libclc: Optimize vload4/8/16 from global for int/uint types.

R600 doesn't seem to support v2/v3 loads, doesn't support constant address
space well currently, and chokes on types that aren't 32-bits in size.

All of those caveats could/should change in the future. For now, the
non-optimized implementations for other sizes/types are left intact.

4 files changed, 208 insertions, 3 deletions

diff --git a/generic/lib/SOURCES b/generic/lib/SOURCES
index ccbb59a..5e0abe1 100644
--- a/generic/lib/SOURCES
+++ b/generic/lib/SOURCES
@@ -26,6 +26,8 @@ shared/clamp.cl
 shared/max.cl
 shared/min.cl
 shared/vload.cl
+shared/vload_if.ll
+shared/vload_impl.ll
 shared/vstore.cl
 workitem/get_global_id.cl
 workitem/get_global_size.cl
diff --git a/generic/lib/shared/vload.cl b/generic/lib/shared/vload.cl
index 24d8240..821641c 100644
--- a/generic/lib/shared/vload.cl
+++ b/generic/lib/shared/vload.cl
@@ -6,7 +6,7 @@
   } \
 \
   _CLC_OVERLOAD _CLC_DEF PRIM_TYPE##3 vload3(size_t offset, const ADDR_SPACE PRIM_TYPE *x) { \
-    return (PRIM_TYPE##3)(x[offset] , x[offset+1], x[offset+2]); \
+    return (PRIM_TYPE##3)(vload2(offset, x), x[offset+2]); \
   } \
 \
   _CLC_OVERLOAD _CLC_DEF PRIM_TYPE##4 vload4(size_t offset, const ADDR_SPACE PRIM_TYPE *x) { \
@@ -27,13 +27,12 @@
     VLOAD_VECTORIZE(SCALAR_GENTYPE, __constant) \
     VLOAD_VECTORIZE(SCALAR_GENTYPE, __global) \
 
+//int/uint are special... see below
 #define VLOAD_TYPES() \
     VLOAD_ADDR_SPACES(char) \
     VLOAD_ADDR_SPACES(uchar) \
     VLOAD_ADDR_SPACES(short) \
     VLOAD_ADDR_SPACES(ushort) \
-    VLOAD_ADDR_SPACES(int) \
-    VLOAD_ADDR_SPACES(uint) \
     VLOAD_ADDR_SPACES(long) \
     VLOAD_ADDR_SPACES(ulong) \
     VLOAD_ADDR_SPACES(float) \
@@ -45,3 +44,53 @@ VLOAD_TYPES()
     VLOAD_ADDR_SPACES(double)
 #endif
 
+VLOAD_VECTORIZE(int, __private)
+VLOAD_VECTORIZE(int, __local)
+VLOAD_VECTORIZE(int, __constant)
+VLOAD_VECTORIZE(uint, __private)
+VLOAD_VECTORIZE(uint, __local)
+VLOAD_VECTORIZE(uint, __constant)
+
+_CLC_OVERLOAD _CLC_DEF int2 vload2(size_t offset, const global int *x) {
+  return (int2)(x[offset] , x[offset+1]);
+}
+_CLC_OVERLOAD _CLC_DEF int3 vload3(size_t offset, const global int *x) {
+  return (int3)(vload2(offset, x), x[offset+2]);
+}
+_CLC_OVERLOAD _CLC_DEF uint2 vload2(size_t offset, const global uint *x) {
+  return (uint2)(x[offset] , x[offset+1]);
+}
+_CLC_OVERLOAD _CLC_DEF uint3 vload3(size_t offset, const global uint *x) {
+  return (uint3)(vload2(offset, x), x[offset+2]);
+}
+        
+/*Note: It is known that R600 doesn't support load <2 x ?> and <3 x ?>... so
+ * they aren't actually overridden here
+ */
+_CLC_DECL int4 __clc_vload4_int__global(size_t offset, const __global int *);
