* config/i386/sse.md (mulv16qi3, mulv2di3): New.

git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@99824 138bc75d-0d04-0410-961f-82ee72b054a4

3 files changed, 120 insertions, 0 deletions

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index dc64af228ba..82233e9ee9b 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,7 @@
+2005-05-17  Richard Henderson  <rth@redhat.com>
+
+	* config/i386/sse.md (mulv16qi3, mulv2di3): New.
+
 2005-05-17  Jakub Jelinek  <jakub@redhat.com>
 
 	PR middle-end/21492
diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
index 76efe5f0586..5ff94ba9814 100644
--- a/gcc/config/i386/sse.md
+++ b/gcc/config/i386/sse.md
@@ -2392,6 +2392,50 @@
   [(set_attr "type" "sseiadd")
    (set_attr "mode" "TI")])
 
+(define_expand "mulv16qi3"
+  [(set (match_operand:V16QI 0 "register_operand" "")
+	(mult:V16QI (match_operand:V16QI 1 "register_operand" "")
+		    (match_operand:V16QI 2 "register_operand" "")))]
+  "TARGET_SSE2"
+{
+  rtx t[12], op0;
+  int i;
+
+  for (i = 0; i < 12; ++i)
+    t[i] = gen_reg_rtx (V16QImode);
+
+  /* Unpack data such that we've got a source byte in each low byte of
+     each word.  We don't care what goes into the high byte of each word.
+     Rather than trying to get zero in there, most convenient is to let
+     it be a copy of the low byte.  */
+  emit_insn (gen_sse2_punpckhbw (t[0], operands[1], operands[1]));
+  emit_insn (gen_sse2_punpckhbw (t[1], operands[2], operands[2]));
+  emit_insn (gen_sse2_punpcklbw (t[2], operands[1], operands[1]));
+  emit_insn (gen_sse2_punpcklbw (t[3], operands[2], operands[2]));
+
+  /* Multiply words.  The end-of-line annotations here give a picture of what
+     the output of that instruction looks like.  Dot means don't care; the 
+     letters are the bytes of the result with A being the most significant.  */
+  emit_insn (gen_mulv8hi3 (gen_lowpart (V8HImode, t[4]), /* .A.B.C.D.E.F.G.H */
+			   gen_lowpart (V8HImode, t[0]),
+			   gen_lowpart (V8HImode, t[1])));
+  emit_insn (gen_mulv8hi3 (gen_lowpart (V8HImode, t[5]), /* .I.J.K.L.M.N.O.P */
+			   gen_lowpart (V8HImode, t[2]),
+			   gen_lowpart (V8HImode, t[3])));
+
+  /* Extract the relevant bytes and merge them back together.  */
+  emit_insn (gen_sse2_punpckhbw (t[6], t[5], t[4]));	/* ..AI..BJ..CK..DL */
+  emit_insn (gen_sse2_punpcklbw (t[7], t[5], t[4]));	/* ..EM..FN..GO..HP */
+  emit_insn (gen_sse2_punpckhbw (t[8], t[7], t[6]));	/* ....AEIM....BFJN */
+  emit_insn (gen_sse2_punpcklbw (t[9], t[7], t[6]));	/* ....CGKO....DHLP */
+  emit_insn (gen_sse2_punpckhbw (t[10], t[9], t[8]));	/* ........ACEGIKMO */
+  emit_insn (gen_sse2_punpcklbw (t[11], t[9], t[8]));	/* ........BDFHJLNP */
+
+  op0 = operands[0];
+  emit_insn (gen_sse2_punpcklbw (op0, t[11], t[10]));	/* ABCDEFGHIJKLMNOP */
+  DONE;
+})
+
 (define_expand "mulv8hi3"
   [(set (match_operand:V8HI 0 "register_operand" "")
 	(mult:V8HI (match_operand:V8HI 1 "nonimmediate_operand" "")
@@ -2536,6 +2580,50 @@
   DONE;
 })
 
+(define_expand "mulv2di3"
+  [(set (match_operand:V2DI 0 "register_operand" "")
+	(mult:V2DI (match_operand:V2DI 1 "nonimmediate_operand" "")
+		   (match_operand:V2DI 2 "nonimmediate_operand" "")))]
+  "TARGET_SSE2"
+{
+  rtx t1, t2, t3, t4, t5, t6, thirtytwo;
+  rtx op0, op1, op2;
+
+  op0 = operands[0];
+  op1 = operands[1];
+  op2 = operands[2];
+  t1 = gen_reg_rtx (V2DImode);
+  t2 = gen_reg_rtx (V2DImode);
+  t3 = gen_reg_rtx (V2DImode);
+  t4 = gen_reg_rtx (V2DImode);
+  t5 = gen_reg_rtx (V2DImode);
+  t6 = gen_reg_rtx (V2DImode);
+  thirtytwo = GEN_INT (32);
+
+  /* Multiply low parts.  */
+  emit_insn (gen_sse2_umulv2siv2di3 (t1, gen_lowpart (V4SImode, op1),
+				     gen_lowpart (V4SImode, op2)));
+
+  /* Shift input vectors left 32 bits so we can multiply high parts.  */
+  emit_insn (gen_lshrv2di3 (t2, op1, thirtytwo));
+  emit_insn (gen_lshrv2di3 (t3, op2, thirtytwo));
+
+  /* Multiply high parts by low parts.  */
+  emit_insn (gen_sse2_umulv2siv2di3 (t4, gen_lowpart (V4SImode, op1),
+				     gen_lowpart (V4SImode, t3)));
+  emit_insn (gen_sse2_umulv2siv2di3 (t5, gen_lowpart (V4SImode, op2),
+				     gen_lowpart (V4SImode, t2)));
+
+  /* Shift them back.  */
+  emit_insn (gen_ashlv2di3 (t4, t4, thirtytwo));
+  emit_insn (gen_ashlv2di3 (t5, t5, thirtytwo));
+
+  /* Add the three parts together.  */
+  emit_insn (gen_addv2di3 (t6, t1, t4));
+  emit_insn (gen_addv2di3 (op0, t6, t5));
+  DONE;
+})
+
 (define_insn "ashr<mode>3"
   [(set (match_operand:SSEMODE24 0 "register_operand" "=x")
 	(ashiftrt:SSEMODE24
diff --git a/gcc/testsuite/gcc.dg/vect/vect-100.c b/gcc/testsuite/gcc.dg/vect/vect-100.c
new file mode 100644
index 00000000000..3b803fc71b6
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/vect-100.c
@@ -0,0 +1,28 @@
+/* Assuming we can vectorize char multiplication, here's an execute test.  */
+
+#include <stdarg.h>
+#include "tree-vect.h"
+
+extern void abort (void);
+void foo()
+{
+  static unsigned char A[256], B[256], C[256];
+  int i;
+
+  for (i = 0; i < 256; ++i)
+    A[i] = B[i] = i;
+
+  for (i = 0; i < 256; ++i)
+    C[i] = A[i] * B[i];
+
+  for (i = 0; i < 256; ++i)
+    if (C[i] != (unsigned char)(i * i))
+      abort ();
+}
+
+int main()
+{
+  check_vect ();
+  foo();
+  return 0;
+}