swr/rast: Convert gather masks to Nx1bit

Simplifies calling code, gets gather function interface closer to llvm's
masked_gather.

Reviewed-by: Bruce Cherniak <[email protected]>

2 files changed, 14 insertions, 40 deletions

diff --git a/src/gallium/drivers/swr/rasterizer/jitter/builder_misc.cpp b/src/gallium/drivers/swr/rasterizer/jitter/builder_misc.cpp
index 0221106664b..04092541e5d 100644
--- a/src/gallium/drivers/swr/rasterizer/jitter/builder_misc.cpp
+++ b/src/gallium/drivers/swr/rasterizer/jitter/builder_misc.cpp
@@ -602,7 +602,7 @@ namespace SwrJit
         if(JM()->mArch.AVX2())
         {
             // force mask to <N x float>, required by vgather
-            Value *mask = BITCAST(vMask, mSimdFP32Ty);
+            Value *mask = BITCAST(VMASK(vMask), mSimdFP32Ty);
 
             vGather = VGATHERPS(vSrc, pBase, vIndices, mask, C(scale));
         }
@@ -617,7 +617,6 @@ namespace SwrJit
             vGather = VUNDEF_F();
             Value *vScaleVec = VIMMED1((uint32_t)scale);
             Value *vOffsets = MUL(vIndices,vScaleVec);
-            Value *mask = MASK(vMask);
             for(uint32_t i = 0; i < mVWidth; ++i)
             {
                 // single component byte index
@@ -627,7 +626,7 @@ namespace SwrJit
                 loadAddress = BITCAST(loadAddress,PointerType::get(mFP32Ty,0));
                 // pointer to the value to load if we're masking off a component
                 Value *maskLoadAddress = GEP(vSrcPtr,{C(0), C(i)});
-                Value *selMask = VEXTRACT(mask,C(i));
+                Value *selMask = VEXTRACT(vMask,C(i));
                 // switch in a safe address to load if we're trying to access a vertex 
                 Value *validAddress = SELECT(selMask, loadAddress, maskLoadAddress);
                 Value *val = LOAD(validAddress);
@@ -648,7 +647,7 @@ namespace SwrJit
         if (JM()->mArch.AVX512F())
         {
             // force mask to <N-bit Integer>, required by vgather2
-            Value *mask = BITCAST(MASK2(vMask), mInt16Ty);
+            Value *mask = BITCAST(vMask, mInt16Ty);
 