+_CLC_DECL int8 __clc_vload8_int__global(size_t offset, const __global int *);
+_CLC_DECL int16 __clc_vload16_int__global(size_t offset, const __global int *);
+
+_CLC_OVERLOAD _CLC_DEF int4 vload4(size_t offset, const global int *x) {
+  return __clc_vload4_int__global(offset, x);
+}
+_CLC_OVERLOAD _CLC_DEF int8 vload8(size_t offset, const global int *x) {
+  return __clc_vload8_int__global(offset, x);
+}
+_CLC_OVERLOAD _CLC_DEF int16 vload16(size_t offset, const global int *x) {
+  return __clc_vload16_int__global(offset, x);
+}
+
+_CLC_DECL uint4 __clc_vload4_uint__global(size_t offset, const __global uint *);
+_CLC_DECL uint8 __clc_vload8_uint__global(size_t offset, const __global uint *);
+_CLC_DECL uint16 __clc_vload16_uint__global(size_t offset, const __global uint *);
+
+_CLC_OVERLOAD _CLC_DEF uint4 vload4(size_t offset, const global uint *x) {
+  return __clc_vload4_uint__global(offset, x);
+}
+_CLC_OVERLOAD _CLC_DEF uint8 vload8(size_t offset, const global uint *x) {
+  return __clc_vload8_uint__global(offset, x);
+}
+_CLC_OVERLOAD _CLC_DEF uint16 vload16(size_t offset, const global uint *x) {
+  return __clc_vload16_uint__global(offset, x);
+}
\ No newline at end of file
diff --git a/generic/lib/shared/vload_if.ll b/generic/lib/shared/vload_if.ll
new file mode 100644
index 0000000..2a0fa99
--- /dev/null
+++ b/generic/lib/shared/vload_if.ll
@@ -0,0 +1,64 @@
+;Start int global vload
+
+declare <2 x i32> @__clc_vload2_impl_int__global(i32 %x, i32 %y)
+define <2 x i32> @__clc_vload2_int__global(i32 %x, i32 %y) nounwind readnone alwaysinline {
+  %call = call <2 x i32> @__clc_vload2_impl_int__global(i32 %x, i32 %y)
+  ret <2 x i32> %call
+}
+
+declare <3 x i32> @__clc_vload3_impl_int__global(i32 %x, i32 %y)
+define <3 x i32> @__clc_vload3_int__global(i32 %x, i32 %y) nounwind readnone alwaysinline {
+  %call = call <3 x i32> @__clc_vload3_impl_int__global(i32 %x, i32 %y)
+  ret <3 x i32> %call
+}
+
+declare <4 x i32> @__clc_vload4_impl_int__global(i32 %x, i32 %y)
+define <4 x i32> @__clc_vload4_int__global(i32 %x, i32 %y) nounwind readnone alwaysinline {
+  %call = call <4 x i32> @__clc_vload4_impl_int__global(i32 %x, i32 %y)
+  ret <4 x i32> %call
+}
+
+declare <8 x i32> @__clc_vload8_impl_int__global(i32 %x, i32 %y)
+define <8 x i32> @__clc_vload8_int__global(i32 %x, i32 %y) nounwind readnone alwaysinline {
+  %call = call <8 x i32> @__clc_vload8_impl_int__global(i32 %x, i32 %y)
+  ret <8 x i32> %call
+}
+
+declare <16 x i32> @__clc_vload16_impl_int__global(i32 %x, i32 %y)
+define <16 x i32> @__clc_vload16_int__global(i32 %x, i32 %y) nounwind readnone alwaysinline {
+  %call = call <16 x i32> @__clc_vload16_impl_int__global(i32 %x, i32 %y)
