Skip to contentSkip to conversation

target-arm improvments for aarch64

PATCH Applied Richard Henderson 2 participants 13 messages Sep 10, 2015 – Sep 11, 2015

lore
Share & tools
lore
#0 lore raw

v3 cover target-arm improvments for aarch64

Updated from v2 based on review from Peter.

I did go ahead with the use of andc for translating ccmp; the result looked pretty good with Haswell's andn instruction.

r~

Richard Henderson (11):
target-arm: Share all common TCG temporaries
target-arm: Introduce DisasCompare
target-arm: Handle always condition codes within arm_test_cc
target-arm: Use setcond and movcond for csel
target-arm: Implement ccmp branchless
target-arm: Implement fcsel with movcond
target-arm: Recognize SXTB, SXTH, SXTW, ASR
target-arm: Recognize UXTB, UXTH, LSR, LSL
target-arm: Eliminate unnecessary zero-extend in disas_bitfield
target-arm: Recognize ROR
target-arm: Use tcg_gen_extrh_i64_i32

 target-arm/translate-a64.c | 340 ++++++++++++++++++++++++++-------------------
 target-arm/translate.c     | 134 +++++++++++-------
 target-arm/translate.h     |  17 +++
 3 files changed, 299 insertions(+), 192 deletions(-)
Signature · 1 line
2.4.3
#1 lore raw

PATCH v3 1/11 target-arm: Share all common TCG temporaries

This is a bug fix for aarch64. At present, we have branches using the 32-bit (translate.c) versions of cpu_[NZCV]F, but we set the flags using the 64-bit (translate-a64.c) versions of cpu_[NZCV]F. From the view of the TCG code generator, these are unrelated variables.

The bug is hard to see because we currently only read these variables from branches, and upon reaching a branch TCG will first spill live variables and then reload the arguments of the branch. Since the 32-bit versions were never live until reaching the branch, we'd re-read the data that had just been spilled from the 64-bit versions.

There is currently no such problem with the cpu_exclusive_* variables, but there's no point in tempting fate.

---
 target-arm/translate-a64.c | 22 ----------------------
 target-arm/translate.c     | 10 +++++-----
 target-arm/translate.h     |  8 ++++++++
 3 files changed, 13 insertions(+), 27 deletions(-)
target-arm/translate-a64.c +0 -22
index 529bb0c..342bdbd 100644
@@ -40,16 +40,9 @@
 
 static TCGv_i64 cpu_X[32];
 static TCGv_i64 cpu_pc;
-static TCGv_i32 cpu_NF, cpu_ZF, cpu_CF, cpu_VF;
 
 /* Load/store exclusive handling */
-static TCGv_i64 cpu_exclusive_addr;
-static TCGv_i64 cpu_exclusive_val;
 static TCGv_i64 cpu_exclusive_high;
-#ifdef CONFIG_USER_ONLY
-static TCGv_i64 cpu_exclusive_test;
-static TCGv_i32 cpu_exclusive_info;
-#endif
 
 static const char *regnames[] = {
     "x0", "x1", "x2", "x3", "x4", "x5", "x6", "x7",
@@ -105,23 +98,8 @@ void a64_translate_init(void)
                                           regnames[i]);
     }
 
-    cpu_NF = tcg_global_mem_new_i32(TCG_AREG0, offsetof(CPUARMState, NF), "NF");
-    cpu_ZF = tcg_global_mem_new_i32(TCG_AREG0, offsetof(CPUARMState, ZF), "ZF");
-    cpu_CF = tcg_global_mem_new_i32(TCG_AREG0, offsetof(CPUARMState, CF), "CF");
-    cpu_VF = tcg_global_mem_new_i32(TCG_AREG0, offsetof(CPUARMState, VF), "VF");
-
-    cpu_exclusive_addr = tcg_global_mem_new_i64(TCG_AREG0,
-        offsetof(CPUARMState, exclusive_addr), "exclusive_addr");
-    cpu_exclusive_val = tcg_global_mem_new_i64(TCG_AREG0,
-        offsetof(CPUARMState, exclusive_val), "exclusive_val");
     cpu_exclusive_high = tcg_global_mem_new_i64(TCG_AREG0,
         offsetof(CPUARMState, exclusive_high), "exclusive_high");
-#ifdef CONFIG_USER_ONLY
-    cpu_exclusive_test = tcg_global_mem_new_i64(TCG_AREG0,
-        offsetof(CPUARMState, exclusive_test), "exclusive_test");
-    cpu_exclusive_info = tcg_global_mem_new_i32(TCG_AREG0,
-        offsetof(CPUARMState, exclusive_info), "exclusive_info");
-#endif
 }
 
 static inline ARMMMUIdx get_a64_user_mem_index(DisasContext *s)
target-arm/translate.c +5 -5
index e27634f..3826a02 100644
@@ -64,12 +64,12 @@ TCGv_ptr cpu_env;
 /* We reuse the same 64-bit temporaries for efficiency.  */
 static TCGv_i64 cpu_V0, cpu_V1, cpu_M0;
 static TCGv_i32 cpu_R[16];
-static TCGv_i32 cpu_CF, cpu_NF, cpu_VF, cpu_ZF;
-static TCGv_i64 cpu_exclusive_addr;
-static TCGv_i64 cpu_exclusive_val;
+TCGv_i32 cpu_CF, cpu_NF, cpu_VF, cpu_ZF;
+TCGv_i64 cpu_exclusive_addr;
+TCGv_i64 cpu_exclusive_val;
 #ifdef CONFIG_USER_ONLY
-static TCGv_i64 cpu_exclusive_test;
-static TCGv_i32 cpu_exclusive_info;
+TCGv_i64 cpu_exclusive_test;
+TCGv_i32 cpu_exclusive_info;
 #endif
 
