Mesa (staging/19.3): llvmpipe: use ppc64le/ppc64 Large code model for JIT-compiled shaders

[GitLab Mirror]

Module: Mesa
Branch: staging/19.3
Commit: 32aba91c07fd2ab05ca4ce2a05c2806d29e18435
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=32aba91c07fd2ab05ca4ce2a05c2806d29e18435

Author: Ben Crocker <bcrocker at redhat.com>
Date:   Wed Nov 13 20:27:24 2019 +0000

llvmpipe: use ppc64le/ppc64 Large code model for JIT-compiled shaders

Large programs, e.g. gnome-shell and firefox, may tax the
addressability of the Medium code model once a (potentially unbounded)
number of dynamically generated JIT-compiled shader programs are
linked in and relocated.  Yet the default code model as of LLVM 8 is
Medium or even Small.

The cost of changing from Medium to Large is negligible:
- an additional 8-byte pointer stored immediately before the shader entrypoint;
- change an add-immediate (addis) instruction to a load (ld).

Testing with WebGL Conformance
(https://www.khronos.org/registry/webgl/sdk/tests/webgl-conformance-tests.html)
yields clean runs with this change (and crashes without it).

Testing with glxgears shows no detectable performance difference.

Bugzilla: https://bugzilla.redhat.com/show_bug.cgi?id=1753327, 1753789, 1543572, 1747110, and 1582226

Closes: https://gitlab.freedesktop.org/mesa/mesa/issues/223

Co-authored by: Nemanja Ivanovic <nemanjai at ca.ibm.com>, Tom Stellard <tstellar at redhat.com>

CC: mesa-stable at lists.freedesktop.org

Signed-off-by: Ben Crocker <bcrocker at redhat.com>
(cherry picked from commit 9c3be6d21fa6a45852045d0286b80fb420f82fe3)
Conflicts resolved Dylan (PIPE_ARCH -> UTIL_ARCH rename)

---

 src/gallium/auxiliary/gallivm/lp_bld_misc.cpp | 16 +++++++++++++++-
 1 file changed, 15 insertions(+), 1 deletion(-)

diff --git a/src/gallium/auxiliary/gallivm/lp_bld_misc.cpp b/src/gallium/auxiliary/gallivm/lp_bld_misc.cpp
index 67ae1fa1e93..ee27f346254 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_misc.cpp
+++ b/src/gallium/auxiliary/gallivm/lp_bld_misc.cpp
@@ -456,7 +456,20 @@ lp_build_create_jit_compiler_for_module(LLVMExecutionEngineRef *OutJIT,
     * when not using MCJIT so no instructions are generated which the old JIT
     * can't handle. Not entirely sure if we really need to do anything yet.
     */
-#if defined(PIPE_ARCH_LITTLE_ENDIAN)  && defined(PIPE_ARCH_PPC_64)
+
+#ifdef PIPE_ARCH_PPC_64
+   /*
+    * Large programs, e.g. gnome-shell and firefox, may tax the addressability
+    * of the Medium code model once dynamically generated JIT-compiled shader
+    * programs are linked in and relocated.  Yet the default code model as of
+    * LLVM 8 is Medium or even Small.
+    * The cost of changing from Medium to Large is negligible:
+    * - an additional 8-byte pointer stored immediately before the shader entrypoint;
+    * - change an add-immediate (addis) instruction to a load (ld).
+    */
+   builder.setCodeModel(CodeModel::Large);
+
+#if PIPE_ARCH_LITTLE_ENDIAN
    /*
     * Versions of LLVM prior to 4.0 lacked a table entry for "POWER8NVL",
     * resulting in (big-endian) "generic" being returned on
@@ -469,6 +482,7 @@ lp_build_create_jit_compiler_for_module(LLVMExecutionEngineRef *OutJIT,
    if (MCPU == "generic")
       MCPU = "pwr8";
 #endif
+#endif
    builder.setMCPU(MCPU);
    if (gallivm_debug & (GALLIVM_DEBUG_IR | GALLIVM_DEBUG_ASM | GALLIVM_DEBUG_DUMP_BC)) {
       debug_printf("llc -mcpu option: %s\n", MCPU.str().c_str());