diff --git a/src/intel/compiler/brw_compile_bs.cpp b/src/intel/compiler/brw_compile_bs.cpp
index 9eeba745160..ab71d3dbfa7 100644
--- a/src/intel/compiler/brw_compile_bs.cpp
+++ b/src/intel/compiler/brw_compile_bs.cpp
@@ -147,6 +147,10 @@ compile_single_bs(const struct brw_compiler *compiler,
    else
       assert(offset == 0);
 
+   if (!prog_offset)
+      prog_data->base.grf_used = MAX2(prog_data->base.grf_used,
+                                      selected->grf_used);
+
    return dispatch_width;
 }
 
diff --git a/src/intel/compiler/brw_compile_cs.cpp b/src/intel/compiler/brw_compile_cs.cpp
index dd86044f5b8..59e26e64da6 100644
--- a/src/intel/compiler/brw_compile_cs.cpp
+++ b/src/intel/compiler/brw_compile_cs.cpp
@@ -263,6 +263,10 @@ brw_compile_cs(const struct brw_compiler *compiler,
          if (stats)
             stats->max_dispatch_width = max_dispatch_width;
          stats = stats ? stats + 1 : NULL;
+
+         prog_data->base.grf_used = MAX2(prog_data->base.grf_used,
+                                         v[simd]->grf_used);
+
          max_dispatch_width = 8u << simd;
       }
    }
diff --git a/src/intel/compiler/brw_compile_fs.cpp b/src/intel/compiler/brw_compile_fs.cpp
index 22fbc31d233..31d9e312139 100644
--- a/src/intel/compiler/brw_compile_fs.cpp
+++ b/src/intel/compiler/brw_compile_fs.cpp
@@ -1606,6 +1606,8 @@ brw_compile_fs(const struct brw_compiler *compiler,
 
          assert(v8->payload().num_regs % reg_unit(devinfo) == 0);
          prog_data->base.dispatch_grf_start_reg = v8->payload().num_regs / reg_unit(devinfo);
+         prog_data->base.grf_used = MAX2(prog_data->base.grf_used,
+                                         v8->grf_used);
 
          const performance &perf = v8->performance_analysis.require();
          throughput = MAX2(throughput, perf.throughput);
@@ -1714,8 +1716,8 @@ brw_compile_fs(const struct brw_compiler *compiler,
             simd32_cfg = v32->cfg;
             assert(v32->payload().num_regs % reg_unit(devinfo) == 0);
             prog_data->dispatch_grf_start_reg_32 = v32->payload().num_regs / reg_unit(devinfo);
-            prog_data->base.grf_used = std::max(prog_data->base.grf_used,
-                                                v32->grf_used);
+            prog_data->base.grf_used = MAX2(prog_data->base.grf_used,
+                                            v32->grf_used);
          }
       }
 
@@ -1735,8 +1737,8 @@ brw_compile_fs(const struct brw_compiler *compiler,
 
             assert(v16->payload().num_regs % reg_unit(devinfo) == 0);
             prog_data->dispatch_grf_start_reg_16 = v16->payload().num_regs / reg_unit(devinfo);
-            prog_data->base.grf_used = std::max(prog_data->base.grf_used,
-                                                v16->grf_used);
+            prog_data->base.grf_used = MAX2(prog_data->base.grf_used,
+                                            v16->grf_used);
          }
       }
 
@@ -1760,6 +1762,8 @@ brw_compile_fs(const struct brw_compiler *compiler,
 
             assert(v16->payload().num_regs % reg_unit(devinfo) == 0);
             prog_data->dispatch_grf_start_reg_16 = v16->payload().num_regs / reg_unit(devinfo);
+            prog_data->base.grf_used = MAX2(prog_data->base.grf_used,
+                                            v16->grf_used);
 
             const performance &perf = v16->performance_analysis.require();
             throughput = MAX2(throughput, perf.throughput);
@@ -1801,6 +1805,8 @@ brw_compile_fs(const struct brw_compiler *compiler,
 
                assert(v32->payload().num_regs % reg_unit(devinfo) == 0);
                prog_data->dispatch_grf_start_reg_32 = v32->payload().num_regs / reg_unit(devinfo);
+               prog_data->base.grf_used = MAX2(prog_data->base.grf_used,
+                                               v32->grf_used);
 
                throughput = MAX2(throughput, perf.throughput);
             }
@@ -1877,6 +1883,8 @@ brw_compile_fs(const struct brw_compiler *compiler,
    if (multi_cfg) {
       assert(vmulti->payload().num_regs % reg_unit(devinfo) == 0);
       prog_data->base.dispatch_grf_start_reg = vmulti->payload().num_regs / reg_unit(devinfo);
+      prog_data->base.grf_used = MAX2(prog_data->base.grf_used,
+                                      vmulti->grf_used);
    }
 