 /* FIXME:  These should be removed.  */
target-arm/translate.h +8 -0
index 9ab978f..679bdbc 100644
@@ -62,7 +62,15 @@ typedef struct DisasContext {
     TCGv_i64 tmp_a64[TMP_A64_MAX];
 } DisasContext;
 
+/* Share the TCG temporaries common between 32 and 64 bit modes.  */
 extern TCGv_ptr cpu_env;
+extern TCGv_i32 cpu_NF, cpu_ZF, cpu_CF, cpu_VF;
+extern TCGv_i64 cpu_exclusive_addr;
+extern TCGv_i64 cpu_exclusive_val;
+#ifdef CONFIG_USER_ONLY
+extern TCGv_i64 cpu_exclusive_test;
+extern TCGv_i32 cpu_exclusive_info;
+#endif
 
 static inline int arm_dc_feature(DisasContext *dc, int feature)
 {
Signature · 1 line
2.4.3
#2 lore raw

PATCH v3 2/11 target-arm: Introduce DisasCompare

In reply to Richard Henderson

Split arm_gen_test_cc into 3 functions, so that it can be reused for non-branch TCG comparisons.

---
 target-arm/translate.c | 115 +++++++++++++++++++++++++++++--------------------
 target-arm/translate.h |   9 ++++
 2 files changed, 78 insertions(+), 46 deletions(-)
target-arm/translate.c +69 -46
index 3826a02..52452cf 100644
@@ -738,81 +738,104 @@ static void gen_thumb2_parallel_addsub(int op1, int op2, TCGv_i32 a, TCGv_i32 b)
 #undef PAS_OP
 
 /*
- * generate a conditional branch based on ARM condition code cc.
+ * Generate a conditional based on ARM condition code cc.
  * This is common between ARM and Aarch64 targets.
  */
-void arm_gen_test_cc(int cc, TCGLabel *label)
+void arm_test_cc(DisasCompare *cmp, int cc)
 {
-    TCGv_i32 tmp;
-    TCGLabel *inv;
+    TCGv_i32 value;
+    TCGCond cond;
+    bool global = true;
 
     switch (cc) {
     case 0: /* eq: Z */
-        tcg_gen_brcondi_i32(TCG_COND_EQ, cpu_ZF, 0, label);
-        break;
     case 1: /* ne: !Z */
-        tcg_gen_brcondi_i32(TCG_COND_NE, cpu_ZF, 0, label);
+        cond = TCG_COND_EQ;
+        value = cpu_ZF;
         break;
+
     case 2: /* cs: C */
-        tcg_gen_brcondi_i32(TCG_COND_NE, cpu_CF, 0, label);
-        break;
     case 3: /* cc: !C */
-        tcg_gen_brcondi_i32(TCG_COND_EQ, cpu_CF, 0, label);
+        cond = TCG_COND_NE;
+        value = cpu_CF;
         break;
+
     case 4: /* mi: N */
-        tcg_gen_brcondi_i32(TCG_COND_LT, cpu_NF, 0, label);
-        break;
     case 5: /* pl: !N */
-        tcg_gen_brcondi_i32(TCG_COND_GE, cpu_NF, 0, label);
+        cond = TCG_COND_LT;
+        value = cpu_NF;
         break;
+
     case 6: /* vs: V */
-        tcg_gen_brcondi_i32(TCG_COND_LT, cpu_VF, 0, label);
-        break;
     case 7: /* vc: !V */
-        tcg_gen_brcondi_i32(TCG_COND_GE, cpu_VF, 0, label);
+        cond = TCG_COND_LT;
+        value = cpu_VF;
         break;
+
     case 8: /* hi: C && !Z */
-        inv = gen_new_label();
-        tcg_gen_brcondi_i32(TCG_COND_EQ, cpu_CF, 0, inv);
-        tcg_gen_brcondi_i32(TCG_COND_NE, cpu_ZF, 0, label);
-        gen_set_label(inv);
-        break;
-    case 9: /* ls: !C || Z */
-        tcg_gen_brcondi_i32(TCG_COND_EQ, cpu_CF, 0, label);
-        tcg_gen_brcondi_i32(TCG_COND_EQ, cpu_ZF, 0, label);
+    case 9: /* ls: !C || Z -> !(C && !Z) */
+        cond = TCG_COND_NE;
+        value = tcg_temp_new_i32();
+        global = false;
+        /* CF is 1 for C, so -CF is an all-bits-set mask for C;
+           ZF is non-zero for !Z; so AND the two subexpressions.  */
+        tcg_gen_neg_i32(value, cpu_CF);
+        tcg_gen_and_i32(value, value, cpu_ZF);
         break;
+
     case 10: /* ge: N == V -> N ^ V == 0 */
-        tmp = tcg_temp_new_i32();
-        tcg_gen_xor_i32(tmp, cpu_VF, cpu_NF);
-        tcg_gen_brcondi_i32(TCG_COND_GE, tmp, 0, label);
-        tcg_temp_free_i32(tmp);
-        break;
     case 11: /* lt: N != V -> N ^ V != 0 */
-        tmp = tcg_temp_new_i32();
-        tcg_gen_xor_i32(tmp, cpu_VF, cpu_NF);
-        tcg_gen_brcondi_i32(TCG_COND_LT, tmp, 0, label);
-        tcg_temp_free_i32(tmp);
+        /* Since we're only interested in the sign bit, == 0 is >= 0.  */
+        cond = TCG_COND_GE;
+        value = tcg_temp_new_i32();
+        global = false;
+        tcg_gen_xor_i32(value, cpu_VF, cpu_NF);
         break;
+
     case 12: /* gt: !Z && N == V */
-        inv = gen_new_label();
-        tcg_gen_brcondi_i32(TCG_COND_EQ, cpu_ZF, 0, inv);
-        tmp = tcg_temp_new_i32();
-        tcg_gen_xor_i32(tmp, cpu_VF, cpu_NF);
-        tcg_gen_brcondi_i32(TCG_COND_GE, tmp, 0, label);
-        tcg_temp_free_i32(tmp);
-        gen_set_label(inv);
-        break;
     case 13: /* le: Z || N != V */
-        tcg_gen_brcondi_i32(TCG_COND_EQ, cpu_ZF, 0, label);
-        tmp = tcg_temp_new_i32();
-        tcg_gen_xor_i32(tmp, cpu_VF, cpu_NF);
-        tcg_gen_brcondi_i32(TCG_COND_LT, tmp, 0, label);
-        tcg_temp_free_i32(tmp);
+        cond = TCG_COND_NE;
+        value = tcg_temp_new_i32();
+        global = false;
+        /* (N == V) is equal to the sign bit of ~(NF ^ VF).  Propagate
+         * the sign bit then AND with ZF to yield the result.  */
+        tcg_gen_xor_i32(value, cpu_VF, cpu_NF);
+        tcg_gen_sari_i32(value, value, 31);
+        tcg_gen_andc_i32(value, cpu_ZF, value);
         break;
+
     default:
         fprintf(stderr, "Bad condition code 0x%x\n", cc);
         abort();
     }
+
+    if (cc & 1) {
+        cond = tcg_invert_cond(cond);
+    }
+
+    cmp->cond = cond;
+    cmp->value = value;
+    cmp->value_global = global;
+}
+
+void arm_free_cc(DisasCompare *cmp)
+{
+    if (!cmp->value_global) {
+        tcg_temp_free_i32(cmp->value);
+    }
+}
+
+void arm_jump_cc(DisasCompare *cmp, TCGLabel *label)
+{
+    tcg_gen_brcondi_i32(cmp->cond, cmp->value, 0, label);
+}
+
+void arm_gen_test_cc(int cc, TCGLabel *label)
+{
+    DisasCompare cmp;
+    arm_test_cc(&cmp, cc);
+    arm_jump_cc(&cmp, label);
+    arm_free_cc(&cmp);
 }
 
