gallivm: Fix Altivec pack intrinsics for little-endian

This patch fixes use of Altivec pack intrinsics on little-endian PowerPC
systems.  Since little-endian operation only affects the load and store
instructions, the semantics of pack (and other) instructions that take
two input vectors implicitly change: the pack instructions still fill
a register placing values from the first operand into the "high" parts
of the register, and values from the second operand into the "low" parts
of the register, but since vector loads and stores perform an endian swap,
the high parts end up at high memory addresses.

To still achieve the desired effect, we have to swap the two inputs to
the pack instruction on little-endian systems.  This is done automatically
by the back-end for instructions generated by LLVM, but needs to be done
manually when emitting intrisincs (which still result in that instruction
being emitted directly).

Signed-off-by: Ulrich Weigand <ulrich.weigand@de.ibm.com>
Signed-off-by: Maarten Lankhorst <dev@mblankhorst.nl>

diff --git a/src/gallium/auxiliary/gallivm/lp_bld_pack.c b/src/gallium/auxiliary/gallivm/lp_bld_pack.c
index a48a9225eb22edb41946016d466f21ff7346b5ca..cdf6d80c2611553a49f269c08753b0ddffe31f79 100644 (file)
--- a/src/gallium/auxiliary/gallivm/lp_bld_pack.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_pack.c
@@ -464,6 +464,7 @@ lp_build_pack2(struct gallivm_state *gallivm,
    if((util_cpu_caps.has_sse2 || util_cpu_caps.has_altivec) &&
        src_type.width * src_type.length >= 128) {
       const char *intrinsic = NULL;
+      boolean swap_intrinsic_operands = FALSE;
 
       switch(src_type.width) {
       case 32:
@@ -482,6 +483,9 @@ lp_build_pack2(struct gallivm_state *gallivm,
            } else {
               intrinsic = "llvm.ppc.altivec.vpkuwus";
            }
+#ifdef PIPE_ARCH_LITTLE_ENDIAN
+           swap_intrinsic_operands = TRUE;
+#endif
          }
          break;
       case 16:
@@ -490,12 +494,18 @@ lp_build_pack2(struct gallivm_state *gallivm,
               intrinsic = "llvm.x86.sse2.packsswb.128";
             } else if (util_cpu_caps.has_altivec) {
               intrinsic = "llvm.ppc.altivec.vpkshss";
+#ifdef PIPE_ARCH_LITTLE_ENDIAN
+              swap_intrinsic_operands = TRUE;
+#endif
             }
          } else {
             if (util_cpu_caps.has_sse2) {
               intrinsic = "llvm.x86.sse2.packuswb.128";
             } else if (util_cpu_caps.has_altivec) {
              intrinsic = "llvm.ppc.altivec.vpkshus";
+#ifdef PIPE_ARCH_LITTLE_ENDIAN
+              swap_intrinsic_operands = TRUE;
+#endif
             }
          }
          break;
@@ -504,7 +514,11 @@ lp_build_pack2(struct gallivm_state *gallivm,
       if (intrinsic) {
          if (src_type.width * src_type.length == 128) {
             LLVMTypeRef intr_vec_type = lp_build_vec_type(gallivm, intr_type);
-            res = lp_build_intrinsic_binary(builder, intrinsic, intr_vec_type, lo, hi);
+            if (swap_intrinsic_operands) {
+               res = lp_build_intrinsic_binary(builder, intrinsic, intr_vec_type, hi, lo);
+            } else {
+               res = lp_build_intrinsic_binary(builder, intrinsic, intr_vec_type, lo, hi);
+            }
             if (dst_vec_type != intr_vec_type) {
                res = LLVMBuildBitCast(builder, res, dst_vec_type, "");
             }
@@ -513,6 +527,8 @@ lp_build_pack2(struct gallivm_state *gallivm,
             int num_split = src_type.width * src_type.length / 128;
             int i;
             int nlen = 128 / src_type.width;
+            int lo_off = swap_intrinsic_operands ? nlen : 0;
+            int hi_off = swap_intrinsic_operands ? 0 : nlen;
             struct lp_type ndst_type = lp_type_unorm(dst_type.width, 128);
             struct lp_type nintr_type = lp_type_unorm(intr_type.width, 128);
             LLVMValueRef tmpres[LP_MAX_VECTOR_WIDTH / 128];
@@ -524,9 +540,9 @@ lp_build_pack2(struct gallivm_state *gallivm,
 
             for (i = 0; i < num_split / 2; i++) {
                tmplo = lp_build_extract_range(gallivm,
-                                              lo, i*nlen*2, nlen);
+                                              lo, i*nlen*2 + lo_off, nlen);
                tmphi = lp_build_extract_range(gallivm,
-                                              lo, i*nlen*2 + nlen, nlen);
+                                              lo, i*nlen*2 + hi_off, nlen);
                tmpres[i] = lp_build_intrinsic_binary(builder, intrinsic,
                                                      nintr_vec_type, tmplo, tmphi);
                if (ndst_vec_type != nintr_vec_type) {
@@ -535,9 +551,9 @@ lp_build_pack2(struct gallivm_state *gallivm,
             }
             for (i = 0; i < num_split / 2; i++) {
                tmplo = lp_build_extract_range(gallivm,
-                                              hi, i*nlen*2, nlen);
+                                              hi, i*nlen*2 + lo_off, nlen);
                tmphi = lp_build_extract_range(gallivm,
-                                              hi, i*nlen*2 + nlen, nlen);
+                                              hi, i*nlen*2 + hi_off, nlen);
                tmpres[i+num_split/2] = lp_build_intrinsic_binary(builder, intrinsic,
                                                                  nintr_vec_type,
                                                                  tmplo, tmphi);