             vGather = VGATHERPS2(vSrc, pBase, vIndices, mask, C((uint32_t)scale));
         }
@@ -689,7 +688,7 @@ namespace SwrJit
         // use avx2 gather instruction if available
         if(JM()->mArch.AVX2())
         {
-            vGather = VGATHERDD(vSrc, pBase, vIndices, vMask, C(scale));
+            vGather = VGATHERDD(vSrc, pBase, vIndices, VMASK(vMask), C(scale));
         }
         else
         {
@@ -702,7 +701,6 @@ namespace SwrJit
             vGather = VUNDEF_I();
             Value *vScaleVec = VIMMED1((uint32_t)scale);
             Value *vOffsets = MUL(vIndices, vScaleVec);
-            Value *mask = MASK(vMask);
             for(uint32_t i = 0; i < mVWidth; ++i)
             {
                 // single component byte index
@@ -712,7 +710,7 @@ namespace SwrJit
                 loadAddress = BITCAST(loadAddress, PointerType::get(mInt32Ty, 0));
                 // pointer to the value to load if we're masking off a component
                 Value *maskLoadAddress = GEP(vSrcPtr, {C(0), C(i)});
-                Value *selMask = VEXTRACT(mask, C(i));
+                Value *selMask = VEXTRACT(vMask, C(i));
                 // switch in a safe address to load if we're trying to access a vertex 
                 Value *validAddress = SELECT(selMask, loadAddress, maskLoadAddress);
                 Value *val = LOAD(validAddress, C(0));
@@ -739,6 +737,7 @@ namespace SwrJit
         // use avx2 gather instruction if available
         if(JM()->mArch.AVX2())
         {
+            vMask = BITCAST(S_EXT(vMask, VectorType::get(mInt64Ty, mVWidth/2)), VectorType::get(mDoubleTy, mVWidth/2));
             vGather = VGATHERPD(vSrc, pBase, vIndices, vMask, C(scale));
         }
         else
@@ -752,7 +751,6 @@ namespace SwrJit
             vGather = UndefValue::get(VectorType::get(mDoubleTy, 4));
             Value *vScaleVec = VECTOR_SPLAT(4, C((uint32_t)scale));
             Value *vOffsets = MUL(vIndices,vScaleVec);
-            Value *mask = MASK(vMask);
             for(uint32_t i = 0; i < mVWidth/2; ++i)
             {
                 // single component byte index
@@ -762,7 +760,7 @@ namespace SwrJit
                 loadAddress = BITCAST(loadAddress,PointerType::get(mDoubleTy,0));
                 // pointer to the value to load if we're masking off a component
                 Value *maskLoadAddress = GEP(vSrcPtr,{C(0), C(i)});
-                Value *selMask = VEXTRACT(mask,C(i));
+                Value *selMask = VEXTRACT(vMask,C(i));
                 // switch in a safe address to load if we're trying to access a vertex
                 Value *validAddress = SELECT(selMask, loadAddress, maskLoadAddress);
                 Value *val = LOAD(validAddress);
@@ -1094,14 +1092,10 @@ namespace SwrJit
         const SWR_FORMAT_INFO &info = GetFormatInfo(format);
         if(info.type[0] == SWR_TYPE_FLOAT && info.bpc[0] == 32)
         {
-            // ensure our mask is the correct type
-            mask = BITCAST(mask, mSimdFP32Ty);
             GATHER4PS(info, pSrcBase, byteOffsets, mask, vGatherComponents, bPackedOutput);
         }
         else
         {
-            // ensure our mask is the correct type
-            mask = BITCAST(mask, mSimdInt32Ty);
             GATHER4DD(info, pSrcBase, byteOffsets, mask, vGatherComponents, bPackedOutput);
         }
     }
diff --git a/src/gallium/drivers/swr/rasterizer/jitter/fetch_jit.cpp b/src/gallium/drivers/swr/rasterizer/jitter/fetch_jit.cpp
index 6c0e658e68f..67a4a040726 100644
--- a/src/gallium/drivers/swr/rasterizer/jitter/fetch_jit.cpp
+++ b/src/gallium/drivers/swr/rasterizer/jitter/fetch_jit.cpp
@@ -1004,10 +1004,6 @@ void FetchJit::JitGatherVertices(const FETCH_COMPILE_STATE &fetchState,
         // blend in any partially OOB indices that have valid elements
         vGatherMask = SELECT(vPartialOOBMask, vElementInBoundsMask, vGatherMask);
         vGatherMask2 = SELECT(vPartialOOBMask2, vElementInBoundsMask, vGatherMask2);
-        Value *pMask = vGatherMask;
-        Value *pMask2 = vGatherMask2;
-        vGatherMask = VMASK(vGatherMask);
-        vGatherMask2 = VMASK(vGatherMask2);
 
         // calculate the actual offsets into the VB
         Value* vOffsets = MUL(vCurIndices, vStride);
@@ -1051,8 +1047,6 @@ void FetchJit::JitGatherVertices(const FETCH_COMPILE_STATE &fetchState,
 
         // blend in any partially OOB indices that have valid elements
         vGatherMask = SELECT(vPartialOOBMask, vElementInBoundsMask, vGatherMask);
-        Value* pMask = vGatherMask;
-        vGatherMask = VMASK(vGatherMask);
 
         // calculate the actual offsets into the VB
         Value* vOffsets = MUL(vCurIndices, vStride);
@@ -1289,9 +1283,7 @@ void FetchJit::JitGatherVertices(const FETCH_COMPILE_STATE &fetchState,
                                 indices = INSERT2_I(indices, vShiftedOffsets,  0);
                                 indices = INSERT2_I(indices, vShiftedOffsets2, 1);
 
-                                Value *mask = VUNDEF2_I();
-                                mask = INSERT2_I(mask, vGatherMask,  0);
-                                mask = INSERT2_I(mask, vGatherMask2, 1);
+                                Value *mask = VSHUFFLE(vGatherMask, vGatherMask2, {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15});
 