 static const uint8_t table_logic_cc[16] = {
target-arm/translate.h +9 -0
index 679bdbc..2f30862 100644
@@ -62,6 +62,12 @@ typedef struct DisasContext {
     TCGv_i64 tmp_a64[TMP_A64_MAX];
 } DisasContext;
 
+typedef struct DisasCompare {
+    TCGCond cond;
+    TCGv_i32 value;
+    bool value_global;
+} DisasCompare;
+
 /* Share the TCG temporaries common between 32 and 64 bit modes.  */
 extern TCGv_ptr cpu_env;
 extern TCGv_i32 cpu_NF, cpu_ZF, cpu_CF, cpu_VF;
@@ -143,6 +149,9 @@ static inline void aarch64_cpu_dump_state(CPUState *cs, FILE *f,
 }
 #endif
 
+void arm_test_cc(DisasCompare *cmp, int cc);
+void arm_free_cc(DisasCompare *cmp);
+void arm_jump_cc(DisasCompare *cmp, TCGLabel *label);
 void arm_gen_test_cc(int cc, TCGLabel *label);
 
 #endif /* TARGET_ARM_TRANSLATE_H */
Signature · 1 line
2.4.3
#3 lore raw

PATCH v3 3/11 target-arm: Handle always condition codes within arm_test_cc

In reply to Richard Henderson

Handling this with TCG_COND_ALWAYS will allow these unlikely cases to be handled without special cases in the rest of the translator. The TCG optimizer ought to be able to reduce these ALWAYS conditions completely.

---
 target-arm/translate.c | 9 +++++++++
 1 file changed, 9 insertions(+)
target-arm/translate.c +9 -0
index 52452cf..3e7c367 100644
@@ -804,6 +804,14 @@ void arm_test_cc(DisasCompare *cmp, int cc)
         tcg_gen_andc_i32(value, cpu_ZF, value);
         break;
 
+    case 14: /* always */
+    case 15: /* always */
+        /* Use the ALWAYS condition, which will fold early.
+         * It doesn't matter what we use for the value.  */
+        cond = TCG_COND_ALWAYS;
+        value = cpu_ZF;
+        goto no_invert;
+
     default:
         fprintf(stderr, "Bad condition code 0x%x\n", cc);
         abort();
@@ -813,6 +821,7 @@ void arm_test_cc(DisasCompare *cmp, int cc)
         cond = tcg_invert_cond(cond);
     }
 
+ no_invert:
     cmp->cond = cond;
     cmp->value = value;
     cmp->value_global = global;
Signature · 1 line
2.4.3
#4 lore raw

PATCH v3 4/11 target-arm: Use setcond and movcond for csel

In reply to Richard Henderson
---
 target-arm/translate-a64.c | 85 ++++++++++++++++++++++++++--------------------
 1 file changed, 49 insertions(+), 36 deletions(-)
target-arm/translate-a64.c +49 -36
index 342bdbd..f941fc3 100644
@@ -167,6 +167,31 @@ void gen_a64_set_pc_im(uint64_t val)
     tcg_gen_movi_i64(cpu_pc, val);
 }
 