+  ret <16 x i32> %call
+}
+
+
+;Start uint global vload
+
+declare <2 x i32> @__clc_vload2_impl_uint__global(i32 %x, i32 %y)
+define <2 x i32> @__clc_vload2_uint__global(i32 %x, i32 %y) nounwind readnone alwaysinline {
+  %call = call <2 x i32> @__clc_vload2_impl_uint__global(i32 %x, i32 %y)
+  ret <2 x i32> %call
+}
+
+declare <3 x i32> @__clc_vload3_impl_uint__global(i32 %x, i32 %y)
+define <3 x i32> @__clc_vload3_uint__global(i32 %x, i32 %y) nounwind readnone alwaysinline {
+  %call = call <3 x i32> @__clc_vload3_impl_uint__global(i32 %x, i32 %y)
+  ret <3 x i32> %call
+}
+
+declare <4 x i32> @__clc_vload4_impl_uint__global(i32 %x, i32 %y)
+define <4 x i32> @__clc_vload4_uint__global(i32 %x, i32 %y) nounwind readnone alwaysinline {
+  %call = call <4 x i32> @__clc_vload4_impl_uint__global(i32 %x, i32 %y)
+  ret <4 x i32> %call
+}
+
+declare <8 x i32> @__clc_vload8_impl_uint__global(i32 %x, i32 %y)
+define <8 x i32> @__clc_vload8_uint__global(i32 %x, i32 %y) nounwind readnone alwaysinline {
+  %call = call <8 x i32> @__clc_vload8_impl_uint__global(i32 %x, i32 %y)
+  ret <8 x i32> %call
+}
+
+declare <16 x i32> @__clc_vload16_impl_uint__global(i32 %x, i32 %y)
+define <16 x i32> @__clc_vload16_uint__global(i32 %x, i32 %y) nounwind readnone alwaysinline {
+  %call = call <16 x i32> @__clc_vload16_impl_uint__global(i32 %x, i32 %y)
+  ret <16 x i32> %call
+}
diff --git a/generic/lib/shared/vload_impl.ll b/generic/lib/shared/vload_impl.ll
new file mode 100644
index 0000000..f1735f3
--- /dev/null
+++ b/generic/lib/shared/vload_impl.ll
@@ -0,0 +1,90 @@
+; This provides optimized implementations of vload4/8/16 for 32-bit int/uint
+
+define <2 x i32> @__clc_vload2_impl_int__global(i32 %offset,  i32 addrspace(1)* nocapture %addr) nounwind readnone alwaysinline {
+  %1 = ptrtoint i32 addrspace(1)* %addr to i32
+  %2 = add i32 %1, %offset
+  %3 = inttoptr i32 %2 to <2 x i32> addrspace(1)*
+  %4 = load <2 x i32> addrspace(1)* %3, align 4, !tbaa !3
+  ret <2 x i32> %4
+}
+
+define <2 x i32> @__clc_vload3_impl_int__global(i32 %offset,  i32 addrspace(1)* nocapture %addr) nounwind readnone alwaysinline {
+  %1 = ptrtoint i32 addrspace(1)* %addr to i32
+  %2 = add i32 %1, %offset
+  %3 = inttoptr i32 %2 to <2 x i32> addrspace(1)*
+  %4 = load <2 x i32> addrspace(1)* %3, align 4, !tbaa !3
+  ret <2 x i32> %4
+}
+
+define <4 x i32> @__clc_vload4_impl_int__global(i32 %offset,  i32 addrspace(1)* nocapture %addr) nounwind readnone alwaysinline {
+  %1 = ptrtoint i32 addrspace(1)* %addr to i32
+  %2 = add i32 %1, %offset
+  %3 = inttoptr i32 %2 to <4 x i32> addrspace(1)*
+  %4 = load <4 x i32> addrspace(1)* %3, align 4, !tbaa !3
+  ret <4 x i32> %4
+}
+
