gallium: add special cases in spe_load_float(), spe_load_int(), added spe_splat()

2 files changed, 40 insertions, 9 deletions

diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c
index 61010e4333..a04cc6c4ff 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c
+++ b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.c
@@ -473,21 +473,48 @@ EMIT_R   (spe_mtspr, 0x10c);
 void
 spe_load_float(struct spe_function *p, unsigned rT, float x)
 {
-   union {
-      float f;
-      unsigned u;
-   } bits;
-   bits.f = x;
-   spe_ilhu(p, rT, bits.u >> 16);
-   spe_iohl(p, rT, bits.u & 0xffff);
+   if (x == 0.0f) {
+      spe_il(p, rT, 0x0);
+   }
+   else if (x == 0.5f) {
+      spe_ilhu(p, rT, 0x3f00);
+   }
+   else if (x == 1.0f) {
+      spe_ilhu(p, rT, 0x3f80);
+   }
+   else if (x == -1.0f) {
+      spe_ilhu(p, rT, 0xbf80);
+   }
+   else {
+      union {
+         float f;
+         unsigned u;
+      } bits;
+      bits.f = x;
+      spe_ilhu(p, rT, bits.u >> 16);
+      spe_iohl(p, rT, bits.u & 0xffff);
+   }
 }
 
 
 void
 spe_load_int(struct spe_function *p, unsigned rT, int i)
 {
-   spe_ilhu(p, rT, i >> 16);
-   spe_iohl(p, rT, i & 0xffff);
+   if (-32768 <= i && i <= 32767) {
+      spe_il(p, rT, i);
+   }
+   else {
+      spe_ilhu(p, rT, i >> 16);
+      spe_iohl(p, rT, i & 0xffff);
+   }
+}
+
+
+void
+spe_splat(struct spe_function *p, unsigned rT, unsigned rA)
+{
+   spe_ila(p, rT, 66051);
+   spe_shufb(p, rT, rA, rA, rT);
 }
 
 
diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h
index dee8c55c4a..d95e5aace3 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h
+++ b/src/gallium/auxiliary/rtasm/rtasm_ppc_spe.h
@@ -292,6 +292,10 @@ spe_load_float(struct spe_function *p, unsigned rT, float x);
 extern void
 spe_load_int(struct spe_function *p, unsigned rT, int i);
 
+/** Replicate word 0 of rA across rT. */
+extern void
+spe_splat(struct spe_function *p, unsigned rT, unsigned rA);
+
 /** Complement/invert all bits in rT. */
 extern void
 spe_complement(struct spe_function *p, unsigned rT);