+typedef struct DisasCompare64 {
+    TCGCond cond;
+    TCGv_i64 value;
+} DisasCompare64;
+
+static void a64_test_cc(DisasCompare64 *c64, int cc)
+{
+    DisasCompare c32;
+
+    arm_test_cc(&c32, cc);
+
+    /* Sign-extend the 32-bit value so that the GE/LT comparisons work
+       * properly.  The NE/EQ comparisons are also fine with this choice.  */
+    c64->cond = c32.cond;
+    c64->value = tcg_temp_new_i64();
+    tcg_gen_ext_i32_i64(c64->value, c32.value);
+
+    arm_free_cc(&c32);
+}
+
+static void a64_free_cc(DisasCompare64 *c64)
+{
+    tcg_temp_free_i64(c64->value);
+}
+
 static void gen_exception_internal(int excp)
 {
     TCGv_i32 tcg_excp = tcg_const_i32(excp);
@@ -3607,7 +3632,8 @@ static void disas_cc(DisasContext *s, uint32_t insn)
 static void disas_cond_select(DisasContext *s, uint32_t insn)
 {
     unsigned int sf, else_inv, rm, cond, else_inc, rn, rd;
-    TCGv_i64 tcg_rd, tcg_src;
+    TCGv_i64 tcg_rd, zero;
+    DisasCompare64 c;
 
     if (extract32(insn, 29, 1) || extract32(insn, 11, 1)) {
         /* S == 1 or op2<1> == 1 */
@@ -3622,48 +3648,35 @@ static void disas_cond_select(DisasContext *s, uint32_t insn)
     rn = extract32(insn, 5, 5);
     rd = extract32(insn, 0, 5);
 
-    if (rd == 31) {
-        /* silly no-op write; until we use movcond we must special-case
-         * this to avoid a dead temporary across basic blocks.
-         */
-        return;
-    }
-
     tcg_rd = cpu_reg(s, rd);
 
-    if (cond >= 0x0e) { /* condition "always" */
-        tcg_src = read_cpu_reg(s, rn, sf);
-        tcg_gen_mov_i64(tcg_rd, tcg_src);
-    } else {
-        /* OPTME: we could use movcond here, at the cost of duplicating
-         * a lot of the arm_gen_test_cc() logic.
-         */
-        TCGLabel *label_match = gen_new_label();
-        TCGLabel *label_continue = gen_new_label();
-
-        arm_gen_test_cc(cond, label_match);
-        /* nomatch: */
-        tcg_src = cpu_reg(s, rm);
+    a64_test_cc(&c, cond);
+    zero = tcg_const_i64(0);
 
+    if (rn == 31 && rm == 31 && (else_inc ^ else_inv)) {
+        /* CSET & CSETM.  */
+        tcg_gen_setcond_i64(tcg_invert_cond(c.cond), tcg_rd, c.value, zero);
+        if (else_inv) {
+            tcg_gen_neg_i64(tcg_rd, tcg_rd);
+        }
+    } else {
+        TCGv_i64 t_true = cpu_reg(s, rn);
+        TCGv_i64 t_false = read_cpu_reg(s, rm, 1);
         if (else_inv && else_inc) {
-            tcg_gen_neg_i64(tcg_rd, tcg_src);
+            tcg_gen_neg_i64(t_false, t_false);
         } else if (else_inv) {
-            tcg_gen_not_i64(tcg_rd, tcg_src);
+            tcg_gen_not_i64(t_false, t_false);
         } else if (else_inc) {
-            tcg_gen_addi_i64(tcg_rd, tcg_src, 1);
-        } else {
-            tcg_gen_mov_i64(tcg_rd, tcg_src);
-        }
-        if (!sf) {
-            tcg_gen_ext32u_i64(tcg_rd, tcg_rd);
+            tcg_gen_addi_i64(t_false, t_false, 1);
         }
-        tcg_gen_br(label_continue);
-        /* match: */
-        gen_set_label(label_match);
-        tcg_src = read_cpu_reg(s, rn, sf);
-        tcg_gen_mov_i64(tcg_rd, tcg_src);
-        /* continue: */
-        gen_set_label(label_continue);
+        tcg_gen_movcond_i64(c.cond, tcg_rd, c.value, zero, t_true, t_false);
+    }
+
+    tcg_temp_free_i64(zero);
+    a64_free_cc(&c);
+
+    if (!sf) {
+        tcg_gen_ext32u_i64(tcg_rd, tcg_rd);
     }
 }
 
Signature · 1 line
2.4.3
#5 lore raw

PATCH v3 5/11 target-arm: Implement ccmp branchless

In reply to Richard Henderson

This can allow much of a ccmp to be elided when particular flags are subsequently dead.