                                 pVtxSrc2[currentVertexElement] = GATHERPS2(gatherSrc16, pStreamBase, indices, mask, 2);
 #else
@@ -1396,18 +1388,10 @@ void FetchJit::JitGatherVertices(const FETCH_COMPILE_STATE &fetchState,
                             // if we need to gather the component
                             if (compCtrl[i] == StoreSrc)
                             {
-                                Value *vMaskLo = VSHUFFLE(pMask, VUNDEF(mInt1Ty, 8), C({ 0, 1, 2, 3 }));
-                                Value *vMaskLo2 = VSHUFFLE(pMask2, VUNDEF(mInt1Ty, 8), C({ 0, 1, 2, 3 }));
-                                Value *vMaskHi = VSHUFFLE(pMask, VUNDEF(mInt1Ty, 8), C({ 4, 5, 6, 7 }));
-                                Value *vMaskHi2 = VSHUFFLE(pMask2, VUNDEF(mInt1Ty, 8), C({ 4, 5, 6, 7 }));
-                                vMaskLo = S_EXT(vMaskLo, VectorType::get(mInt64Ty, 4));
-                                vMaskLo2 = S_EXT(vMaskLo2, VectorType::get(mInt64Ty, 4));
-                                vMaskHi = S_EXT(vMaskHi, VectorType::get(mInt64Ty, 4));
-                                vMaskHi2 = S_EXT(vMaskHi2, VectorType::get(mInt64Ty, 4));
-                                vMaskLo = BITCAST(vMaskLo, VectorType::get(mDoubleTy, 4));
-                                vMaskLo2 = BITCAST(vMaskLo2, VectorType::get(mDoubleTy, 4));
-                                vMaskHi = BITCAST(vMaskHi, VectorType::get(mDoubleTy, 4));
-                                vMaskHi2 = BITCAST(vMaskHi2, VectorType::get(mDoubleTy, 4));
+                                Value *vMaskLo = VSHUFFLE(vGatherMask, VUNDEF(mInt1Ty, 8), C({ 0, 1, 2, 3 }));
+                                Value *vMaskLo2 = VSHUFFLE(vGatherMask2, VUNDEF(mInt1Ty, 8), C({ 0, 1, 2, 3 }));
+                                Value *vMaskHi = VSHUFFLE(vGatherMask, VUNDEF(mInt1Ty, 8), C({ 4, 5, 6, 7 }));
+                                Value *vMaskHi2 = VSHUFFLE(vGatherMask2, VUNDEF(mInt1Ty, 8), C({ 4, 5, 6, 7 }));
 
                                 Value *vOffsetsLo = VEXTRACTI128(vOffsets, C(0));
                                 Value *vOffsetsLo2 = VEXTRACTI128(vOffsets2, C(0));
@@ -1483,12 +1467,8 @@ void FetchJit::JitGatherVertices(const FETCH_COMPILE_STATE &fetchState,
                             // if we need to gather the component
                             if (compCtrl[i] == StoreSrc)
                             {
-                                Value *vMaskLo = VSHUFFLE(pMask, VUNDEF(mInt1Ty, 8), C({0, 1, 2, 3}));
-                                Value *vMaskHi = VSHUFFLE(pMask, VUNDEF(mInt1Ty, 8), C({4, 5, 6, 7}));
-                                vMaskLo = S_EXT(vMaskLo, VectorType::get(mInt64Ty, 4));
-                                vMaskHi = S_EXT(vMaskHi, VectorType::get(mInt64Ty, 4));
-                                vMaskLo = BITCAST(vMaskLo, VectorType::get(mDoubleTy, 4));
-                                vMaskHi = BITCAST(vMaskHi, VectorType::get(mDoubleTy, 4));
+                                Value *vMaskLo = VSHUFFLE(vGatherMask, VUNDEF(mInt1Ty, 8), C({0, 1, 2, 3}));
+                                Value *vMaskHi = VSHUFFLE(vGatherMask, VUNDEF(mInt1Ty, 8), C({4, 5, 6, 7}));
 
                                 Value *vOffsetsLo = VEXTRACTI128(vOffsets, C(0));
                                 Value *vOffsetsHi = VEXTRACTI128(vOffsets, C(1));