gallivm: Always use floating-point operators for floating-point types

This fixes the assert added in LLVM 2.8:
assert(getType()->isIntOrIntVectorTy() &&
       "Tried to create an integer operation on a non-integer type!")

But it also fixes some subtle bugs, since we should've been doing this
since LLVM 2.6 anyway.

Includes a modified patch from steckdenis@yahoo.fr for the
FNeg instructions in emit_fetch(); thanks for pointing those out.

http://bugs.freedesktop.org/29404
http://bugs.freedesktop.org/29407

Signed-off-by: José Fonseca <jfonseca@vmware.com>

7 files changed, 137 insertions, 92 deletions

diff --git a/src/gallium/auxiliary/gallivm/lp_bld_arit.c b/src/gallium/auxiliary/gallivm/lp_bld_arit.c
index 98e8e49..d2dde41 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_arit.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_arit.c
@@ -173,9 +173,15 @@ lp_build_comp(struct lp_build_context *bld,
    }
 
    if(LLVMIsConstant(a))
-      return LLVMConstSub(bld->one, a);
+      if (type.floating)
+          return LLVMConstFSub(bld->one, a);
+      else
+          return LLVMConstSub(bld->one, a);
    else
-      return LLVMBuildSub(bld->builder, bld->one, a, "");
+      if (type.floating)
+         return LLVMBuildFSub(bld->builder, bld->one, a, "");
+      else
+         return LLVMBuildSub(bld->builder, bld->one, a, "");
 }
 
 
@@ -220,9 +226,15 @@ lp_build_add(struct lp_build_context *bld,
    }
 
    if(LLVMIsConstant(a) && LLVMIsConstant(b))
-      res = LLVMConstAdd(a, b);
+      if (type.floating)
+         res = LLVMConstFAdd(a, b);
+      else
+         res = LLVMConstAdd(a, b);
    else
-      res = LLVMBuildAdd(bld->builder, a, b, "");
+      if (type.floating)
+         res = LLVMBuildFAdd(bld->builder, a, b, "");
+      else
+         res = LLVMBuildAdd(bld->builder, a, b, "");
 
    /* clamp to ceiling of 1.0 */
    if(bld->type.norm && (bld->type.floating || bld->type.fixed))
@@ -256,9 +268,16 @@ lp_build_sum_vector(struct lp_build_context *bld,
 
    for (i = 1; i < type.length; i++) {
       index = LLVMConstInt(LLVMInt32Type(), i, 0);
-      res = LLVMBuildAdd(bld->builder, res,
-                         LLVMBuildExtractElement(bld->builder, a, index, ""),
-                         "");
+      if (type.floating)
+         res = LLVMBuildFAdd(bld->builder, res,
+                            LLVMBuildExtractElement(bld->builder,
+                                                    a, index, ""),
+                            "");
+      else
+         res = LLVMBuildAdd(bld->builder, res,
+                            LLVMBuildExtractElement(bld->builder,
+                                                    a, index, ""),
+                            "");
    }
 
    return res;
@@ -306,9 +325,15 @@ lp_build_sub(struct lp_build_context *bld,
    }
 
    if(LLVMIsConstant(a) && LLVMIsConstant(b))
-      res = LLVMConstSub(a, b);
+      if (type.floating)
+         res = LLVMConstFSub(a, b);
+      else
+         res = LLVMConstSub(a, b);
    else
-      res = LLVMBuildSub(bld->builder, a, b, "");
+      if (type.floating)
+         res = LLVMBuildFSub(bld->builder, a, b, "");
+      else
+         res = LLVMBuildSub(bld->builder, a, b, "");
 
    if(bld->type.norm && (bld->type.floating || bld->type.fixed))
       res = lp_build_max_simple(bld, res, bld->zero);
@@ -442,7 +467,10 @@ lp_build_mul(struct lp_build_context *bld,
       shift = NULL;
 
