[libclc] Move rotate to CLC library; optimize (#125713)

This commit moves the rotate builtin to the CLC library.

It also optimizes rotate(x, n) to generate the @llvm.fshl(x, x, n)
intrinsic, for both scalar and vector types. The previous implementation
was too cautious in its handling of the shift amount; the OpenCL rules
state that the shift amount is always treated as an unsigned value
modulo the bitwidth.

Author: frasercrmck

libclc/clc/include/clc/integer/clc_rotate.h

@@ -0,0 +1,12 @@
+#ifndef __CLC_INTEGER_CLC_ROTATE_H__
+#define __CLC_INTEGER_CLC_ROTATE_H__
+
+#define __CLC_FUNCTION __clc_rotate
+#define __CLC_BODY <clc/shared/binary_decl.inc>
+
+#include <clc/integer/gentype.inc>
+
+#undef __CLC_BODY
+#undef __CLC_FUNCTION
+
+#endif // __CLC_INTEGER_CLC_ROTATE_H__

libclc/clc/lib/clspv/SOURCES

@@ -7,6 +7,7 @@
 ../generic/integer/clc_mul_hi.cl
 ../generic/integer/clc_popcount.cl
 ../generic/integer/clc_rhadd.cl
+../generic/integer/clc_rotate.cl
 ../generic/integer/clc_sub_sat.cl
 ../generic/integer/clc_upsample.cl
 ../generic/math/clc_ceil.cl

libclc/clc/lib/generic/SOURCES

@@ -13,6 +13,7 @@ integer/clc_mul24.cl
 integer/clc_mul_hi.cl
 integer/clc_popcount.cl
 integer/clc_rhadd.cl
+integer/clc_rotate.cl
 integer/clc_sub_sat.cl
 integer/clc_upsample.cl
 math/clc_ceil.cl

libclc/clc/lib/generic/integer/clc_rotate.cl

@@ -0,0 +1,5 @@
+#include <clc/internal/clc.h>
+#include <clc/utils.h>
+
+#define __CLC_BODY <clc_rotate.inc>
+#include <clc/integer/gentype.inc>

libclc/clc/lib/generic/integer/clc_rotate.inc

@@ -0,0 +1,22 @@
+#define __CLC_AS_GENTYPE(x) __CLC_XCONCAT(__clc_as_, __CLC_GENTYPE)(x)
+#define __CLC_AS_U_GENTYPE(x) __CLC_XCONCAT(__clc_as_, __CLC_U_GENTYPE)(x)
+
+// The rotate(A, B) builtin left-shifts corresponding to the usual OpenCL shift
+// modulo rules. These rules state that A is left-shifted by the log2(N) least
+// significant bits in B when viewed as an unsigned integer value. Thus we don't
+// have to worry about signed shift amounts, and can perform the computation in
+// unsigned types.
+_CLC_OVERLOAD _CLC_DEF __CLC_GENTYPE __clc_rotate(__CLC_GENTYPE x,
+                                                  __CLC_GENTYPE n) {
+  __CLC_U_GENTYPE x_as_u = __CLC_AS_U_GENTYPE(x);
+  __CLC_U_GENTYPE mask = (__CLC_U_GENTYPE)(__CLC_GENSIZE - 1);
+
+  __CLC_U_GENTYPE lshift_amt = __CLC_AS_U_GENTYPE(n) & mask;
+
+  __CLC_U_GENTYPE rshift_amt =
+      (((__CLC_U_GENTYPE)__CLC_GENSIZE - lshift_amt) & mask);
+
+  __CLC_U_GENTYPE result = (x_as_u << lshift_amt) | (x_as_u >> rshift_amt);
+
+  return __CLC_AS_GENTYPE(result);
+}

libclc/clc/lib/spirv/SOURCES

@@ -11,6 +11,7 @@
 ../generic/integer/clc_mul_hi.cl
 ../generic/integer/clc_popcount.cl
 ../generic/integer/clc_rhadd.cl
+../generic/integer/clc_rotate.cl
 ../generic/integer/clc_sub_sat.cl
 ../generic/integer/clc_upsample.cl
 ../generic/math/clc_ceil.cl

libclc/generic/lib/integer/rotate.cl

@@ -1,4 +1,7 @@
 #include <clc/clc.h>
+#include <clc/integer/clc_rotate.h>
+
+#define FUNCTION rotate
+#define __CLC_BODY <clc/shared/binary_def.inc>
 
-#define __CLC_BODY <rotate.inc>
 #include <clc/integer/gentype.inc>