sf.net #572: gas+x86+SSE faulty optimisations

- with '-vec 2' for intra-expression vectorisation of horizontal add optimisation was generating bad AST resulting in bad assembly

changelog.txt

@@ -112,6 +112,7 @@ Version 1.08.0
 - gfxlib2: character data was incorrectly stored for values >= 128 causing incorrect values returned from SCREEN() function
 - LLVM backend: escape procedure names on win32 targets to avoid having llvm mangle the names also
 - sf.net #572: don't call vectorize operations for anything but gas+x86+sse backend / options
+- sf.net #572: gas+x86+SSE faulty optimisations with '-vec 2' for intra-expression vectorisation of horizontal add optimisation was generating bad AST resulting in bad assembly
 
 
 Version 1.07.0

src/compiler/ast-vectorize.bas

@@ -574,30 +574,56 @@ private function astIntraTreeVectorize _
 	if( n->class = AST_NODECLASS_BOP ) then
 		if( n->op.op = AST_OP_ADD ) then
 
+			'' test if nodes can be merged / vectorized
+			'' careful, maxVectorWidth & vectorWidth are shared
 			maxVectorWidth = 4
 			vectorWidth = 0
+
 			if( hMergeNode( n->l, n->r, FALSE ) ) then
+
+				'' go ahead and do the merge / vectorize
+				'' 
 				maxVectorWidth = 4
 				vectorWidth = 0
 				hMergeNode( n->l, n->r, TRUE )
 
-				'' check for multiple HADDs
+				'' n    = AST_OP_ADD, can be removed
+				'' n->l = vectorized node or an existing AST_OP_HADD
+				'' n->r = can be discarded
+
+				assert( n->l )
+
 				l = n->l
+
+				'' check for multiple AST_OP_HADD's
 				if( l->class = AST_NODECLASS_UOP ) then
 					if( l->op.op = AST_OP_HADD ) then
+
+						'' replace the callers node
 						*n = *l
+
+						assert( n->l )
+						assert( n->l->vector <> 0 )
+
+						'' copy the new value of the vector to AST_OP_HADD node
+						n->vector = n->l->vector
+
 						'' remove this node
 						astDelNode( l )
-						n->vector = 0
 						return TRUE
 					end if
 				end if
 
+				'' n    = AST_OP_ADD, can be replaced
+				'' n->l = vectorized node
+				'' n->r = can be discarded
+				assert( n->l )
+
 				astDelTree( n->r )
 				n->r = NULL
 				n->class = AST_NODECLASS_UOP
 				n->op.op = AST_OP_HADD
-				n->vector = 0
+				n->vector = n->l->vector
 
 				return TRUE
 			end if

src/compiler/emit_SSE.bas

@@ -1094,6 +1094,7 @@ private sub _emitADDF_SSE _
 	else
 		'' This should never happen due to IR_OPT_FPUCONV
 		outp " implement 'add integer to float'"
+		assert( 0 )
 	end if
 end sub
 
@@ -1174,6 +1175,7 @@ private sub _emitSUBF_SSE _
 	else
 		'' This should never happen due to IR_OPT_FPUCONV
 		outp " implement 'subtract integer from float'"
+		assert( 0 )
 	end if
 end sub
 
@@ -1253,6 +1255,7 @@ private sub _emitMULF_SSE _
 	else
 		'' This should never happen due to IR_OPT_FPUCONV
 		outp " implement 'multiply float by integer'"
+		assert( 0 )
 	end if
 
 end sub
@@ -1333,6 +1336,7 @@ private sub _emitDIVF_SSE _
 	else
 		'' This should never happen due to IR_OPT_FPUCONV
 		outp " implement 'divide float by integer'"
+		assert( 0 )
 	end if
 
 end sub
@@ -1764,6 +1768,8 @@ private sub _emitHADDF_SSE _
 			outp "addps " + dst + COMMA + "xmm7"
 			outp "pshufd xmm7" + COMMA + dst + COMMA + "0x01"
 			outp "addss " + dst + COMMA + "xmm7"
+		else
+			assert( 0 )
 		end if
 	end if
 

tests/optimizations/vector.bas

@@ -157,6 +157,262 @@ SUITE( fbc_tests.optimizations.vector )
 