    if(LLVMIsConstant(a) && LLVMIsConstant(b)) {
-      res =  LLVMConstMul(a, b);
+      if (type.floating)
+         res = LLVMConstFMul(a, b);
+      else
+         res = LLVMConstMul(a, b);
       if(shift) {
          if(type.sign)
             res = LLVMConstAShr(res, shift);
@@ -451,7 +479,10 @@ lp_build_mul(struct lp_build_context *bld,
       }
    }
    else {
-      res = LLVMBuildMul(bld->builder, a, b, "");
+      if (type.floating)
+         res = LLVMBuildFMul(bld->builder, a, b, "");
+      else
+         res = LLVMBuildMul(bld->builder, a, b, "");
       if(shift) {
          if(type.sign)
             res = LLVMBuildAShr(bld->builder, res, shift, "");
@@ -481,7 +512,10 @@ lp_build_mul_imm(struct lp_build_context *bld,
       return a;
 
    if(b == -1)
-      return LLVMBuildNeg(bld->builder, a, "");
+      if (bld->type.floating)
+         return LLVMBuildFNeg(bld->builder, a, "");
+      else
+         return LLVMBuildNeg(bld->builder, a, "");
 
    if(b == 2 && bld->type.floating)
       return lp_build_add(bld, a, a);
@@ -714,7 +748,12 @@ LLVMValueRef
 lp_build_negate(struct lp_build_context *bld,
                 LLVMValueRef a)
 {
-   return LLVMBuildNeg(bld->builder, a, "");
+   if (bld->type.floating)
+      a = LLVMBuildFNeg(bld->builder, a, "");
+   else
+      a = LLVMBuildNeg(bld->builder, a, "");
+
+   return a;
 }
 
 
@@ -1033,7 +1072,7 @@ lp_build_iround(struct lp_build_context *bld,
       half = LLVMBuildOr(bld->builder, sign, half, "");
       half = LLVMBuildBitCast(bld->builder, half, vec_type, "");
 
-      res = LLVMBuildAdd(bld->builder, a, half, "");
+      res = LLVMBuildFAdd(bld->builder, a, half, "");
    }
 
    res = LLVMBuildFPToSI(bld->builder, res, int_vec_type, "");
@@ -1082,7 +1121,7 @@ lp_build_ifloor(struct lp_build_context *bld,
       offset = LLVMBuildAnd(bld->builder, offset, sign, "");
       offset = LLVMBuildBitCast(bld->builder, offset, vec_type, "ifloor.offset");
 
-      res = LLVMBuildAdd(bld->builder, a, offset, "ifloor.res");
+      res = LLVMBuildFAdd(bld->builder, a, offset, "ifloor.res");
    }
 
    /* round to nearest (toward zero) */
@@ -1132,7 +1171,7 @@ lp_build_iceil(struct lp_build_context *bld,
       offset = LLVMBuildAnd(bld->builder, offset, sign, "");
       offset = LLVMBuildBitCast(bld->builder, offset, vec_type, "iceil.offset");
 
-      res = LLVMBuildAdd(bld->builder, a, offset, "iceil.res");
+      res = LLVMBuildFAdd(bld->builder, a, offset, "iceil.res");
    }
 
    /* round to nearest (toward zero) */
@@ -1197,9 +1236,9 @@ lp_build_rcp(struct lp_build_context *bld,
 
       rcp_a = lp_build_intrinsic_unary(bld->builder, "llvm.x86.sse.rcp.ps", lp_build_vec_type(type), a);
 
-      res = LLVMBuildMul(bld->builder, a, rcp_a, "");
-      res = LLVMBuildSub(bld->builder, two, res, "");
-      res = LLVMBuildMul(bld->builder, res, rcp_a, "");
+      res = LLVMBuildFMul(bld->builder, a, rcp_a, "");
+      res = LLVMBuildFSub(bld->builder, two, res, "");
+      res = LLVMBuildFMul(bld->builder, res, rcp_a, "");
 
       return rcp_a;
 #else
@@ -1282,7 +1321,7 @@ lp_build_sin(struct lp_build_context *bld,
     */
    
    LLVMValueRef FOPi = lp_build_const_v4sf(1.27323954473516);
-   LLVMValueRef scale_y = LLVMBuildMul(b, x_abs, FOPi, "scale_y");
+   LLVMValueRef scale_y = LLVMBuildFMul(b, x_abs, FOPi, "scale_y");
 