    /* When the caller compiles a repclear or fast clear shader, they
diff --git a/src/intel/compiler/brw_compile_gs.cpp b/src/intel/compiler/brw_compile_gs.cpp
index f205bb4fcf2..14a577bb4ce 100644
--- a/src/intel/compiler/brw_compile_gs.cpp
+++ b/src/intel/compiler/brw_compile_gs.cpp
@@ -358,6 +358,7 @@ brw_compile_gs(const struct brw_compiler *compiler,
       assert(v.payload().num_regs % reg_unit(compiler->devinfo) == 0);
       prog_data->base.base.dispatch_grf_start_reg =
          v.payload().num_regs / reg_unit(compiler->devinfo);
+      prog_data->base.base.grf_used = v.grf_used;
 
       brw_generator g(compiler, &params->base,
                      &prog_data->base.base, MESA_SHADER_GEOMETRY);
diff --git a/src/intel/compiler/brw_compile_mesh.cpp b/src/intel/compiler/brw_compile_mesh.cpp
index 8316e090357..59e38b37259 100644
--- a/src/intel/compiler/brw_compile_mesh.cpp
+++ b/src/intel/compiler/brw_compile_mesh.cpp
@@ -495,6 +495,8 @@ brw_compile_task(const struct brw_compiler *compiler,
 
    fs_visitor *selected = v[selected_simd].get();
    prog_data->base.prog_mask = 1 << selected_simd;
+   prog_data->base.base.grf_used = MAX2(prog_data->base.base.grf_used,
+                                        selected->grf_used);
 
    if (unlikely(debug_enabled)) {
       fprintf(stderr, "Task Output ");
@@ -1816,6 +1818,8 @@ brw_compile_mesh(const struct brw_compiler *compiler,
 
    fs_visitor *selected = v[selected_simd].get();
    prog_data->base.prog_mask = 1 << selected_simd;
+   prog_data->base.base.grf_used = MAX2(prog_data->base.base.grf_used,
+                                        selected->grf_used);
 
    if (unlikely(debug_enabled)) {
       if (params->tue_map) {
diff --git a/src/intel/compiler/brw_compile_tcs.cpp b/src/intel/compiler/brw_compile_tcs.cpp
index 23d70651fcf..b0380b4523f 100644
--- a/src/intel/compiler/brw_compile_tcs.cpp
+++ b/src/intel/compiler/brw_compile_tcs.cpp
@@ -287,6 +287,7 @@ brw_compile_tcs(const struct brw_compiler *compiler,
 
    assert(v.payload().num_regs % reg_unit(devinfo) == 0);
    prog_data->base.base.dispatch_grf_start_reg = v.payload().num_regs / reg_unit(devinfo);
+   prog_data->base.base.grf_used = v.grf_used;
 
    brw_generator g(compiler, &params->base,
                   &prog_data->base.base, MESA_SHADER_TESS_CTRL);
diff --git a/src/intel/compiler/brw_compile_tes.cpp b/src/intel/compiler/brw_compile_tes.cpp
index c0cc7ab58b9..96af0d405ee 100644
--- a/src/intel/compiler/brw_compile_tes.cpp
+++ b/src/intel/compiler/brw_compile_tes.cpp
@@ -164,7 +164,7 @@ brw_compile_tes(const struct brw_compiler *compiler,
 
    assert(v.payload().num_regs % reg_unit(devinfo) == 0);
    prog_data->base.base.dispatch_grf_start_reg = v.payload().num_regs / reg_unit(devinfo);
-
+   prog_data->base.base.grf_used = v.grf_used;
    prog_data->base.dispatch_mode = INTEL_DISPATCH_MODE_SIMD8;
 
    brw_generator g(compiler, &params->base,
diff --git a/src/intel/compiler/brw_compile_vs.cpp b/src/intel/compiler/brw_compile_vs.cpp
index 3e6e5befefb..83c990668d9 100644
--- a/src/intel/compiler/brw_compile_vs.cpp
+++ b/src/intel/compiler/brw_compile_vs.cpp
@@ -161,6 +161,7 @@ brw_compile_vs(const struct brw_compiler *compiler,
    assert(v.payload().num_regs % reg_unit(compiler->devinfo) == 0);
    prog_data->base.base.dispatch_grf_start_reg =
       v.payload().num_regs / reg_unit(compiler->devinfo);
+   prog_data->base.base.grf_used = v.grf_used;
 
    brw_generator g(compiler, &params->base,
                   &prog_data->base.base,
diff --git a/src/intel/compiler/brw_compiler.h b/src/intel/compiler/brw_compiler.h
index 6880ca0406b..ce6c5f5fd5b 100644
--- a/src/intel/compiler/brw_compiler.h
+++ b/src/intel/compiler/brw_compiler.h
@@ -548,6 +548,9 @@ struct brw_stage_prog_data {
     */
    unsigned dispatch_grf_start_reg;
 
+   /** Number of GRF registers used. */
+   unsigned grf_used;
+
    bool use_alt_mode; /**< Use ALT floating point mode?  Otherwise, IEEE. */
 
    /* 32-bit identifiers for all push/pull parameters.  These can be anything
diff --git a/src/intel/compiler/intel_clc.c b/src/intel/compiler/intel_clc.c
index dd2ee81974a..22754240f9a 100644
--- a/src/intel/compiler/intel_clc.c
+++ b/src/intel/compiler/intel_clc.c
@@ -169,6 +169,7 @@ print_cs_prog_data_fields(FILE *fp, const char *prefix, const char *pad,
    PROG_DATA_FIELD("%u", base.const_data_offset);
    PROG_DATA_FIELD("%u", base.num_relocs);
    fprintf(fp, "%s.base.relocs = %s_relocs,\n", pad, prefix);
+   PROG_DATA_FIELD("%u", base.grf_used);
    PROG_DATA_FIELD("%u", base.printf_info_count);
    fprintf(fp, "%s.base.printf_info = (u_printf_info *)%s_printfs,\n", pad, prefix);
    assert(!cs_prog_data->base.has_ubo_pull);