Outsmart the LLVM optimizer (halide#8073)

The old definitions of bool_1, bool_2, bool_3 in simd_op_check_x86 (etc) all referred to the same entry in in_f32; as of llvm/llvm-project#76367, the LLVM optimizer is smart enough to realize that (eg) bool1 != bool2 by construction, and optimizes away the code that tests their conditions, such as the one for andps and orps. Initing them from different locations is enough to outsmart the compiler.

(bug was only noticed in the x86 test, but I updated the other tests to guard against future improvements there too.)

test/correctness/simd_op_check_arm.cpp

@@ -37,7 +37,7 @@ class SimdOpCheckARM : public SimdOpCheckTest {
         Expr u32_1 = in_u32(x), u32_2 = in_u32(x + 16), u32_3 = in_u32(x + 32);
         Expr i64_1 = in_i64(x), i64_2 = in_i64(x + 16), i64_3 = in_i64(x + 32);
         Expr u64_1 = in_u64(x), u64_2 = in_u64(x + 16), u64_3 = in_u64(x + 32);
-        Expr bool_1 = (f32_1 > 0.3f), bool_2 = (f32_1 < -0.3f), bool_3 = (f32_1 != -0.34f);
+        Expr bool_1 = (f32_1 > 0.3f), bool_2 = (f32_2 < -0.3f), bool_3 = (f32_3 != -0.34f);
 
         // Table copied from the Cortex-A9 TRM.
 

test/correctness/simd_op_check_hvx.cpp

@@ -45,7 +45,7 @@ class SimdOpCheckHVX : public SimdOpCheckTest {
         Expr u32_1 = in_u32(x), u32_2 = in_u32(x + 16), u32_3 = in_u32(x + 32);
         Expr i64_1 = in_i64(x), i64_2 = in_i64(x + 16), i64_3 = in_i64(x + 32);
         Expr u64_1 = in_u64(x), u64_2 = in_u64(x + 16), u64_3 = in_u64(x + 32);
-        Expr bool_1 = (f32_1 > 0.3f), bool_2 = (f32_1 < -0.3f), bool_3 = (f32_1 != -0.34f);
+        Expr bool_1 = (f32_1 > 0.3f), bool_2 = (f32_2 < -0.3f), bool_3 = (f32_3 != -0.34f);
 
         constexpr int hvx_width = 128;
 

test/correctness/simd_op_check_powerpc.cpp

@@ -36,7 +36,7 @@ class SimdOpCheckPowerPC : public SimdOpCheckTest {
         Expr u32_1 = in_u32(x), u32_2 = in_u32(x + 16), u32_3 = in_u32(x + 32);
         Expr i64_1 = in_i64(x), i64_2 = in_i64(x + 16), i64_3 = in_i64(x + 32);
         Expr u64_1 = in_u64(x), u64_2 = in_u64(x + 16), u64_3 = in_u64(x + 32);
-        // Expr bool_1 = (f32_1 > 0.3f), bool_2 = (f32_1 < -0.3f), bool_3 = (f32_1 != -0.34f);
+        // Expr bool_1 = (f32_1 > 0.3f), bool_2 = (f32_2 < -0.3f), bool_3 = (f32_3 != -0.34f);
 
         // Basic AltiVec SIMD instructions.
         for (int w = 1; w <= 4; w++) {

test/correctness/simd_op_check_wasm.cpp

@@ -37,7 +37,7 @@ class SimdOpCheckWASM : public SimdOpCheckTest {
         Expr u32_1 = in_u32(x), u32_2 = in_u32(x + 16), u32_3 = in_u32(x + 32);
         Expr i64_1 = in_i64(x), i64_2 = in_i64(x + 16), i64_3 = in_i64(x + 32);
         Expr u64_1 = in_u64(x), u64_2 = in_u64(x + 16), u64_3 = in_u64(x + 32);
-        Expr bool_1 = (f32_1 > 0.3f), bool_2 = (f32_1 < -0.3f), bool_3 = (f32_1 != -0.34f);
+        Expr bool_1 = (f32_1 > 0.3f), bool_2 = (f32_2 < -0.3f), bool_3 = (f32_3 != -0.34f);
 
         check("f32.sqrt", 1, sqrt(f32_1));
         check("f32.min", 1, min(f32_1, f32_2));

test/correctness/simd_op_check_x86.cpp

@@ -57,7 +57,7 @@ class SimdOpCheckX86 : public SimdOpCheckTest {
         Expr u32_1 = in_u32(x), u32_2 = in_u32(x + 16), u32_3 = in_u32(x + 32);
         Expr i64_1 = in_i64(x), i64_2 = in_i64(x + 16), i64_3 = in_i64(x + 32);
         Expr u64_1 = in_u64(x), u64_2 = in_u64(x + 16), u64_3 = in_u64(x + 32);
-        Expr bool_1 = (f32_1 > 0.3f), bool_2 = (f32_1 < -0.3f), bool_3 = (f32_1 != -0.34f);
+        Expr bool_1 = (f32_1 > 0.3f), bool_2 = (f32_2 < -0.3f), bool_3 = (f32_3 != -0.34f);
 
         // MMX and SSE1 (in 64 and 128 bits)
         for (int w = 1; w <= 4; w++) {