    /*
     * store the integer part of y in mm0
@@ -1356,9 +1395,9 @@ lp_build_sin(struct lp_build_context *bld,
     * xmm2 = _mm_mul_ps(y, xmm2);
     * xmm3 = _mm_mul_ps(y, xmm3);
     */
-   LLVMValueRef xmm1 = LLVMBuildMul(b, y_2, DP1, "xmm1");
-   LLVMValueRef xmm2 = LLVMBuildMul(b, y_2, DP2, "xmm2");
-   LLVMValueRef xmm3 = LLVMBuildMul(b, y_2, DP3, "xmm3");
+   LLVMValueRef xmm1 = LLVMBuildFMul(b, y_2, DP1, "xmm1");
+   LLVMValueRef xmm2 = LLVMBuildFMul(b, y_2, DP2, "xmm2");
+   LLVMValueRef xmm3 = LLVMBuildFMul(b, y_2, DP3, "xmm3");
 
    /*
     * x = _mm_add_ps(x, xmm1);
@@ -1366,16 +1405,16 @@ lp_build_sin(struct lp_build_context *bld,
     * x = _mm_add_ps(x, xmm3);
     */ 
 
-   LLVMValueRef x_1 = LLVMBuildAdd(b, x_abs, xmm1, "x_1");
-   LLVMValueRef x_2 = LLVMBuildAdd(b, x_1, xmm2, "x_2");
-   LLVMValueRef x_3 = LLVMBuildAdd(b, x_2, xmm3, "x_3");
+   LLVMValueRef x_1 = LLVMBuildFAdd(b, x_abs, xmm1, "x_1");
+   LLVMValueRef x_2 = LLVMBuildFAdd(b, x_1, xmm2, "x_2");
+   LLVMValueRef x_3 = LLVMBuildFAdd(b, x_2, xmm3, "x_3");
 
    /*
     * Evaluate the first polynom  (0 <= x <= Pi/4)
     *
     * z = _mm_mul_ps(x,x);
     */
-   LLVMValueRef z = LLVMBuildMul(b, x_3, x_3, "z");
+   LLVMValueRef z = LLVMBuildFMul(b, x_3, x_3, "z");
 
    /*
     * _PS_CONST(coscof_p0,  2.443315711809948E-005);
@@ -1390,12 +1429,12 @@ lp_build_sin(struct lp_build_context *bld,
     * y = *(v4sf*)_ps_coscof_p0;
     * y = _mm_mul_ps(y, z);
     */
-   LLVMValueRef y_3 = LLVMBuildMul(b, z, coscof_p0, "y_3");
-   LLVMValueRef y_4 = LLVMBuildAdd(b, y_3, coscof_p1, "y_4");
-   LLVMValueRef y_5 = LLVMBuildMul(b, y_4, z, "y_5");
-   LLVMValueRef y_6 = LLVMBuildAdd(b, y_5, coscof_p2, "y_6");
-   LLVMValueRef y_7 = LLVMBuildMul(b, y_6, z, "y_7");
-   LLVMValueRef y_8 = LLVMBuildMul(b, y_7, z, "y_8");
+   LLVMValueRef y_3 = LLVMBuildFMul(b, z, coscof_p0, "y_3");
+   LLVMValueRef y_4 = LLVMBuildFAdd(b, y_3, coscof_p1, "y_4");
+   LLVMValueRef y_5 = LLVMBuildFMul(b, y_4, z, "y_5");
+   LLVMValueRef y_6 = LLVMBuildFAdd(b, y_5, coscof_p2, "y_6");
+   LLVMValueRef y_7 = LLVMBuildFMul(b, y_6, z, "y_7");
+   LLVMValueRef y_8 = LLVMBuildFMul(b, y_7, z, "y_8");
 