---
 target-arm/translate-a64.c | 74 ++++++++++++++++++++++++++++++++++++----------
 1 file changed, 58 insertions(+), 16 deletions(-)
target-arm/translate-a64.c +58 -16
index f941fc3..7ce286e 100644
@@ -3570,8 +3570,9 @@ static void disas_adc_sbc(DisasContext *s, uint32_t insn)
 static void disas_cc(DisasContext *s, uint32_t insn)
 {
     unsigned int sf, op, y, cond, rn, nzcv, is_imm;
-    TCGLabel *label_continue = NULL;
+    TCGv_i32 tcg_t0, tcg_t1, tcg_t2;
     TCGv_i64 tcg_tmp, tcg_y, tcg_rn;
+    DisasCompare c;
 
     if (!extract32(insn, 29, 1)) {
         unallocated_encoding(s);
@@ -3589,19 +3590,13 @@ static void disas_cc(DisasContext *s, uint32_t insn)
     rn = extract32(insn, 5, 5);
     nzcv = extract32(insn, 0, 4);
 
-    if (cond < 0x0e) { /* not always */
-        TCGLabel *label_match = gen_new_label();
-        label_continue = gen_new_label();
-        arm_gen_test_cc(cond, label_match);
-        /* nomatch: */
-        tcg_tmp = tcg_temp_new_i64();
-        tcg_gen_movi_i64(tcg_tmp, nzcv << 28);
-        gen_set_nzcv(tcg_tmp);
-        tcg_temp_free_i64(tcg_tmp);
-        tcg_gen_br(label_continue);
-        gen_set_label(label_match);
-    }
-    /* match, or condition is always */
+    /* Set T0 = !COND.  */
+    tcg_t0 = tcg_temp_new_i32();
+    arm_test_cc(&c, cond);
+    tcg_gen_setcondi_i32(tcg_invert_cond(c.cond), tcg_t0, c.value, 0);
+    arm_free_cc(&c);
+
+    /* Load the arguments for the new comparison.  */
     if (is_imm) {
         tcg_y = new_tmp_a64(s);
         tcg_gen_movi_i64(tcg_y, y);
@@ -3610,6 +3605,7 @@ static void disas_cc(DisasContext *s, uint32_t insn)
     }
     tcg_rn = cpu_reg(s, rn);
 
+    /* Set the flags for the new comparison.  */
     tcg_tmp = tcg_temp_new_i64();
     if (op) {
         gen_sub_CC(sf, tcg_tmp, tcg_rn, tcg_y);
@@ -3618,9 +3614,55 @@ static void disas_cc(DisasContext *s, uint32_t insn)
     }
     tcg_temp_free_i64(tcg_tmp);
 
-    if (cond < 0x0e) { /* continue */
-        gen_set_label(label_continue);
+    /* If COND was false, force the flags to #nzcv.  Compute two masks
+     * to help with this: T1 = (COND ? 0 : -1), T2 = (COND ? -1 : 0).
+     * For tcg hosts that support ANDC, we can make do with just T1.
+     * In either case, allow the tcg optimizer to delete any unused mask.
+     */
+    tcg_t1 = tcg_temp_new_i32();
+    tcg_t2 = tcg_temp_new_i32();
+    tcg_gen_neg_i32(tcg_t1, tcg_t0);
+    tcg_gen_subi_i32(tcg_t2, tcg_t0, 1);
+
+    if (nzcv & 8) { /* N */
+        tcg_gen_or_i32(cpu_NF, cpu_NF, tcg_t1);
+    } else {
+        if (TCG_TARGET_HAS_andc_i32) {
+            tcg_gen_andc_i32(cpu_NF, cpu_NF, tcg_t1);
+        } else {
+            tcg_gen_and_i32(cpu_NF, cpu_NF, tcg_t2);
+        }
+    }
+    if (nzcv & 4) { /* Z */
+        if (TCG_TARGET_HAS_andc_i32) {
+            tcg_gen_andc_i32(cpu_ZF, cpu_ZF, tcg_t1);
+        } else {
+            tcg_gen_and_i32(cpu_ZF, cpu_ZF, tcg_t2);
+        }
+    } else {
+        tcg_gen_or_i32(cpu_ZF, cpu_ZF, tcg_t0);
+    }
+    if (nzcv & 2) { /* C */
+        tcg_gen_or_i32(cpu_CF, cpu_CF, tcg_t0);
+    } else {
+        if (TCG_TARGET_HAS_andc_i32) {
+            tcg_gen_andc_i32(cpu_CF, cpu_CF, tcg_t1);
+        } else {
+            tcg_gen_and_i32(cpu_CF, cpu_CF, tcg_t2);
+        }
+    }
+    if (nzcv & 1) { /* V */
+        tcg_gen_or_i32(cpu_VF, cpu_VF, tcg_t1);
+    } else {
+        if (TCG_TARGET_HAS_andc_i32) {
+            tcg_gen_andc_i32(cpu_VF, cpu_VF, tcg_t1);
+        } else {
+            tcg_gen_and_i32(cpu_VF, cpu_VF, tcg_t2);
+        }
     }
+    tcg_temp_free_i32(tcg_t0);
+    tcg_temp_free_i32(tcg_t1);
+    tcg_temp_free_i32(tcg_t2);
 }
 
 /* C3.5.6 Conditional select
Signature · 1 line
2.4.3
#6 lore raw

PATCH v3 6/11 target-arm: Implement fcsel with movcond

In reply to Richard Henderson
---
 target-arm/translate-a64.c | 45 +++++++++++++++++----------------------------
 1 file changed, 17 insertions(+), 28 deletions(-)
target-arm/translate-a64.c +17 -28
index 7ce286e..bb0d752 100644
@@ -4205,20 +4205,6 @@ static void disas_fp_ccomp(DisasContext *s, uint32_t insn)
     }
 }
 