+define <8 x i32> @__clc_vload8_impl_int__global(i32 %offset,  i32 addrspace(1)* nocapture %addr) nounwind readnone alwaysinline {
+  %1 = ptrtoint i32 addrspace(1)* %addr to i32
+  %2 = add i32 %1, %offset
+  %3 = inttoptr i32 %2 to <8 x i32> addrspace(1)*
+  %4 = load <8 x i32> addrspace(1)* %3, align 4, !tbaa !3
+  ret <8 x i32> %4
+}
+
+define <16 x i32> @__clc_vload16_impl_int__global(i32 %offset,  i32 addrspace(1)* nocapture %addr) nounwind readnone alwaysinline {
+  %1 = ptrtoint i32 addrspace(1)* %addr to i32
+  %2 = add i32 %1, %offset
+  %3 = inttoptr i32 %2 to <16 x i32> addrspace(1)*
+  %4 = load <16 x i32> addrspace(1)* %3, align 4, !tbaa !3
+  ret <16 x i32> %4
+}
+
+define <2 x i32> @__clc_vload2_impl_uint__global(i32 %offset,  i32 addrspace(1)* nocapture %addr) nounwind readnone alwaysinline {
+  %1 = ptrtoint i32 addrspace(1)* %addr to i32
+  %2 = add i32 %1, %offset
+  %3 = inttoptr i32 %2 to <2 x i32> addrspace(1)*
+  %4 = load <2 x i32> addrspace(1)* %3, align 4, !tbaa !3
+  ret <2 x i32> %4
+}
+
+define <2 x i32> @__clc_vload3_impl_uint__global(i32 %offset,  i32 addrspace(1)* nocapture %addr) nounwind readnone alwaysinline {
+  %1 = ptrtoint i32 addrspace(1)* %addr to i32
+  %2 = add i32 %1, %offset
+  %3 = inttoptr i32 %2 to <2 x i32> addrspace(1)*
+  %4 = load <2 x i32> addrspace(1)* %3, align 4, !tbaa !3
+  ret <2 x i32> %4
+}
+
+define <4 x i32> @__clc_vload4_impl_uint__global(i32 %offset,  i32 addrspace(1)* nocapture %addr) nounwind readnone alwaysinline {
+  %1 = ptrtoint i32 addrspace(1)* %addr to i32
+  %2 = add i32 %1, %offset
+  %3 = inttoptr i32 %2 to <4 x i32> addrspace(1)*
+  %4 = load <4 x i32> addrspace(1)* %3, align 4, !tbaa !3
+  ret <4 x i32> %4
+}
+
+define <8 x i32> @__clc_vload8_impl_uint__global(i32 %offset,  i32 addrspace(1)* nocapture %addr) nounwind readnone alwaysinline {
+  %1 = ptrtoint i32 addrspace(1)* %addr to i32
+  %2 = add i32 %1, %offset
+  %3 = inttoptr i32 %2 to <8 x i32> addrspace(1)*
+  %4 = load <8 x i32> addrspace(1)* %3, align 4, !tbaa !3
+  ret <8 x i32> %4
+}
+
+define <16 x i32> @__clc_vload16_impl_uint__global(i32 %offset,  i32 addrspace(1)* nocapture %addr) nounwind readnone alwaysinline {
+  %1 = ptrtoint i32 addrspace(1)* %addr to i32
+  %2 = add i32 %1, %offset
+  %3 = inttoptr i32 %2 to <16 x i32> addrspace(1)*
+  %4 = load <16 x i32> addrspace(1)* %3, align 4, !tbaa !3
+  ret <16 x i32> %4
+}
+
+
+!1 = metadata !{metadata !"char", metadata !5}
+!2 = metadata !{metadata !"short", metadata !5}
+!3 = metadata !{metadata !"int", metadata !5}
+!4 = metadata !{metadata !"long", metadata !5}
+!5 = metadata !{metadata !"omnipotent char", metadata !6}
+!6 = metadata !{metadata !"Simple C/C++ TBAA"}
+