 
    /*
@@ -1404,10 +1443,10 @@ lp_build_sin(struct lp_build_context *bld,
     * y = _mm_add_ps(y, *(v4sf*)_ps_1);
     */ 
    LLVMValueRef half = lp_build_const_v4sf(0.5);
-   LLVMValueRef tmp = LLVMBuildMul(b, z, half, "tmp");
-   LLVMValueRef y_9 = LLVMBuildSub(b, y_8, tmp, "y_8");
+   LLVMValueRef tmp = LLVMBuildFMul(b, z, half, "tmp");
+   LLVMValueRef y_9 = LLVMBuildFSub(b, y_8, tmp, "y_8");
    LLVMValueRef one = lp_build_const_v4sf(1.0);
-   LLVMValueRef y_10 = LLVMBuildAdd(b, y_9, one, "y_9");
+   LLVMValueRef y_10 = LLVMBuildFAdd(b, y_9, one, "y_9");
 
    /*
     * _PS_CONST(sincof_p0, -1.9515295891E-4);
@@ -1431,13 +1470,13 @@ lp_build_sin(struct lp_build_context *bld,
     * y2 = _mm_add_ps(y2, x);
     */
 
-   LLVMValueRef y2_3 = LLVMBuildMul(b, z, sincof_p0, "y2_3");
-   LLVMValueRef y2_4 = LLVMBuildAdd(b, y2_3, sincof_p1, "y2_4");
-   LLVMValueRef y2_5 = LLVMBuildMul(b, y2_4, z, "y2_5");
-   LLVMValueRef y2_6 = LLVMBuildAdd(b, y2_5, sincof_p2, "y2_6");
-   LLVMValueRef y2_7 = LLVMBuildMul(b, y2_6, z, "y2_7");
-   LLVMValueRef y2_8 = LLVMBuildMul(b, y2_7, x_3, "y2_8");
-   LLVMValueRef y2_9 = LLVMBuildAdd(b, y2_8, x_3, "y2_9");
+   LLVMValueRef y2_3 = LLVMBuildFMul(b, z, sincof_p0, "y2_3");
+   LLVMValueRef y2_4 = LLVMBuildFAdd(b, y2_3, sincof_p1, "y2_4");
+   LLVMValueRef y2_5 = LLVMBuildFMul(b, y2_4, z, "y2_5");
+   LLVMValueRef y2_6 = LLVMBuildFAdd(b, y2_5, sincof_p2, "y2_6");
+   LLVMValueRef y2_7 = LLVMBuildFMul(b, y2_6, z, "y2_7");
+   LLVMValueRef y2_8 = LLVMBuildFMul(b, y2_7, x_3, "y2_8");
+   LLVMValueRef y2_9 = LLVMBuildFAdd(b, y2_8, x_3, "y2_9");
 
    /*
     * select the correct result from the two polynoms
@@ -1493,7 +1532,7 @@ lp_build_cos(struct lp_build_context *bld,
     */
    
    LLVMValueRef FOPi = lp_build_const_v4sf(1.27323954473516);
-   LLVMValueRef scale_y = LLVMBuildMul(b, x_abs, FOPi, "scale_y");
+   LLVMValueRef scale_y = LLVMBuildFMul(b, x_abs, FOPi, "scale_y");
 
    /*
     * store the integer part of y in mm0
@@ -1573,9 +1612,9 @@ lp_build_cos(struct lp_build_context *bld,
     * xmm2 = _mm_mul_ps(y, xmm2);
     * xmm3 = _mm_mul_ps(y, xmm3);
     */
-   LLVMValueRef xmm1 = LLVMBuildMul(b, y_2, DP1, "xmm1");
-   LLVMValueRef xmm2 = LLVMBuildMul(b, y_2, DP2, "xmm2");
-   LLVMValueRef xmm3 = LLVMBuildMul(b, y_2, DP3, "xmm3");
+   LLVMValueRef xmm1 = LLVMBuildFMul(b, y_2, DP1, "xmm1");
+   LLVMValueRef xmm2 = LLVMBuildFMul(b, y_2, DP2, "xmm2");
+   LLVMValueRef xmm3 = LLVMBuildFMul(b, y_2, DP3, "xmm3");
 
    /*
     * x = _mm_add_ps(x, xmm1);
@@ -1583,16 +1622,16 @@ lp_build_cos(struct lp_build_context *bld,
     * x = _mm_add_ps(x, xmm3);
     */ 
 