-/* copy src FP register to dst FP register; type specifies single or double */
-static void gen_mov_fp2fp(DisasContext *s, int type, int dst, int src)
-{
-    if (type) {
-        TCGv_i64 v = read_fp_dreg(s, src);
-        write_fp_dreg(s, dst, v);
-        tcg_temp_free_i64(v);
-    } else {
-        TCGv_i32 v = read_fp_sreg(s, src);
-        write_fp_sreg(s, dst, v);
-        tcg_temp_free_i32(v);
-    }
-}
-
 /* C3.6.24 Floating point conditional select
  *   31  30  29 28       24 23  22  21 20  16 15  12 11 10 9    5 4    0
  * +---+---+---+-----------+------+---+------+------+-----+------+------+
@@ -4228,7 +4214,8 @@ static void gen_mov_fp2fp(DisasContext *s, int type, int dst, int src)
 static void disas_fp_csel(DisasContext *s, uint32_t insn)
 {
     unsigned int mos, type, rm, cond, rn, rd;
-    TCGLabel *label_continue = NULL;
+    TCGv_i64 t_true, t_false, t_zero;
+    DisasCompare64 c;
 
     mos = extract32(insn, 29, 3);
     type = extract32(insn, 22, 2); /* 0 = single, 1 = double */
@@ -4246,21 +4233,23 @@ static void disas_fp_csel(DisasContext *s, uint32_t insn)
         return;
     }
 
-    if (cond < 0x0e) { /* not always */
-        TCGLabel *label_match = gen_new_label();
-        label_continue = gen_new_label();
-        arm_gen_test_cc(cond, label_match);
-        /* nomatch: */
-        gen_mov_fp2fp(s, type, rd, rm);
-        tcg_gen_br(label_continue);
-        gen_set_label(label_match);
-    }
+    /* Zero extend sreg inputs to 64 bits now.  */
+    t_true = tcg_temp_new_i64();
+    t_false = tcg_temp_new_i64();
+    read_vec_element(s, t_true, rn, 0, type ? MO_64 : MO_32);
+    read_vec_element(s, t_false, rm, 0, type ? MO_64 : MO_32);
 
-    gen_mov_fp2fp(s, type, rd, rn);
+    a64_test_cc(&c, cond);
+    t_zero = tcg_const_i64(0);
+    tcg_gen_movcond_i64(c.cond, t_true, c.value, t_zero, t_true, t_false);
+    tcg_temp_free_i64(t_zero);
+    tcg_temp_free_i64(t_false);
+    a64_free_cc(&c);
 
-    if (cond < 0x0e) { /* continue */
-        gen_set_label(label_continue);
-    }
+    /* Note that sregs write back zeros to the high bits,
+       and we've already done the zero-extension.  */
+    write_fp_dreg(s, rd, t_true);
+    tcg_temp_free_i64(t_true);
 }
 
 /* C3.6.25 Floating-point data-processing (1 source) - single precision */
Signature · 1 line
2.4.3
#7 lore raw

PATCH v3 7/11 target-arm: Recognize SXTB, SXTH, SXTW, ASR

In reply to Richard Henderson

These are all special case aliases of SBFM.

---
 target-arm/translate-a64.c | 24 +++++++++++++++++++++++-
 1 file changed, 23 insertions(+), 1 deletion(-)
target-arm/translate-a64.c +23 -1
index bb0d752..0572635 100644
@@ -3017,7 +3017,28 @@ static void disas_bitfield(DisasContext *s, uint32_t insn)
     tcg_rd = cpu_reg(s, rd);
     tcg_tmp = read_cpu_reg(s, rn, sf);
 
-    /* OPTME: probably worth recognizing common cases of ext{8,16,32}{u,s} */
+    /* Recognize the common aliases.  */
+    if (opc == 0) { /* SBFM */
+        if (ri == 0) {
+            if (si == 7) { /* SXTB */
+                tcg_gen_ext8s_i64(tcg_rd, tcg_tmp);
+                goto done;
+            } else if (si == 15) { /* SXTH */
+                tcg_gen_ext16s_i64(tcg_rd, tcg_tmp);
+                goto done;
+            } else if (si == 31) { /* SXTW */
+                tcg_gen_ext32s_i64(tcg_rd, tcg_tmp);
+                goto done;
+            }
+        }
+        if (si == 63 || (si == 31 && ri <= si)) { /* ASR */
+            if (si == 31) {
+                tcg_gen_ext32s_i64(tcg_tmp, tcg_tmp);
+            }
+            tcg_gen_sari_i64(tcg_rd, tcg_tmp, ri);
+            goto done;
+        }
+    }
 
     if (opc != 1) { /* SBFM or UBFM */
         tcg_gen_movi_i64(tcg_rd, 0);
@@ -3042,6 +3063,7 @@ static void disas_bitfield(DisasContext *s, uint32_t insn)
         tcg_gen_sari_i64(tcg_rd, tcg_rd, 64 - (pos + len));
     }
 
+ done:
     if (!sf) { /* zero extend final result */
         tcg_gen_ext32u_i64(tcg_rd, tcg_rd);
     }
Signature · 1 line
2.4.3
#8 lore raw

PATCH v3 8/11 target-arm: Recognize UXTB, UXTH, LSR, LSL

In reply to Richard Henderson

These are all special case aliases of UBFM.