 	END_TEST
 
+	TEST( fvec3_intra )
+
+		dim as FVec3 a3, b3, c3
+
+		#macro check( result, expr )
+			c3.x = expr
+
+			CU_ASSERT_EQUAL( c3.x,  result )
+			CU_ASSERT_EQUAL( c3.y,  0f  )
+			CU_ASSERT_EQUAL( c3.z,  0f  )
+		#endmacro
+
+		a3 = type(1f,  2f,  4f)
+		b3 = type(8f, 16f, 32f)
+
+		c3.y = 0.0f
+		c3.z = 0.0f
+
+		check(  1f, a3.x )
+		check(  2f, a3.y )
+		check(  3f, a3.x + a3.y )
+		check(  3f, a3.y + a3.x )
+		check(  4f, a3.z )
+		check(  5f, a3.x + a3.z )
+		check(  5f, a3.z + a3.x )
+		check(  6f, a3.y + a3.z )
+		check(  6f, a3.z + a3.y )
+		check(  7f, a3.x + a3.y + a3.z )
+		check(  7f, a3.y + a3.x + a3.z )
+		check(  7f, a3.x + a3.z + a3.y )
+		check(  7f, a3.z + a3.x + a3.y )
+		check(  7f, a3.y + a3.z + a3.x )
+		check(  7f, a3.z + a3.y + a3.x )
+
+		'' mix in another source
+
+		check( 17f, b3.y + a3.x )
+		check( 18f, a3.y + b3.y )
+		check( 19f, b3.y + a3.x + a3.y )
+		check( 19f, a3.x + b3.y + a3.y )
+		check( 19f, a3.x + a3.y + b3.y )
+		check( 19f, b3.y + a3.y + a3.x )
+		check( 19f, a3.y + b3.y + a3.x )
+		check( 19f, a3.y + a3.x + b3.y )
+		check( 20f, b3.y + a3.z )
+		check( 20f, a3.z + b3.y )
+		check( 21f, b3.y + a3.x + a3.z )
+		check( 21f, a3.x + b3.y + a3.z )
+		check( 21f, a3.x + a3.z + b3.y )
+		check( 21f, b3.y + a3.z + a3.x )
+		check( 21f, a3.z + b3.y + a3.x )
+		check( 21f, a3.z + a3.x + b3.y )
+		check( 22f, b3.y + a3.y + a3.z )
+		check( 22f, a3.y + b3.y + a3.z )
+		check( 22f, a3.y + a3.z + b3.y )
+		check( 22f, b3.y + a3.z + a3.y )
+		check( 22f, a3.z + b3.y + a3.y )
+		check( 22f, a3.z + a3.y + b3.y )
+		check( 23f, b3.y + a3.x + a3.y + a3.z )
+		check( 23f, a3.x + b3.y + a3.y + a3.z )
+		check( 23f, a3.x + a3.y + b3.y + a3.z )
+		check( 23f, a3.x + a3.y + a3.z + b3.y )
+		check( 23f, b3.y + a3.y + a3.x + a3.z )
+		check( 23f, a3.y + b3.y + a3.x + a3.z )
+		check( 23f, a3.y + a3.x + b3.y + a3.z )
+		check( 23f, a3.y + a3.x + a3.z + b3.y )
+		check( 23f, b3.y + a3.x + a3.z + a3.y )
+		check( 23f, a3.x + b3.y + a3.z + a3.y )
+		check( 23f, a3.x + a3.z + b3.y + a3.y )
+		check( 23f, a3.x + a3.z + a3.y + b3.y )
+		check( 23f, b3.y + a3.z + a3.x + a3.y )
+		check( 23f, a3.z + b3.y + a3.x + a3.y )
+		check( 23f, a3.z + a3.x + b3.y + a3.y )
+		check( 23f, a3.z + a3.x + a3.y + b3.y )
+		check( 23f, b3.y + a3.y + a3.z + a3.x )
+		check( 23f, a3.y + b3.y + a3.z + a3.x )
+		check( 23f, a3.y + a3.z + b3.y + a3.x )
+		check( 23f, a3.y + a3.z + a3.x + b3.y )
+		check( 23f, b3.y + a3.z + a3.y + a3.x )
+		check( 23f, a3.z + b3.y + a3.y + a3.x )