-   LLVMValueRef x_1 = LLVMBuildAdd(b, x_abs, xmm1, "x_1");
-   LLVMValueRef x_2 = LLVMBuildAdd(b, x_1, xmm2, "x_2");
-   LLVMValueRef x_3 = LLVMBuildAdd(b, x_2, xmm3, "x_3");
+   LLVMValueRef x_1 = LLVMBuildFAdd(b, x_abs, xmm1, "x_1");
+   LLVMValueRef x_2 = LLVMBuildFAdd(b, x_1, xmm2, "x_2");
+   LLVMValueRef x_3 = LLVMBuildFAdd(b, x_2, xmm3, "x_3");
 
    /*
     * Evaluate the first polynom  (0 <= x <= Pi/4)
     *
     * z = _mm_mul_ps(x,x);
     */
-   LLVMValueRef z = LLVMBuildMul(b, x_3, x_3, "z");
+   LLVMValueRef z = LLVMBuildFMul(b, x_3, x_3, "z");
 
    /*
     * _PS_CONST(coscof_p0,  2.443315711809948E-005);
@@ -1607,12 +1646,12 @@ lp_build_cos(struct lp_build_context *bld,
     * y = *(v4sf*)_ps_coscof_p0;
     * y = _mm_mul_ps(y, z);
     */
-   LLVMValueRef y_3 = LLVMBuildMul(b, z, coscof_p0, "y_3");
-   LLVMValueRef y_4 = LLVMBuildAdd(b, y_3, coscof_p1, "y_4");
-   LLVMValueRef y_5 = LLVMBuildMul(b, y_4, z, "y_5");
-   LLVMValueRef y_6 = LLVMBuildAdd(b, y_5, coscof_p2, "y_6");
-   LLVMValueRef y_7 = LLVMBuildMul(b, y_6, z, "y_7");
-   LLVMValueRef y_8 = LLVMBuildMul(b, y_7, z, "y_8");
+   LLVMValueRef y_3 = LLVMBuildFMul(b, z, coscof_p0, "y_3");
+   LLVMValueRef y_4 = LLVMBuildFAdd(b, y_3, coscof_p1, "y_4");
+   LLVMValueRef y_5 = LLVMBuildFMul(b, y_4, z, "y_5");
+   LLVMValueRef y_6 = LLVMBuildFAdd(b, y_5, coscof_p2, "y_6");
+   LLVMValueRef y_7 = LLVMBuildFMul(b, y_6, z, "y_7");
+   LLVMValueRef y_8 = LLVMBuildFMul(b, y_7, z, "y_8");
 
 
    /*
@@ -1621,10 +1660,10 @@ lp_build_cos(struct lp_build_context *bld,
     * y = _mm_add_ps(y, *(v4sf*)_ps_1);
     */ 
    LLVMValueRef half = lp_build_const_v4sf(0.5);
-   LLVMValueRef tmp = LLVMBuildMul(b, z, half, "tmp");
-   LLVMValueRef y_9 = LLVMBuildSub(b, y_8, tmp, "y_8");
+   LLVMValueRef tmp = LLVMBuildFMul(b, z, half, "tmp");
+   LLVMValueRef y_9 = LLVMBuildFSub(b, y_8, tmp, "y_8");
    LLVMValueRef one = lp_build_const_v4sf(1.0);
-   LLVMValueRef y_10 = LLVMBuildAdd(b, y_9, one, "y_9");
+   LLVMValueRef y_10 = LLVMBuildFAdd(b, y_9, one, "y_9");
 
    /*
     * _PS_CONST(sincof_p0, -1.9515295891E-4);
@@ -1648,13 +1687,13 @@ lp_build_cos(struct lp_build_context *bld,
     * y2 = _mm_add_ps(y2, x);
     */
 