---
 target-arm/translate-a64.c | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)
target-arm/translate-a64.c +17 -0
index 0572635..cf68754 100644
@@ -3038,6 +3038,23 @@ static void disas_bitfield(DisasContext *s, uint32_t insn)
             tcg_gen_sari_i64(tcg_rd, tcg_tmp, ri);
             goto done;
         }
+    } else if (opc == 2) { /* UBFM */
+        if (ri == 0) { /* UXTB, UXTH, plus non-canonical AND */
+            tcg_gen_andi_i64(tcg_rd, tcg_tmp, bitmask64(si + 1));
+            return;
+        }
+        if (si == 63 || (si == 31 && ri <= si)) { /* LSR */
+            if (si == 31) {
+                tcg_gen_ext32u_i64(tcg_tmp, tcg_tmp);
+            }
+            tcg_gen_shri_i64(tcg_rd, tcg_tmp, ri);
+            return;
+        }
+        if (si + 1 == ri && si != bitsize - 1) { /* LSL */
+            int shift = bitsize - 1 - si;
+            tcg_gen_shli_i64(tcg_rd, tcg_tmp, shift);
+            goto done;
+        }
     }
 
     if (opc != 1) { /* SBFM or UBFM */
Signature · 1 line
2.4.3
#9 lore raw

PATCH v3 9/11 target-arm: Eliminate unnecessary zero-extend in disas_bitfield

In reply to Richard Henderson

For !SF, this initial ext32u can't be optimized away by the current TCG code generator. (It would require backward bit liveness propagation.)

But since the range of bits for !SF are already constrained by unallocated_encoding, we'll never reference the high bits anyway.

---
 target-arm/translate-a64.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
target-arm/translate-a64.c +5 -1
index cf68754..56aa043 100644
@@ -3015,7 +3015,11 @@ static void disas_bitfield(DisasContext *s, uint32_t insn)
     }
 
     tcg_rd = cpu_reg(s, rd);
-    tcg_tmp = read_cpu_reg(s, rn, sf);
+
+    /* Suppress the zero-extend for !sf.  Since RI and SI are constrained
+       to be smaller than bitsize, we'll never reference data outside the
+       low 32-bits anyway.  */
+    tcg_tmp = read_cpu_reg(s, rn, 1);
 
     /* Recognize the common aliases.  */
     if (opc == 0) { /* SBFM */
Signature · 1 line
2.4.3
#10 lore raw

PATCH v3 10/11 target-arm: Recognize ROR

In reply to Richard Henderson
---
 target-arm/translate-a64.c | 33 +++++++++++++++++++++------------
 1 file changed, 21 insertions(+), 12 deletions(-)
target-arm/translate-a64.c +21 -12
index 56aa043..0bb12c2 100644
@@ -3117,17 +3117,7 @@ static void disas_extract(DisasContext *s, uint32_t insn)
 
         tcg_rd = cpu_reg(s, rd);
 
-        if (imm) {
-            /* OPTME: we can special case rm==rn as a rotate */
-            tcg_rm = read_cpu_reg(s, rm, sf);
-            tcg_rn = read_cpu_reg(s, rn, sf);
-            tcg_gen_shri_i64(tcg_rm, tcg_rm, imm);
-            tcg_gen_shli_i64(tcg_rn, tcg_rn, bitsize - imm);
-            tcg_gen_or_i64(tcg_rd, tcg_rm, tcg_rn);
-            if (!sf) {
-                tcg_gen_ext32u_i64(tcg_rd, tcg_rd);
-            }
-        } else {
+        if (unlikely(imm == 0)) {
             /* tcg shl_i32/shl_i64 is undefined for 32/64 bit shifts,
              * so an extract from bit 0 is a special case.
              */
@@ -3136,8 +3126,27 @@ static void disas_extract(DisasContext *s, uint32_t insn)
             } else {
                 tcg_gen_ext32u_i64(tcg_rd, cpu_reg(s, rm));
             }
+        } else if (rm == rn) { /* ROR */
+            tcg_rm = cpu_reg(s, rm);
+            if (sf) {
+                tcg_gen_rotri_i64(tcg_rd, tcg_rm, imm);
+            } else {
+                TCGv_i32 tmp = tcg_temp_new_i32();
+                tcg_gen_extrl_i64_i32(tmp, tcg_rm);
+                tcg_gen_rotri_i32(tmp, tmp, imm);
+                tcg_gen_extu_i32_i64(tcg_rd, tmp);
+                tcg_temp_free_i32(tmp);
+            }
+        } else {
+            tcg_rm = read_cpu_reg(s, rm, sf);
+            tcg_rn = read_cpu_reg(s, rn, sf);
+            tcg_gen_shri_i64(tcg_rm, tcg_rm, imm);
+            tcg_gen_shli_i64(tcg_rn, tcg_rn, bitsize - imm);
+            tcg_gen_or_i64(tcg_rd, tcg_rm, tcg_rn);
+            if (!sf) {
+                tcg_gen_ext32u_i64(tcg_rd, tcg_rd);
+            }
         }
-
     }
 }
 
Signature · 1 line
2.4.3
#11 lore raw

PATCH v3 11/11 target-arm: Use tcg_gen_extrh_i64_i32

In reply to Richard Henderson

Usually, eliminate an operation from the translator by combining a shift with an extract.

In the case of gen_set_NZ64, we don't need a boolean value for cpu_ZF, merely a non-zero value. Given that we can extract both halves of a 64-bit input in one call, this simplifies the code.

