[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[PATCH v2 37/81] tcg/arm: Add full [US]XT[BH] into {s}extract
From: |
Richard Henderson |
Subject: |
[PATCH v2 37/81] tcg/arm: Add full [US]XT[BH] into {s}extract |
Date: |
Tue, 7 Jan 2025 00:00:28 -0800 |
The armv6 uxt and sxt opcodes have a 2-bit rotate field
which supports extractions from ofs = {0,8,16,24}.
Special case ofs = 0, len <= 8 as AND.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
---
tcg/arm/tcg-target-has.h | 17 +++++++++++++
tcg/arm/tcg-target.c.inc | 54 +++++++++++++++++++++++++++++++++++-----
2 files changed, 65 insertions(+), 6 deletions(-)
diff --git a/tcg/arm/tcg-target-has.h b/tcg/arm/tcg-target-has.h
index 316185500d..d6ca35ed1a 100644
--- a/tcg/arm/tcg-target-has.h
+++ b/tcg/arm/tcg-target-has.h
@@ -82,4 +82,21 @@ extern bool use_neon_instructions;
#define TCG_TARGET_HAS_cmpsel_vec 0
#define TCG_TARGET_HAS_tst_vec 1
+static inline bool
+tcg_target_extract_valid(TCGType type, unsigned ofs, unsigned len)
+{
+ if (use_armv7_instructions) {
+ return true; /* SBFX or UBFX */
+ }
+ switch (len) {
+ case 8: /* SXTB or UXTB */
+ case 16: /* SXTH or UXTH */
+ return (ofs % 8) == 0;
+ }
+ return false;
+}
+
+#define TCG_TARGET_extract_valid tcg_target_extract_valid
+#define TCG_TARGET_sextract_valid tcg_target_extract_valid
+
#endif
diff --git a/tcg/arm/tcg-target.c.inc b/tcg/arm/tcg-target.c.inc
index 9cfb733a14..12dad7307f 100644
--- a/tcg/arm/tcg-target.c.inc
+++ b/tcg/arm/tcg-target.c.inc
@@ -1036,19 +1036,61 @@ static void tcg_out_deposit(TCGContext *s, ARMCond
cond, TCGReg rd,
static void tcg_out_extract(TCGContext *s, ARMCond cond, TCGReg rd,
TCGReg rn, int ofs, int len)
{
- /* ubfx */
- tcg_out32(s, 0x07e00050 | (cond << 28) | (rd << 12) | rn
- | (ofs << 7) | ((len - 1) << 16));
+ /* According to gcc, AND can be faster. */
+ if (ofs == 0 && len <= 8) {
+ tcg_out_dat_imm(s, cond, ARITH_AND, rd, rn,
+ encode_imm_nofail((1 << len) - 1));
+ return;
+ }
+
+ if (use_armv7_instructions) {
+ /* ubfx */
+ tcg_out32(s, 0x07e00050 | (cond << 28) | (rd << 12) | rn
+ | (ofs << 7) | ((len - 1) << 16));
+ return;
+ }
+
+ assert(ofs % 8 == 0);
+ switch (len) {
+ case 8:
+ /* uxtb */
+ tcg_out32(s, 0x06ef0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
+ break;
+ case 16:
+ /* uxth */
+ tcg_out32(s, 0x06ff0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
+ break;
+ default:
+ g_assert_not_reached();
+ }
}
static void tcg_out_sextract(TCGContext *s, ARMCond cond, TCGReg rd,
TCGReg rn, int ofs, int len)
{
- /* sbfx */
- tcg_out32(s, 0x07a00050 | (cond << 28) | (rd << 12) | rn
- | (ofs << 7) | ((len - 1) << 16));
+ if (use_armv7_instructions) {
+ /* sbfx */
+ tcg_out32(s, 0x07a00050 | (cond << 28) | (rd << 12) | rn
+ | (ofs << 7) | ((len - 1) << 16));
+ return;
+ }
+
+ assert(ofs % 8 == 0);
+ switch (len) {
+ case 8:
+ /* sxtb */
+ tcg_out32(s, 0x06af0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
+ break;
+ case 16:
+ /* sxth */
+ tcg_out32(s, 0x06bf0070 | (cond << 28) | (rd << 12) | (ofs << 7) | rn);
+ break;
+ default:
+ g_assert_not_reached();
+ }
}
+
static void tcg_out_ld32u(TCGContext *s, ARMCond cond,
TCGReg rd, TCGReg rn, int32_t offset)
{
--
2.43.0
- Re: [PATCH v2 35/81] tcg/aarch64: Provide TCG_TARGET_{s}extract_valid, (continued)
- [PATCH v2 36/81] tcg/aarch64: Expand extract with offset 0 with andi, Richard Henderson, 2025/01/07
- [PATCH v2 38/81] tcg/loongarch64: Fold the ext{8, 16, 32}[us] cases into {s}extract, Richard Henderson, 2025/01/07
- [PATCH v2 12/81] target/i386: Use tcg_op_deposit_valid, Richard Henderson, 2025/01/07
- [PATCH v2 27/81] tcg: Pass type and flags to tcg_target_op_def, Richard Henderson, 2025/01/07
- [PATCH v2 30/81] tcg: Drop implementation checks from tcg-opc.h, Richard Henderson, 2025/01/07
- [PATCH v2 37/81] tcg/arm: Add full [US]XT[BH] into {s}extract,
Richard Henderson <=
- [PATCH v2 39/81] tcg/mips: Fold the ext{8, 16, 32}[us] cases into {s}extract, Richard Henderson, 2025/01/07
- [PATCH v2 41/81] tcg/riscv64: Fold the ext{8, 16, 32}[us] cases into {s}extract, Richard Henderson, 2025/01/07
- [PATCH v2 56/81] tcg/optimize: Fold andc with immediate to and, Richard Henderson, 2025/01/07
- [PATCH v2 55/81] tcg: Merge INDEX_op_and_{i32,i64}, Richard Henderson, 2025/01/07