[Mesa-dev] [PATCH] gallivm: fix SCALED -> NORM conversions

[Jose Fonseca]

Looks good to me AFAICT. Thanks for the update.

Jose
________________________________________
From: sroland at vmware.com <sroland at vmware.com>
Sent: 18 June 2014 00:36
To: Jose Fonseca; mesa-dev at lists.freedesktop.org
Cc: Roland Scheidegger
Subject: [PATCH] gallivm: fix SCALED -> NORM conversions

From: Roland Scheidegger <sroland at vmware.com>

Such conversions (which are most likely rather pointless in practice) were
resulting in shifts with negative shift counts and shifts with counts the same
as the bit width. This was always undefined in llvm, the code generated was
rather horrendous but happened to work.
So make sure such shifts are filtered out and replaced with something that
works (the generated code is still just as horrendous as before).

This fixes lp_test_format, https://bugs.freedesktop.org/show_bug.cgi?id=73846.

v2: prettify by using build context shift helpers.
---
 src/gallium/auxiliary/gallivm/lp_bld_conv.c | 39 +++++++++++++++++------------
 1 file changed, 23 insertions(+), 16 deletions(-)

diff --git a/src/gallium/auxiliary/gallivm/lp_bld_conv.c b/src/gallium/auxiliary/gallivm/lp_bld_conv.c
index d3bf621..1424447 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_conv.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_conv.c
@@ -792,29 +792,23 @@ lp_build_conv(struct gallivm_state *gallivm,
       unsigned dst_shift = lp_const_shift(dst_type);
       unsigned src_offset = lp_const_offset(src_type);
       unsigned dst_offset = lp_const_offset(dst_type);
+      struct lp_build_context bld;
+      lp_build_context_init(&bld, gallivm, tmp_type);

       /* Compensate for different offsets */
-      if (dst_offset > src_offset && src_type.width > dst_type.width) {
+      /* sscaled -> unorm and similar would cause negative shift count, skip */
+      if (dst_offset > src_offset && src_type.width > dst_type.width && src_shift > 0) {
          for (i = 0; i < num_tmps; ++i) {
             LLVMValueRef shifted;
-            LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type, src_shift - 1);
-            if(src_type.sign)
-               shifted = LLVMBuildAShr(builder, tmp[i], shift, "");
-            else
-               shifted = LLVMBuildLShr(builder, tmp[i], shift, "");

+            shifted = lp_build_shr_imm(&bld, tmp[i], src_shift - 1);
             tmp[i] = LLVMBuildSub(builder, tmp[i], shifted, "");
          }
       }

       if(src_shift > dst_shift) {
-         LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type,
-                                                     src_shift - dst_shift);
          for(i = 0; i < num_tmps; ++i)
-            if(src_type.sign)
-               tmp[i] = LLVMBuildAShr(builder, tmp[i], shift, "");
-            else
-               tmp[i] = LLVMBuildLShr(builder, tmp[i], shift, "");
+            tmp[i] = lp_build_shr_imm(&bld, tmp[i], src_shift - dst_shift);
       }
    }

@@ -900,14 +894,27 @@ lp_build_conv(struct gallivm_state *gallivm,
        unsigned dst_shift = lp_const_shift(dst_type);
        unsigned src_offset = lp_const_offset(src_type);
        unsigned dst_offset = lp_const_offset(dst_type);
+       struct lp_build_context bld;
+       lp_build_context_init(&bld, gallivm, tmp_type);

        if (src_shift < dst_shift) {
           LLVMValueRef pre_shift[LP_MAX_VECTOR_LENGTH];
-          LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type, dst_shift - src_shift);

-          for (i = 0; i < num_tmps; ++i) {
-             pre_shift[i] = tmp[i];
-             tmp[i] = LLVMBuildShl(builder, tmp[i], shift, "");
+          if (dst_shift - src_shift < dst_type.width) {
+             for (i = 0; i < num_tmps; ++i) {
+                pre_shift[i] = tmp[i];
+                tmp[i] = lp_build_shl_imm(&bld, tmp[i], dst_shift - src_shift);
+             }
+          }
+          else {
+             /*
+              * This happens for things like sscaled -> unorm conversions. Shift
+              * counts equal to bit width cause undefined results, so hack around it.
+              */
+             for (i = 0; i < num_tmps; ++i) {
+                pre_shift[i] = tmp[i];
+                tmp[i] = lp_build_zero(gallivm, dst_type);
+             }
           }

           /* Compensate for different offsets */
--
1.9.1