---
 target-arm/translate-a64.c | 34 +++++++++-------------------------
 1 file changed, 9 insertions(+), 25 deletions(-)
target-arm/translate-a64.c +9 -25
index 0bb12c2..8248edd 100644
@@ -529,13 +529,8 @@ static TCGv_ptr get_fpstatus_ptr(void)
  */
 static inline void gen_set_NZ64(TCGv_i64 result)
 {
-    TCGv_i64 flag = tcg_temp_new_i64();
-
-    tcg_gen_setcondi_i64(TCG_COND_NE, flag, result, 0);
-    tcg_gen_extrl_i64_i32(cpu_ZF, flag);
-    tcg_gen_shri_i64(flag, result, 32);
-    tcg_gen_extrl_i64_i32(cpu_NF, flag);
-    tcg_temp_free_i64(flag);
+    tcg_gen_extr_i64_i32(cpu_ZF, cpu_NF, result);
+    tcg_gen_or_i32(cpu_ZF, cpu_ZF, cpu_NF);
 }
 
 /* Set NZCV as for a logical operation: NZ as per result, CV cleared. */
@@ -545,7 +540,7 @@ static inline void gen_logic_CC(int sf, TCGv_i64 result)
         gen_set_NZ64(result);
     } else {
         tcg_gen_extrl_i64_i32(cpu_ZF, result);
-        tcg_gen_extrl_i64_i32(cpu_NF, result);
+        tcg_gen_mov_i32(cpu_NF, cpu_ZF);
     }
     tcg_gen_movi_i32(cpu_CF, 0);
     tcg_gen_movi_i32(cpu_VF, 0);
@@ -571,8 +566,7 @@ static void gen_add_CC(int sf, TCGv_i64 dest, TCGv_i64 t0, TCGv_i64 t1)
         tcg_gen_xor_i64(tmp, t0, t1);
         tcg_gen_andc_i64(flag, flag, tmp);
         tcg_temp_free_i64(tmp);
-        tcg_gen_shri_i64(flag, flag, 32);
-        tcg_gen_extrl_i64_i32(cpu_VF, flag);
+        tcg_gen_extrh_i64_i32(cpu_VF, flag);
 
         tcg_gen_mov_i64(dest, result);
         tcg_temp_free_i64(result);
@@ -620,8 +614,7 @@ static void gen_sub_CC(int sf, TCGv_i64 dest, TCGv_i64 t0, TCGv_i64 t1)
         tcg_gen_xor_i64(tmp, t0, t1);
         tcg_gen_and_i64(flag, flag, tmp);
         tcg_temp_free_i64(tmp);
-        tcg_gen_shri_i64(flag, flag, 32);
-        tcg_gen_extrl_i64_i32(cpu_VF, flag);
+        tcg_gen_extrh_i64_i32(cpu_VF, flag);
         tcg_gen_mov_i64(dest, result);
         tcg_temp_free_i64(flag);
         tcg_temp_free_i64(result);
@@ -680,8 +673,7 @@ static void gen_adc_CC(int sf, TCGv_i64 dest, TCGv_i64 t0, TCGv_i64 t1)
         tcg_gen_xor_i64(vf_64, result, t0);
         tcg_gen_xor_i64(tmp, t0, t1);
         tcg_gen_andc_i64(vf_64, vf_64, tmp);
-        tcg_gen_shri_i64(vf_64, vf_64, 32);
-        tcg_gen_extrl_i64_i32(cpu_VF, vf_64);
+        tcg_gen_extrh_i64_i32(cpu_VF, vf_64);
 
         tcg_gen_mov_i64(dest, result);
 
@@ -7775,10 +7767,8 @@ static void handle_2misc_narrow(DisasContext *s, bool scalar,
             } else {
                 TCGv_i32 tcg_lo = tcg_temp_new_i32();
                 TCGv_i32 tcg_hi = tcg_temp_new_i32();
-                tcg_gen_extrl_i64_i32(tcg_lo, tcg_op);
+                tcg_gen_extr_i64_i32(tcg_lo, tcg_hi, tcg_op);
                 gen_helper_vfp_fcvt_f32_to_f16(tcg_lo, tcg_lo, cpu_env);
-                tcg_gen_shri_i64(tcg_op, tcg_op, 32);
-                tcg_gen_extrl_i64_i32(tcg_hi, tcg_op);
                 gen_helper_vfp_fcvt_f32_to_f16(tcg_hi, tcg_hi, cpu_env);
                 tcg_gen_deposit_i32(tcg_res[pass], tcg_lo, tcg_hi, 16, 16);
                 tcg_temp_free_i32(tcg_lo);
@@ -8684,16 +8674,10 @@ static void handle_3rd_wide(DisasContext *s, int is_q, int is_u, int size,
     }
 }
 
-static void do_narrow_high_u32(TCGv_i32 res, TCGv_i64 in)
-{
-    tcg_gen_shri_i64(in, in, 32);
-    tcg_gen_extrl_i64_i32(res, in);
-}
-
 static void do_narrow_round_high_u32(TCGv_i32 res, TCGv_i64 in)
 {
     tcg_gen_addi_i64(in, in, 1U << 31);
-    do_narrow_high_u32(res, in);
+    tcg_gen_extrh_i64_i32(res, in);
 }
 
 static void handle_3rd_narrowing(DisasContext *s, int is_q, int is_u, int size,
@@ -8712,7 +8696,7 @@ static void handle_3rd_narrowing(DisasContext *s, int is_q, int is_u, int size,
               gen_helper_neon_narrow_round_high_u8 },
             { gen_helper_neon_narrow_high_u16,
               gen_helper_neon_narrow_round_high_u16 },
-            { do_narrow_high_u32, do_narrow_round_high_u32 },
+            { tcg_gen_extrh_i64_i32, do_narrow_round_high_u32 },
         };
         NeonGenNarrowFn *gennarrow = narrowfns[size][is_u];
 
Signature · 1 line
2.4.3
#12 lore raw
In reply to Richard Henderson
Quoted 7 lines · Richard Henderson

Updated from v2 based on review from Peter.

I did go ahead with the use of andc for translating ccmp; the result looked pretty good with Haswell's andn instruction.

r~

Applied to target-arm.next, thanks.

-- PMM