add dumbest possible 32-bit SkJumper backend

Everything uses a ton of stack, nothing tail calls, and for now this is
non-Windows only.  But, it does run faster than the portable serial code.

On my trashcan, running `monobench SkRasterPipeline_compile`:
   - Normal 64-bit AVX build:  43.6ns
   - Before this CL, 32-bit:  707.9ns
   - This CL:                 147.5ns

Change-Id: I4a8929570ace47193ed8925c58b70bb22d6b1447
Reviewed-on: https://skia-review.googlesource.com/20964
Reviewed-by: Mike Reed <reed@google.com>
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>

1 files changed, 19 insertions, 0 deletions

diff --git a/src/jumper/SkJumper.cpp b/src/jumper/SkJumper.cpp
index 8c9eb73325..aa4b794347 100644
--- a/src/jumper/SkJumper.cpp
+++ b/src/jumper/SkJumper.cpp
@@ -155,6 +155,14 @@ extern "C" {
     #define M(st) StageFn ASM(st,ssse3_lowp);
         LOWP_STAGES(M)
     #undef M
+
+#elif defined(__i386__)
+    StartPipelineFn ASM(start_pipeline,sse2);
+    StageFn ASM(just_return,sse2);
+    #define M(st) StageFn ASM(st,sse2);
+        SK_RASTER_PIPELINE_STAGES(M)
+    #undef M
+
 #endif
 
     // Portable, single-pixel stages.
@@ -256,6 +264,17 @@ static SkJumper_Engine choose_engine() {
         #undef M
         };
     }
+
+#elif defined(__i386__)
+    if (1 && SkCpu::Supports(SkCpu::SSE2)) {
+        return {
+        #define M(stage) ASM(stage, sse2),
+            { SK_RASTER_PIPELINE_STAGES(M) },
+            M(start_pipeline) M(just_return)
+        #undef M
+        };
+    }
+
 #endif
     return kPortable;
 }