-   LLVMValueRef y2_3 = LLVMBuildMul(b, z, sincof_p0, "y2_3");
-   LLVMValueRef y2_4 = LLVMBuildAdd(b, y2_3, sincof_p1, "y2_4");
-   LLVMValueRef y2_5 = LLVMBuildMul(b, y2_4, z, "y2_5");
-   LLVMValueRef y2_6 = LLVMBuildAdd(b, y2_5, sincof_p2, "y2_6");
-   LLVMValueRef y2_7 = LLVMBuildMul(b, y2_6, z, "y2_7");
-   LLVMValueRef y2_8 = LLVMBuildMul(b, y2_7, x_3, "y2_8");
-   LLVMValueRef y2_9 = LLVMBuildAdd(b, y2_8, x_3, "y2_9");
+   LLVMValueRef y2_3 = LLVMBuildFMul(b, z, sincof_p0, "y2_3");
+   LLVMValueRef y2_4 = LLVMBuildFAdd(b, y2_3, sincof_p1, "y2_4");
+   LLVMValueRef y2_5 = LLVMBuildFMul(b, y2_4, z, "y2_5");
+   LLVMValueRef y2_6 = LLVMBuildFAdd(b, y2_5, sincof_p2, "y2_6");
+   LLVMValueRef y2_7 = LLVMBuildFMul(b, y2_6, z, "y2_7");
+   LLVMValueRef y2_8 = LLVMBuildFMul(b, y2_7, x_3, "y2_8");
+   LLVMValueRef y2_9 = LLVMBuildFAdd(b, y2_8, x_3, "y2_9");
 
    /*
     * select the correct result from the two polynoms
@@ -1829,7 +1868,7 @@ lp_build_exp2_approx(struct lp_build_context *bld,
       ipart = lp_build_floor(bld, x);
 
       /* fpart = x - ipart */
-      fpart = LLVMBuildSub(bld->builder, x, ipart, "");
+      fpart = LLVMBuildFSub(bld->builder, x, ipart, "");
    }
 
    if(p_exp2_int_part || p_exp2) {
@@ -1844,7 +1883,7 @@ lp_build_exp2_approx(struct lp_build_context *bld,
       expfpart = lp_build_polynomial(bld, fpart, lp_build_exp2_polynomial,
                                      Elements(lp_build_exp2_polynomial));
 
-      res = LLVMBuildMul(bld->builder, expipart, expfpart, "");
+      res = LLVMBuildFMul(bld->builder, expipart, expfpart, "");
    }
 
    if(p_exp2_int_part)
@@ -1957,9 +1996,9 @@ lp_build_log2_approx(struct lp_build_context *bld,
                                     Elements(lp_build_log2_polynomial));
 
       /* This effectively increases the polynomial degree by one, but ensures that log2(1) == 0*/
-      logmant = LLVMBuildMul(bld->builder, logmant, LLVMBuildSub(bld->builder, mant, bld->one, ""), "");
+      logmant = LLVMBuildFMul(bld->builder, logmant, LLVMBuildFSub(bld->builder, mant, bld->one, ""), "");
 
-      res = LLVMBuildAdd(bld->builder, logmant, logexp, "");
+      res = LLVMBuildFAdd(bld->builder, logmant, logexp, "");
    }
 
    if(p_exp) {
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_conv.c b/src/gallium/auxiliary/gallivm/lp_bld_conv.c
index 77012f1..8b47731 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_conv.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_conv.c
@@ -117,8 +117,8 @@ lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder,
    scale = (double)mask/ubound;
    bias = (double)((unsigned long long)1 << (mantissa - n));
 
-   res = LLVMBuildMul(builder, src, lp_build_const_vec(src_type, scale), "");
-   res = LLVMBuildAdd(builder, res, lp_build_const_vec(src_type, bias), "");
+   res = LLVMBuildFMul(builder, src, lp_build_const_vec(src_type, scale), "");
+   res = LLVMBuildFAdd(builder, res, lp_build_const_vec(src_type, bias), "");
    res = LLVMBuildBitCast(builder, res, int_vec_type, "");
 
    if(dst_width > n) {
@@ -175,6 +175,8 @@ lp_build_unsigned_norm_to_float(LLVMBuilderRef builder,
    double scale;
    double bias;
 
+   assert(dst_type.floating);
+
    mantissa = lp_mantissa(dst_type);
 
    n = MIN2(mantissa, src_width);
@@ -199,8 +201,8 @@ lp_build_unsigned_norm_to_float(LLVMBuilderRef builder,
 
    res = LLVMBuildBitCast(builder, res, vec_type, "");
 