+		check( 23f, a3.z + a3.y + b3.y + a3.x )
+		check( 23f, a3.z + a3.y + a3.x + b3.y )
+
+	END_TEST
+
+	TEST( fvec4_intra )
+
+		dim as FVec4 a4, b4, c4
+
+		#macro check( result, expr )
+			c4.x = expr
+
+			CU_ASSERT_EQUAL( c4.x,  result )
+			CU_ASSERT_EQUAL( c4.y,  0f  )
+			CU_ASSERT_EQUAL( c4.z,  0f  )
+			CU_ASSERT_EQUAL( c4.w,  0f  )
+		#endmacro
+
+		#macro check3( r, a, b, c )
+			check( r, a + b + c )
+			check( r, b + a + c )
+			check( r, a + c + b )
+			check( r, c + a + b )
+			check( r, b + c + a )
+			check( r, c + b + a )
+		#endmacro
+
+		#macro check4( r, p, a, b, c )
+			check( r, p + a + b + c )
+			check( r, p + b + a + c )
+			check( r, p + a + c + b )
+			check( r, p + c + a + b )
+			check( r, p + b + c + a )
+			check( r, p + c + b + a )
+		#endmacro
+
+		a4 = type( 1f,  2f,  4f,   8f)
+		b4 = type(16f, 32f, 64f, 128f)
+
+		c4.y = 0.0f
+		c4.z = 0.0f
+		c4.w = 0.0f
+
+		check (  1f, a4.x )
+		check (  2f, a4.y )
+		check (  3f, a4.x + a4.y )
+		check (  3f, a4.y + a4.x )
+		check (  4f, a4.z )
+		check (  5f, a4.x + a4.z )
+		check (  5f, a4.z + a4.x )
+		check (  6f, a4.y + a4.z )
+		check (  6f, a4.z + a4.y )
+		check3(  7f, a4.x, a4.y, a4.z )
+		check (  8f, a4.w )
+		check (  9f, a4.x + a4.w )
+		check (  9f, a4.w + a4.x )
+		check ( 10f, a4.w + a4.y )
+		check ( 10f, a4.y + a4.w )
+		check3( 11f, a4.x, a4.y, a4.w )
+		check ( 12f, a4.z + a4.w )
+		check ( 12f, a4.w + a4.z )
+		check3( 13f, a4.x, a4.z, a4.w )
+		check3( 14f, a4.y, a4.z, a4.w )
+		check4( 15f, a4.x, a4.y, a4.z, a4.w )
+		check4( 15f, a4.y, a4.x, a4.z, a4.w )
+		check4( 15f, a4.z, a4.x, a4.y, a4.w )
+		check4( 15f, a4.w, a4.x, a4.y, a4.z )
+
+		'' mix in another source
+
+		check ( 33f, b4.y + a4.x )
+		check ( 33f, a4.x + b4.y )
+
+		check ( 34f, b4.y + a4.y )
+		check ( 34f, a4.y + b4.y )
+
+		check ( 35f, b4.y + a4.x + a4.y )
+		check ( 35f, a4.x + b4.y + a4.y )
+		check ( 35f, a4.x + a4.y + b4.y )
+
+		check ( 35f, b4.y + a4.y + a4.x )
+		check ( 35f, a4.y + b4.y + a4.x )
+		check ( 35f, a4.y + a4.x + b4.y )
+
+		check ( 36f, b4.y + a4.z )
+		check ( 36f, a4.z + b4.y )
+
+		check ( 37f, b4.y + a4.x + a4.z )
+		check ( 37f, a4.x + b4.y + a4.z )
+		check ( 37f, a4.x + a4.z + b4.y )
+
+		check ( 37f, b4.y + a4.z + a4.x )
+		check ( 37f, a4.z + b4.y + a4.x )
+		check ( 37f, a4.z + a4.x + b4.y )
+
+		check ( 38f, b4.y + a4.y + a4.z )
+		check ( 38f, a4.y + b4.y + a4.z )
+		check ( 38f, a4.y + a4.z + b4.y )
+
+		check ( 38f, b4.y + a4.z + a4.y )
+		check ( 38f, a4.z + b4.y + a4.y )
+		check ( 38f, a4.z + a4.y + b4.y )
+
+		check3( 39f, b4.y + a4.x, a4.y, a4.z )
+		check3( 39f, a4.x, b4.y + a4.y, a4.z )
+		check3( 39f, a4.x, a4.y, b4.y + a4.z )
+		check3( 39f, a4.x, a4.y, a4.z + b4.y )
+
+		check ( 40f, b4.y + a4.w )
+		check ( 40f, a4.w + b4.y )
+
+		check ( 41f, b4.y + a4.x + a4.w )
+		check ( 41f, a4.x + b4.y + a4.w )
+		check ( 41f, a4.x + a4.w + b4.y )
+
+		check ( 41f, b4.y + a4.w + a4.x )
+		check ( 41f, a4.w + b4.y + a4.x )
+		check ( 41f, a4.w + a4.x + b4.y )
+
+		check ( 42f, b4.y + a4.w + a4.y )
+		check ( 42f, a4.w + b4.y + a4.y )
+		check ( 42f, a4.w + a4.y + b4.y )
+
+		check ( 42f, b4.y + a4.y + a4.w )
+		check ( 42f, a4.y + b4.y + a4.w )
+		check ( 42f, a4.y + a4.w + b4.y )
+
+		check3( 43f, b4.y + a4.x, a4.y, a4.w )
+		check3( 43f, a4.x, b4.y + a4.y, a4.w )
+		check3( 43f, a4.x, a4.y, b4.y + a4.w )
+		check3( 43f, a4.x, a4.y, a4.w + b4.y )
+
+		check ( 44f, b4.y + a4.z + a4.w )
+		check ( 44f, a4.z + b4.y + a4.w )
+		check ( 44f, a4.z + a4.w + b4.y )
+
+		check ( 44f, b4.y + a4.w + a4.z )
+		check ( 44f, a4.w + b4.y + a4.z )
+		check ( 44f, a4.w + a4.z + b4.y )
+
+		check3( 45f, b4.y + a4.x, a4.z, a4.w )
+		check3( 45f, a4.x, b4.y + a4.z, a4.w )
+		check3( 45f, a4.x, a4.z, b4.y + a4.w )
+		check3( 45f, a4.x, a4.z, a4.w + b4.y )
+
+		check3( 46f, b4.y + a4.y, a4.z, a4.w )
+		check3( 46f, a4.y, b4.y + a4.z, a4.w )
+		check3( 46f, a4.y, a4.z, b4.y + a4.w )
+		check3( 46f, a4.y, a4.z, a4.w + b4.y )
+
+		check4( 47f, b4.y + a4.x, a4.y, a4.z, a4.w )
+		check4( 47f, a4.x, b4.y + a4.y, a4.z, a4.w )
+		check4( 47f, a4.x, a4.y, b4.y + a4.z, a4.w )
+		check4( 47f, a4.x, a4.y, a4.z, b4.y + a4.w )
+		check4( 47f, a4.x, a4.y, a4.z, a4.w + b4.y )
+
+		check4( 47f, b4.y + a4.y, a4.x, a4.z, a4.w )
+		check4( 47f, a4.y, b4.y + a4.x, a4.z, a4.w )
+		check4( 47f, a4.y, a4.x, b4.y + a4.z, a4.w )
+		check4( 47f, a4.y, a4.x, a4.z, b4.y + a4.w )
+		check4( 47f, a4.y, a4.x, a4.z, a4.w + b4.y )
+
+		check4( 47f, b4.y + a4.z, a4.x, a4.y, a4.w )
+		check4( 47f, a4.z, b4.y + a4.x, a4.y, a4.w )
+		check4( 47f, a4.z, a4.x, b4.y + a4.y, a4.w )
+		check4( 47f, a4.z, a4.x, a4.y, b4.y + a4.w )
+		check4( 47f, a4.z, a4.x, a4.y, a4.w + b4.y )
+
+		check4( 47f, b4.y + a4.w, a4.x, a4.y, a4.z )
+		check4( 47f, a4.w, b4.y + a4.x, a4.y, a4.z )
+		check4( 47f, a4.w, a4.x, b4.y + a4.y, a4.z )
+		check4( 47f, a4.w, a4.x, a4.y, b4.y + a4.z )
+		check4( 47f, a4.w, a4.x, a4.y, a4.z + b4.y )
+
+	END_TEST
+
 	''
 	'' simple arithmetic (double precision)
 	''