[Qemu-devel] [PATCH 2/4] target-mips: Use TCG registers for the FPU.

[Richard Henderson <rth@twiddle.net>]

With normal FP, this doesn't have much affect on the generated code,
because most of the FP operations are not CONST/PURE, and so we spill
registers in about the same frequency as the explicit load/stores.

But with Loongson multimedia instructions, which are all integral and
whose helpers are in fact CONST+PURE, this greatly improves the code.

On a 64-bit host, rather than over-use the deposit operation, we
create TCG registers for both the 64-bit FPU register as a whole
and the two 32-bit halves.  We only ever reference the whole register
or the two half registers in any one TB, so there's no problem with
aliasing.

On a 32-bit host, we only create the 64-bit FPU registers, and then
directly reference the internal 32-bit TCG register halves as needed.

Signed-off-by: Richard Henderson <rth@twiddle.net>
---
 target-mips/translate.c |  123 ++++++++++++++++++++++++++++++++++++-----------
 1 files changed, 95 insertions(+), 28 deletions(-)

diff --git a/target-mips/translate.c b/target-mips/translate.c
index d2aeff0..b6a7aeb 100644
--- a/target-mips/translate.c
+++ b/target-mips/translate.c
@@ -486,6 +486,14 @@ static TCGv cpu_dspctrl, btarget, bcond;
 static TCGv_i32 hflags;
 static TCGv_i32 fpu_fcr0, fpu_fcr31;
 
+/* FPU registers.  These alias, but we'll only use one or the other in any
+   one TB based on MIPS_HFLAG_F64.  */
+#if TCG_TARGET_REG_BITS == 64
+static TCGv_i32 fpu_f32[32];
+static TCGv_i32 fpu_fh32[32];
+#endif
+static TCGv_i64 fpu_f64[32];
+
 static uint32_t gen_opc_hflags[OPC_BUF_SIZE];
 
 #include "gen-icount.h"
@@ -555,6 +563,14 @@ static const char *fregnames[] =
       "f16", "f17", "f18", "f19", "f20", "f21", "f22", "f23",
       "f24", "f25", "f26", "f27", "f28", "f29", "f30", "f31", };
 
+#if TCG_TARGET_REG_BITS == 64
+static const char *fhregnames[] =
+    { "fh0",  "fh1",  "fh2",  "fh3",  "fh4",  "fh5",  "fh6",  "fh7",
+      "fh8",  "fh9",  "fh10", "fh11", "fh12", "fh13", "fh14", "fh15",
+      "fh16", "fh17", "fh18", "fh19", "fh20", "fh21", "fh22", "fh23",
+      "fh24", "fh25", "fh26", "fh27", "fh28", "fh29", "fh30", "fh31", };
+#endif
+
 #ifdef MIPS_DEBUG_DISAS
 #define MIPS_DEBUG(fmt, ...)                         \
         qemu_log_mask(CPU_LOG_TB_IN_ASM,                \
@@ -647,55 +663,91 @@ static inline void gen_store_srsgpr (int from, int to)
 }
 
 /* Floating point register moves. */
-static inline void gen_load_fpr32 (DisasContext *ctx, TCGv_i32 t, int reg)
+static void gen_load_fpr32 (DisasContext *ctx, TCGv_i32 t, int reg)
 {
-    tcg_gen_ld_i32(t, cpu_env, offsetof(CPUState, active_fpu.fpr[reg].w[FP_ENDIAN_IDX]));
+#if TCG_TARGET_REG_BITS == 32
+    tcg_gen_mov_i32(t, TCGV_LOW(fpu_f64[reg]));
+#else
+    if (ctx->hflags & MIPS_HFLAG_F64) {
+        tcg_gen_trunc_i64_i32(t, fpu_f64[reg]);
+    } else {
+        tcg_gen_mov_i32(t, fpu_f32[reg]);
+    }
+#endif
 }
 
-static inline void gen_store_fpr32 (DisasContext *ctx, TCGv_i32 t, int reg)
+static void gen_store_fpr32 (DisasContext *ctx, TCGv_i32 t, int reg)
 {
-    tcg_gen_st_i32(t, cpu_env, offsetof(CPUState, active_fpu.fpr[reg].w[FP_ENDIAN_IDX]));
+#if TCG_TARGET_REG_BITS == 32
+    tcg_gen_mov_i32(TCGV_LOW(fpu_f64[reg]), t);
+#else
+    if (ctx->hflags & MIPS_HFLAG_F64) {
+        TCGv_i64 t64 = MAKE_TCGV_I64(GET_TCGV_I32(t));
+        tcg_gen_deposit_i64(fpu_f64[reg], fpu_f64[reg], t64, 0, 32);
+    } else {
+        tcg_gen_mov_i32(fpu_f32[reg], t);
+    }
+#endif
 }
 
-static inline void gen_load_fpr32h (DisasContext *ctx, TCGv_i32 t, int reg)
+static void gen_load_fpr32h (DisasContext *ctx, TCGv_i32 t, int reg)
 {
-    tcg_gen_ld_i32(t, cpu_env, offsetof(CPUState, active_fpu.fpr[reg].w[!FP_ENDIAN_IDX]));
+#if TCG_TARGET_REG_BITS == 32
+    tcg_gen_mov_i32(t, TCGV_HIGH(fpu_f64[reg]));
+#else
+    if (ctx->hflags & MIPS_HFLAG_F64) {
+        TCGv_i64 t64 = tcg_temp_new_i64();
+        tcg_gen_shri_i64(t64, fpu_f64[reg], 32);
+        tcg_gen_trunc_i64_i32(t, t64);
+        tcg_temp_free_i64(t64);
+    } else {
+        tcg_gen_mov_i32(t, fpu_fh32[reg]);
+    }
+#endif
 }
 
-static inline void gen_store_fpr32h (DisasContext *ctx, TCGv_i32 t, int reg)
+static void gen_store_fpr32h (DisasContext *ctx, TCGv_i32 t, int reg)
 {
-    tcg_gen_st_i32(t, cpu_env, offsetof(CPUState, active_fpu.fpr[reg].w[!FP_ENDIAN_IDX]));
+#if TCG_TARGET_REG_BITS == 32
+    tcg_gen_mov_i32(TCGV_HIGH(fpu_f64[reg]), t);
+#else
+    if (ctx->hflags & MIPS_HFLAG_F64) {
+        TCGv_i64 t64 = MAKE_TCGV_I64(GET_TCGV_I32(t));
+        tcg_gen_deposit_i64(fpu_f64[reg], fpu_f64[reg], t64, 32, 32);
+    } else {
+        tcg_gen_mov_i32(fpu_fh32[reg], t);
+    }
+#endif
 }
 
-static inline void gen_load_fpr64 (DisasContext *ctx, TCGv_i64 t, int reg)
+static void gen_load_fpr64 (DisasContext *ctx, TCGv_i64 t, int reg)
 {
     if (ctx->hflags & MIPS_HFLAG_F64) {
-        tcg_gen_ld_i64(t, cpu_env, offsetof(CPUState, active_fpu.fpr[reg].d));
+        tcg_gen_mov_i64(t, fpu_f64[reg]);
     } else {
-        TCGv_i32 t0 = tcg_temp_new_i32();
-        TCGv_i32 t1 = tcg_temp_new_i32();
-        gen_load_fpr32(ctx, t0, reg & ~1);
-        gen_load_fpr32(ctx, t1, reg | 1);
-        tcg_gen_concat_i32_i64(t, t0, t1);
-        tcg_temp_free_i32(t0);
-        tcg_temp_free_i32(t1);
+#if TCG_TARGET_REG_BITS == 32
+        tcg_gen_concat32_i64(t, fpu_f64[reg & ~1], fpu_f64[reg | 1]);
+#else
+        tcg_gen_concat_i32_i64(t, fpu_f32[reg & ~1], fpu_f32[reg | 1]);
+#endif
     }
 }
 
-static inline void gen_store_fpr64 (DisasContext *ctx, TCGv_i64 t, int reg)
+static void gen_store_fpr64 (DisasContext *ctx, TCGv_i64 t, int reg)
 {
     if (ctx->hflags & MIPS_HFLAG_F64) {
-        tcg_gen_st_i64(t, cpu_env, offsetof(CPUState, active_fpu.fpr[reg].d));
+        tcg_gen_mov_i64(fpu_f64[reg], t);
     } else {
-        TCGv_i64 t0 = tcg_temp_new_i64();
-        TCGv_i32 t1 = tcg_temp_new_i32();
-        tcg_gen_trunc_i64_i32(t1, t);
-        gen_store_fpr32(ctx, t1, reg & ~1);
-        tcg_gen_shri_i64(t0, t, 32);
-        tcg_gen_trunc_i64_i32(t1, t0);
-        gen_store_fpr32(ctx, t1, reg | 1);
-        tcg_temp_free_i32(t1);
-        tcg_temp_free_i64(t0);
+#if TCG_TARGET_REG_BITS == 32
+        tcg_gen_mov_i32(TCGV_LOW(fpu_f64[reg & ~1]), TCGV_LOW(t));
+        tcg_gen_mov_i32(TCGV_LOW(fpu_f64[reg | 1]), TCGV_HIGH(t));
+#else
+        TCGv_i64 t64 = tcg_temp_new_i64();
+        tcg_gen_shri_i64(t64, t, 32);
+        tcg_gen_trunc_i64_i32(fpu_f32[reg | 1], t64);
+        tcg_temp_free_i64(t64);
+        tcg_gen_trunc_i64_i32(fpu_f32[reg & ~1], t);
+#endif
     }
 }
 
@@ -12681,6 +12733,21 @@ static void mips_tcg_init(void)
                                        offsetof(CPUState, active_fpu.fcr31),
                                        "fcr31");
 
+#if TCG_TARGET_REG_BITS == 64
+    for (i = 0; i < 32; i++) {
+        int off = offsetof(CPUState, active_fpu.fpr[i].w[FP_ENDIAN_IDX]);
+        fpu_f32[i] = tcg_global_mem_new_i32(TCG_AREG0, off, fregnames[i]);
+    }
+    for (i = 0; i < 32; i++) {
+        int off = offsetof(CPUState, active_fpu.fpr[i].w[!FP_ENDIAN_IDX]);
+        fpu_fh32[i] = tcg_global_mem_new_i32(TCG_AREG0, off, fhregnames[i]);
+    }
+#endif
+    for (i = 0; i < 32; i++) {
+        int off = offsetof(CPUState, active_fpu.fpr[i].w[FP_ENDIAN_IDX]);
+        fpu_f64[i] = tcg_global_mem_new_i64(TCG_AREG0, off, fregnames[i]);
+    }
+
     /* register helpers */
 #define GEN_HELPER 2
 #include "helper.h"
-- 
1.7.7.4