-   res = LLVMBuildSub(builder, res, bias_, "");
-   res = LLVMBuildMul(builder, res, lp_build_const_vec(dst_type, scale), "");
+   res = LLVMBuildFSub(builder, res, bias_, "");
+   res = LLVMBuildFMul(builder, res, lp_build_const_vec(dst_type, scale), "");
 
    return res;
 }
@@ -296,7 +298,7 @@ lp_build_conv(LLVMBuilderRef builder,
          if (dst_scale != 1.0) {
             LLVMValueRef scale = lp_build_const_vec(tmp_type, dst_scale);
             for(i = 0; i < num_tmps; ++i)
-               tmp[i] = LLVMBuildMul(builder, tmp[i], scale, "");
+               tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, "");
          }
 
          /* Use an equally sized integer for intermediate computations */
@@ -391,7 +393,7 @@ lp_build_conv(LLVMBuilderRef builder,
           if (src_scale != 1.0) {
              LLVMValueRef scale = lp_build_const_vec(tmp_type, 1.0/src_scale);
              for(i = 0; i < num_tmps; ++i)
-                tmp[i] = LLVMBuildMul(builder, tmp[i], scale, "");
+                tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, "");
           }
       }
     }
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c b/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c
index 0f01fc1..247cb83 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c
@@ -240,7 +240,7 @@ lp_build_unpack_arith_rgba_aos(LLVMBuilderRef builder,
     */
 
    if (normalized)
-      scaled = LLVMBuildMul(builder, casted, LLVMConstVector(scales, 4), "");
+      scaled = LLVMBuildFMul(builder, casted, LLVMConstVector(scales, 4), "");
    else
       scaled = casted;
 
@@ -322,7 +322,7 @@ lp_build_pack_rgba_aos(LLVMBuilderRef builder,
    }
 
    if (normalized)
-      scaled = LLVMBuildMul(builder, unswizzled, LLVMConstVector(scales, 4), "");
+      scaled = LLVMBuildFMul(builder, unswizzled, LLVMConstVector(scales, 4), "");
    else
       scaled = unswizzled;
 
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c
index 9f40592..c724a44 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_format_soa.c
@@ -197,7 +197,7 @@ lp_build_unpack_rgba_soa(LLVMBuilderRef builder,
             if (format_desc->channel[chan].normalized) {
                double scale = 1.0 / ((1 << (format_desc->channel[chan].size - 1)) - 1);
                LLVMValueRef scale_val = lp_build_const_vec(type, scale);
-               input = LLVMBuildMul(builder, input, scale_val, "");
+               input = LLVMBuildFMul(builder, input, scale_val, "");
             }
          }
          else {
@@ -227,7 +227,7 @@ lp_build_unpack_rgba_soa(LLVMBuilderRef builder,
             double scale = 1.0 / ((1 << (format_desc->channel[chan].size/2)) - 1);
             LLVMValueRef scale_val = lp_build_const_vec(type, scale);
             input = LLVMBuildSIToFP(builder, input, lp_build_vec_type(type), "");
-            input = LLVMBuildMul(builder, input, scale_val, "");
+            input = LLVMBuildFMul(builder, input, scale_val, "");
          }
          else {
             /* FIXME */
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_quad.c b/src/gallium/auxiliary/gallivm/lp_bld_quad.c
index ca36046..7b10889 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_quad.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_quad.c
@@ -85,7 +85,7 @@ lp_build_scalar_ddx(struct lp_build_context *bld,
    LLVMValueRef idx_right = LLVMConstInt(LLVMInt32Type(), LP_BLD_QUAD_TOP_RIGHT, 0);
    LLVMValueRef a_left  = LLVMBuildExtractElement(bld->builder, a, idx_left, "");
    LLVMValueRef a_right = LLVMBuildExtractElement(bld->builder, a, idx_right, "");
-   return LLVMBuildSub(bld->builder, a_right, a_left, "");
+   return lp_build_sub(bld, a_right, a_left);
 }
 
