perf: fix code bugs with SSE2 compiler

mastercoms · mastercoms · commit 57fcefde82b0 · 2025-02-22T20:56:45.000-05:00
with vectorization, more math optimizations, and alignment,
the compiler is more sensitive to some forms of code bugs
diff --git a/src/mathlib/mathlib_base.cpp b/src/mathlib/mathlib_base.cpp
@@ -1479,7 +1479,7 @@ float SmoothCurve( float x )
 inline float MovePeak( float x, float flPeakPos )
 {
 	// Todo: make this higher-order?
-	if( x < flPeakPos )
+	if ( (x < flPeakPos || flPeakPos == 1) && flPeakPos != 0 )
 		return x * 0.5f / flPeakPos;
 	else
 		return 0.5 + 0.5 * (x - flPeakPos) / (1 - flPeakPos);
diff --git a/src/public/mathlib/mathlib.h b/src/public/mathlib/mathlib.h
@@ -954,7 +954,7 @@ void BuildGammaTable( float gamma, float texGamma, float brightness, int overbri
 // convert texture to linear 0..1 value
 inline float TexLightToLinear( int c, int exponent )
 {
-	extern float power2_n[256]; 
+	extern ALIGN128 float power2_n[256]; 
 	Assert( exponent >= -128 && exponent <= 127 );
 	return ( float )c * power2_n[exponent+128];
 }

Original file line number	Diff line number	Diff line change
`@@ -1479,7 +1479,7 @@ float SmoothCurve( float x )`
`1479`	`1479`	`inline float MovePeak( float x, float flPeakPos )`
`1480`	`1480`	`{`
`1481`	`1481`	`// Todo: make this higher-order?`
`1482`		`- if( x < flPeakPos )`
	`1482`	`+ if ( (x < flPeakPos \|\| flPeakPos == 1) && flPeakPos != 0 )`
`1483`	`1483`	`return x * 0.5f / flPeakPos;`
`1484`	`1484`	`else`
`1485`	`1485`	`return 0.5 + 0.5 * (x - flPeakPos) / (1 - flPeakPos);`
Original file line number	Diff line number	Diff line change
`@@ -954,7 +954,7 @@ void BuildGammaTable( float gamma, float texGamma, float brightness, int overbri`
`954`	`954`	`// convert texture to linear 0..1 value`
`955`	`955`	`inline float TexLightToLinear( int c, int exponent )`
`956`	`956`	`{`
`957`		`- extern float power2_n[256];`
	`957`	`+ extern ALIGN128 float power2_n[256];`
`958`	`958`	`Assert( exponent >= -128 && exponent <= 127 );`
`959`	`959`	`return ( float )c * power2_n[exponent+128];`
`960`	`960`	`}`