0150513[ ]+addi[ ]+a0,a0,-255
[^:]+:[ ]+0010051b[ ]+addiw[ ]+a0,zero,1
[^:]+:[ ]+151a[ ]+slli[ ]+a0,a0,0x26
[^:]+:[ ]+1565[ ]+addi[ ]+a0,a0,-7
[^:]+:[ ]+0536[ ]+slli[ ]+a0,a0,0xd
[^:]+:[ ]+34550513[ ]+addi[ ]+a0,a0,837
[^:]+:[ ]+0532[ ]+slli[ ]+a0,a0,0xc
[^:]+:[ ]+0505[ ]+addi[ ]+a0,a0,1
[^:]+:[ ]+01fc4537[ ]+lui[ ]+a0,0x1fc4
[^:]+:[ ]+c915051b[ ]+addiw[ ]+a0,a0,-879
[^:]+:[ ]+0536[ ]+slli[ ]+a0,a0,0xd
[^:]+:[ ]+1565[ ]+addi[ ]+a0,a0,-7
[^:]+:[ ]+0536[ ]+slli[ ]+a0,a0,0xd
[^:]+:[ ]+34550513[ ]+addi[ ]+a0,a0,837 # .*
[^:]+:[ ]+0532[ ]+slli[ ]+a0,a0,0xc
[^:]+:[ ]+0505[ ]+addi[ ]+a0,a0,1

View File

@ -0,0 +1,9 @@
target:
li a0, 0x8001
li a0, 0x1f01
li a0, 0x12345001
li a0, 0xf2345001
li a0, 0xf12345001
li a0, 0xff00ff00ff001f01
li a0, 0x7ffffffff2345001
li a0, 0x7f0f243ff2345001