 
@@ -97,5 +97,5 @@ lp_build_scalar_ddy(struct lp_build_context *bld,
    LLVMValueRef idx_bottom = LLVMConstInt(LLVMInt32Type(), LP_BLD_QUAD_BOTTOM_LEFT, 0);
    LLVMValueRef a_top    = LLVMBuildExtractElement(bld->builder, a, idx_top, "");
    LLVMValueRef a_bottom = LLVMBuildExtractElement(bld->builder, a, idx_bottom, "");
-   return LLVMBuildSub(bld->builder, a_bottom, a_top, "");
+   return lp_build_sub(bld, a_bottom, a_top);
 }
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
index 1a20d74..955d328 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
@@ -888,17 +888,17 @@ lp_build_lod_selector(struct lp_build_sample_context *bld,
          /* Compute rho = max of all partial derivatives scaled by texture size.
           * XXX this could be vectorized somewhat
           */
-         rho = LLVMBuildMul(bld->builder,
+         rho = LLVMBuildFMul(bld->builder,
                             lp_build_max(float_bld, dsdx, dsdy),
                             lp_build_int_to_float(float_bld, width), "");
          if (dims > 1) {
             LLVMValueRef max;
-            max = LLVMBuildMul(bld->builder,
+            max = LLVMBuildFMul(bld->builder,
                                lp_build_max(float_bld, dtdx, dtdy),
                                lp_build_int_to_float(float_bld, height), "");
             rho = lp_build_max(float_bld, rho, max);
             if (dims > 2) {
-               max = LLVMBuildMul(bld->builder,
+               max = LLVMBuildFMul(bld->builder,
                                   lp_build_max(float_bld, drdx, drdy),
                                   lp_build_int_to_float(float_bld, depth), "");
                rho = lp_build_max(float_bld, rho, max);
@@ -912,12 +912,12 @@ lp_build_lod_selector(struct lp_build_sample_context *bld,
          if (lod_bias) {
             lod_bias = LLVMBuildExtractElement(bld->builder, lod_bias,
                                                index0, "");
-            lod = LLVMBuildAdd(bld->builder, lod, lod_bias, "shader_lod_bias");
+            lod = LLVMBuildFAdd(bld->builder, lod, lod_bias, "shader_lod_bias");
          }
       }
 
       /* add sampler lod bias */
-      lod = LLVMBuildAdd(bld->builder, lod, sampler_lod_bias, "sampler_lod_bias");
+      lod = LLVMBuildFAdd(bld->builder, lod, sampler_lod_bias, "sampler_lod_bias");
 
       /* clamp lod */
       lod = lp_build_clamp(float_bld, lod, min_lod, max_lod);
@@ -2029,6 +2029,8 @@ lp_build_sample_soa(LLVMBuilderRef builder,
       debug_printf("Sample from %s\n", util_format_name(fmt));
    }
 
+   assert(type.floating);
+
    /* Setup our build context */
    memset(&bld, 0, sizeof bld);
    bld.builder = builder;
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c
index 42d796c..becbd3b 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_tgsi_soa.c
@@ -506,6 +506,7 @@ emit_fetch(
    const unsigned chan_index )
 {
    const struct tgsi_full_src_register *reg = &inst->Src[src_op];
+   const struct lp_type type = bld->base.type;
    const unsigned swizzle =
       tgsi_util_get_full_src_register_swizzle(reg, chan_index);
    LLVMValueRef res;
@@ -612,11 +613,12 @@ emit_fetch(
    case TGSI_UTIL_SIGN_SET:
       /* TODO: Use bitwese OR for floating point */
       res = lp_build_abs( &bld->base, res );
-      res = LLVMBuildNeg( bld->base.builder, res, "" );
-      break;
-
+      /* fall through */
    case TGSI_UTIL_SIGN_TOGGLE:
-      res = LLVMBuildNeg( bld->base.builder, res, "" );
+      if (type.floating)
+         res = LLVMBuildFNeg( bld->base.builder, res, "" );
+      else
+         res = LLVMBuildNeg( bld->base.builder, res, "" );
       break;
 
    case TGSI_UTIL_SIGN_KEEP: