fix alignment

camel-cdr · Aug 19, 2024 · 0d8eee1 · 0d8eee1
1 parent 72b0b92
commit 0d8eee1
Show file tree

Hide file tree

Showing 5 changed files with 15 additions and 3 deletions.
diff --git a/bench/mandelbrot.S b/bench/mandelbrot.S
@@ -46,7 +46,7 @@ mandelbrot_rvv(size_t width, size_t maxIter, uint32_t *res)
 
 #if HAS_F16
 .global MX(mandelbrot_rvv_f16_) # generated by clang
-.p2align 1
+.balign 2
 MX(rvv_f16_m1p5):
 	.half 0xbe00 # half -1.5
 MX(rvv_f16_m1):
@@ -249,7 +249,7 @@ MX(rvv_f32_13):
 
 #if MX_N == 2 && HAS_E64
 
-.p2align 3
+.balign 8
 .global MX(mandelbrot_rvv_f64_) # generated by clang
 MX(rvv_f64_m1p5):
 	.quad 0xbff8000000000000 # double -1.5

diff --git a/instructions/rvv/Makefile b/instructions/rvv/Makefile
@@ -6,7 +6,7 @@ all: rvv
 
 rvv: gen.S main.c config.h
 	m4 gen.S > main.S
-	${CC} ${CFLAGS} -s main.S main.c -o $@
+	${CC} ${CFLAGS} main.S main.c -o $@
 	rm main.S
 
 clean:

diff --git a/instructions/rvv/gen.S b/instructions/rvv/gen.S
@@ -350,36 +350,43 @@ define(`m_format',`$2 shift(shift($*))')
 define(`m_gen_strings', `.string "m_format(,shift(shift(shift($*))))"')
 
 
+.balign 8
 define(`m_LMUL', 1)
 .global bench_m`'m_LMUL
 	bench_m`'m_LMUL:
 	m_bench_all(`gen_function_pointers')
 
+.balign 8
 define(`m_LMUL', 2)
 .global bench_m`'m_LMUL
 	bench_m`'m_LMUL:
 	m_bench_all(`gen_function_pointers')
 
+.balign 8
 define(`m_LMUL', 4)
 .global bench_m`'m_LMUL
 	bench_m`'m_LMUL:
 	m_bench_all(`gen_function_pointers')
 
+.balign 8
 define(`m_LMUL', 8)
 .global bench_m`'m_LMUL
 	bench_m`'m_LMUL:
 	m_bench_all(`gen_function_pointers')
 
 pushdef(`m_ifmx',`m_ifmx_t($*)')
 
+.balign 8
 .global bench_types
 	bench_types:
 	m_bench_all(`gen_types')
 
+.balign 8
 .global bench_names
 	bench_names:
 	m_bench_all(`gen_strings')
 
+.balign 8
 define(`m_1p', 1+)
 .global bench_count
 bench_count:

diff --git a/instructions/scalar/main.S b/instructions/scalar/main.S
@@ -297,6 +297,7 @@ fabs.d  fa0,fa0
 #define defptr .dword
 #endif
 
+.balign 8
 .global benchmarks
 benchmarks:
 .macro gen_function_pointers name setup code:vararg
@@ -310,6 +311,7 @@ defptr 0 # zero termination
 	.string "\code"
 .endm
 
+.balign 8
 .global benchmark_names
 benchmark_names:
 m_benchmarks_all gen_strings

diff --git a/instructions/xtheadvector/main.S b/instructions/xtheadvector/main.S
@@ -272,6 +272,7 @@ vmseq.vx v8, v0, t0
 
 
 
+.balign 8
 .global benchmarks
 benchmarks:
 .macro gen_function_pointers name type setup code:vararg
@@ -280,6 +281,7 @@ benchmarks:
 m_benchmarks_all gen_function_pointers
 .quad 0 # zero termination
 
+.balign 8
 .global benchmark_types
 benchmark_types:
 .macro gen_types name type setup code:vararg
@@ -299,6 +301,7 @@ m_benchmarks_all gen_types
 
 m_benchmarks_all gen_strings
 
+.balign 8
 .global benchmark_names
 benchmark_names:
 m_benchmarks_all gen_string_arr